Experimenting with a gemm based on the cuda steel utils

.circleci/config.yml

@@ -18,17 +18,16 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "26.0.0"
-    resource_class: m4pro.medium
+      xcode: "16.2.0"
+    resource_class: m2pro.medium
     steps:
       - checkout
       - run:
           name: Install
           command: |
-            xcodebuild -downloadComponent MetalToolchain
-            brew install python@3.10
+            brew install python@3.9
             brew install doxygen
-            python3.10 -m venv env
+            python3.9 -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
@@ -90,7 +89,6 @@ jobs:
           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
@@ -120,7 +118,7 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "26.0.0"
+        default: "16.2.0"
       macosx_deployment_target:
         type: string
         default: ""
@@ -128,19 +126,18 @@ jobs:
       xcode: << parameters.xcode_version >>
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m4pro.medium
+    resource_class: m2pro.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 venv --python 3.9
             uv pip install \
               nanobind==2.4.0 \
               cmake \
@@ -199,7 +196,7 @@ jobs:
           name: Run Python tests with JIT
           command: |
             CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e . -v
+              uv pip install -e .
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               uv run --no-project python -m xmlrunner discover \
@@ -225,20 +222,15 @@ jobs:
             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`" \
+            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               uv pip install -e ".[dev]" -v
       - run:
           name: Run Python tests
@@ -246,23 +238,12 @@ jobs:
             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 --max-size 400MB
             ccache --cleanup
       - save_cache:
           key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
@@ -273,10 +254,10 @@ jobs:
     parameters:
       python_version:
         type: string
-        default: "3.10"
+        default: "3.9"
       xcode_version:
         type: string
-        default: "26.0.0"
+        default: "16.2.0"
       build_env:
         type: string
         default: ""
@@ -285,7 +266,7 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: m4pro.medium
+    resource_class: m2pro.medium
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
     steps:
@@ -293,15 +274,11 @@ jobs:
       - 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
+            brew install python@<< parameters.python_version >>
+            brew install openmpi
+            python<< parameters.python_version >> -m venv env
+            source env/bin/activate
+            pip install --upgrade pip
             pip install --upgrade cmake
             pip install nanobind==2.4.0
             pip install --upgrade setuptools
@@ -311,29 +288,29 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            conda activate env
+            source env/bin/activate
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
-            conda activate env
+            source env/bin/activate
             pip install typing_extensions
             python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
-            conda activate env
+            source env/bin/activate
             python setup.py clean --all
             << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
       - when:
           condition:
-            equal: ["3.10", << parameters.python_version >>]
+            equal: ["3.9", << parameters.python_version >>]
           steps:
             - run:
                 name: Build common package
                 command: |
-                  conda activate env
+                  source env/bin/activate
                   python setup.py clean --all
                   << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
@@ -342,7 +319,7 @@ jobs:
             - run:
                 name: Upload package
                 command: |
-                  conda activate env
+                  source env/bin/activate
                   twine upload dist/*
       - store_artifacts:
           path: dist/
@@ -351,7 +328,7 @@ jobs:
     parameters:
       python_version:
         type: string
-        default: "3.10"
+        default: "3.9"
       build_env:
         type: string
         default: ""
@@ -387,7 +364,7 @@ jobs:
             bash python/scripts/repair_linux.sh
       - when:
           condition:
-            equal: ["3.10", << parameters.python_version >>]
+            equal: ["3.9", << parameters.python_version >>]
           steps:
             - run:
                 name: Build common package
@@ -415,7 +392,7 @@ jobs:
         default: ""
     machine:
       image: ubuntu-2204:current
-      resource_class: xlarge
+      resource_class: large
     steps:
       - checkout
       - run:
@@ -462,7 +439,7 @@ workflows:
       - mac_build_and_test:
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test
       - cuda_build_and_test:
           matrix:
@@ -484,10 +461,71 @@ workflows:
               ignore: /.*/
           matrix:
             parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["26.0.0"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
       - build_documentation:
           filters:
             tags:
@@ -503,7 +541,7 @@ workflows:
               ignore: /.*/
           matrix:
             parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               build_env: ["PYPI_RELEASE=1"]
       - build_cuda_release:
           filters:
@@ -529,7 +567,7 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test:
           requires: [ hold ]
       - cuda_build_and_test:
@@ -546,13 +584,59 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["26.0.0"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
       - build_linux_release:
           matrix:
             parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
       - build_cuda_release
 
   build_dev_release:
@@ -564,14 +648,75 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["26.0.0"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
       - build_linux_release:
           matrix:
             parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               build_env: ["DEV_RELEASE=1"]
       - build_cuda_release:
           matrix:

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

@@ -1,24 +0,0 @@
-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 }}
-      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

.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

@@ -1,38 +0,0 @@
-name: 'Build Documentation'
-description: 'Build documentation on a mac'
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup machine
-      uses: ./.github/actions/setup-macos
-
-    - name: Install dependencies
-      shell: sh
-      run: |
-        brew install doxygen
-        uv pip install --upgrade pip cmake
-        uv pip install -r docs/requirements.txt
-        uv pip install . -v
-  
-    - name: Build documentation
-      shell: bash
-      run: |
-        source .venv/bin/activate
-        cd docs
-        doxygen
-        make html O=-W
-    
-    - name: Create artifact tar
-      shell: sh
-      run: tar -cf artifact.tar --cd docs/build/html -L .
-
-    # Do it manually because upload-pages-artifact requires gtar
-    - name: Upload artifact
-      id: upload-artifact
-      uses: actions/upload-artifact@v5
-      with:
-        name: github-pages
-        path: artifact.tar
-        retention-days: 1
-        if-no-files-found: error

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

@@ -1,78 +0,0 @@
-name: 'Build and Test on Linux'
-description: 'Build and test MLX on Linux'
-
-inputs:
-  build-type:
-    description: 'Build type'
-    required: false
-    default: 'debug'
-    type: choice
-    options:
-      - debug
-      - release
-  run-tests:
-    description: 'Whether to run tests'
-    required: false
-    default: 'true'
-    type: boolean
-
-runs:
-  using: "composite"
-  steps:
-    - name: Set DEBUG
-      shell: sh
-      if: inputs.build-type == 'debug'
-      run: echo "DEBUG=1" >> $GITHUB_ENV
-
-    - name: Install Python package
-      shell: sh
-      env:
-        CMAKE_ARGS: "-DCMAKE_COMPILE_WARNING_AS_ERROR=ON"
-      run: pip install -e ".[dev]" -v
-    
-    - 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

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

@@ -1,22 +0,0 @@
-name: 'Build macOS release'
-description: 'Build MLX releases macOS'
-
-inputs:
-  macos-target:
-    description: 'macOS build target'
-    required: false
-    default: '15.0'
-
-runs:
-  using: "composite"
-  steps:
-    - name: Build Python package(s)
-      shell: bash
-      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

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

@@ -1,124 +0,0 @@
-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
-      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
-
-    - name: Generate package stubs
-      shell: bash
-      run: |
-        uv pip install typing_extensions
-        uv run --no-project setup.py generate_stubs
-
-    - name: Run Python tests
-      if: inputs.run-tests == 'true'
-      shell: bash
-      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
-        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
-      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
-    
-    - name: Build CPP only
-      if: inputs.build-type == 'debug'
-      shell: bash
-      run: |
-        mkdir -p build
-        cd build
-        cmake ..
-        make -j $(sysctl -n hw.ncpu)
-    
-    - name: Run CPP tests
-      if: ${{ inputs.build-type == 'debug' && inputs.run-tests == 'true' }}
-      shell: bash
-      env:
-        DEVICE: gpu
-        METAL_DEVICE_WRAPPER_TYPE: 1
-        METAL_DEBUG_ERROR_MODE: 0
-      run: ./build/tests/tests
-    
-    - name: Build small binary with JIT
-      if: inputs.build-jit == 'true'
-      shell: bash
-      run: |
-        mkdir -p build
-        cd build
-        cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
-          -DBUILD_SHARED_LIBS=ON \
-          -DMLX_BUILD_CPU=OFF \
-          -DMLX_BUILD_SAFETENSORS=OFF \
-          -DMLX_BUILD_GGUF=OFF \
-          -DMLX_METAL_JIT=ON
-        make -j $(sysctl -n hw.ncpu)
-    
-    - name: Run Python tests with JIT
-      if: ${{ inputs.build-jit == 'true' && inputs.run-tests == 'true' }}
-      shell: bash
-      env:
-        LOW_MEMORY: 1
-        DEVICE: gpu
-        METAL_DEVICE_WRAPPER_TYPE: 1
-        METAL_DEBUG_ERROR_MODE: 0
-      run: |
-        CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-          uv pip install -e . -v
-        uv run -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

@@ -1,83 +0,0 @@
-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'
-    required: false
-    default: 'linux'
-    type: choice
-    options:
-      - linux
-      - cuda
-  python-version:
-    description: 'Version of python to set up'
-    required: false
-    default: '3.10'
-
-runs:
-  using: "composite"
-  steps:
-    - name: Free disk space
-      shell: sh
-      if: inputs.runner-type == 'linux'
-      run: sudo rm -rf "$AGENT_TOOLSDIRECTORY"
-
-    - 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
-
-    - uses: actions/setup-python@v6
-      with:
-        python-version: ${{ inputs.python-version }}
-        cache: 'pip'
-
-    - name: setup python venv
-      shell: bash
-      run: |
-        python -m venv .venv
-        source .venv/bin/activate
-        echo PATH=$PATH >> $GITHUB_ENV
-        pip install --upgrade pip cmake
-
-    - 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'
-      env:
-        TZ: Etc/UTC
-      shell: bash ## Specific to Ubuntu 22.04 & Architecture x86_64
-      run: |
-        wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/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 -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
-
-    - name: Package and Driver Report
-      if: inputs.runner-type == 'cuda'
-      shell: bash
-      run: |
-        sudo apt-get install -y ubuntu-drivers-common dkms
-        echo "NVIDIA Driver Packages Available:"
-        sudo ubuntu-drivers list --gpgpu
-        echo "NVIDIA Driver Version:"
-        cat /proc/driver/nvidia/version || echo "nvidia driver not found"
-        echo "Installed NVIDIA and CUDA packages:"
-        dpkg -l | egrep "cuda|nvidia" -i
-        echo "DKMS Status:"
-        dkms status || echo "dkms not found"
-        echo "NVIDIA-SMI Status:"
-        nvidia-smi || echo "nvidia-smi not found"

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

@@ -1,31 +0,0 @@
-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
-    default: '3.10'
-
-runs:
-  using: "composite"
-  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
-      with:
-          python-version: ${{ inputs.python-version }}
-          activate-environment: true

.github/dependabot.yml

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

.github/workflows/documentation.yml

@@ -1,28 +0,0 @@
-name: Documentation
-
-on:
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-jobs:
-  build:
-    runs-on: [self-hosted, macos]
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/build-docs
-      
-  deploy:
-    needs: build
-    permissions:
-      pages: write
-      id-token: write
-    runs-on: ubuntu-latest
-    environment:
-      name: github-pages
-      url: ${{ steps.deployment.outputs.page_url }}
-    steps:
-      - name: Deploy to GitHub Pages
-        id: deployment
-        uses: actions/deploy-pages@v4

.github/workflows/nightly.yml

@@ -1,93 +0,0 @@
-name: Nightly Build
-
-on:
-  schedule:
-    - cron: 33 6 * * 1-5
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-jobs:
-  build_linux_release:
-    strategy:
-      fail-fast: false
-      matrix:
-        python_version: ["3.10", "3.14"]
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-      - uses: ./.github/actions/build-linux
-        with:
-          build-type: release
-          run-tests: false
-      - name: Upload mlx artifacts
-        uses: actions/upload-artifact@v5
-        with:
-          name: linux-wheels-${{ matrix.python_version }}
-          path: wheelhouse/mlx-*.whl
-          retention-days: 7
-      - name: Upload mlx-cpu artifacts
-        if: matrix.python_version == '3.10'
-        uses: actions/upload-artifact@v5
-        with:
-          name: mlx-cpu
-          path: wheelhouse/mlx_cpu-*.whl
-          retention-days: 7
-  
-  build_linux_with_tests:
-    strategy:
-      fail-fast: false
-      matrix:
-        python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-        with:
-          python-version: ${{ matrix.python_version }}
-      - uses: ./.github/actions/build-linux
-
-  build_mac_release:
-    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: ./.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
-        with:
-          runner-type: 'cuda'
-      - uses: ./.github/actions/build-cuda
-
-  build_cuda_release:
-    runs-on: ubuntu-22-large
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-        with:
-          runner-type: 'cuda'
-      - 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:
-          name: mlx-cuda
-          path: wheelhouse/mlx_cuda-*.whl
-          retention-days: 7
-

.github/workflows/pull_request.yml

@@ -1,46 +1,20 @@
-name: Build and Test
-
-on: pull_request  
-
-permissions:
-  contents: read
+on:
+  pull_request:
+    branches:
+      - main
 
 jobs:
   check_lint:
-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-latest
     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
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v4
         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
+          python-version: 3.8
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install pre-commit black isort clang-format
+      - name: Run lint
+        run: |
+          pre-commit run --all-files

.github/workflows/release.yml

@@ -1,188 +0,0 @@
-name: PyPI Release
-
-on:
-  push:
-    tags:
-      - 'v*'
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-jobs:
-  build_documentation:
-    runs-on: [self-hosted, macos]
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/build-docs
-    
-  deploy_documentation:
-    needs: build_documentation
-    permissions:
-      pages: write
-      id-token: write
-    runs-on: ubuntu-latest
-    environment:
-      name: github-pages
-      url: ${{ steps.deployment.outputs.page_url }}
-    steps:
-      - name: Deploy to GitHub Pages
-        id: deployment
-        uses: actions/deploy-pages@v4
-
-  build_linux_release:
-    strategy:
-      matrix:
-        python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-    runs-on: ubuntu-22.04
-    env:
-      PYPI_RELEASE: 1
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-        with:
-          python-version: ${{ matrix.python_version }}
-      - uses: ./.github/actions/build-linux
-        with:
-          build-type: release
-          run-tests: false
-      - name: Upload MLX artifacts
-        uses: actions/upload-artifact@v5
-        with:
-          name: linux-wheels-${{ matrix.python_version }}
-          path: wheelhouse/mlx-*.whl
-      - name: Upload CPU artifacts
-        if: matrix.python_version == '3.10'
-        uses: actions/upload-artifact@v5
-        with:
-          name: mlx-cpu
-          path: wheelhouse/mlx_cpu-*.whl
-  
-  build_mac_release:
-    strategy:
-      matrix:
-        python-version: ["3.10", "3.11", "3.12", "3.13"]
-        # TODO: 3.14 had issues finding a compatible tensorflow
-    runs-on: [self-hosted, macos]
-    env:
-      PYPI_RELEASE: 1
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-macos
-        with:
-          python-version: ${{ matrix.python-version }}
-      - uses: ./.github/actions/build-macos
-        with:
-          build-type: release
-      - name: Upload MLX artifacts
-        uses: actions/upload-artifact@v5
-        with:
-          name: mac-wheels-${{ matrix.python-version }}
-          path: dist/mlx-*.whl
-      - name: Upload Metal artifacts
-        if: matrix.python-version == '3.10'
-        uses: actions/upload-artifact@v5
-        with:
-          name: mlx-metal
-          path: dist/mlx_metal-*.whl
-
-  build_cuda_release:
-    runs-on: ubuntu-22-large
-    env:
-      PYPI_RELEASE: 1
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-        with:
-          runner-type: 'cuda'
-      - 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:
-          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]
-    permissions:
-      id-token: write
-    environment:
-      name: pypi
-      url: https://pypi.org/p/mlx
-    steps:
-      - uses: actions/download-artifact@v6
-        with:
-          pattern: linux-wheels-*
-          merge-multiples: true
-          path: artifacts
-      - uses: actions/download-artifact@v6
-        with:
-          pattern: mac-wheels-*
-          merge-multiples: true
-          path: artifacts
-      - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1
-  
-  pypi-publish-cuda:
-    name: Upload CUDA release to PyPI
-    runs-on: ubuntu-latest
-    needs: build_cuda_release
-    permissions:
-      id-token: write
-    environment:
-      name: pypi
-      url: https://pypi.org/p/mlx-cuda
-    steps:
-      - uses: actions/download-artifact@v6
-        with:
-          name: mlx-cuda
-          path: artifacts
-      - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1
-
-  pypi-publish-cpu:
-    name: Upload CPU release to PyPI
-    runs-on: ubuntu-latest
-    needs: build_linux_release
-    permissions:
-      id-token: write
-    environment:
-      name: pypi
-      url: https://pypi.org/p/mlx-cpu
-    steps:
-      - uses: actions/download-artifact@v6
-        with:
-          name: mlx-cpu
-          path: artifacts
-      - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1
-
-  pypi-publish-metal:
-    name: Upload Metal release to PyPI
-    runs-on: ubuntu-latest
-    needs: build_mac_release
-    permissions:
-      id-token: write
-    environment:
-      name: pypi
-      url: https://pypi.org/p/mlx-metal
-    steps:
-      - uses: actions/download-artifact@v6
-        with:
-          name: mlx-metal
-          path: artifacts
-      - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1

.pre-commit-config.yaml

@@ -1,10 +1,4 @@
 repos:
--   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v6.0.0
-    hooks:
-    -   id: check-yaml
-    # -   id: end-of-file-fixer
-    # -   id: trailing-whitespace
 -   repo: https://github.com/pre-commit/mirrors-clang-format
     rev: v19.1.7
     hooks:

ACKNOWLEDGMENTS.md

@@ -19,17 +19,12 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
-- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer, and the `ReLU²` activation function.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>
 
-# Organizations
-
-MLX has received contributions from the following companies:
-- NVIDIA Corporation & Affiliates
-
 # Third-Party Software
 
 MLX leverages several third-party software, listed here together with

CMakeLists.txt

@@ -26,7 +26,6 @@ set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)
-set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # ----------------------------- Configuration -----------------------------
 option(MLX_BUILD_TESTS "Build tests for mlx" ON)
@@ -88,26 +87,22 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-# Supress warnings: note: parameter passing for argument of type
-# ‘std::pair<float, float>’ when C++17 is enabled changed to match C++14 in GCC
-# 10.1
-target_compile_options(mlx PRIVATE -Wno-psabi)
+if(MLX_BUILD_METAL)
+  set(METAL_LIB "-framework Metal")
+  set(FOUNDATION_LIB "-framework Foundation")
+  set(QUARTZ_LIB "-framework QuartzCore")
+endif()
 
 if(MLX_BUILD_CUDA)
   enable_language(CUDA)
 endif()
 
-if(MLX_BUILD_METAL)
-  find_library(METAL_LIB Metal)
-  find_library(FOUNDATION_LIB Foundation)
-  find_library(QUARTZ_LIB QuartzCore)
-  if(METAL_LIB)
-    message(STATUS "Metal found ${METAL_LIB}")
-  else()
-    message(
-      FATAL_ERROR
-        "Metal not found. Set MLX_BUILD_METAL=OFF to build without GPU")
-  endif()
+if(MLX_BUILD_METAL AND NOT METAL_LIB)
+  message(STATUS "Metal not found. Unable to build GPU")
+  set(MLX_BUILD_METAL OFF)
+  set(MLX_METAL_DEBUG OFF)
+elseif(MLX_BUILD_METAL)
+  message(STATUS "Building METAL sources")
 
   if(MLX_METAL_DEBUG)
     add_compile_definitions(MLX_METAL_DEBUG)
@@ -116,8 +111,7 @@ if(MLX_BUILD_METAL)
   # Throw an error if xcrun not found
   execute_process(
     COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
-    OUTPUT_VARIABLE MACOS_SDK_VERSION
-    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
+    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
 
   if(${MACOS_SDK_VERSION} LESS 14.0)
     message(
@@ -146,12 +140,6 @@ if(MLX_BUILD_METAL)
   target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
 endif()
 
-if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
-  # With newer clang/gcc versions following libs are implicitly linked, but when
-  # building on old distributions they need to be explicitly listed.
-  target_link_libraries(mlx PRIVATE dl pthread)
-endif()
-
 if(WIN32)
   if(MSVC)
     # GGUF does not build with MSVC.
@@ -179,7 +167,7 @@ if(MLX_BUILD_CPU)
     message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
     set(MLX_BUILD_ACCELERATE ON)
   else()
-    message(STATUS "Accelerate not found, using default backend.")
+    message(STATUS "Accelerate or arm neon not found, using default backend.")
     set(MLX_BUILD_ACCELERATE OFF)
   endif()
 

README.md

@@ -110,7 +110,7 @@ Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. If you find
 MLX useful in your research and wish to cite it, please use the following
 BibTex entry:
 
-```text
+```
 @software{mlx2023,
   author = {Awni Hannun and Jagrit Digani and Angelos Katharopoulos and Ronan Collobert},
   title = {{MLX}: Efficient and flexible machine learning on Apple silicon},

benchmarks/python/blas/bench_gemm.py

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

benchmarks/python/blas/bench_gemv.py

@@ -196,7 +196,7 @@ def bench_with_out_len(ax, out_vec_len, in_vector_lens, dtype, transpose):
 
 
 for transpose in (False, True):
-    for dtype in ("float32", "float16", "complex64"):
+    for dtype in ("float32", "float16"):
         fig, axs = plt.subplots(
             len(in_vec_sizes), 2, figsize=(8.5, 11), layout="constrained"
         )
@@ -215,7 +215,7 @@ for transpose in (False, True):
         fig.suptitle(f"{device_name}: {dtype} {op_name}")
         fig.savefig(
             os.path.join(
-                results_dir, f"{device_name.replace(' ', '_')}_{dtype}_{op_name}.pdf"
+                results_dir, f'{device_name.replace(" ", "_")}_{dtype}_{op_name}.pdf'
             )
         )
         plt.close(fig)

cmake/FindNCCL.cmake

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

docs/src/dev/custom_metal_kernels.rst

@@ -127,8 +127,7 @@ relying on a copy from ``ensure_row_contiguous``:
       name="myexp_strided",
       input_names=["inp"],
       output_names=["out"],
-      source=source,
-      ensure_row_contiguous=False,
+      source=source
   )
 
   def exp_elementwise(a: mx.array):
@@ -139,6 +138,7 @@ relying on a copy from ``ensure_row_contiguous``:
           threadgroup=(256, 1, 1),
           output_shapes=[a.shape],
           output_dtypes=[a.dtype],
+          ensure_row_contiguous=False,
       )
       return outputs[0]
 

docs/src/index.rst

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

docs/src/install.rst

@@ -16,7 +16,7 @@ silicon computer is
 To install from PyPI your system must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
-- Using a native Python >= 3.10
+- Using a native Python >= 3.9
 - macOS >= 13.5
 
 .. note::
@@ -39,7 +39,7 @@ requirements:
 - Nvidia driver >= 550.54.14
 - CUDA toolkit >= 12.0
 - Linux distribution with glibc >= 2.35
-- Python >= 3.10
+- Python >= 3.9
 
 
 CPU-only (Linux)
@@ -55,7 +55,7 @@ To install the CPU-only package from PyPi your system must meet the following
 requirements:
 
 - Linux distribution with glibc >= 2.35
-- Python >= 3.10
+- Python >= 3.9
 
 
 Troubleshooting
@@ -271,7 +271,7 @@ and the CUDA toolkit. For example on Ubuntu, run the following:
    dpkg -i cuda-keyring_1.1-1_all.deb
    apt-get update -y
    apt-get -y install cuda-toolkit-12-9
-   apt-get install libblas-dev liblapack-dev liblapacke-dev libcudnn9-dev-cuda-12 -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
 
 
 When building either the Python or C++ APIs make sure to pass the cmake flag

docs/src/python/cuda.rst

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

docs/src/python/fast.rst

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

docs/src/python/nn/functions.rst

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

docs/src/python/nn/layers.rst

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

docs/src/python/ops.rst

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

docs/src/usage/compile.rst

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

docs/src/usage/distributed.rst

@@ -184,7 +184,7 @@ almost identical to the example above:
 
     def step(model, x, y):
         loss, grads = loss_grad_fn(model, x, y)
-        grads = mx.nn.average_gradients(grads)  # <---- This line was added
+        grads = mlx.nn.average_gradients(grads) # <---- This line was added
         optimizer.update(model, grads)
         return loss
 

docs/src/usage/export.rst

@@ -164,11 +164,11 @@ to export a function which can be used for inputs with variable shapes:
 
 .. code-block:: python
 
-  mx.export_function("fun.mlxfn", mx.abs, mx.array([0.0]), shapeless=True)
+  mx.export_function("fun.mlxfn", mx.abs, mx.array(0.0), shapeless=True)
   imported_abs = mx.import_function("fun.mlxfn")
 
   # Ok
-  out, = imported_abs(mx.array([-1.0]))
+  out, = imported_abs(mx.array(-1.0))
 
   # Also ok
   out, = imported_abs(mx.array([-1.0, -2.0]))

docs/src/usage/indexing.rst

@@ -107,20 +107,8 @@ same array:
   >>> a
   array([1, 2, 0], dtype=int32)
 
-Note that unlike NumPy, slicing an array creates a copy, not a view. So
-mutating it does not mutate the original array:
 
-.. code-block:: shell
-
-  >>> a = mx.array([1, 2, 3])
-  >>> b = a[:]
-  >>> b[2] = 0
-  >>> b
-  array([1, 2, 0], dtype=int32)
-  >>> a
-  array([1, 2, 3], dtype=int32)
-
-Also unlike NumPy, updates to the same location are nondeterministic:
+Note, unlike NumPy, updates to the same location are nondeterministic:
 
 .. code-block:: shell
 

mlx/allocator.h

@@ -14,7 +14,7 @@ class Buffer {
   void* ptr_;
 
  public:
-  explicit Buffer(void* ptr) : ptr_(ptr) {};
+  Buffer(void* ptr) : ptr_(ptr) {};
 
   // Get the raw data pointer from the buffer
   void* raw_ptr();

mlx/array.cpp

@@ -64,7 +64,7 @@ array array::unsafe_weak_copy(const array& other) {
       other.strides(),
       other.flags(),
       [](auto) {});
-  cpy.array_desc_->offset = other.array_desc_->offset;
+  cpy.array_desc_->data_ptr = other.array_desc_->data_ptr;
   return cpy;
 }
 
@@ -141,7 +141,7 @@ bool array::is_tracer() const {
 
 void array::set_data(allocator::Buffer buffer, Deleter d) {
   array_desc_->data = std::make_shared<Data>(buffer, d);
-  array_desc_->offset = 0;
+  array_desc_->data_ptr = buffer.raw_ptr();
   array_desc_->data_size = size();
   array_desc_->flags.contiguous = true;
   array_desc_->flags.row_contiguous = true;
@@ -156,7 +156,7 @@ void array::set_data(
     Flags flags,
     Deleter d) {
   array_desc_->data = std::make_shared<Data>(buffer, d);
-  array_desc_->offset = 0;
+  array_desc_->data_ptr = buffer.raw_ptr();
   array_desc_->data_size = data_size;
   array_desc_->strides = std::move(strides);
   array_desc_->flags = flags;
@@ -172,8 +172,9 @@ void array::copy_shared_buffer(
   array_desc_->strides = strides;
   array_desc_->flags = flags;
   array_desc_->data_size = data_size;
-  array_desc_->offset =
-      sizeof(char) * itemsize() * offset + other.array_desc_->offset;
+  auto char_offset = sizeof(char) * itemsize() * offset;
+  array_desc_->data_ptr = static_cast<void*>(
+      static_cast<char*>(other.array_desc_->data_ptr) + char_offset);
 }
 
 void array::copy_shared_buffer(const array& other) {
@@ -240,8 +241,8 @@ array::ArrayDesc::ArrayDesc(
     std::vector<array> inputs)
     : shape(std::move(shape)),
       dtype(dtype),
-      primitive(std::move(primitive)),
       status(Status::unscheduled),
+      primitive(std::move(primitive)),
       inputs(std::move(inputs)) {
   init();
 }

mlx/array.h

@@ -349,23 +349,15 @@ class array {
     return array_desc_->data;
   }
 
-  // Return a raw pointer to the arrays data. This function may do a copy if
-  // the underlying buffer is not accessible on the CPU. When accessing the
-  // data for GPU kernels, be sure to use the correct method / function for the
-  // given backend to access the GPU pointer.
+  // Return a raw pointer to the arrays data
   template <typename T>
   T* data() {
-    return reinterpret_cast<T*>(
-        (static_cast<char*>(buffer().raw_ptr()) + array_desc_->offset));
+    return static_cast<T*>(array_desc_->data_ptr);
   }
 
   template <typename T>
   const T* data() const {
-    return const_cast<array&>(*this).data<T>();
-  }
-
-  int64_t offset() const {
-    return array_desc_->offset;
+    return static_cast<T*>(array_desc_->data_ptr);
   }
 
   enum Status {
@@ -469,8 +461,8 @@ class array {
     // can share the underlying data buffer.
     std::shared_ptr<Data> data;
 
-    // Offset from beginning of data pointer
-    int64_t offset{0};
+    // Properly offset data pointer
+    void* data_ptr{nullptr};
 
     // The size in elements of the data buffer the array accesses
     size_t data_size;

mlx/backend/common/binary.h

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

mlx/backend/common/broadcasting.cpp

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

mlx/backend/common/compiled.cpp

@@ -114,9 +114,7 @@ void compiled_allocate_outputs(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
     const std::function<bool(size_t)>& is_constant,
-    bool contiguous,
-    const std::function<allocator::Buffer(size_t)>&
-        mallocfn /* = allocator::malloc */) {
+    bool contiguous) {
   if (contiguous) {
     int o = 0;
     Strides strides;
@@ -142,7 +140,7 @@ void compiled_allocate_outputs(
     }
     for (; o < outputs.size(); ++o) {
       outputs[o].set_data(
-          mallocfn(data_size * outputs[o].itemsize()),
+          allocator::malloc(data_size * outputs[o].itemsize()),
           data_size,
           strides,
           flags);
@@ -165,7 +163,7 @@ void compiled_allocate_outputs(
       }
     }
     for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(mallocfn(outputs[o].nbytes()));
+      outputs[o].set_data(allocator::malloc(outputs[o].nbytes()));
     }
   }
 }

mlx/backend/common/compiled.h

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

mlx/backend/common/copy.h

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

mlx/backend/common/matmul.h

@@ -13,7 +13,7 @@ inline std::tuple<Shape, Strides, Strides> collapse_batches(
     const array& a,
     const array& b) {
   if (a.ndim() == 2) {
-    return {Shape{1}, Strides{0}, Strides{0}};
+    return {{1}, {0}, {0}};
   }
 
   Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
@@ -38,7 +38,7 @@ inline std::tuple<Shape, Strides, Strides> collapse_batches(
 inline std::tuple<Shape, Strides, Strides, Strides>
 collapse_batches(const array& a, const array& b, const array& c) {
   if (a.ndim() == 2) {
-    return {Shape{1}, Strides{0}, Strides{0}, Strides{0}};
+    return {{1}, {0}, {0}, {0}};
   }
 
   Shape A_bshape{a.shape().begin(), a.shape().end() - 2};

mlx/backend/common/slicing.cpp

@@ -45,7 +45,7 @@ void slice(
     const Shape& start_indices,
     const Shape& strides) {
   if (out.size() == 0) {
-    out.set_data(allocator::malloc(0));
+    out.set_data(nullptr);
     return;
   }
 

mlx/backend/common/ternary.h

@@ -11,8 +11,6 @@ namespace mlx::core {
 enum class TernaryOpType {
   ScalarScalarScalar,
   VectorVectorVector,
-  VectorVectorScalar,
-  VectorScalarVector,
   General,
 };
 
@@ -27,14 +25,6 @@ get_ternary_op_type(const array& a, const array& b, const array& c) {
       (a.flags().col_contiguous && b.flags().col_contiguous &&
        c.flags().col_contiguous)) {
     topt = TernaryOpType::VectorVectorVector;
-  } else if (
-      b.data_size() == 1 && a.flags().row_contiguous &&
-      c.flags().row_contiguous) {
-    topt = TernaryOpType::VectorScalarVector;
-  } else if (
-      c.data_size() == 1 && a.flags().row_contiguous &&
-      b.flags().row_contiguous) {
-    topt = TernaryOpType::VectorVectorScalar;
   } else {
     topt = TernaryOpType::General;
   }
@@ -46,8 +36,7 @@ inline void set_ternary_op_output_data(
     const array& b,
     const array& c,
     array& out,
-    TernaryOpType topt,
-    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
+    TernaryOpType topt) {
   auto maybe_donate = [&out](const array& x) {
     if (is_donatable(x, out)) {
       out.copy_shared_buffer(x);
@@ -58,25 +47,24 @@ inline void set_ternary_op_output_data(
 
   switch (topt) {
     case TernaryOpType::ScalarScalarScalar:
-      out.set_data(mallocfn(out.itemsize()), 1, b.strides(), b.flags());
+      out.set_data(
+          allocator::malloc(out.itemsize()), 1, b.strides(), b.flags());
       break;
     case TernaryOpType::VectorVectorVector:
       if (!(maybe_donate(a) || maybe_donate(b) || maybe_donate(c))) {
         out.set_data(
-            mallocfn(out.itemsize() * b.data_size()),
+            allocator::malloc(out.itemsize() * b.data_size()),
             b.data_size(),
             b.strides(),
             b.flags());
       }
       break;
-    case TernaryOpType::VectorVectorScalar:
-    case TernaryOpType::VectorScalarVector:
     case TernaryOpType::General:
       // Try to donate an input which is row_contiguous
       if (!((a.flags().row_contiguous && maybe_donate(a)) ||
             (b.flags().row_contiguous && maybe_donate(b)) ||
             (c.flags().row_contiguous && maybe_donate(c)))) {
-        out.set_data(mallocfn(out.nbytes()));
+        out.set_data(allocator::malloc(out.nbytes()));
       }
       break;
   }

mlx/backend/common/unary.h

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

mlx/backend/common/utils.cpp

@@ -228,4 +228,31 @@ std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
       std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
 }
 
+array swapaxes_in_eval(const array& x, int axis1, int axis2) {
+  int ndim = x.ndim();
+  if (axis1 < 0) {
+    axis1 += ndim;
+  }
+  if (axis2 < 0) {
+    axis2 += ndim;
+  }
+
+  auto shape = x.shape();
+  std::swap(shape[axis1], shape[axis2]);
+  auto strides = x.strides();
+  std::swap(strides[axis1], strides[axis2]);
+
+  auto [data_size, row_contiguous, col_contiguous] =
+      check_contiguity(shape, strides);
+  bool contiguous = data_size == x.data_size();
+
+  array out(std::move(shape), x.dtype(), nullptr, {});
+  out.copy_shared_buffer(
+      x,
+      std::move(strides),
+      {contiguous, row_contiguous, col_contiguous},
+      x.data_size());
+  return out;
+}
+
 } // namespace mlx::core

mlx/backend/common/utils.h

@@ -196,6 +196,9 @@ void shared_buffer_reshape(
     const Strides& out_strides,
     array& out);
 
+// Like the swapaxes op but safe to call in eval_gpu.
+array swapaxes_in_eval(const array& x, int axis1, int axis2);
+
 template <typename T>
 inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
   vec.erase(std::next(vec.begin(), index));

mlx/backend/cpu/compiled.cpp

@@ -15,7 +15,6 @@
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
-#include "mlx/version.h"
 
 namespace mlx::core {
 
@@ -95,11 +94,7 @@ void* compile(
     kernel_file_name = kernel_name;
   }
 
-  auto output_dir =
-      std::filesystem::temp_directory_path() / "mlx" / version() / "cpu";
-  if (!std::filesystem::exists(output_dir)) {
-    std::filesystem::create_directories(output_dir);
-  }
+  auto output_dir = std::filesystem::temp_directory_path();
 
   std::string shared_lib_name = "lib" + kernel_file_name + ".so";
   auto shared_lib_path = (output_dir / shared_lib_name).string();
@@ -162,12 +157,10 @@ inline void build_kernel(
 #endif
 
   // Start the kernel
-  os << "void " << kernel_name
-     << "(int* shape, int64_t** strides, void** args) {" << std::endl;
+  os << "void " << kernel_name << "(void** args) {" << std::endl;
 
   // Add the input arguments
   int cnt = 0;
-  int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
     // Skip constants from the input list
     if (is_constant(i)) {
@@ -182,8 +175,8 @@ inline void build_kernel(
        << "];" << std::endl;
     // Scalars and contiguous need no strides
     if (!is_scalar(x) && !contiguous) {
-      os << "  const int64_t* " << xname << "_strides = strides["
-         << strides_index++ << "];" << std::endl;
+      os << "  const size_t* " << xname << "_strides = (size_t*)args[" << cnt++
+         << "];" << std::endl;
     }
   }
 
@@ -193,8 +186,10 @@ inline void build_kernel(
     os << "  " << tstr << "* " << namer.get_name(x) << " = (" << tstr
        << "*)args[" << cnt++ << "];" << std::endl;
   }
-  // Add output size
-  if (contiguous) {
+  // Add output strides and shape to extract the indices.
+  if (!contiguous) {
+    os << "  const int* shape = (int*)args[" << cnt++ << "];" << std::endl;
+  } else {
     os << "  const size_t size = (size_t)args[" << cnt++ << "];" << std::endl;
   }
 
@@ -293,8 +288,17 @@ void Compiled::eval_cpu(
   auto [contiguous, shape, strides] =
       compiled_collapse_contiguous_dims(inputs, outputs[0], is_constant_);
 
+  // Force allocating shape/strides on heap so we can take their data() first
+  // and then std::move them.
+  // TODO: Refactor code to avoid heap allocation.
+  shape.grow();
+  for (auto& s : strides) {
+    s.grow();
+  }
+
   // Collect function input arguments.
   std::vector<void*> args;
+  int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
     if (is_constant_(i)) {
       continue;
@@ -302,6 +306,9 @@ void Compiled::eval_cpu(
     const auto& x = inputs[i];
     encoder.set_input_array(x);
     args.push_back((void*)x.data<void>());
+    if (!contiguous && !is_scalar(x)) {
+      args.push_back(strides[strides_index++].data());
+    }
   }
 
   // Get the kernel name from the lib
@@ -336,20 +343,16 @@ void Compiled::eval_cpu(
     args.push_back(x.data<void>());
     encoder.set_output_array(x);
   }
-  if (contiguous) {
+  if (!contiguous) {
+    args.push_back((void*)shape.data());
+  } else {
     args.push_back((void*)outputs[0].data_size());
   }
-  auto fun = reinterpret_cast<void (*)(int*, int64_t**, void**)>(fn_ptr);
+  auto fun = (void (*)(void**))fn_ptr;
   encoder.dispatch([fun,
                     args = std::move(args),
                     strides = std::move(strides),
-                    shape = std::move(shape)]() mutable {
-    SmallVector<int64_t*> strides_ptrs;
-    for (auto& s : strides) {
-      strides_ptrs.push_back(s.data());
-    }
-    fun(shape.data(), strides_ptrs.data(), args.data());
-  });
+                    shape = std::move(shape)]() mutable { fun(args.data()); });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/conv.cpp

@@ -996,6 +996,131 @@ void explicit_gemm_conv_1D_cpu(
   encoder.add_temporaries(std::move(temps));
 }
 
+void explicit_gemm_conv_2D_cpu(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding_lo,
+    const std::vector<int>& padding_hi,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    Stream stream) {
+  const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
+  const int iH = in.shape(1); // Input spatial dim
+  const int iW = in.shape(2); // Input spatial dim
+  const int oH = out.shape(1); // Output spatial dim
+  const int oW = out.shape(2); // Output spatial dim
+  const int O = wt.shape(0); // Out channels
+  const int C = wt.shape(3); // In channels
+  const int wH = wt.shape(1); // Weight spatial dim
+  const int wW = wt.shape(2); // Weight spatial dim
+
+  auto conv_dtype = out.dtype();
+  auto& encoder = cpu::get_command_encoder(stream);
+
+  // Pad input
+  Shape padded_shape = {
+      N,
+      iH + padding_lo[0] + padding_hi[0],
+      iW + padding_lo[1] + padding_hi[1],
+      C};
+  array in_padded(padded_shape, conv_dtype, nullptr, {});
+
+  // Fill with zeros
+  std::vector<array> temps;
+  temps.push_back(array(0, conv_dtype));
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
+
+  // Pick input slice from padded
+  size_t data_offset = padding_lo[0] * in_padded.strides()[1] +
+      padding_lo[1] * in_padded.strides()[2];
+  array in_padded_slice(in.shape(), in_padded.dtype(), nullptr, {});
+  in_padded_slice.copy_shared_buffer(
+      in_padded,
+      in_padded.strides(),
+      in_padded.flags(),
+      in_padded_slice.size(),
+      data_offset);
+  temps.push_back(in_padded_slice);
+
+  // Copy input values into the slice
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+
+  // Make strided view
+  Shape strided_shape = {N, oH, oW, wH, wW, C};
+
+  Strides strided_strides = {
+      in_padded.strides()[0],
+      in_padded.strides()[1] * wt_strides[0],
+      in_padded.strides()[2] * wt_strides[1],
+      in_padded.strides()[1],
+      in_padded.strides()[2],
+      in_padded.strides()[3]};
+  auto flags = in_padded.flags();
+
+  array in_strided_view(strided_shape, in_padded.dtype(), nullptr, {});
+  in_strided_view.copy_shared_buffer(
+      in_padded, strided_strides, flags, in_strided_view.size(), 0);
+
+  // Materialize strided view
+  Shape strided_reshape = {N * oH * oW, wH * wW * C};
+  array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
+  temps.push_back(in_strided);
+
+  // Check wt dtype and prepare
+  auto gemm_wt = wt;
+  auto gemm_out = out;
+
+  if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
+    auto ctype =
+        wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
+    gemm_wt = array(wt.shape(), float32, nullptr, {});
+    copy_cpu(wt, gemm_wt, ctype, stream);
+    temps.push_back(gemm_wt);
+  }
+
+  if (out.dtype() != float32) {
+    gemm_out = array(out.shape(), float32, nullptr, {});
+    gemm_out.set_data(allocator::malloc(gemm_out.nbytes()));
+    temps.push_back(gemm_out);
+  }
+
+  encoder.set_input_array(in_strided);
+  encoder.set_input_array(gemm_wt);
+  encoder.set_output_array(gemm_out);
+
+  encoder.dispatch([in_strided_ptr = in_strided.data<float>(),
+                    gemm_wt_ptr = gemm_wt.data<float>(),
+                    gemm_out_ptr = gemm_out.data<float>(),
+                    strided_reshape = std::move(strided_reshape),
+                    O]() {
+    // Perform gemm
+    cblas_sgemm(
+        CblasRowMajor,
+        CblasNoTrans, // no trans A
+        CblasTrans, // transB
+        strided_reshape[0], // M
+        O, // N
+        strided_reshape[1], // K
+        1.0f, // alpha
+        in_strided_ptr,
+        strided_reshape[1], // lda
+        gemm_wt_ptr,
+        strided_reshape[1], // ldb
+        0.0f, // beta
+        gemm_out_ptr,
+        O // ldc
+    );
+  });
+
+  // Copy results if needed
+  if (out.dtype() != float32) {
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
+  }
+  encoder.add_temporaries(std::move(temps));
+}
+
 void explicit_gemm_conv_ND_cpu(
     const array& in,
     const array& wt,

mlx/backend/cpu/eig.cpp

@@ -46,6 +46,7 @@ void eig_impl(
     int info;
     {
       T work;
+      int iwork;
       geev<T>(
           &jobl,
           &jobr,

mlx/backend/cpu/gemms/bnns.cpp

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

mlx/backend/cpu/gemms/cblas.cpp

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

mlx/backend/cpu/lapack.h

@@ -47,7 +47,7 @@ INSTANTIATE_LAPACK_REAL(orgqr)
 INSTANTIATE_LAPACK_REAL(syevd)
 INSTANTIATE_LAPACK_REAL(geev)
 INSTANTIATE_LAPACK_REAL(potrf)
-INSTANTIATE_LAPACK_REAL(gesdd)
+INSTANTIATE_LAPACK_REAL(gesvdx)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)

mlx/backend/cpu/masked_mm.cpp

@@ -215,18 +215,18 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   encoder.set_input_array(a);
   encoder.set_input_array(b);
-  const void* a_mask_ptr = nullptr;
-  const void* b_mask_ptr = nullptr;
-  const void* out_mask_ptr = nullptr;
+  const void* a_mask_ptr;
+  const void* b_mask_ptr;
+  const void* out_mask_ptr;
   Shape a_mask_shape;
   Shape b_mask_shape;
   Shape out_mask_shape;
   Strides a_mask_strides;
   Strides b_mask_strides;
   Strides out_mask_strides;
-  bool a_mask_bool = false;
-  bool b_mask_bool = false;
-  bool out_mask_bool = false;
+  bool a_mask_bool;
+  bool b_mask_bool;
+  bool out_mask_bool;
   if (has_op_mask) {
     auto& a_mask = inputs[inputs.size() - 2];
     auto& b_mask = inputs[inputs.size() - 1];
@@ -423,6 +423,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& rhs_indices = inputs[3];
 
   auto batch_shape = get_batch_dims(out.shape());
+  int batch_ndim = batch_shape.size();
 
   auto batch_shape_A = get_batch_dims(a.shape());
   auto batch_strides_A = get_batch_dims(a.strides());

mlx/backend/cpu/matmul.cpp

@@ -91,6 +91,7 @@ void matmul_general(
   auto [b_transposed, ldb, b] = check_transpose(b_pre);
   size_t M = a.shape(-2);
   size_t N = b.shape(-1);
+  size_t K = a.shape(-1);
   if (M == 0 || N == 0) {
     return;
   }
@@ -107,9 +108,6 @@ void matmul_general(
   } else if (out.dtype() == float64) {
     matmul_dispatch<double>(
         a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
-  } else if (out.dtype() == complex64) {
-    matmul_dispatch<complex64_t>(
-        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else {
     throw std::runtime_error("[Matmul::eval_cpu] Invalid type.");
   }
@@ -130,6 +128,10 @@ void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
 }
 
 void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  if (out.dtype() != float32) {
+    throw std::runtime_error(
+        "[AddMM::eval_cpu] Currently only supports float32.");
+  }
   if (out.size() == 0) {
     out.set_data(allocator::malloc(out.nbytes()));
     return;

mlx/backend/cpu/primitives.cpp

@@ -333,7 +333,7 @@ void Reshape::eval_cpu(const std::vector<array>& inputs, array& out) {
 
 void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
   if (out.size() == 0) {
-    out.set_data(allocator::malloc(0));
+    out.set_data(nullptr);
     return;
   }
   auto& in = inputs[0];
@@ -361,7 +361,7 @@ void DynamicSliceUpdate::eval_cpu(
     const std::vector<array>& inputs,
     array& out) {
   if (out.size() == 0) {
-    out.set_data(allocator::malloc(0));
+    out.set_data(nullptr);
     return;
   }
 
@@ -396,7 +396,7 @@ void DynamicSliceUpdate::eval_cpu(
 void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   if (out.size() == 0) {
-    out.set_data(allocator::malloc(0));
+    out.set_data(nullptr);
     return;
   }
 

mlx/backend/cpu/quantized.cpp

@@ -1,11 +1,10 @@
 // Copyright © 2023 Apple Inc.
 
-#include "mlx/backend/common/unary.h"
+#include <cassert>
+
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
-#include "mlx/backend/cpu/unary.h"
-#include "mlx/backend/cpu/unary_ops.h"
 #include "mlx/fast_primitives.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
@@ -14,35 +13,6 @@ namespace mlx::core {
 
 namespace {
 
-const static float MXFP4_LUT[16] = {
-    +0.0f,
-    +0.5f,
-    +1.0f,
-    +1.5f,
-    +2.0f,
-    +3.0f,
-    +4.0f,
-    +6.0f,
-    -0.0f,
-    -0.5f,
-    -1.0f,
-    -1.5f,
-    -2.0f,
-    -3.0f,
-    -4.0f,
-    -6.0f};
-
-template <typename T>
-static inline T dequantize_scale(uint8_t s) {
-  using FOrI = union {
-    bfloat16_t f;
-    uint16_t i;
-  };
-  FOrI out;
-  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
-  return static_cast<T>(out.f);
-}
-
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
   return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -437,229 +407,6 @@ void _qmm_dispatch(
   }
 }
 
-template <typename T>
-void mxfp4_qmm(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = get_pack_factor(4, 8);
-  constexpr int packs_in_group = group_size / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint8_t* w_local = (const uint8_t*)w;
-    const uint8_t* scales_local = scales;
-
-    std::fill(result, result + N, 0);
-
-    for (int k = 0; k < K; k++) {
-      T* result_local = result;
-      T xi = *x++;
-
-      for (int n = 0; n < N; n += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
-        for (int ng = 0; ng < packs_in_group; ng++) {
-          uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-          for (int p = 0; p < pack_factor; p++) {
-            (*result_local++) +=
-                xi * scale * static_cast<T>(MXFP4_LUT[wi & 0xf]);
-            wi >>= 4;
-          }
-        }
-      }
-    }
-
-    result += N;
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_t(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = get_pack_factor(4, 8);
-  constexpr int packs_in_group = group_size / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint8_t* w_local = (const uint8_t*)w;
-    const uint8_t* scales_local = scales;
-
-    for (int n = 0; n < N; n++) {
-      const T* x_local = x;
-      T sum = 0;
-      for (int k = 0; k < K; k += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
-
-        T gsum = 0;
-        for (int kw = 0; kw < packs_in_group; kw++) {
-          uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-          for (int p = 0; p < pack_factor; p++) {
-            gsum += (*x_local++) * static_cast<T>(MXFP4_LUT[wi & 0xf]);
-            wi >>= 4;
-          }
-        }
-        sum += scale * gsum;
-      }
-      *result = sum;
-      result++;
-    }
-
-    x += K;
-  }
-}
-
-template <int S>
-simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
-  if constexpr (S == 8) {
-    constexpr std::array<uint32_t, 8> shifts_ = {{0, 4, 8, 12, 16, 20, 24, 28}};
-    auto shifts(*(simd::Simd<uint32_t, S>*)&shifts_);
-    auto wi = simd::Simd<uint32_t, S>(*w);
-    wi = wi >> shifts;
-    wi = wi & 0xf;
-    simd::Simd<float, S> w_out;
-    for (int i = 0; i < S; ++i) {
-      w_out[i] = MXFP4_LUT[wi[i]];
-    }
-    return w_out;
-  } else {
-    // Appease compiler.. but should never get here
-    throw std::runtime_error("Unsupported combination for simd qmm.");
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_t_simd(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = 32 / 4;
-  constexpr int packs_in_group = group_size / pack_factor;
-  constexpr int S = simd::max_size<T>;
-  static_assert(
-      S % pack_factor == 0, "SIMD size must be divisible by pack factor");
-  constexpr int packs_per_simd = S / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint32_t* w_local = w;
-    const uint8_t* scales_local = scales;
-
-    for (int n = 0; n < N; n++) {
-      simd::Simd<float, S> acc(0);
-      auto x_local = x;
-      for (int k = 0; k < K; k += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
-
-        simd::Simd<float, S> g_acc(0);
-        for (int kw = 0; kw < packs_in_group; kw += packs_per_simd) {
-          // Extract bits
-          auto wf = mxfp4_extract_bits_simd<S>(w_local);
-          w_local += packs_per_simd;
-          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
-          g_acc = g_acc + x_simd * wf;
-          x_local += S;
-        }
-        acc = acc + scale * g_acc;
-      }
-
-      *result = T(simd::sum(acc));
-      result++;
-    }
-    x += K;
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_dispatch_transpose(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K,
-    bool transposed_w) {
-  if (transposed_w) {
-    // the simd size must be a multiple of the number of elements per word
-    if constexpr (simd::max_size<T> % 8 == 0) {
-      mxfp4_qmm_t_simd<T>(result, x, w, scales, M, N, K);
-    } else {
-      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
-    }
-  } else {
-    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_dispatch_typed(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    bool transposed_w) {
-  int K = x.shape(-1);
-  int M = x.ndim() > 1 ? x.shape(-2) : 1;
-  int N = out.shape(-1);
-  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
-  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
-  int batch_size = x.size() / (K * M);
-
-  auto out_ptr = out.data<T>();
-  auto x_ptr = x.data<T>();
-  auto w_ptr = w.data<uint32_t>();
-  auto scales_ptr = scales.data<uint8_t>();
-  for (int i = 0; i < batch_size; i++) {
-    mxfp4_qmm_dispatch_transpose<T>(
-        out_ptr + i * M * N,
-        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
-        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
-        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
-        M,
-        N,
-        K,
-        transposed_w);
-  }
-}
-
-void mxfp4_qmm_dispatch(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    bool transposed_w) {
-  switch (x.dtype()) {
-    case bfloat16:
-      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
-      break;
-    case float16:
-      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
-      break;
-    case float32:
-      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
-      break;
-    default:
-      throw std::invalid_argument(
-          "[quantized_matmul] only floating types are supported");
-  }
-}
-
 template <typename T>
 void _bs_qmm_dispatch_typed(
     array& out,
@@ -766,106 +513,41 @@ void _bs_qmm_dispatch(
   }
 }
 
-template <typename T>
-void mxfp4_bs_qmm_dispatch_typed(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& lhs_indices,
-    const array& rhs_indices,
-    bool transposed_w) {
-  int K = x.shape(-1);
-  int M = x.shape(-2);
-  int N = out.shape(-1);
-
-  int w_els = w.shape(-1) * w.shape(-2);
-  int g_els = scales.shape(-1) * scales.shape(-2);
-
-  auto out_ptr = out.data<T>();
-  auto x_ptr = x.data<T>();
-  auto w_ptr = w.data<uint32_t>();
-  auto scales_ptr = scales.data<uint8_t>();
-  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
-  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
-
-  for (int i = 0; i < lhs_indices.size(); i++) {
-    int x_idx = lhs_indices_ptr[elem_to_loc(
-        i, lhs_indices.shape(), lhs_indices.strides())];
-    int w_idx = rhs_indices_ptr[elem_to_loc(
-        i, rhs_indices.shape(), rhs_indices.strides())];
-    mxfp4_qmm_dispatch_transpose<T>(
-        out_ptr + i * M * N,
-        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
-        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
-        scales_ptr +
-            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
-        M,
-        N,
-        K,
-        transposed_w);
-  }
-}
-
-void mxfp4_bs_qmm_dispatch(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& lhs_indices,
-    const array& rhs_indices,
-    bool transposed_w) {
-  switch (x.dtype()) {
-    case float32:
-      mxfp4_bs_qmm_dispatch_typed<float>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    case float16:
-      mxfp4_bs_qmm_dispatch_typed<float16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    case bfloat16:
-      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    default:
-      throw std::invalid_argument(
-          "[quantized_matmul] only floating types are supported");
-  }
-}
-
 } // namespace
 
 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 4);
+
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
+  auto& biases_pre = inputs[3];
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
+  std::vector<array> temps;
+  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, arr_cpy, CopyType::General, s);
-      encoder.add_temporary(arr_cpy);
-      return arr_cpy;
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy_cpu(arr, temps.back(), CopyType::General, s);
+      return temps.back();
     }
   };
 
   auto x = ensure_row_contiguous(x_pre);
   auto w = ensure_row_contiguous(w_pre);
   auto scales = ensure_row_contiguous(scales_pre);
+  auto biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_output_array(out);
-  if (mode_ == QuantizationMode::Affine) {
-    auto biases = ensure_row_contiguous(inputs[3]);
   encoder.set_input_array(biases);
+  encoder.set_output_array(out);
   encoder.dispatch([out = array::unsafe_weak_copy(out),
                     x = array::unsafe_weak_copy(x),
                     w = array::unsafe_weak_copy(w),
@@ -876,54 +558,48 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
                     transpose_ = transpose_]() mutable {
     _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
   });
-  } else {
-    encoder.dispatch([out = array::unsafe_weak_copy(out),
-                      x = array::unsafe_weak_copy(x),
-                      w = array::unsafe_weak_copy(w),
-                      scales = array::unsafe_weak_copy(scales),
-                      transpose_ = transpose_]() mutable {
-      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
-    });
-  }
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 6);
+
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& lhs_indices = inputs[inputs.size() - 2];
-  auto& rhs_indices = inputs[inputs.size() - 1];
+  auto& biases_pre = inputs[3];
+  auto& lhs_indices = inputs[4];
+  auto& rhs_indices = inputs[5];
 
-  auto& encoder = cpu::get_command_encoder(stream());
+  std::vector<array> temps;
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &encoder](const array& arr) {
+                                          &temps](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, arr_cpy, CopyType::General, s);
-      encoder.add_temporary(arr_cpy);
-      return arr_cpy;
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy_cpu(arr, temps.back(), CopyType::General, s);
+      return temps.back();
     }
   };
 
   auto x = ensure_row_contiguous_last_dims(x_pre);
   auto w = ensure_row_contiguous_last_dims(w_pre);
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
+  auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
+  encoder.set_input_array(biases);
   encoder.set_input_array(lhs_indices);
   encoder.set_input_array(rhs_indices);
   encoder.set_output_array(out);
-  if (mode_ == QuantizationMode::Affine) {
-    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
-    encoder.set_input_array(biases);
   encoder.dispatch([out = array::unsafe_weak_copy(out),
                     x = array::unsafe_weak_copy(x),
                     w = array::unsafe_weak_copy(w),
@@ -946,18 +622,6 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
         bits_,
         transpose_);
   });
-  } else {
-    encoder.dispatch([out = array::unsafe_weak_copy(out),
-                      x = array::unsafe_weak_copy(x),
-                      w = array::unsafe_weak_copy(w),
-                      scales = array::unsafe_weak_copy(scales),
-                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
-                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
-                      transpose_ = transpose_]() mutable {
-      mxfp4_bs_qmm_dispatch(
-          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
-    });
-  }
 }
 
 template <typename T, typename U>
@@ -1041,7 +705,7 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::Quantize::eval_cpu(
+void fast::AffineQuantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto ensure_row_contiguous = [s = stream()](const array& arr) {
@@ -1100,47 +764,7 @@ void fast::Quantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
-    }
-  });
-}
-
-void fast::ConvertFP8::eval_cpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  auto& in = inputs[0];
-  auto& out = outputs[0];
-  set_unary_output_data(in, out);
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.set_input_array(in);
-  encoder.set_output_array(out);
-  encoder.dispatch([in = array::unsafe_weak_copy(in),
-                    out = array::unsafe_weak_copy(out),
-                    to_fp8 = to_fp8_]() mutable {
-    if (to_fp8) {
-      switch (in.dtype()) {
-        case float16:
-          unary_op<float16_t, uint8_t>(in, out, detail::ToFP8());
-          break;
-        case bfloat16:
-          unary_op<bfloat16_t, uint8_t>(in, out, detail::ToFP8());
-          break;
-        default:
-          unary_op<float, uint8_t>(in, out, detail::ToFP8());
-          break;
-      }
-    } else {
-      switch (out.dtype()) {
-        case float16:
-          unary_op<uint8_t, float16_t>(in, out, detail::FromFP8());
-          break;
-        case bfloat16:
-          unary_op<uint8_t, bfloat16_t>(in, out, detail::FromFP8());
-          break;
-        default:
-          unary_op<uint8_t, float>(in, out, detail::FromFP8());
-          break;
-      }
+          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
     }
   });
 }

mlx/backend/cpu/simd/accelerate_simd.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include <arm_neon.h>
 #include <simd/math.h>
 #include <simd/vector.h>
 
@@ -10,7 +9,7 @@
 
 #include "mlx/backend/cpu/simd/base_simd.h"
 
-// There seems to be a bug in simd/base_simd.h
+// There seems to be a bug in sims/base.h
 // __XROS_2_0 is not defined, the expression evaluates
 // to true instead of false setting the SIMD library
 // higher than it should be even on macOS < 15
@@ -201,15 +200,6 @@ SIMD_DEFAULT_COMPARISONS(<=)
 SIMD_DEFAULT_COMPARISONS(==)
 SIMD_DEFAULT_COMPARISONS(!=)
 
-template <typename T, int N>
-Simd<T, N> clz(Simd<T, N> x) {
-  auto a = *(uint32x4_t*)(&x);
-  auto b = *((uint32x4_t*)(&x) + 1);
-  a = vclzq_u32(a);
-  b = vclzq_u32(b);
-  return asd::make_uint8(a, b);
-}
-
 template <typename T, int N>
 Simd<T, N> atan2(Simd<T, N> a, Simd<T, N> b) {
   return asd::atan2(a.value, b.value);
@@ -244,7 +234,6 @@ Simd<T, N> remainder(Simd<T, N> a, Simd<T, N> b) {
 
 template <typename MaskT, typename T1, typename T2, int N>
 Simd<T1, N> select(Simd<MaskT, N> mask, Simd<T1, N> x, Simd<T2, N> y) {
-  static_assert(std::is_same_v<MaskT, bool>);
   if constexpr (sizeof(T1) == 1) {
     return asd::bitselect(y.value, x.value, asd::convert<char>(mask.value));
   } else if constexpr (sizeof(T1) == 2) {
@@ -262,13 +251,9 @@ Simd<T, N> pow(Simd<T, N> base, Simd<T, N> exp) {
     return asd::pow(base.value, exp.value);
   } else {
     Simd<T, N> res = 1;
-    // Raising an integer to a negative power is undefined
-    if (any(exp < 0)) {
-      return 0;
-    }
-    while (any(exp > 0)) {
-      res = select((exp & 1) != 0, res * base, res);
-      base = select(exp > 0, base * base, base);
+    while (any(exp)) {
+      res = select(exp & 1, res * base, res);
+      base = select(exp, base * base, base);
       exp = exp >> 1;
     }
     return res;

mlx/backend/cpu/simd/base_simd.h

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

mlx/backend/cpu/sort.cpp

@@ -15,18 +15,6 @@ namespace mlx::core {
 
 namespace {
 
-// NaN-aware comparator that places NaNs at the end
-template <typename T>
-bool nan_aware_less(T a, T b) {
-  if constexpr (std::is_floating_point_v<T> || std::is_same_v<T, complex64_t>) {
-    if (std::isnan(a))
-      return false;
-    if (std::isnan(b))
-      return true;
-  }
-  return a < b;
-}
-
 template <typename T>
 struct StridedIterator {
   using iterator_category = std::random_access_iterator_tag;
@@ -39,7 +27,7 @@ struct StridedIterator {
   StridedIterator() = default;
 
   explicit StridedIterator(T* ptr, int64_t stride, difference_type offset = 0)
-      : stride_(stride), ptr_(ptr + offset * stride) {}
+      : ptr_(ptr + offset * stride), stride_(stride) {}
 
   explicit StridedIterator(array& arr, int axis, difference_type offset = 0)
       : StridedIterator(arr.data<T>(), arr.strides()[axis], offset) {}
@@ -142,7 +130,7 @@ void sort(array& out, int axis) {
     StridedIterator st(data_ptr, axis_stride, 0);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::stable_sort(st, ed, nan_aware_less<T>);
+    std::stable_sort(st, ed);
     src_it.step();
   }
 }
@@ -196,15 +184,6 @@ void argsort(const array& in, array& out, int axis) {
     std::stable_sort(st, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
-
-      // Handle NaNs (place them at the end)
-      if (std::is_floating_point<T>::value) {
-        if (std::isnan(v1))
-          return false;
-        if (std::isnan(v2))
-          return true;
-      }
-
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }
@@ -240,7 +219,7 @@ void partition(array& out, int axis, int kth) {
     StridedIterator md(data_ptr, axis_stride, kth);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::nth_element(st, md, ed, nan_aware_less<T>);
+    std::nth_element(st, md, ed);
   }
 }
 
@@ -297,15 +276,6 @@ void argpartition(const array& in, array& out, int axis, int kth) {
     std::nth_element(st, md, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
-
-      // Handle NaNs (place them at the end)
-      if (std::is_floating_point<T>::value) {
-        if (std::isnan(v1))
-          return false;
-        if (std::isnan(v2))
-          return true;
-      }
-
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }

mlx/backend/cpu/svd.cpp

@@ -81,26 +81,40 @@ void svd_impl(
     // Vᵀ of shape N x N. (M x M in lapack).
     const int ldvt = M;
 
-    auto jobz = (u_ptr) ? "A" : "N";
+    auto job_u = (u_ptr) ? "V" : "N";
+    auto job_vt = (u_ptr) ? "V" : "N";
+    static constexpr auto range = "A";
 
+    // Will contain the number of singular values after the call has returned.
+    int ns = 0;
     T workspace_dimension = 0;
 
     // Will contain the indices of eigenvectors that failed to converge (not
     // used here but required by lapack).
-    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
+    auto iwork = array::Data{allocator::malloc(sizeof(int) * 12 * K)};
 
     static const int lwork_query = -1;
 
+    static const int ignored_int = 0;
+    static const T ignored_float = 0;
+
     int info;
 
     // Compute workspace size.
-    gesdd<T>(
-        /* jobz = */ jobz,
+    gesvdx<T>(
+        /* jobu = */ job_u,
+        /* jobvt = */ job_vt,
+        /* range = */ range,
         // M and N are swapped since lapack expects column-major.
         /* m = */ &N,
         /* n = */ &M,
         /* a = */ nullptr,
         /* lda = */ &lda,
+        /* vl = */ &ignored_float,
+        /* vu = */ &ignored_float,
+        /* il = */ &ignored_int,
+        /* iu = */ &ignored_int,
+        /* ns = */ &ns,
         /* s = */ nullptr,
         /* u = */ nullptr,
         /* ldu = */ &ldu,
@@ -122,13 +136,20 @@ void svd_impl(
 
     // Loop over matrices.
     for (int i = 0; i < num_matrices; i++) {
-      gesdd<T>(
-          /* jobz = */ jobz,
+      gesvdx<T>(
+          /* jobu = */ job_u,
+          /* jobvt = */ job_vt,
+          /* range = */ range,
           // M and N are swapped since lapack expects column-major.
           /* m = */ &N,
           /* n = */ &M,
           /* a = */ in_ptr + M * N * i,
           /* lda = */ &lda,
+          /* vl = */ &ignored_float,
+          /* vu = */ &ignored_float,
+          /* il = */ &ignored_int,
+          /* iu = */ &ignored_int,
+          /* ns = */ &ns,
           /* 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,
@@ -146,6 +167,13 @@ void svd_impl(
         ss << "svd_impl: sgesvdx_ failed with code " << info;
         throw std::runtime_error(ss.str());
       }
+
+      if (ns != K) {
+        std::stringstream ss;
+        ss << "svd_impl: expected " << K << " singular values, but " << ns
+           << " were computed.";
+        throw std::runtime_error(ss.str());
+      }
     }
   });
   encoder.add_temporary(in);

mlx/backend/cpu/unary.h

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

mlx/backend/cpu/unary_ops.h

@@ -77,8 +77,7 @@ struct Real {
 struct Sigmoid {
   template <int N, typename T>
   Simd<T, N> operator()(Simd<T, N> x) {
-    auto y = 1.0f / (1.0f + simd::exp(simd::abs(x)));
-    return simd::select(x < Simd<T, N>{0}, y, Simd<T, N>{1} - y);
+    return 1.0f / (1.0f + simd::exp(-x));
   }
   SINGLE()
 };
@@ -108,73 +107,4 @@ struct Square {
   SINGLE()
 };
 
-template <int N>
-Simd<float, N> fp32_from_bits(Simd<uint32_t, N> x) {
-  return *(Simd<float, N>*)(&x);
-}
-template <int N>
-Simd<uint32_t, N> fp32_to_bits(Simd<float, N> x) {
-  return *(Simd<uint32_t, N>*)(&x);
-}
-
-struct ToFP8 {
-  template <typename T, int N>
-  Simd<uint8_t, N> operator()(Simd<T, N> f) {
-    uint32_t fp8_max = 543 << 21;
-    auto denorm_mask = Simd<uint32_t, N>(141 << 23);
-    Simd<uint32_t, N> f_bits;
-    Simd<float, N> f32 = f;
-    f_bits = fp32_to_bits(f32);
-    Simd<uint8_t, N> result = 0u;
-    auto sign = f_bits & 0x80000000;
-    f_bits = f_bits ^ sign;
-
-    auto f_bits_low =
-        fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask));
-    auto result_low = Simd<uint8_t, N>(f_bits_low - denorm_mask);
-
-    auto mant_odd = Simd<uint8_t, N>((f_bits >> 20) & 1);
-    auto f_bits_high = f_bits + (((uint32_t)(7 - 127) << 23) + 0x7FFFF);
-    f_bits_high = f_bits_high + Simd<uint32_t, N>(mant_odd);
-
-    auto result_high = Simd<uint8_t, N>(f_bits_high >> 20);
-    result = select(f_bits < (121 << 23), result_low, result_high);
-
-    auto result_sat = Simd<uint8_t, N>(0x7E);
-    result = select(f_bits >= fp8_max, result_sat, result);
-    return result | Simd<uint8_t, N>(sign >> 24);
-  }
-
-  template <typename T>
-  uint8_t operator()(T x) {
-    return (*this)(Simd<T, 1>(x)).value;
-  }
-};
-
-struct FromFP8 {
-  template <int N>
-  Simd<float, N> operator()(Simd<uint8_t, N> x) {
-    auto w = Simd<uint32_t, N>(x) << 24;
-    auto sign = w & 0x80000000;
-    auto nonsign = w & 0x7FFFFFFF;
-
-    auto renorm_shift = clz(nonsign);
-    renorm_shift = simd::select(
-        renorm_shift > Simd<uint32_t, N>{4},
-        renorm_shift - Simd<uint32_t, N>{4},
-        Simd<uint32_t, N>{0});
-
-    Simd<int32_t, N> inf_nan_mask =
-        (Simd<int32_t, N>(nonsign + 0x01000000) >> 8) & 0x7F800000;
-    auto zero_mask = Simd<int32_t, N>(nonsign - 1) >> 31;
-    auto result = sign |
-        ((((nonsign << renorm_shift >> 4) + ((0x78 - renorm_shift) << 23)) |
-          inf_nan_mask) &
-         ~zero_mask);
-    return fp32_from_bits(result);
-  }
-  float operator()(uint8_t x) {
-    return (*this)(Simd<uint8_t, 1>(x)).value;
-  }
-};
 } // namespace mlx::core::detail

mlx/backend/cuda/CMakeLists.txt

@@ -8,6 +8,7 @@ target_sources(
   PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/arange.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cu
@@ -16,23 +17,18 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_dynamic.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_input.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/conv/gemm_conv.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/conv/gemm_grouped_conv.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/cudnn_utils.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/custom_kernel.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/gemv.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/steel_gemm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/jit_module.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/kernel_utils.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/layer_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/logsumexp.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
@@ -50,21 +46,12 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/sort.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/fp_quantize.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/convert_fp8.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/binary)
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/unary)
-
-# fp4 is not available on < 12.8
-if(CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 12.8.0)
-  target_include_directories(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/quantized/)
-endif()
-
 if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.9.0)
   target_sources(
     mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_9.cu)
@@ -162,7 +149,7 @@ 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 v1.12.1
   GIT_SHALLOW TRUE
   EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)
@@ -178,6 +165,7 @@ target_link_libraries(mlx PRIVATE CUDNN::cudnn_all)
 # Suppress nvcc warnings on MLX headers.
 target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
                                    --diag_suppress=997>)
+
 # Install CCCL headers for JIT.
 install(DIRECTORY ${cccl_SOURCE_DIR}/include/cuda
         DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/cccl)

mlx/backend/cuda/allocator.cpp

@@ -1,7 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/allocator.h"
-#include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/utils.h"
 #include "mlx/utils.h"
 
@@ -31,20 +30,8 @@ SmallSizePool::SmallSizePool() {
   next_free_ = buffer_;
 
   CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
-
-  int device_count = 0;
-  CHECK_CUDA_ERROR(cudaGetDeviceCount(&device_count));
-  for (int i = 0; i < device_count; ++i) {
-#if CUDART_VERSION >= 13000
-    cudaMemLocation loc;
-    loc.type = cudaMemLocationTypeDevice;
-    loc.id = i;
-#else
-    int loc = i;
-#endif // CUDART_VERSION >= 13000
   CHECK_CUDA_ERROR(
-        cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetAccessedBy, loc));
-  }
+      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, 0));
 
   auto curr = next_free_;
   for (size_t i = 1; i < num_blocks; ++i) {
@@ -68,7 +55,6 @@ CudaBuffer* SmallSizePool::malloc() {
   next_free_ = next_free_->next;
   b->buf.data = static_cast<char*>(data_) + i * small_block_size;
   b->buf.size = small_block_size;
-  b->buf.device = -1;
   return &b->buf;
 }
 
@@ -90,22 +76,14 @@ CudaAllocator::CudaAllocator()
           page_size,
           [](CudaBuffer* buf) { return buf->size; },
           [this](CudaBuffer* buf) { cuda_free(buf); }) {
+  // TODO: Set memory limit for multi-device.
   size_t free, total;
   CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
-  memory_limit_ = total * 0.95;
+  memory_limit_ = total * 0.8;
   max_pool_size_ = memory_limit_;
-
-  int device_count = 0;
-  CHECK_CUDA_ERROR(cudaGetDeviceCount(&device_count));
-  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);
-  }
 }
 
-Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
+Buffer CudaAllocator::malloc(size_t size) {
   // Find available buffer from cache.
   auto orig_size = size;
   std::unique_lock lock(mutex_);
@@ -132,17 +110,8 @@ 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;
-      if (device == -1) {
-        err = cudaMallocManaged(&buf->data, size);
-      } else {
-        err = cudaMallocAsync(&buf->data, size, stream);
-      }
+      buf = new CudaBuffer{nullptr, size};
+      cudaError_t err = cudaMallocManaged(&buf->data, size);
       if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
         throw std::runtime_error(fmt::format(
             "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
@@ -160,14 +129,6 @@ Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
   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);
-}
-
 void CudaAllocator::free(Buffer buffer) {
   auto* buf = static_cast<CudaBuffer*>(buffer.ptr());
   if (!buf) {
@@ -195,12 +156,8 @@ size_t CudaAllocator::size(Buffer buffer) const {
 void CudaAllocator::cuda_free(CudaBuffer* buf) {
   if (scalar_pool_.in_pool(buf)) {
     scalar_pool_.free(buf);
-  } else {
-    if (buf->device >= 0) {
-      cudaFreeAsync(buf->data, free_streams_[buf->device]);
   } else {
     cudaFree(buf->data);
-    }
     delete buf;
   }
 }
@@ -251,16 +208,6 @@ CudaAllocator& allocator() {
   return *allocator_;
 }
 
-Buffer malloc_async(size_t size, cudaStream_t stream) {
-  auto buffer = allocator().malloc_async(size, stream);
-  if (size && !buffer.ptr()) {
-    std::ostringstream msg;
-    msg << "[malloc_async] Unable to allocate " << size << " bytes.";
-    throw std::runtime_error(msg.str());
-  }
-  return buffer;
-}
-
 } // namespace cu
 
 namespace allocator {
@@ -273,19 +220,7 @@ void* Buffer::raw_ptr() {
   if (!ptr_) {
     return nullptr;
   }
-  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;
-  }
-  return cbuf.data;
+  return static_cast<cu::CudaBuffer*>(ptr_)->data;
 }
 
 } // namespace allocator

mlx/backend/cuda/allocator.h

@@ -4,9 +4,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/buffer_cache.h"
-#include "mlx/backend/cuda/cuda_utils.h"
 
-#include <cuda_runtime.h>
 #include <mutex>
 #include <set>
 #include <utility>
@@ -19,7 +17,6 @@ using allocator::Buffer;
 struct CudaBuffer {
   void* data;
   size_t size;
-  int device; // -1 for managed
 };
 
 class SmallSizePool {
@@ -48,7 +45,6 @@ class SmallSizePool {
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
-  Buffer malloc_async(size_t size, cudaStream_t stream);
   void free(Buffer buffer) override;
   size_t size(Buffer buffer) const override;
 
@@ -62,7 +58,6 @@ class CudaAllocator : public allocator::Allocator {
   void clear_cache();
 
  private:
-  Buffer malloc_impl(size_t size, cudaStream_t stream);
   void cuda_free(CudaBuffer* buf);
 
   CudaAllocator();
@@ -74,12 +69,9 @@ class CudaAllocator : public allocator::Allocator {
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
-  std::vector<cudaStream_t> free_streams_;
   SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();
 
-Buffer malloc_async(size_t size, cudaStream_t stream);
-
 } // namespace mlx::core::cu

mlx/backend/cuda/arange.cu

@@ -6,33 +6,23 @@
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
-#include <cooperative_groups.h>
 #include <nvtx3/nvtx3.hpp>
+#include <thrust/device_ptr.h>
+#include <thrust/transform.h>
 
 namespace mlx::core {
 
 namespace cu {
 
-namespace cg = cooperative_groups;
+template <typename T>
+struct Arange {
+  const T start;
+  const T step;
 
-template <typename T, typename IdxT, int N_WRITES>
-__global__ void arange(T* out, IdxT size, T start, T step) {
-  IdxT index = cg::this_grid().thread_rank();
-
-  if ((index + 1) * N_WRITES > size) {
-    for (IdxT i = index * N_WRITES; i < size; ++i) {
-      out[i] = start + i * step;
-    }
-  } else {
-    AlignedVector<T, N_WRITES> out_vec;
-#pragma unroll
-    for (int i = 0; i < N_WRITES; ++i) {
-      out_vec[i] = start + (index * N_WRITES + i) * step;
-    }
-
-    store_vector<N_WRITES>(out, index, out_vec);
-  }
+  __device__ T operator()(uint32_t i) const {
+    return start + i * step;
   }
+};
 
 } // namespace cu
 
@@ -41,27 +31,24 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   if (out.size() == 0) {
     return;
   }
+  out.set_data(allocator::malloc(out.nbytes()));
+
   auto& encoder = cu::get_command_encoder(stream());
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
   encoder.set_output_array(out);
 
+  auto capture = encoder.capture_context();
   dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
     using CTYPE = MLX_GET_TYPE(type_tag);
     using OutType = cuda_type_t<CTYPE>;
-    constexpr int N_WRITES = 16 / sizeof(OutType);
-    dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
-      using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-      auto [num_blocks, block_dims] = get_launch_args(out, large(), N_WRITES);
-      encoder.add_kernel_node(
-          cu::arange<OutType, IdxT, N_WRITES>,
-          num_blocks,
-          block_dims,
-          0,
-          gpu_ptr<OutType>(out),
-          out.data_size(),
-          static_cast<CTYPE>(start_),
-          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));
-    });
+    CTYPE step =
+        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
+    thrust::transform(
+        cu::thrust_policy(encoder.stream()),
+        thrust::counting_iterator<uint32_t>(0),
+        thrust::counting_iterator<uint32_t>(out.data_size()),
+        thrust::device_pointer_cast(out.data<OutType>()),
+        cu::Arange<OutType>{
+            static_cast<OutType>(start_), static_cast<OutType>(step)});
   });
 }
 

mlx/backend/cuda/arg_reduce.cu

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

mlx/backend/cuda/binary.cu

@@ -99,89 +99,39 @@ __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
   }
 }
 
-template <
-    typename Op,
-    typename In,
-    typename Out,
-    typename IdxT,
-    int NDIM,
-    int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
 __global__ void binary_g_nd(
     const In* a,
     const In* b,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[NDIM - 1];
-  auto a_stride_x = a_strides[NDIM - 1];
-  auto b_stride_x = b_strides[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
     auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
-      index_rest * shape_x, shape.data(), a_strides.data(), b_strides.data());
-  auto a_vec =
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(a_vec[i], b_vec[i]);
+        index, shape.data(), a_strides.data(), b_strides.data());
+    out[index] = Op{}(a[a_idx], b[b_idx]);
   }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT>
 __global__ void binary_g(
     const In* a,
     const In* b,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto a_stride_x = a_strides[ndim - 1];
-  auto b_stride_x = b_strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
     auto [a_idx, b_idx] = elem_to_loc(
-      index_rest * shape_x,
-      shape.data(),
-      a_strides.data(),
-      b_strides.data(),
-      ndim);
-  auto a_vec =
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(a_vec[i], b_vec[i]);
+        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
+    out[index] = Op{}(a[a_idx], b[b_idx]);
   }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -259,61 +209,39 @@ void binary_op_gpu_inplace(
                 auto& a_strides = strides[0];
                 auto& b_strides = strides[1];
                 int ndim = shape.size();
-                int work_per_thread = 1;
-                auto dim0 = ndim > 0 ? shape.back() : 1;
-                auto rest = out.size() / dim0;
-                if (dim0 >= 4) {
-                  work_per_thread = 4;
-                }
-                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-                auto block_dims = get_block_dims(dim0, rest, 1);
-                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::binary_g_nd<
-                        Op,
-                        InType,
-                        OutType,
-                        IdxT,
-                        dims_constant(),
-                        1>;
-                    if (work_per_thread == 4) {
-                      kernel = cu::binary_g_nd<
-                          Op,
-                          InType,
-                          OutType,
-                          IdxT,
-                          dims_constant(),
-                          4>;
-                    }
+                    auto [num_blocks, block_dims] =
+                        get_launch_args(out, large());
                     encoder.add_kernel_node(
-                        kernel,
-                        {num_blocks_x, num_blocks_y},
+                        cu::binary_g_nd<
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
+                        num_blocks,
                         block_dims,
                         0,
-                        gpu_ptr<InType>(a),
-                        gpu_ptr<InType>(b),
-                        gpu_ptr<OutType>(out),
-                        rest,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out.data<OutType>(),
+                        out.size(),
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT, 1>;
-                  if (work_per_thread == 4) {
-                    kernel = cu::binary_g<Op, InType, OutType, IdxT, 4>;
-                  }
+                  auto [num_blocks, block_dims] = get_launch_args(out, large());
                   encoder.add_kernel_node(
-                      kernel,
-                      {num_blocks_x, num_blocks_y},
+                      cu::binary_g<Op, InType, OutType, IdxT>,
+                      num_blocks,
                       block_dims,
                       0,
-                      gpu_ptr<InType>(a),
-                      gpu_ptr<InType>(b),
-                      gpu_ptr<OutType>(out),
-                      rest,
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out.data<OutType>(),
+                      out.size(),
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),
@@ -339,9 +267,9 @@ void binary_op_gpu_inplace(
                 num_blocks,
                 block_dims,
                 0,
-                gpu_ptr<InType>(a),
-                gpu_ptr<InType>(b),
-                gpu_ptr<OutType>(out),
+                a.data<InType>(),
+                b.data<InType>(),
+                out.data<OutType>(),
                 out.data_size());
           });
         }
@@ -365,11 +293,7 @@ void binary_op_gpu(
   auto& a = inputs[0];
   auto& b = inputs[1];
   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);
   binary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 
@@ -380,4 +304,54 @@ void binary_op_gpu(
     binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
   }
 
+BINARY_GPU(Add)
+BINARY_GPU(ArcTan2)
+BINARY_GPU(Divide)
+BINARY_GPU(Remainder)
+BINARY_GPU(Greater)
+BINARY_GPU(GreaterEqual)
+BINARY_GPU(Less)
+BINARY_GPU(LessEqual)
+BINARY_GPU(LogicalAnd)
+BINARY_GPU(LogicalOr)
+BINARY_GPU(LogAddExp)
+BINARY_GPU(Maximum)
+BINARY_GPU(Minimum)
+BINARY_GPU(Multiply)
+BINARY_GPU(NotEqual)
+BINARY_GPU(Power)
+BINARY_GPU(Subtract)
+
+void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Equal::eval_gpu");
+  auto& s = out.primitive().stream();
+  if (equal_nan_) {
+    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
+  } else {
+    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
+  }
+}
+
+void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
+  auto& s = out.primitive().stream();
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Or:
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Xor:
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::RightShift:
+      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
+      break;
+  }
+}
+
 } // namespace mlx::core

mlx/backend/cuda/binary/CMakeLists.txt

@@ -1,21 +0,0 @@
-target_sources(
-  mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/add.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctan2.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/bitwise_binary.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/divide.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_and.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_or.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log_add_exp.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/minimum.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/maximum.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/multiply.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/power.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/remainder.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/not_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/subtract.cu)

mlx/backend/cuda/binary/add.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Add)
-} // namespace mlx::core

mlx/backend/cuda/binary/arctan2.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(ArcTan2)
-} // namespace mlx::core

mlx/backend/cuda/binary/bitwise_binary.cu

@@ -1,27 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
-  auto& s = out.primitive().stream();
-  switch (op_) {
-    case BitwiseBinary::And:
-      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
-      break;
-  }
-}
-} // namespace mlx::core

mlx/backend/cuda/binary/divide.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Divide)
-} // namespace mlx::core

mlx/backend/cuda/binary/equal.cu

@@ -1,15 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Equal::eval_gpu");
-  auto& s = out.primitive().stream();
-  if (equal_nan_) {
-    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
-  } else {
-    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
-  }
-}
-} // namespace mlx::core

mlx/backend/cuda/binary/greater.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Greater)
-} // namespace mlx::core

mlx/backend/cuda/binary/greater_equal.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(GreaterEqual)
-} // namespace mlx::core

mlx/backend/cuda/binary/less.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Less)
-} // namespace mlx::core

mlx/backend/cuda/binary/less_equal.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LessEqual)
-} // namespace mlx::core

mlx/backend/cuda/binary/log_add_exp.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogAddExp)
-} // namespace mlx::core

mlx/backend/cuda/binary/logical_and.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogicalAnd)
-} // namespace mlx::core

mlx/backend/cuda/binary/logical_or.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogicalOr)
-} // namespace mlx::core

mlx/backend/cuda/binary/maximum.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Maximum)
-} // namespace mlx::core

mlx/backend/cuda/binary/minimum.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Minimum)
-} // namespace mlx::core

mlx/backend/cuda/binary/multiply.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Multiply)
-} // namespace mlx::core

mlx/backend/cuda/binary/not_equal.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(NotEqual)
-} // namespace mlx::core

mlx/backend/cuda/binary/power.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Power)
-} // namespace mlx::core

mlx/backend/cuda/binary/remainder.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Remainder)
-} // namespace mlx::core

mlx/backend/cuda/binary/subtract.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Subtract)
-} // namespace mlx::core

mlx/backend/cuda/binary_two.cu

@@ -127,99 +127,45 @@ binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   }
 }
 
-template <
-    typename Op,
-    typename In,
-    typename Out,
-    typename IdxT,
-    int NDIM,
-    int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
 __global__ void binary_two_g_nd(
     const In* a,
     const In* b,
     Out* out_a,
     Out* out_b,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[NDIM - 1];
-  auto a_stride_x = a_strides[NDIM - 1];
-  auto b_stride_x = b_strides[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
     auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
-      index_rest * shape_x, shape.data(), a_strides.data(), b_strides.data());
-  auto a_vec =
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec_a;
-  AlignedVector<Out, N_READS> out_vec_b;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    auto out = Op{}(a_vec[i], b_vec[i]);
-    out_vec_a[i] = out[0];
-    out_vec_b[i] = out[1];
+        index, shape.data(), a_strides.data(), b_strides.data());
+    auto out = Op{}(a[a_idx], b[b_idx]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
   }
-  store_vector(out_a + shape_x * index_rest, index_x, out_vec_a, shape_x);
-  store_vector(out_b + shape_x * index_rest, index_x, out_vec_b, shape_x);
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT>
 __global__ void binary_two_g(
     const In* a,
     const In* b,
     Out* out_a,
     Out* out_b,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto a_stride_x = a_strides[ndim - 1];
-  auto b_stride_x = b_strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
     auto [a_idx, b_idx] = elem_to_loc(
-      index_rest * shape_x,
-      shape.data(),
-      a_strides.data(),
-      b_strides.data(),
-      ndim);
-  auto a_vec =
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec_a;
-  AlignedVector<Out, N_READS> out_vec_b;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    auto out = Op{}(a_vec[i], b_vec[i]);
-    out_vec_a[i] = out[0];
-    out_vec_b[i] = out[1];
+        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
+    auto out = Op{}(a[a_idx], b[b_idx]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
   }
-  store_vector(out_a + shape_x * index_rest, index_x, out_vec_a, shape_x);
-  store_vector(out_b + shape_x * index_rest, index_x, out_vec_b, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -245,18 +191,14 @@ void binary_two_op_gpu_inplace(
   auto& out_a = outputs[0];
   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);
+  set_binary_op_output_data(a, b, out_b, bopt);
 
   if (out_a.size() == 0) {
     return;
   }
 
+  auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(a);
   encoder.set_input_array(b);
   encoder.set_output_array(out_a);
@@ -283,64 +225,42 @@ void binary_two_op_gpu_inplace(
                 auto& a_strides = strides[0];
                 auto& b_strides = strides[1];
                 int ndim = shape.size();
-                int work_per_thread = 1;
-                auto dim0 = ndim > 0 ? shape.back() : 1;
-                auto rest = out_a.size() / dim0;
-                if (dim0 >= 4) {
-                  work_per_thread = 4;
-                }
-                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-                auto block_dims = get_block_dims(dim0, rest, 1);
-                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::binary_two_g_nd<
-                        Op,
-                        InType,
-                        OutType,
-                        IdxT,
-                        dims_constant(),
-                        1>;
-                    if (work_per_thread == 4) {
-                      kernel = cu::binary_two_g_nd<
-                          Op,
-                          InType,
-                          OutType,
-                          IdxT,
-                          dims_constant(),
-                          4>;
-                    }
+                    auto [num_blocks, block_dims] =
+                        get_launch_args(out_a, large());
                     encoder.add_kernel_node(
-                        kernel,
-                        {num_blocks_x, num_blocks_y},
+                        cu::binary_two_g_nd<
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
+                        num_blocks,
                         block_dims,
                         0,
-                        gpu_ptr<InType>(a),
-                        gpu_ptr<InType>(b),
-                        gpu_ptr<OutType>(out_a),
-                        gpu_ptr<OutType>(out_b),
-                        rest,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out_a.data<OutType>(),
+                        out_b.data<OutType>(),
+                        out_a.size(),
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 1>;
-                  if (work_per_thread == 4) {
-                    kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 4>;
-                  }
+                  auto [num_blocks, block_dims] =
+                      get_launch_args(out_a, large());
                   encoder.add_kernel_node(
-                      kernel,
-                      {num_blocks_x, num_blocks_y},
+                      cu::binary_two_g<Op, InType, OutType, IdxT>,
+                      num_blocks,
                       block_dims,
                       0,
-                      gpu_ptr<InType>(a),
-                      gpu_ptr<InType>(b),
-                      gpu_ptr<OutType>(out_a),
-                      gpu_ptr<OutType>(out_b),
-                      rest,
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out_a.data<OutType>(),
+                      out_b.data<OutType>(),
+                      out_a.size(),
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),
@@ -370,10 +290,10 @@ void binary_two_op_gpu_inplace(
                 num_blocks,
                 block_dims,
                 0,
-                gpu_ptr<InType>(a),
-                gpu_ptr<InType>(b),
-                gpu_ptr<OutType>(out_a),
-                gpu_ptr<OutType>(out_b),
+                a.data<InType>(),
+                b.data<InType>(),
+                out_a.data<OutType>(),
+                out_b.data<OutType>(),
                 out_a.data_size());
           });
         }

mlx/backend/cuda/compiled.cpp

@@ -267,8 +267,7 @@ void Compiled::eval_gpu(
       }
     }
 
-    return std::make_tuple(
-        false, std::move(builder.os), std::move(kernel_names));
+    return std::make_pair(std::move(builder.os), std::move(kernel_names));
   });
 
   // Collapse contiguous dims to route to a faster kernel if possible. Also
@@ -293,13 +292,8 @@ void Compiled::eval_gpu(
     }
   }
 
-  auto& encoder = cu::get_command_encoder(s);
-
   // Put outputs.
-  compiled_allocate_outputs(
-      inputs, outputs, is_constant_, contiguous, [&](auto n) {
-        return cu::malloc_async(n, encoder.stream());
-      });
+  compiled_allocate_outputs(inputs, outputs, is_constant_, contiguous);
   for (auto& x : outputs) {
     args.append(x);
   }
@@ -329,6 +323,7 @@ void Compiled::eval_gpu(
     kernel_name += fmt::format(
         "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
   }
+  auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
     encoder.set_input_array(in);
   }
@@ -336,9 +331,9 @@ void Compiled::eval_gpu(
     encoder.set_output_array(out);
   }
 
-  auto [kernel, max_block_dims] = mod.get_kernel_and_dims(kernel_name);
+  auto kernel = mod.get_kernel(kernel_name);
   auto [num_blocks, block_dims] =
-      get_launch_args(outputs[0], large, work_per_thread, max_block_dims);
+      get_launch_args(outputs[0], large, work_per_thread);
   encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 

mlx/backend/cuda/conv.cpp

@@ -1,12 +1,18 @@
 // Copyright © 2025 Apple Inc.
 
-#include "mlx/backend/cuda/conv/conv.h"
-#include "mlx/backend/cuda/cudnn_utils.h"
 #include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/config.h"
 #include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
+// cudnn_frontend.h redefines this macro.
+#undef CHECK_CUDA_ERROR
+
+#include <cudnn_frontend.h>
+#include <cudnn_frontend_find_plan.h>
+#include <fmt/format.h>
 #include <nvtx3/nvtx3.hpp>
 
 #include <cassert>
@@ -15,6 +21,9 @@ namespace mlx::core {
 
 namespace {
 
+// Not all engines support it so can not use this API now.
+#define MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API 0
+
 // Alias for better readability.
 #define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
 #define CONV_BACKWARD_INPUT \
@@ -22,9 +31,6 @@ namespace {
 #define CONV_BACKWARD_WEIGHT \
   CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
 
-// Custom placeholder representing fallback kernel.
-#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
-
 struct ConvCacheKey {
   int device_id;
   cudnnDataType_t cudnn_dtype;
@@ -44,13 +50,203 @@ struct ConvCacheKey {
 auto& conv_cache() {
   static LRUBytesKeyCache<
       ConvCacheKey,
-      std::pair<
-          cudnnBackendDescriptorType_t,
-          std::optional<cudnn_frontend::ExecutionPlan>>>
-      cache("MLX_CUDA_CONV_CACHE_SIZE", /* default_capacity */ 128);
+      std::pair<cudnnBackendDescriptorType_t, cudnn_frontend::ExecutionPlan>>
+      cache(/* capacity */ 128);
   return cache;
 }
 
+template <typename T, typename Vec>
+inline SmallVector<T> convert_vector(const Vec& vec) {
+  return SmallVector<T>(vec.begin(), vec.end());
+}
+
+template <typename T, template <typename U> class Vec>
+inline std::array<T, MAX_NDIM> fixed_vector(const Vec<T>& vec) {
+  if (vec.size() > MAX_NDIM) {
+    throw std::runtime_error(
+        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
+  }
+  std::array<T, MAX_NDIM> result = {};
+  std::copy_n(vec.begin(), vec.size(), result.begin());
+  return result;
+}
+
+auto nhwc_to_nchw(const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  shape.insert(shape.begin() + 1, shape.back());
+  shape.erase(shape.end() - 1);
+  auto strides = convert_vector<int64_t>(x.strides());
+  strides.insert(strides.begin() + 1, strides.back());
+  strides.erase(strides.end() - 1);
+  return std::make_tuple(std::move(shape), std::move(strides));
+}
+
+inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return CUDNN_DATA_INT8;
+    case int32:
+      return CUDNN_DATA_INT32;
+    case uint8:
+      return CUDNN_DATA_UINT8;
+    case float16:
+      return CUDNN_DATA_HALF;
+    case bfloat16:
+      return CUDNN_DATA_BFLOAT16;
+    case float32:
+      return CUDNN_DATA_FLOAT;
+    case float64:
+      return CUDNN_DATA_DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+  }
+}
+
+inline uint8_t get_alignment(const array& x) {
+  uint8_t alignment = 1;
+  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
+  for (; alignment < 32; alignment *= 2) {
+    if (address % (alignment * 2)) {
+      return alignment;
+    }
+  }
+  return alignment;
+}
+
+inline cudnn_frontend::Tensor build_tensor(int64_t id, const array& x) {
+  auto [shape, strides] = nhwc_to_nchw(x);
+  return cudnn_frontend::TensorBuilder()
+      .setDim(shape.size(), shape.data())
+      .setStrides(strides.size(), strides.data())
+      .setId(id)
+      .setAlignment(get_alignment(x))
+      .setDataType(dtype_to_cudnn_type(x.dtype()))
+      .build();
+}
+
+cudnn_frontend::EngineConfigList get_engine_configs(
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph,
+    bool use_fallback = false) {
+  cudnn_frontend::GeneratorSource source;
+  if (use_fallback) {
+    source = [&backend_type](cudnn_frontend::OperationGraph& op_graph) {
+      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
+                          .setOperationGraph(op_graph)
+                          .setOperation(backend_type)
+                          .build();
+      return fallback.getFallbackList();
+    };
+  } else {
+    source = [](cudnn_frontend::OperationGraph& op_graph) {
+      auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
+                            .setOperationGraph(op_graph)
+                            .setHeurMode(CUDNN_HEUR_MODE_A)
+                            .build();
+      return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
+    };
+  }
+
+  cudnn_frontend::EngineConfigGenerator generator(1, &source);
+  auto configs = generator.generate_engine_config(op_graph);
+
+  cudnn_frontend::EngineConfigList filtered_configs;
+  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
+    if (cudnn_frontend::hasNumericalNote<
+            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
+      return true;
+    }
+    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
+        dtype == float32 && !env::enable_tf32()) {
+      return true;
+    }
+    return false;
+  });
+  return filtered_configs;
+}
+
+bool execute_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    array& x,
+    array& w,
+    array& y) {
+  int workspace_size = plan.getWorkspaceSize();
+  array workspace(allocator::malloc(workspace_size), {workspace_size}, uint8);
+
+  int64_t uids[3] = {'x', 'w', 'y'};
+  void* data_ptrs[3] = {
+      x.data<void>(),
+      w.data<void>(),
+      y.data<void>(),
+  };
+
+  auto variantPack = cudnn_frontend::VariantPackBuilder()
+                         .setWorkspacePointer(workspace.data<void>())
+                         .setDataPointers(3, data_ptrs)
+                         .setUids(3, uids)
+                         .build();
+
+  auto handle = encoder.device().cudnn_handle();
+  cudnnSetStream(handle, encoder.stream());
+
+#if CUDNN_VERSION >= 90500 && MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API
+  cudaGraph_t graph;
+  cudaGraphCreate(&graph, 0);
+  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
+      &graph, [](cudaGraph_t* p) { cudaGraphDestroy(*p); });
+  if (cudnnBackendPopulateCudaGraph(
+          handle, plan.get_raw_desc(), variantPack.get_raw_desc(), graph) !=
+      CUDNN_STATUS_SUCCESS) {
+    return false;
+  }
+  encoder.add_graph_node(graph);
+#else
+  auto capture = encoder.capture_context();
+  if (cudnnBackendExecute(
+          handle, plan.get_raw_desc(), variantPack.get_raw_desc()) !=
+      CUDNN_STATUS_SUCCESS) {
+    // Discard the captured graph when failed.
+    capture.discard = true;
+    return false;
+  }
+#endif
+
+  encoder.add_temporary(workspace);
+  return true;
+}
+
+bool try_engines(
+    cu::CommandEncoder& encoder,
+    const ConvCacheKey& cache_key,
+    cudnnBackendDescriptorType_t backend_type,
+    cudnn_frontend::EngineConfigList& configs,
+    const std::string& op_graph_tag,
+    array& x,
+    array& w,
+    array& y) {
+  for (auto& config : configs) {
+    try {
+      auto plan = cudnn_frontend::ExecutionPlanBuilder()
+                      .setHandle(encoder.device().cudnn_handle())
+                      .setEngineConfig(config, op_graph_tag)
+                      .build();
+      if (execute_plan(encoder, plan, x, w, y)) {
+        conv_cache().emplace(
+            cache_key, std::make_pair(backend_type, std::move(plan)));
+        return true;
+      }
+    } catch (cudnn_frontend::cudnnException& error) {
+      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
+        throw;
+      }
+    }
+  }
+  return false;
+}
+
 auto get_conv_op_settings(
     cudnnBackendDescriptorType_t backend_type,
     array& x,
@@ -95,7 +291,7 @@ auto get_conv_op_settings(
   }
 }
 
-std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
+std::optional<cudnn_frontend::OperationGraph> build_op_graph(
     cu::CommandEncoder& encoder,
     cudnnBackendDescriptorType_t backend_type,
     Dtype dtype,
@@ -121,9 +317,9 @@ std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
                          .build();
 
     auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .setxDesc(build_cudnn_tensor_nchw('x', x))
-                  .setwDesc(build_cudnn_tensor_nchw('w', w))
-                  .setyDesc(build_cudnn_tensor_nchw('y', y))
+                  .setxDesc(build_tensor('x', x))
+                  .setwDesc(build_tensor('w', w))
+                  .setyDesc(build_tensor('y', y))
                   .setcDesc(conv_desc)
                   .build();
 
@@ -140,42 +336,6 @@ std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
   }
 }
 
-// Transpose from (C_out, H, W, C_in / groups) to (C_in, H, W, C_out / groups).
-array group_transpose(
-    const array& x,
-    int groups,
-    int group_dim,
-    int axis1,
-    int axis2,
-    Stream s) {
-  if (groups == 1) {
-    return swapaxes_in_eval(x, axis1, axis2);
-  }
-  int ndim = x.ndim();
-  if (group_dim < 0) {
-    group_dim += ndim;
-  }
-  if (axis1 < 0) {
-    axis1 += ndim;
-  }
-  if (axis2 < 0) {
-    axis2 += ndim;
-  }
-  if (group_dim <= axis1) {
-    axis1 += 1;
-  }
-  if (group_dim <= axis2) {
-    axis2 += 1;
-  }
-  auto shape = x.shape();
-  shape.insert(shape.begin() + group_dim, groups);
-  shape[group_dim + 1] = shape[group_dim + 1] / groups;
-  array x_trans = reshape_in_eval(x, std::move(shape), s);
-  x_trans = swapaxes_in_eval(x_trans, axis1, axis2);
-  x_trans = flatten_in_eval(x_trans, group_dim, group_dim + 1, s);
-  return x_trans;
-}
-
 // Do necessary transposes and copies to prepare the inputs and outputs for
 // building the cuDNN conv op. It is safe to be called multiple times in one
 // eval_gpu, with cost of possible redundant copies.
@@ -185,14 +345,13 @@ std::tuple<array, array, array> prepare_args(
     array in,
     array wt,
     array out,
-    int groups,
     Stream s) {
   // Transpose the args depending on the backend type.
   // TODO: Handle groups.
   if (backend_type == CONV_BACKWARD_INPUT) {
-    wt = group_transpose(wt, groups, 0, 0, -1, s);
+    wt = swapaxes_in_eval(wt, 0, -1);
   } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    in = group_transpose(in, groups, -1, 0, -1, s);
+    in = swapaxes_in_eval(in, 0, -1);
     wt = swapaxes_in_eval(wt, 0, -1);
     // Create a contiguous array that shares the data with |out|, but with dim
     // C_in and C_out swapped.
@@ -270,26 +429,27 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   if (out_.size() == 0) {
     return;
   }
-  auto& s = stream();
-  auto& encoder = cu::get_command_encoder(s);
 
   assert(inputs.size() == 2);
   array in = inputs[0];
   array wt = inputs[1];
   array out = out_;
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(allocator::malloc(out.nbytes()));
   Dtype dtype = out.dtype();
 
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
   // Search cache.
   ConvCacheKey cache_key{
       encoder.device().cuda_device(),
       dtype_to_cudnn_type(dtype),
-      vector_key(in.shape()),
-      vector_key(wt.shape()),
-      vector_key(kernel_strides_),
-      vector_key(padding_lo_),
-      vector_key(padding_hi_),
-      vector_key(kernel_dilation_),
+      fixed_vector(in.shape()),
+      fixed_vector(wt.shape()),
+      fixed_vector(kernel_strides_),
+      fixed_vector(padding_lo_),
+      fixed_vector(padding_hi_),
+      fixed_vector(kernel_dilation_),
       groups_,
       flip_,
       get_alignment(in),
@@ -297,30 +457,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       get_alignment(out)};
   if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
     auto& [backend_type, plan] = it->second;
-    if (plan) {
-      // Run cached plan.
-      std::tie(in, wt, out) =
-          prepare_args(encoder, backend_type, in, wt, out, groups_, s);
+    std::tie(in, wt, out) = prepare_args(encoder, backend_type, in, wt, out, s);
     register_args(encoder, backend_type, in, wt, out, out_);
     auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-      if (!encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+    if (!execute_plan(encoder, plan, x, w, y)) {
       throw std::runtime_error("[conv] Cached plan failed to execute.");
     }
-    } else {
-      // Run fallback kernel.
-      gemm_conv(
-          encoder,
-          in,
-          wt,
-          out,
-          kernel_strides_,
-          padding_lo_,
-          kernel_dilation_,
-          input_dilation_,
-          groups_,
-          flip_,
-          s);
-    }
     return;
   }
 
@@ -348,7 +490,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   std::optional<cudnn_frontend::OperationGraph> op_graph;
   for (auto try_backend : try_backends) {
     auto [in_copy, wt_copy, out_copy] =
-        prepare_args(encoder, try_backend, in, wt, out, groups_, s);
+        prepare_args(encoder, try_backend, in, wt, out, s);
     auto [x, w, y] = dispatch_args(try_backend, in_copy, wt_copy, out_copy);
     auto [stride, padding_lo, padding_hi, dilation] = get_conv_op_settings(
         try_backend,
@@ -360,7 +502,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_hi_,
         kernel_dilation_,
         input_dilation_);
-    op_graph = build_conv_op_graph(
+    op_graph = build_op_graph(
         encoder,
         try_backend,
         dtype,
@@ -379,38 +521,26 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       break;
     }
   }
+  if (!op_graph) {
+    throw std::runtime_error("[conv] Can not build op graph.");
+  }
 
-  if (op_graph) {
-    // Find a plan for the graph and execute it.
-    auto plan = find_cudnn_plan_from_op_graph(
-        encoder.device().cudnn_handle(), backend_type, dtype, *op_graph);
-    if (plan) {
-      // Setup inputs and outputs.
+  // Get ready to execute the graph.
   register_args(encoder, backend_type, in, wt, out, out_);
 
+  // Try to run plans based on heuristics.
+  auto configs = get_engine_configs(backend_type, dtype, *op_graph);
+  auto tag = op_graph->getTag();
   auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-      if (encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
-        conv_cache().emplace(
-            cache_key, std::make_pair(backend_type, std::move(*plan)));
+  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
     return;
   }
+  // Then try fallback plans.
+  configs = get_engine_configs(backend_type, dtype, *op_graph);
+  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
+    return;
   }
-  }
-
-  // Use fallback kernel for settings not supported by cuDNN.
-  gemm_conv(
-      encoder,
-      in,
-      wt,
-      out,
-      kernel_strides_,
-      padding_lo_,
-      kernel_dilation_,
-      input_dilation_,
-      groups_,
-      flip_,
-      s);
-  conv_cache().emplace(cache_key, std::make_pair(CONV_FALLBACK, std::nullopt));
+  throw std::runtime_error("[conv] Unable to find a working engine.");
 }
 
 } // namespace mlx::core

mlx/backend/cuda/conv/conv.h

@@ -1,126 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/gpu/copy.h"
-
-namespace mlx::core {
-
-template <int NDIM>
-struct ConvParams {
-  int N; // Batch size
-  int C; // In channels
-  int O; // Out channels
-  int strides[NDIM];
-  int padding[NDIM];
-  int kernel_dilation[NDIM];
-  int input_dilation[NDIM];
-  int groups;
-  bool flip;
-  int in_spatial_dims[NDIM];
-  int wt_spatial_dims[NDIM];
-  int out_spatial_dims[NDIM];
-  int64_t in_strides[NDIM + 2];
-
-  ConvParams(
-      const array& in,
-      const array& wt,
-      const array& out,
-      const std::vector<int>& strides,
-      const std::vector<int>& padding,
-      const std::vector<int>& kernel_dilation,
-      const std::vector<int>& input_dilation,
-      int groups,
-      bool flip)
-      : N(in.shape(0)),
-        C(in.shape(-1)),
-        O(wt.shape(0)),
-        groups(groups),
-        flip(flip) {
-    std::copy_n(strides.begin(), NDIM, this->strides);
-    std::copy_n(padding.begin(), NDIM, this->padding);
-    std::copy_n(kernel_dilation.begin(), NDIM, this->kernel_dilation);
-    std::copy_n(input_dilation.begin(), NDIM, this->input_dilation);
-    std::copy_n(in.shape().begin() + 1, NDIM, this->in_spatial_dims);
-    std::copy_n(wt.shape().begin() + 1, NDIM, this->wt_spatial_dims);
-    std::copy_n(out.shape().begin() + 1, NDIM, this->out_spatial_dims);
-    std::copy_n(in.strides().begin(), NDIM + 2, this->in_strides);
-  }
-};
-
-void gemm_grouped_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s);
-
-void gemm_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    bool flip,
-    Stream s);
-
-inline void gemm_conv(
-    cu::CommandEncoder& encoder,
-    array in,
-    array wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s) {
-  if (!in.flags().row_contiguous) {
-    in = contiguous_copy_gpu(in, s);
-    encoder.add_temporary(in);
-  }
-  if (!wt.flags().row_contiguous) {
-    wt = contiguous_copy_gpu(wt, s);
-    encoder.add_temporary(wt);
-  }
-
-  if (groups == 1) {
-    gemm_conv(
-        encoder,
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        flip,
-        s);
-  } else {
-    gemm_grouped_conv(
-        encoder,
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        groups,
-        flip,
-        s);
-  }
-}
-
-} // namespace mlx::core

mlx/backend/cuda/conv/gemm_conv.cu

@@ -1,217 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/conv/conv.h"
-#include "mlx/backend/cuda/gemms/cublas_gemm.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-
-#include <cooperative_groups.h>
-
-namespace mlx::core {
-
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-template <typename T, int NDIM>
-__global__ void naive_unfold_nd(
-    const T* in,
-    T* out,
-    int filter_size,
-    int out_pixels,
-    const __grid_constant__ ConvParams<NDIM> params) {
-  auto block = cg::this_thread_block();
-  auto tid = block.group_index();
-  auto lid = block.thread_index();
-
-  int index_batch = tid.z / out_pixels; // [0, N)
-  int index_out_spatial = tid.z % out_pixels; // [0, H_out * W_out)
-  int index_wt_spatial =
-      tid.x * block.dim_threads().x + lid.x; // [0, H_wt * W_wt)
-
-  if (index_wt_spatial >= filter_size / params.C) {
-    return;
-  }
-
-  in += tid.y; // [0, C)
-  out += tid.z * filter_size + index_wt_spatial * params.C + tid.y;
-
-  bool valid = index_batch < params.N;
-
-  // Get the coordinates in input.
-  int index_in[NDIM] = {};
-#pragma unroll
-  for (int i = NDIM - 1; i >= 0; --i) {
-    int index_out = index_out_spatial % params.out_spatial_dims[i];
-    int index_wt = index_wt_spatial % params.wt_spatial_dims[i];
-
-    if (params.flip) {
-      index_wt = params.wt_spatial_dims[i] - index_wt - 1;
-    }
-
-    int index = index_out * params.strides[i] - params.padding[i] +
-        index_wt * params.kernel_dilation[i];
-    int index_max =
-        1 + params.input_dilation[i] * (params.in_spatial_dims[i] - 1);
-
-    valid &= (index >= 0) && (index < index_max) &&
-        (index % params.input_dilation[i] == 0);
-
-    index_in[i] = index / params.input_dilation[i];
-
-    index_out_spatial /= params.out_spatial_dims[i];
-    index_wt_spatial /= params.wt_spatial_dims[i];
-  }
-
-  if (valid) {
-    int in_offset = index_batch * params.in_strides[0];
-#pragma unroll
-    for (int i = 0; i < NDIM; ++i) {
-      in_offset += index_in[i] * params.in_strides[i + 1];
-    }
-    *out = in[in_offset];
-  } else {
-    *out = T{0};
-  }
-}
-
-} // namespace cu
-
-template <int NDIM>
-array unfold_inputs_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    int mat_M,
-    int mat_K,
-    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()));
-  encoder.add_temporary(unfolded);
-
-  int filter_size = params.C;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    filter_size *= params.wt_spatial_dims[i];
-  }
-
-  int out_pixels = 1;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    out_pixels *= params.out_spatial_dims[i];
-  }
-
-  int wt_spatial_size = mat_K / params.C;
-  dim3 block_dims;
-  block_dims.x = std::min(std::max(wt_spatial_size, 32), 1024);
-  dim3 num_blocks;
-  num_blocks.x = cuda::ceil_div(wt_spatial_size, block_dims.x);
-  num_blocks.y = params.C;
-  num_blocks.z = mat_M;
-
-  encoder.set_input_array(in);
-  encoder.set_output_array(unfolded);
-  dispatch_float_types(in.dtype(), "unfold", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    encoder.add_kernel_node(
-        cu::naive_unfold_nd<DataType, NDIM>,
-        num_blocks,
-        block_dims,
-        0,
-        gpu_ptr<DataType>(in),
-        gpu_ptr<DataType>(unfolded),
-        filter_size,
-        out_pixels,
-        params);
-  });
-
-  return unfolded;
-}
-
-template <int NDIM>
-void gemm_conv_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    ConvParams<NDIM>& params,
-    Stream s) {
-  // Get gemm shapes.
-  int mat_M = out.size() / params.O; // N * H_out * W_out
-  int mat_K = wt.size() / params.O; // C * H_wt * W_wt
-  int mat_N = params.O; // O
-
-  // Unfold input to (N * H_out * W_out, C * H_wt * W_wt) for gemm.
-  array in_unfolded =
-      unfold_inputs_nd<NDIM>(encoder, in, mat_M, mat_K, mat_N, params);
-
-  // Reshape weight to (C * H_wt * W_wt, O) for gemm.
-  array wt_reshaped({mat_K, mat_N}, wt.dtype(), nullptr, {});
-  wt_reshaped.copy_shared_buffer(
-      wt,
-      {1, mat_K},
-      {false, false, /* col_contiguous */ true},
-      wt.data_size());
-
-  // Single batch.
-  Shape batch_shape{1};
-  Strides a_batch_strides{0};
-  Strides b_batch_strides{0};
-
-  // Run matmul.
-  CublasGemm gemm(
-      encoder.device(),
-      in.dtype(),
-      false, // a_transposed
-      mat_M, // a_rows
-      mat_K, // a_cols
-      mat_K, // lda
-      true, // b_transposed
-      mat_K, // b_rows
-      mat_N, // b_cols
-      mat_K, // ldb
-      batch_shape.back(),
-      a_batch_strides.back(),
-      b_batch_strides.back());
-  gemm.run(
-      encoder,
-      out,
-      in_unfolded,
-      wt_reshaped,
-      batch_shape,
-      a_batch_strides,
-      b_batch_strides);
-}
-
-void gemm_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    bool flip,
-    Stream s) {
-  int conv_ndim = in.ndim() - 2;
-  if (conv_ndim < 1 || conv_ndim > 3) {
-    throw std::runtime_error(
-        fmt::format("[conv] Unsupported gemm_conv for {}D conv.", conv_ndim));
-  }
-  dispatch_1_2_3(conv_ndim, [&](auto ndim_constant) {
-    ConvParams<ndim_constant()> params(
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        1, // groups
-        flip);
-    gemm_conv_nd<ndim_constant()>(encoder, in, wt, out, params, s);
-  });
-}
-
-} // namespace mlx::core

mlx/backend/cuda/conv/gemm_grouped_conv.cu

@@ -1,231 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/conv/conv.h"
-#include "mlx/backend/cuda/gemms/cublas_gemm.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-
-#include <cooperative_groups.h>
-
-namespace mlx::core {
-
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-template <typename T, int NDIM>
-__global__ void naive_grouped_unfold_transpose_nd(
-    const T* in,
-    T* out,
-    int filter_size,
-    int out_pixels,
-    const __grid_constant__ ConvParams<NDIM> params) {
-  auto block = cg::this_thread_block();
-  auto tid = block.group_index();
-  auto lid = block.thread_index();
-
-  int index_batch = tid.z / out_pixels; // [0, N)
-  int index_out_spatial = tid.z % out_pixels; // [0, H_out * W_out)
-  int index_wt_spatial =
-      tid.x * block.dim_threads().x + lid.x; // [0, H_wt * W_wt)
-
-  if (index_wt_spatial >= filter_size / params.C) {
-    return;
-  }
-
-  in += tid.y; // [0, C)
-  out += tid.z * filter_size + tid.y * (filter_size / params.C);
-
-  bool valid = index_batch < params.N;
-
-  // Get the coordinates in input.
-  int index_in[NDIM] = {};
-  int wt_stride = 1;
-#pragma unroll
-  for (int i = NDIM - 1; i >= 0; --i) {
-    int index_out = index_out_spatial % params.out_spatial_dims[i];
-    int index_wt = index_wt_spatial % params.wt_spatial_dims[i];
-    out += index_wt * wt_stride;
-
-    if (params.flip) {
-      index_wt = params.wt_spatial_dims[i] - index_wt - 1;
-    }
-
-    int index = index_out * params.strides[i] - params.padding[i] +
-        index_wt * params.kernel_dilation[i];
-    int index_max =
-        1 + params.input_dilation[i] * (params.in_spatial_dims[i] - 1);
-
-    valid &= (index >= 0) && (index < index_max) &&
-        (index % params.input_dilation[i] == 0);
-
-    index_in[i] = index / params.input_dilation[i];
-
-    index_out_spatial /= params.out_spatial_dims[i];
-    index_wt_spatial /= params.wt_spatial_dims[i];
-    wt_stride *= params.wt_spatial_dims[i];
-  }
-
-  if (valid) {
-    int in_offset = index_batch * params.in_strides[0];
-#pragma unroll
-    for (int i = 0; i < NDIM; ++i) {
-      in_offset += index_in[i] * params.in_strides[i + 1];
-    }
-    *out = in[in_offset];
-  } else {
-    *out = T{0};
-  }
-}
-
-} // namespace cu
-
-template <int NDIM>
-array grouped_unfold_transpose_inputs_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    int mat_M,
-    int mat_K,
-    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()));
-  encoder.add_temporary(unfolded);
-
-  int filter_size = params.C;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    filter_size *= params.wt_spatial_dims[i];
-  }
-
-  int out_pixels = 1;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    out_pixels *= params.out_spatial_dims[i];
-  }
-
-  int wt_spatial_size = (mat_K * params.groups) / params.C;
-  dim3 block_dims;
-  block_dims.x = std::min(std::max(wt_spatial_size, 32), 1024);
-  dim3 num_blocks;
-  num_blocks.x = cuda::ceil_div(wt_spatial_size, block_dims.x);
-  num_blocks.y = params.C;
-  num_blocks.z = mat_M;
-
-  encoder.set_input_array(in);
-  encoder.set_output_array(unfolded);
-  dispatch_float_types(in.dtype(), "unfold", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    encoder.add_kernel_node(
-        cu::naive_grouped_unfold_transpose_nd<DataType, NDIM>,
-        num_blocks,
-        block_dims,
-        0,
-        gpu_ptr<DataType>(in),
-        gpu_ptr<DataType>(unfolded),
-        filter_size,
-        out_pixels,
-        params);
-  });
-
-  return unfolded;
-}
-
-template <int NDIM>
-void gemm_grouped_conv_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    ConvParams<NDIM>& params,
-    Stream s) {
-  // Get gemm shapes.
-  int C_per_group = params.C / params.groups;
-  int O_per_group = params.O / params.groups;
-  int mat_M = out.size() / params.O; // N * H_out * W_out
-  int mat_K = wt.size() / params.O; // C_per_group * H_wt * W_wt
-  int mat_N = O_per_group; // O_per_group
-
-  // Unfold input to (N * H_out * W_out, C * H_wt * W_wt) for gemm.
-  array in_unfolded = grouped_unfold_transpose_inputs_nd<NDIM>(
-      encoder, in, mat_M, mat_K, mat_N, params);
-
-  // Reshape weight to (O, C_per_group, H_wt * W_wt) for gemm.
-  int wt_spatial_size = (wt.size() / wt.shape(0)) / wt.shape(-1);
-  array wt_view(
-      {params.O, C_per_group, wt_spatial_size}, wt.dtype(), nullptr, {});
-  wt_view.copy_shared_buffer(
-      wt, {wt.strides(0), 1, C_per_group}, wt.flags(), wt.size());
-  array wt_reshaped = contiguous_copy_gpu(wt_view, s);
-
-  // Batch with size of groups.
-  Shape batch_shape{params.groups};
-  Strides a_batch_strides{mat_K};
-  Strides b_batch_strides{mat_N * mat_K};
-
-  // Run matmul.
-  CublasGemm gemm(
-      encoder.device(),
-      in.dtype(),
-      false, // a_transposed
-      mat_M, // a_rows
-      mat_K, // a_cols
-      mat_K * params.groups, // lda
-      true, // b_transposed
-      mat_K, // b_rows
-      mat_N, // b_cols
-      mat_K, // ldb
-      batch_shape.back(),
-      a_batch_strides.back(),
-      b_batch_strides.back());
-  gemm.set_out(
-      out.dtype(),
-      false, // out_transposed
-      mat_M, // out_rows
-      mat_N, // out_cols
-      mat_N * params.groups, // out_ld
-      params.groups, // batch_count
-      mat_N); // batch_stride
-  gemm.run(
-      encoder,
-      out,
-      in_unfolded,
-      wt_reshaped,
-      batch_shape,
-      a_batch_strides,
-      b_batch_strides);
-}
-
-void gemm_grouped_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s) {
-  int conv_ndim = in.ndim() - 2;
-  if (conv_ndim < 1 || conv_ndim > 3) {
-    throw std::runtime_error(
-        fmt::format("[conv] Unsupported gemm_conv for {}D conv.", conv_ndim));
-  }
-  dispatch_1_2_3(conv_ndim, [&](auto ndim_constant) {
-    ConvParams<ndim_constant()> params(
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        groups,
-        flip);
-    gemm_grouped_conv_nd<ndim_constant()>(encoder, in, wt, out, params, s);
-  });
-}
-
-} // namespace mlx::core

mlx/backend/cuda/copy.cu

@@ -5,22 +5,6 @@
 
 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());
-  });
-  if (donated && in.dtype() == out.dtype()) {
-    // If the output has the same type as the input then there is nothing to
-    // copy, just use the buffer.
-    return;
-  }
-  if (ctype == CopyType::GeneralGeneral) {
-    ctype = CopyType::General;
-  }
-  copy_gpu_inplace(in, out, ctype, s);
-}
-
 void copy_gpu_inplace(
     const array& in,
     array& out,
@@ -31,8 +15,8 @@ void copy_gpu_inplace(
     int64_t offset_out,
     CopyType ctype,
     const Stream& s,
-    std::optional<array> dynamic_offset_in,
-    std::optional<array> dynamic_offset_out) {
+    const std::optional<array>& dynamic_offset_in,
+    const std::optional<array>& dynamic_offset_out) {
   if (out.size() == 0) {
     return;
   }
@@ -60,16 +44,6 @@ void copy_gpu_inplace(
           strides_vec[0]);
     } else {
       if (dynamic_offset_in || dynamic_offset_out) {
-        if (!dynamic_offset_in) {
-          dynamic_offset_in = array(0, int64);
-          encoder.add_temporary(*dynamic_offset_in);
-        }
-        if (!dynamic_offset_out) {
-          dynamic_offset_out = array(0, int64);
-          encoder.add_temporary(*dynamic_offset_out);
-        }
-        encoder.set_input_array(*dynamic_offset_in);
-        encoder.set_input_array(*dynamic_offset_out);
         copy_general_dynamic(
             encoder,
             ctype,
@@ -80,8 +54,8 @@ void copy_gpu_inplace(
             shape_collapsed,
             strides_vec[0],
             strides_vec[1],
-            *dynamic_offset_in,
-            *dynamic_offset_out);
+            dynamic_offset_in ? *dynamic_offset_in : array(0, int64),
+            dynamic_offset_out ? *dynamic_offset_out : array(0, int64));
       } else {
         copy_general(
             encoder,
@@ -103,31 +77,11 @@ void fill_gpu(const array& in, array& out, const Stream& s) {
   if (out.size() == 0) {
     return;
   }
+  out.set_data(allocator::malloc(out.nbytes()));
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
   encoder.set_input_array(in);
   encoder.set_output_array(out);
   copy_contiguous(encoder, CopyType::Scalar, in, out, 0, 0);
 }
 
-void reshape_gpu(const array& in, array& out, Stream s) {
-  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
-  if (copy_necessary) {
-    auto& encoder = cu::get_command_encoder(s);
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
-    copy_gpu_inplace(
-        in,
-        out,
-        in.shape(),
-        in.strides(),
-        make_contiguous_strides(in.shape()),
-        0,
-        0,
-        CopyType::General,
-        s);
-  } else {
-    shared_buffer_reshape(in, out_strides, out);
-  }
-}
-
 } // namespace mlx::core

mlx/backend/cuda/copy/copy_contiguous.cu

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

mlx/backend/cuda/copy/copy_general.cu

@@ -10,80 +10,37 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
+template <typename In, typename Out, typename IdxT, int NDIM>
 __global__ void copy_gg_nd(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_out) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[NDIM - 1];
-  auto in_stride_x = strides_in[NDIM - 1];
-  auto out_stride_x = strides_out[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
     auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
-      index_rest * shape_x,
-      shape.data(),
-      strides_in.data(),
-      strides_out.data());
-
-  auto in_vec =
-      load_vector<N_READS>(in + idx_in, index_x, shape_x, in_stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
+        index, shape.data(), strides_in.data(), strides_out.data());
+    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
   }
-  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
-template <typename In, typename Out, typename IdxT, int N_READS>
+template <typename In, typename Out, typename IdxT>
 __global__ void copy_gg(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides strides_in,
     const __grid_constant__ Strides strides_out,
     int ndim) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto in_stride_x = strides_in[ndim - 1];
-  auto out_stride_x = strides_out[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
     auto [idx_in, idx_out] = elem_to_loc(
-      index_rest * shape_x,
-      shape.data(),
-      strides_in.data(),
-      strides_out.data(),
-      ndim);
-
-  auto in_vec =
-      load_vector<N_READS>(in + idx_in, index_x, shape_x, in_stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
+        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
+    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
   }
-  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
 } // namespace cu
@@ -106,58 +63,39 @@ void copy_general(
             using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
             using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
             using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-            const InType* in_ptr = gpu_ptr<InType>(in) + offset_in;
-            OutType* out_ptr = gpu_ptr<OutType>(out) + offset_out;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
             int ndim = shape.size();
             size_t data_size = 1;
             for (auto& s : shape)
               data_size *= s;
-
-            int work_per_thread = 1;
-            auto dim0 = ndim > 0 ? shape.back() : 1;
-            auto rest = data_size / dim0;
-            if (dim0 >= 4) {
-              work_per_thread = 4;
-            }
-
-            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-            auto block_dims = get_block_dims(dim0, rest, 1);
-            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto ndim_constant) {
-                auto kernel =
-                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 1>;
-                if (work_per_thread == 4) {
-                  kernel =
-                      cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 4>;
-                }
+                auto [num_blocks, block_dims] =
+                    get_launch_args(data_size, shape, out.strides(), large());
                 encoder.add_kernel_node(
-                    kernel,
-                    {num_blocks_x, num_blocks_y},
+                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>,
+                    num_blocks,
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    rest,
+                    data_size,
                     const_param<ndim_constant()>(shape),
                     const_param<ndim_constant()>(strides_in),
                     const_param<ndim_constant()>(strides_out));
               });
             } else { // ndim >= 4
-              auto kernel = cu::copy_gg<InType, OutType, IdxT, 1>;
-              if (work_per_thread == 4) {
-                kernel = cu::copy_gg<InType, OutType, IdxT, 4>;
-              }
+              auto [num_blocks, block_dims] =
+                  get_launch_args(data_size, shape, out.strides(), large());
               encoder.add_kernel_node(
-                  kernel,
-                  {num_blocks_x, num_blocks_y},
+                  cu::copy_gg<InType, OutType, IdxT>,
+                  num_blocks,
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  rest,
+                  data_size,
                   const_param(shape),
                   const_param(strides_in),
                   const_param(strides_out),

mlx/backend/cuda/copy/copy_general_dynamic.cu

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

mlx/backend/cuda/copy/copy_general_input.cu

@@ -10,67 +10,33 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
+template <typename In, typename Out, typename IdxT, int NDIM>
 __global__ void copy_g_nd(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
-    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    IdxT idx_in = elem_to_loc_nd<NDIM>(index, shape.data(), strides_in.data());
+    out[index] = CastOp<In, Out>{}(in[idx_in]);
+  }
 }
 
-  auto shape_x = shape[NDIM - 1];
-  auto stride_x = strides[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  auto idx =
-      elem_to_loc_nd<NDIM>(index_rest * shape_x, shape.data(), strides.data());
-  auto in_vec =
-      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
-  }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
-}
-
-template <typename In, typename Out, typename IdxT, int N_READS>
+template <typename In, typename Out, typename IdxT>
 __global__ void copy_g(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
-    const __grid_constant__ Strides strides,
+    const __grid_constant__ Strides strides_in,
     int ndim) {
-  auto block = cg::this_thread_block();
-  auto grid = cg::this_grid();
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    IdxT idx_in = elem_to_loc(index, shape.data(), strides_in.data(), ndim);
+    out[index] = CastOp<In, Out>{}(in[idx_in]);
   }
-
-  auto shape_x = shape[ndim - 1];
-  auto stride_x = strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  auto idx =
-      elem_to_loc(index_rest * shape_x, shape.data(), strides.data(), ndim);
-  auto in_vec =
-      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
-  }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 } // namespace cu
@@ -92,52 +58,33 @@ void copy_general_input(
             using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
             using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
             using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-            const InType* in_ptr = gpu_ptr<InType>(in) + offset_in;
-            OutType* out_ptr = gpu_ptr<OutType>(out) + offset_out;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
             int ndim = shape.size();
-            int work_per_thread = 1;
-            auto dim0 = ndim > 0 ? shape.back() : 1;
-            auto rest = out.size() / dim0;
-            if (dim0 >= 4) {
-              work_per_thread = 4;
-            }
-            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-            auto block_dims = get_block_dims(dim0, rest, 1);
-            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                auto kernel =
-                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 1>;
-                if (work_per_thread == 4) {
-                  kernel =
-                      cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 4>;
-                }
+                auto [num_blocks, block_dims] = get_launch_args(out, large());
                 encoder.add_kernel_node(
-                    kernel,
-                    {num_blocks_x, num_blocks_y},
+                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>,
+                    num_blocks,
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    rest,
+                    out.size(),
                     const_param<dims_constant()>(shape),
                     const_param<dims_constant()>(strides_in));
               });
             } else { // ndim >= 4
-              auto kernel = cu::copy_g<InType, OutType, IdxT, 1>;
-              if (work_per_thread == 4) {
-                kernel = cu::copy_g<InType, OutType, IdxT, 4>;
-              }
+              auto [num_blocks, block_dims] = get_launch_args(out, large());
               encoder.add_kernel_node(
-                  kernel,
-                  {num_blocks_x, num_blocks_y},
+                  cu::copy_g<InType, OutType, IdxT>,
+                  num_blocks,
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  rest,
+                  out.size(),
                   const_param(shape),
                   const_param(strides_in),
                   ndim);