implement batch rope for Metal

* Add type annotations to mlx methods

* Missing list_or_scalar

.circleci/config.yml

@@ -7,15 +7,9 @@ parameters:
   nightly_build:
     type: boolean
     default: false
-  weekly_build:
-    type: boolean
-    default: false
   test_release:
     type: boolean
     default: false
-  linux_release:
-    type: boolean
-    default: false
 
 jobs:
   build_documentation:
@@ -24,13 +18,14 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "16.2.0"
-    resource_class: m2pro.medium
+      xcode: "26.0.0"
+    resource_class: m4pro.medium
     steps:
       - checkout
       - run:
           name: Install
           command: |
+            xcodebuild -downloadComponent MetalToolchain
             brew install python@3.9
             brew install doxygen
             python3.9 -m venv env
@@ -38,7 +33,7 @@ jobs:
             pip install --upgrade pip
             pip install --upgrade cmake
             pip install -r docs/requirements.txt
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install . -v
+            pip install . -v
       - when:
           condition:
             not: << parameters.upload-docs >>
@@ -70,9 +65,9 @@ jobs:
                  git push -f origin gh-pages
 
   linux_build_and_test:
-    docker:
-      - image: cimg/python:3.9
-
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
@@ -84,37 +79,37 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
             sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
             sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
+            curl -LsSf https://astral.sh/uv/install.sh | sh
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py develop
+            uv venv
+            uv pip install cmake
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+              uv pip install -e ".[dev]" -v
       - run:
           name: Generate package stubs
           command: |
-            echo "stubs"
-            pip install typing_extensions
-            python setup.py generate_stubs 
+            uv pip install typing_extensions
+            uv run --no-project setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
-            python3 -m unittest discover python/tests -v
+            source .venv/bin/activate
+            python -m unittest discover python/tests -v
             mpirun --bind-to none -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build CPP only
           command: |
-            mkdir -p build && cd build 
+            source .venv/bin/activate
+            mkdir -p build && cd build
             cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
             make -j `nproc`
       - run:
@@ -125,7 +120,7 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       macosx_deployment_target:
         type: string
         default: ""
@@ -133,57 +128,56 @@ jobs:
       xcode: << parameters.xcode_version >>
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
-            brew install python@3.9
-            brew install openmpi
-            python3.9 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
-            pip install torch
-            pip install tensorflow
-            pip install unittest-xml-reporting
+            xcodebuild -downloadComponent MetalToolchain
+            HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
+              brew install openmpi uv
       - run:
           name: Install Python package
           command: |
-            source env/bin/activate
-            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-            CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              pip install -e . -v
+            uv venv --python 3.9
+            uv pip install \
+              nanobind==2.4.0 \
+              cmake \
+              numpy \
+              torch \
+              tensorflow \
+              unittest-xml-reporting
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+              uv pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
-            source env/bin/activate
-            pip install typing_extensions
-            python setup.py generate_stubs 
+            uv pip install typing_extensions
+            uv run --no-project setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
             mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build example extension
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             cd examples/extensions
-            pip install -r requirements.txt
-            python setup.py build_ext -j8
+            uv pip install -r requirements.txt
+            uv run --no-project setup.py build_ext --inplace
+            uv run --no-project python test.py
       - store_test_results:
           path: test-results
       - run:
           name: Build CPP only
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
       - run:
           name: Run CPP tests
@@ -192,7 +186,7 @@ jobs:
       - run:
           name: Build small binary
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             cd build/
             cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
               -DBUILD_SHARED_LIBS=ON \
@@ -204,37 +198,74 @@ jobs:
       - run:
           name: Run Python tests with JIT
           command: |
-            source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              pip install -e . -v
+            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
+              uv pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
-              python -m xmlrunner discover -v python/tests -o test-results/gpu_jit
+              uv run --no-project python -m xmlrunner discover \
+                -v python/tests \
+                -o test-results/gpu_jit
 
   cuda_build_and_test:
+    parameters:
+      image_date:
+        type: string
+        default: "2023.11.1"
     machine:
-      image: linux-cuda-12:default
+      image: "linux-cuda-12:<< parameters.image_date >>"
       resource_class: gpu.nvidia.small.gen2
     steps:
       - checkout
+      - restore_cache:
+          keys:
+            - cuda-<< parameters.image_date >>-{{ arch }}-
+      - run:
+          name: Install dependencies
+          command: |
+            sudo apt-get update
+            sudo apt-get install libcudnn9-dev-cuda-12
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install libnccl2 libnccl-dev
+            curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
+            sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
+            rm -rf ccache-4.11.3-linux-x86_64
+            curl -LsSf https://astral.sh/uv/install.sh | sh
       - run:
           name: Install Python package
           command: |
-            sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-            python -m venv env
-            source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              pip install -e ".[dev]"
+            uv venv
+            uv pip install cmake
+            DEBUG=1 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              uv pip install -e ".[dev]" -v
       - run:
           name: Run Python tests
           command: |
-            source env/bin/activate
+            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 }}
+          paths:
+            - /home/circleci/.cache/ccache
 
   build_release:
     parameters:
@@ -243,7 +274,7 @@ jobs:
         default: "3.9"
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       build_env:
         type: string
         default: ""
@@ -252,7 +283,7 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
     steps:
@@ -260,11 +291,15 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            brew install python@<< parameters.python_version >>
-            brew install openmpi
-            python<< parameters.python_version >> -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
+            xcodebuild -downloadComponent MetalToolchain
+            mkdir -p ~/miniconda3
+            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
+            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+            rm ~/miniconda3/miniconda.sh
+            source ~/miniconda3/bin/activate
+            conda init --all
+            conda create -n env python=<< parameters.python_version >> -y
+            conda activate env
             pip install --upgrade cmake
             pip install nanobind==2.4.0
             pip install --upgrade setuptools
@@ -274,30 +309,38 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            source env/bin/activate
+            conda activate env
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
-            source env/bin/activate
+            conda activate env
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
-            source env/bin/activate
-            << parameters.build_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              python -m build -w
+            conda activate env
+            python setup.py clean --all
+            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
+      - when:
+          condition:
+            equal: ["3.9", << parameters.python_version >>]
+          steps:
+            - run:
+                name: Build common package
+                command: |
+                  conda activate env
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
           condition: << parameters.build_env >>
           steps:
             - run:
                 name: Upload package
                 command: |
-                  source env/bin/activate
+                  conda activate env
                   twine upload dist/*
       - store_artifacts:
           path: dist/
@@ -307,52 +350,100 @@ jobs:
       python_version:
         type: string
         default: "3.9"
-      extra_env:
+      build_env:
         type: string
-        default: "DEV_RELEASE=1"
-    docker:
-      - image: ubuntu:20.04
+        default: ""
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
           name: Build wheel
           command: |
             PYTHON=python<< parameters.python_version >>
-            apt-get update
-            apt-get upgrade -y
-            DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata
-            apt-get install -y apt-utils
-            apt-get install -y software-properties-common
-            add-apt-repository -y ppa:deadsnakes/ppa
-            apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            apt-get install -y build-essential git
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
+            sudo apt-get update
+            TZ=Etc/UTC sudo apt-get -y install tzdata
+            sudo add-apt-repository -y ppa:deadsnakes/ppa
+            sudo apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
             $PYTHON -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              pip install . -v
+            << parameters.build_env >> pip install ".[dev]" -v
             pip install typing_extensions
-            python setup.py generate_stubs 
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python -m build --wheel
-            auditwheel show dist/*
-            auditwheel repair dist/* --plat manylinux_2_31_x86_64
+            python setup.py generate_stubs
+            python setup.py clean --all
+            MLX_BUILD_STAGE=1 << parameters.build_env >> python -m build -w
+            bash python/scripts/repair_linux.sh
+      - when:
+          condition:
+            equal: ["3.9", << parameters.python_version >>]
+          steps:
+            - run:
+                name: Build common package
+                command: |
+                  source env/bin/activate
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 \
+                    python -m build -w
+                  auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_x86_64
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload packages
+                command: |
+                  source env/bin/activate
+                  twine upload wheelhouse/*.whl
+      - store_artifacts:
+          path: wheelhouse/
+
+  build_cuda_release:
+    parameters:
+      build_env:
+        type: string
+        default: ""
+    machine:
+      image: ubuntu-2204:current
+      resource_class: xlarge
+    steps:
+      - checkout
       - run:
-          name: Upload package
+          name: Build wheel
           command: |
-            source env/bin/activate
-            twine upload wheelhouse/*
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
+            sudo apt-get update
+            sudo apt-get install cuda-toolkit-12-9 libcudnn9-dev-cuda-12
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install zip
+            pip install auditwheel
+            pip install patchelf
+            pip install build
+            pip install twine
+            export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
+            export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
+            << parameters.build_env >> MLX_BUILD_STAGE=2 \
+              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              python -m build -w
+            bash python/scripts/repair_cuda.sh
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload package
+                command: |
+                  twine upload wheelhouse/*.whl
       - store_artifacts:
           path: wheelhouse/
 
@@ -364,22 +455,23 @@ workflows:
             pattern: "^(?!pull/)[-\\w]+$"
             value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
         - not: << pipeline.parameters.test_release >>
     jobs:
       - mac_build_and_test:
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test
-      - cuda_build_and_test 
+      - cuda_build_and_test:
+          matrix:
+            parameters:
+              image_date: ["2023.11.1", "2025.05.1"]
       - build_documentation 
 
   build_pypi_release:
     when:
       and:
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
         - not: << pipeline.parameters.test_release >>
     jobs:
       - build_release:
@@ -393,68 +485,7 @@ workflows:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
+              xcode_version: ["26.0.0"]
       - build_documentation:
           filters:
             tags:
@@ -462,6 +493,25 @@ workflows:
             branches:
               ignore: /.*/
           upload-docs: true
+      - build_linux_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              build_env: ["PYPI_RELEASE=1"]
+      - build_cuda_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              build_env: ["PYPI_RELEASE=1"]
 
   prb:
     when:
@@ -477,11 +527,14 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test:
           requires: [ hold ]
       - cuda_build_and_test:
           requires: [ hold ]
+          matrix:
+            parameters:
+              image_date: ["2023.11.1", "2025.05.1"]
   nightly_build:
     when:
       and:
@@ -493,58 +546,18 @@ workflows:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-  weekly_build:
+              xcode_version: ["26.0.0"]
+      - build_linux_release:
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+      - build_cuda_release
+
+  build_dev_release:
     when:
       and:
         - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.weekly_build >>
+        - << pipeline.parameters.test_release >>
     jobs:
       - build_release:
           matrix:
@@ -552,76 +565,13 @@ workflows:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-  linux_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.linux_release >>
-    jobs:
+              xcode_version: ["26.0.0"]
       - build_linux_release:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]
+              build_env: ["DEV_RELEASE=1"]
+      - build_cuda_release:
+          matrix:
+            parameters:
+              build_env: ["DEV_RELEASE=1"]

ACKNOWLEDGMENTS.md

@@ -19,11 +19,17 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
 
 <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

@@ -41,7 +41,9 @@ option(MLX_BUILD_GGUF "Include support for GGUF format" ON)
 option(MLX_BUILD_SAFETENSORS "Include support for safetensors format" ON)
 option(MLX_BUILD_BLAS_FROM_SOURCE "Build OpenBLAS from source code" OFF)
 option(MLX_METAL_JIT "Use JIT compilation for Metal kernels" OFF)
+option(MLX_USE_CCACHE "Use CCache for compilation cache when available" ON)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
+option(USE_SYSTEM_FMT "Use system's provided fmt library" OFF)
 
 # --------------------- Processor tests -------------------------
 message(
@@ -64,10 +66,17 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
       message(WARNING "Building for x86_64 arch is not officially supported.")
     endif()
   endif()
-
 else()
   set(MLX_BUILD_METAL OFF)
-  message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
+endif()
+
+if(MLX_USE_CCACHE)
+  find_program(CCACHE_PROGRAM ccache)
+  if(CCACHE_PROGRAM)
+    set(CMAKE_C_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
+    set(CMAKE_CXX_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
+    set(CMAKE_CUDA_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
+  endif()
 endif()
 
 # ----------------------------- Lib -----------------------------
@@ -131,6 +140,12 @@ elseif(MLX_BUILD_METAL)
   target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
 endif()
 
+if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+  # With newer clang/gcc versions following libs are implicitly linked, but when
+  # building on old distributions they need to be explicitly listed.
+  target_link_libraries(mlx PRIVATE dl pthread)
+endif()
+
 if(WIN32)
   if(MSVC)
     # GGUF does not build with MSVC.
@@ -234,12 +249,16 @@ target_include_directories(
 # Do not add mlx_EXPORTS define for shared library.
 set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")
 
-FetchContent_Declare(
-  fmt
-  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
-  GIT_TAG 10.2.1
-  EXCLUDE_FROM_ALL)
-FetchContent_MakeAvailable(fmt)
+if(USE_SYSTEM_FMT)
+  find_package(fmt REQUIRED)
+else()
+  FetchContent_Declare(
+    fmt
+    GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+    GIT_TAG 10.2.1
+    EXCLUDE_FROM_ALL)
+  FetchContent_MakeAvailable(fmt)
+endif()
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:fmt::fmt-header-only>)
 
 if(MLX_BUILD_PYTHON_BINDINGS)

README.md

@@ -11,10 +11,10 @@ brought to you by Apple machine learning research.
 
 Some key features of MLX include:
 
- - **Familiar APIs**: MLX has a Python API that closely follows NumPy.  MLX
+ - **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
    also has fully featured C++, [C](https://github.com/ml-explore/mlx-c), and
    [Swift](https://github.com/ml-explore/mlx-swift/) APIs, which closely mirror
-   the Python API.  MLX has higher-level packages like `mlx.nn` and
+   the Python API. MLX has higher-level packages like `mlx.nn` and
    `mlx.optimizers` with APIs that closely follow PyTorch to simplify building
    more complex models.
 
@@ -68,18 +68,23 @@ in the documentation.
 
 ## Installation
 
-MLX is available on [PyPI](https://pypi.org/project/mlx/). To install the Python API, run:
+MLX is available on [PyPI](https://pypi.org/project/mlx/). To install MLX on
+macOS, run:
 
-**With `pip`**:
-
-```
+```bash
 pip install mlx
 ```
 
-**With `conda`**:
+To install the CUDA backend on Linux, run:
 
+```bash
+pip install mlx[cuda]
 ```
-conda install -c conda-forge mlx
+
+To install a CPU-only Linux package, run:
+
+```bash
+pip install mlx[cpu]
 ```
 
 Checkout the

benchmarks/cpp/single_ops.cpp

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

benchmarks/python/comparative/bench_torch.py

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

benchmarks/python/single_ops.py

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

cmake/FindNCCL.cmake

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

docs/requirements.txt

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

docs/src/conf.py

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

docs/src/dev/custom_metal_kernels.rst

@@ -127,7 +127,8 @@ relying on a copy from ``ensure_row_contiguous``:
       name="myexp_strided",
       input_names=["inp"],
       output_names=["out"],
-      source=source
+      source=source,
+      ensure_row_contiguous=False,
   )
 
   def exp_elementwise(a: mx.array):
@@ -138,7 +139,6 @@ 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/dev/extensions.rst

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

docs/src/index.rst

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

docs/src/install.rst

@@ -13,7 +13,7 @@ silicon computer is
 
     pip install mlx
 
-To install from PyPI you must meet the following requirements:
+To install from PyPI your system must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
 - Using a native Python >= 3.9
@@ -23,12 +23,39 @@ To install from PyPI you must meet the following requirements:
     MLX is only available on devices running macOS >= 13.5
     It is highly recommended to use macOS 14 (Sonoma)
 
+CUDA
+^^^^
 
-MLX is also available on conda-forge. To install MLX with conda do:
+MLX has a CUDA backend which you can install with:
 
 .. code-block:: shell
 
-   conda install conda-forge::mlx
+    pip install mlx[cuda]
+
+To install the CUDA package from PyPi your system must meet the following
+requirements:
+
+- Nvidia architecture >= SM 7.0 (Volta)
+- Nvidia driver >= 550.54.14
+- CUDA toolkit >= 12.0
+- Linux distribution with glibc >= 2.35
+- Python >= 3.9
+
+
+CPU-only (Linux)
+^^^^^^^^^^^^^^^^
+
+For a CPU-only version of MLX that runs on Linux use:
+
+.. code-block:: shell
+
+    pip install mlx[cpu]
+
+To install the CPU-only package from PyPi your system must meet the following
+requirements:
+
+- Linux distribution with glibc >= 2.35
+- Python >= 3.9
 
 
 Troubleshooting
@@ -65,6 +92,8 @@ Build Requirements
 Python API
 ^^^^^^^^^^
 
+.. _python install:
+
 To build and install the MLX python library from source, first, clone MLX from
 `its GitHub repo <https://github.com/ml-explore/mlx>`_:
 
@@ -76,20 +105,20 @@ Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
+  pip install .
 
 For developing, install the package with development dependencies, and use an
 editable install:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
+  pip install -e ".[dev]"
 
 Once the development dependencies are installed, you can build faster with:
 
 .. code-block:: shell
 
- CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
+ python setup.py build_ext --inplace
 
 Run the tests with:
 
@@ -107,6 +136,8 @@ IDE:
 C++ API
 ^^^^^^^
 
+.. _cpp install:
+
 Currently, MLX must be built and installed from source.
 
 Similarly to the python library, to build and install the MLX C++ library start
@@ -185,6 +216,7 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
+
 Binary Size Minimization
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -213,6 +245,50 @@ be anwywhere from a few hundred millisecond to a few seconds depending on the
 application. Once a kernel is compiled, it will be cached by the system. The
 Metal kernel cache persists across reboots.
 
+Linux
+^^^^^
+
+To build from source on Linux (CPU only), install the BLAS and LAPACK headers.
+For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   apt-get update -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+
+From here follow the instructions to install either the :ref:`Python <python
+install>` or :ref:`C++ <cpp install>` APIs.
+
+CUDA
+^^^^
+
+To build from source on Linux with CUDA, install the BLAS and LAPACK headers
+and the CUDA toolkit. For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+   dpkg -i cuda-keyring_1.1-1_all.deb
+   apt-get update -y
+   apt-get -y install cuda-toolkit-12-9
+   apt-get install libblas-dev liblapack-dev liblapacke-dev libcudnn9-dev-cuda-12 -y
+
+
+When building either the Python or C++ APIs make sure to pass the cmake flag
+``MLX_BUILD_CUDA=ON``. For example, to build the Python API run:
+
+.. code-block:: shell
+
+  CMAKE_ARGS="-DMLX_BUILD_CUDA=ON" pip install -e ".[dev]"
+
+To build the C++ package run:
+
+.. code-block:: shell
+
+   mkdir -p build && cd build
+   cmake .. -DMLX_BUILD_CUDA=ON && make -j
+
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 

docs/src/python/cuda.rst

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

docs/src/python/fast.rst

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

docs/src/python/optimizers.rst

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

docs/src/python/optimizers/common_optimizers.rst

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

docs/src/usage/compile.rst

@@ -225,7 +225,7 @@ In some cases returning updated state can be pretty inconvenient. Hence,
   def fun(x, y):
       z = x + y
       state.append(z)
-      return mx.exp(z), state
+      return mx.exp(z)
 
   fun(mx.array(1.0), mx.array(2.0))
   # Prints [array(3, dtype=float32)]

docs/src/usage/export.rst

@@ -7,17 +7,17 @@ Exporting Functions
 
 MLX has an API to export and import functions to and from a file. This lets you
 run computations written in one MLX front-end (e.g. Python) in another MLX
-front-end (e.g. C++). 
+front-end (e.g. C++).
 
 This guide walks through the basics of the MLX export API with some examples.
 To see the full list of functions check-out the :ref:`API documentation
 <export>`.
 
-Basics of Exporting 
+Basics of Exporting
 -------------------
 
 Let's start with a simple example:
- 
+
 .. code-block:: python
 
   def fun(x, y):
@@ -67,7 +67,7 @@ specified as variable positional arguments or as a tuple of arrays:
 
   x = mx.array(1.0)
   y = mx.array(1.0)
-   
+
   # Both arguments to fun are positional
   mx.export_function("add.mlxfn", fun, x, y)
 
@@ -133,7 +133,7 @@ parameters are also saved to the ``model.mlxfn`` file.
    For enclosed arrays inside an exported function, be extra careful to ensure
    they are evaluated. The computation graph that gets exported will include
    the computation that produces enclosed inputs.
-  
+
    If the above example was missing ``mx.eval(model.parameters()``, the
    exported function would include the random initialization of the
    :obj:`mlx.nn.Module` parameters.
@@ -150,8 +150,8 @@ parameters, pass them as inputs to the ``call`` wrapper:
      # Set the model's parameters to the input parameters
      model.update(tree_unflatten(list(params.items())))
      return model(x)
- 
-   params = dict(tree_flatten(model.parameters()))
+
+   params = tree_flatten(model.parameters(), destination={})
    mx.export_function("model.mlxfn", call, (mx.zeros(4),), params)
 
 
@@ -169,8 +169,8 @@ to export a function which can be used for inputs with variable shapes:
 
   # Ok
   out, = imported_abs(mx.array(-1.0))
-  
-  # Also ok 
+
+  # Also ok
   out, = imported_abs(mx.array([-1.0, -2.0]))
 
 With ``shapeless=False`` (which is the default), the second call to
@@ -197,7 +197,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
   def fun(x, y=None):
       constant = mx.array(3.0)
       if y is not None:
-        x += y 
+        x += y
       return x + constant
 
   with mx.exporter("fun.mlxfn", fun) as exporter:
@@ -215,7 +215,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
   print(out)
 
 In the above example the function constant data, (i.e. ``constant``), is only
-saved once. 
+saved once.
 
 Transformations with Imported Functions
 ---------------------------------------
@@ -238,7 +238,7 @@ on imported functions just like regular Python functions:
   # Prints: array(1, dtype=float32)
   print(dfdx(x))
 
-  # Compile the imported function 
+  # Compile the imported function
   mx.compile(imported_fun)
   # Prints: array(0, dtype=float32)
   print(compiled_fun(x)[0])
@@ -275,7 +275,7 @@ Import and run the function in C++ with only a few lines of code:
   // Prints: array(2, dtype=float32)
   std::cout << outputs[0] << std::endl;
 
-Imported functions can be transformed in C++ just like in Python. Use 
+Imported functions can be transformed in C++ just like in Python. Use
 ``std::vector<mx::array>`` for positional arguments and ``std::map<std::string,
 mx::array>`` for keyword arguments when calling imported functions in C++.
 

docs/src/usage/indexing.rst

@@ -107,8 +107,20 @@ 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:
 
-Note, unlike NumPy, updates to the same location are nondeterministic:
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> b = a[:]
+  >>> b[2] = 0
+  >>> b
+  array([1, 2, 0], dtype=int32)
+  >>> a
+  array([1, 2, 3], dtype=int32)
+
+Also unlike NumPy, updates to the same location are nondeterministic:
 
 .. code-block:: shell
 

examples/extensions/axpby/axpby.cpp

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

examples/extensions/axpby/axpby.h

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

examples/extensions/requirements.txt

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

examples/extensions/test.py

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

mlx/array.h

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

mlx/backend/common/compiled.cpp

@@ -14,6 +14,8 @@ void print_constant(std::ostream& os, const array& x) {
       return print_float_constant<float16_t>(os, x);
     case bfloat16:
       return print_float_constant<bfloat16_t>(os, x);
+    case float64:
+      return print_float_constant<double>(os, x);
     case complex64:
       return print_complex_constant<complex64_t>(os, x);
     case int8:
@@ -50,6 +52,8 @@ std::string get_type_string(Dtype d) {
       return "float16_t";
     case bfloat16:
       return "bfloat16_t";
+    case float64:
+      return "double";
     case complex64:
       return "complex64_t";
     case bool_:

mlx/backend/common/compiled.h

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

mlx/backend/common/matmul.h

@@ -12,16 +12,11 @@ namespace mlx::core {
 inline std::tuple<Shape, Strides, Strides> collapse_batches(
     const array& a,
     const array& b) {
-  // Get and check the shape for the batched dims
-  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
-  Shape B_bshape{b.shape().begin(), b.shape().end() - 2};
-  if (A_bshape != B_bshape) {
-    std::ostringstream msg;
-    msg << "[matmul] Got matrices with incorrectly broadcasted shapes: " << "A "
-        << a.shape() << ", B " << b.shape() << ".";
-    throw std::runtime_error(msg.str());
+  if (a.ndim() == 2) {
+    return {{1}, {0}, {0}};
   }
 
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
   Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
   Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
 
@@ -42,17 +37,11 @@ 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) {
-  // Get and check the shape for the batched dims
-  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
-  Shape B_bshape{b.shape().begin(), b.shape().end() - 2};
-  Shape C_bshape{c.shape().begin(), c.shape().end() - 2};
-  if (A_bshape != B_bshape || A_bshape != C_bshape) {
-    std::ostringstream msg;
-    msg << "[addmm] Got matrices with incorrectly broadcasted shapes: " << "A "
-        << a.shape() << ", B " << b.shape() << ", B " << c.shape() << ".";
-    throw std::runtime_error(msg.str());
+  if (a.ndim() == 2) {
+    return {{1}, {0}, {0}, {0}};
   }
 
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
   Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
   Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
   Strides C_bstride{c.strides().begin(), c.strides().end() - 2};

mlx/backend/common/reduce.cpp

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

mlx/backend/common/reduce.h

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

mlx/backend/common/utils.cpp

@@ -1,14 +1,20 @@
 // Copyright © 2023-2024 Apple Inc.
 
+#include <dlfcn.h>
+
 #include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive) {
-  std::ostringstream op_t;
-  primitive->print(op_t);
-  return op_t.str();
+std::filesystem::path current_binary_dir() {
+  static std::filesystem::path binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path();
+  }();
+  return binary_dir;
 }
 
 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
@@ -199,12 +205,15 @@ Dims get_2d_grid_dims_common(
       }
     }
   }
-  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX || divisor > 1) {
+  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX) {
     throw std::runtime_error("Unable to safely factor shape.");
   }
   if (grid_y > grid_x) {
     std::swap(grid_x, grid_y);
   }
+  if (divisor > 1) {
+    grid_x = ((grid_x + divisor - 1) / divisor) * divisor;
+  }
   return std::make_tuple(
       static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
 }

mlx/backend/common/utils.h

@@ -2,6 +2,7 @@
 
 #pragma once
 
+#include <filesystem>
 #include <tuple>
 #include <vector>
 
@@ -9,7 +10,8 @@
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive);
+// Return the directory that contains current shared library.
+std::filesystem::path current_binary_dir();
 
 inline int64_t
 elem_to_loc(int elem, const Shape& shape, const Strides& strides) {
@@ -195,7 +197,7 @@ void shared_buffer_reshape(
     array& out);
 
 template <typename T>
-inline std::vector<T> remove_index(std::vector<T> vec, size_t index) {
+inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
   vec.erase(std::next(vec.begin(), index));
   return vec;
 }

mlx/backend/cpu/cholesky.cpp

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

mlx/backend/cpu/compiled.cpp

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

mlx/backend/cpu/conv.cpp

@@ -883,7 +883,7 @@ void explicit_gemm_conv_1D_cpu(
   // Fill with zeros
   std::vector<array> temps;
   temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = padding_lo[0] * in_padded.strides()[1];
@@ -895,7 +895,7 @@ void explicit_gemm_conv_1D_cpu(
       in_padded_slice.size(),
       data_offset);
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -920,7 +920,7 @@ void explicit_gemm_conv_1D_cpu(
   // Materialize strided view
   Shape strided_reshape = {N * oH, wH * C};
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -938,13 +938,13 @@ void explicit_gemm_conv_1D_cpu(
         wt.size(),
         0);
     gemm_wt = array(wt_transpose.shape(), float32, nullptr, {});
-    copy(wt_transpose, gemm_wt, CopyType::General, stream);
+    copy_cpu(wt_transpose, gemm_wt, CopyType::General, stream);
     temps.push_back(gemm_wt);
   } else if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
@@ -991,7 +991,7 @@ void explicit_gemm_conv_1D_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }
@@ -1029,7 +1029,7 @@ void explicit_gemm_conv_2D_cpu(
   // Fill with zeros
   std::vector<array> temps;
   temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = padding_lo[0] * in_padded.strides()[1] +
@@ -1044,7 +1044,7 @@ void explicit_gemm_conv_2D_cpu(
   temps.push_back(in_padded_slice);
 
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
 
   // Make strided view
   Shape strided_shape = {N, oH, oW, wH, wW, C};
@@ -1065,7 +1065,7 @@ void explicit_gemm_conv_2D_cpu(
   // Materialize strided view
   Shape strided_reshape = {N * oH * oW, wH * wW * C};
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -1076,7 +1076,7 @@ void explicit_gemm_conv_2D_cpu(
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
@@ -1116,7 +1116,7 @@ void explicit_gemm_conv_2D_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }
@@ -1156,7 +1156,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Fill with zeros
   std::vector<array> temps = {array(0, conv_dtype)};
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = 0;
@@ -1173,7 +1173,7 @@ void explicit_gemm_conv_ND_cpu(
       data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -1212,7 +1212,7 @@ void explicit_gemm_conv_ND_cpu(
   }
 
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -1223,13 +1223,13 @@ void explicit_gemm_conv_ND_cpu(
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
   if (flip) {
     auto gemm_wt_ = array(gemm_wt.shape(), float32, nullptr, {});
-    copy(gemm_wt, gemm_wt_, CopyType::Vector, stream);
+    copy_cpu(gemm_wt, gemm_wt_, CopyType::Vector, stream);
     temps.push_back(gemm_wt_);
 
     // Calculate the total size of the spatial dimensions
@@ -1284,7 +1284,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }

mlx/backend/cpu/copy.cpp

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

mlx/backend/cpu/copy.h

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

mlx/backend/cpu/distributed.cpp

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

mlx/backend/cpu/eig.cpp

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

mlx/backend/cpu/eigh.cpp

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

mlx/backend/cpu/hadamard.cpp

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

mlx/backend/cpu/indexing.cpp

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

mlx/backend/cpu/inverse.cpp

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

mlx/backend/cpu/jit_compiler.cpp

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

mlx/backend/cpu/lapack.h

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

mlx/backend/cpu/logsumexp.cpp

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

mlx/backend/cpu/luf.cpp

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

mlx/backend/cpu/masked_mm.cpp

@@ -6,6 +6,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/gemm.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
@@ -52,6 +53,58 @@ inline void mask_matrix(
   }
 }
 
+template <typename T>
+inline void segmented_mm(
+    const T* a,
+    const T* b,
+    const uint32_t* segments,
+    T* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides,
+    size_t num_segments,
+    const Shape& segments_shape,
+    const Strides& segments_strides) {
+  int ndim = a_shape.size();
+  Shape a_copy = a_shape;
+  Shape b_copy = b_shape;
+  int32_t M = a_copy[ndim - 2];
+  int32_t N = b_copy[ndim - 1];
+  for (int i = 0; i < num_segments; i++) {
+    uint32_t k_start =
+        segments[elem_to_loc(2 * i, segments_shape, segments_strides)];
+    uint32_t k_end =
+        segments[elem_to_loc(2 * i + 1, segments_shape, segments_strides)];
+    if (k_end <= k_start) {
+      std::fill_n(out + i * M * N, M * N, T(0));
+      continue;
+    }
+    a_copy[ndim - 1] = k_end - k_start;
+    b_copy[ndim - 2] = k_end - k_start;
+    matmul<T>(
+        a + k_start * a_strides[ndim - 1],
+        b + k_start * b_strides[ndim - 2],
+        out + i * M * N,
+        a_transposed,
+        b_transposed,
+        lda,
+        ldb,
+        N,
+        1.0,
+        0.0,
+        1,
+        a_copy,
+        a_strides,
+        b_copy,
+        b_strides);
+  }
+}
+
 } // namespace
 
 void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -71,21 +124,20 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
         if (!expand_all && stx == arr.shape(-1) && sty == 1) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(false, stx, arr_copy, true);
           }
           return std::make_tuple(false, stx, arr, false);
         } else if (!expand_all && stx == 1 && sty == arr.shape(-2)) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(true, sty, arr_copy, true);
           }
           return std::make_tuple(true, sty, arr, false);
         } else {
-          array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy(arr, arr_copy, CopyType::General, s);
           int64_t stx = arr.shape(-1);
+          array arr_copy = contiguous_copy_cpu(arr, s);
           return std::make_tuple(false, stx, arr_copy, true);
         }
       };
@@ -333,7 +385,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       int64_t stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -437,4 +489,121 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   encoder.add_temporaries(std::move(temps));
 }
 
+void SegmentedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto check_transpose = [&s, &encoder](const array& x) {
+    auto stx = x.strides()[x.ndim() - 2];
+    auto sty = x.strides()[x.ndim() - 1];
+    if (stx == x.shape(-1) && sty == 1) {
+      return std::make_tuple(false, stx, x);
+    } else if (stx == 1 && sty == x.shape(-2)) {
+      return std::make_tuple(true, sty, x);
+    } else {
+      array xc(x.shape(), x.dtype(), nullptr, {});
+      copy_cpu(x, xc, CopyType::General, s);
+      encoder.add_temporary(xc);
+      int64_t stx = x.shape(-1);
+      return std::make_tuple(false, stx, xc);
+    }
+  };
+
+  auto [a_transposed, lda, a] = check_transpose(inputs[0]);
+  auto [b_transposed, ldb, b] = check_transpose(inputs[1]);
+  auto& segments = inputs[2];
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(segments);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    segments = array::unsafe_weak_copy(segments),
+                    out_ptr = out.data<void>(),
+                    a_transposed = a_transposed,
+                    b_transposed = b_transposed,
+                    lda = lda,
+                    ldb = ldb]() {
+    switch (a.dtype()) {
+      case float64:
+        segmented_mm<double>(
+            a.data<double>(),
+            b.data<double>(),
+            segments.data<uint32_t>(),
+            static_cast<double*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float32:
+        segmented_mm<float>(
+            a.data<float>(),
+            b.data<float>(),
+            segments.data<uint32_t>(),
+            static_cast<float*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float16:
+        segmented_mm<float16_t>(
+            a.data<float16_t>(),
+            b.data<float16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<float16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case bfloat16:
+        segmented_mm<bfloat16_t>(
+            a.data<bfloat16_t>(),
+            b.data<bfloat16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<bfloat16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      default:
+        throw std::invalid_argument(
+            "Segmented mm supports only real float types.");
+    }
+  });
+}
+
 } // namespace mlx::core

mlx/backend/cpu/matmul.cpp

@@ -81,7 +81,7 @@ void matmul_general(
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, stream);
+      copy_cpu(arr, temps.back(), CopyType::General, stream);
       stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -142,7 +142,7 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   CopyType ctype = c.data_size() == 1
       ? CopyType::Scalar
       : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-  copy(c, out, ctype, stream());
+  copy_cpu(c, out, ctype, stream());
   if (inputs[0].shape(-1) == 0) {
     return;
   }

mlx/backend/cpu/primitives.cpp

@@ -22,7 +22,7 @@ void reshape(const array& in, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
     out.set_data(allocator::malloc(out.nbytes()));
-    copy_inplace(in, out, CopyType::General, out.primitive().stream());
+    copy_cpu_inplace(in, out, CopyType::General, out.primitive().stream());
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -175,7 +175,7 @@ void AsType::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -198,7 +198,7 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
     size_t data_offset = strides[axis_] * sizes[i];
     out_slice.copy_shared_buffer(
         out, strides, flags, out_slice.size(), data_offset);
-    copy_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
+    copy_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
   }
 }
 
@@ -211,7 +211,7 @@ void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
        (allow_col_major_ && in.flags().col_contiguous))) {
     out.copy_shared_buffer(in);
   } else {
-    copy(in, out, CopyType::General, stream());
+    copy_cpu(in, out, CopyType::General, stream());
   }
 }
 
@@ -235,7 +235,7 @@ void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   } else {
     ctype = CopyType::General;
   }
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -251,7 +251,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
 
   // Fill output with val
-  copy(val, out, CopyType::Scalar, stream());
+  copy_cpu(val, out, CopyType::Scalar, stream());
 
   // Find offset for start of input values
   size_t data_offset = 0;
@@ -266,7 +266,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
       out, out.strides(), out.flags(), out_slice.size(), data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
+  copy_cpu_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
 }
 
 void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -340,7 +340,7 @@ void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc(out.nbytes()));
   auto [in_offset, donated] =
       compute_dynamic_offset(inputs[1], in.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ in,
       /* array& dst = */ out,
       /* const Shape& data_shape = */ out.shape(),
@@ -372,11 +372,11 @@ void DynamicSliceUpdate::eval_cpu(
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   auto [out_offset, donated] =
       compute_dynamic_offset(inputs[2], out.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -412,14 +412,14 @@ void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   // Calculate out strides, initial offset and if copy needs to be made
   auto [data_offset, out_strides] =
       prepare_slice(out, start_indices_, strides_);
 
   // Do copy
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -456,9 +456,9 @@ void View::eval_cpu(const std::vector<array>& inputs, array& out) {
     if (in.dtype() == bool_) {
       auto in_tmp = array(in.shape(), uint8, nullptr, {});
       in_tmp.copy_shared_buffer(in);
-      copy_inplace(in_tmp, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in_tmp, tmp, CopyType::General, stream());
     } else {
-      copy_inplace(in, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in, tmp, CopyType::General, stream());
     }
 
     auto flags = out.flags();

mlx/backend/cpu/qrf.cpp

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

mlx/backend/cpu/quantized.cpp

@@ -1,7 +1,5 @@
 // Copyright © 2023 Apple Inc.
 
-#include <cassert>
-
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
@@ -13,6 +11,35 @@ namespace mlx::core {
 
 namespace {
 
+const static float MXFP4_LUT[16] = {
+    +0.0f,
+    +0.5f,
+    +1.0f,
+    +1.5f,
+    +2.0f,
+    +3.0f,
+    +4.0f,
+    +6.0f,
+    -0.0f,
+    -0.5f,
+    -1.0f,
+    -1.5f,
+    -2.0f,
+    -3.0f,
+    -4.0f,
+    -6.0f};
+
+template <typename T>
+static inline T dequantize_scale(uint8_t s) {
+  using FOrI = union {
+    bfloat16_t f;
+    uint16_t i;
+  };
+  FOrI out;
+  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
+  return static_cast<T>(out.f);
+}
+
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
   return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -407,6 +434,231 @@ 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 bytes_per_pack = get_bytes_per_pack(4);
+  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 bytes_per_pack = get_bytes_per_pack(4);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      const T* x_local = x;
+      T sum = 0;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        T gsum = 0;
+        for (int kw = 0; kw < packs_in_group; kw++) {
+          uint8_t wi = *w_local++;
+#pragma clang loop unroll(full)
+          for (int p = 0; p < pack_factor; p++) {
+            gsum += (*x_local++) * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+        sum += scale * gsum;
+      }
+      *result = sum;
+      result++;
+    }
+
+    x += K;
+  }
+}
+
+template <int S>
+simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
+  if constexpr (S == 8) {
+    constexpr std::array<uint32_t, 8> shifts_ = {{0, 4, 8, 12, 16, 20, 24, 28}};
+    auto shifts(*(simd::Simd<uint32_t, S>*)&shifts_);
+    auto wi = simd::Simd<uint32_t, S>(*w);
+    wi = wi >> shifts;
+    wi = wi & 0xf;
+    simd::Simd<float, S> w_out;
+    for (int i = 0; i < S; ++i) {
+      w_out[i] = MXFP4_LUT[wi[i]];
+    }
+    return w_out;
+  } else {
+    // Appease compiler.. but should never get here
+    throw std::runtime_error("Unsupported combination for simd qmm.");
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t_simd(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = 32 / 4;
+  constexpr int packs_in_group = group_size / pack_factor;
+  constexpr int S = simd::max_size<T>;
+  static_assert(
+      S % pack_factor == 0, "SIMD size must be divisible by pack factor");
+  constexpr int packs_per_simd = S / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint32_t* w_local = w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      simd::Simd<float, S> acc(0);
+      auto x_local = x;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        simd::Simd<float, S> g_acc(0);
+        for (int kw = 0; kw < packs_in_group; kw += packs_per_simd) {
+          // Extract bits
+          auto wf = mxfp4_extract_bits_simd<S>(w_local);
+          w_local += packs_per_simd;
+          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
+          g_acc = g_acc + x_simd * wf;
+          x_local += S;
+        }
+        acc = acc + scale * g_acc;
+      }
+
+      *result = T(simd::sum(acc));
+      result++;
+    }
+    x += K;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_transpose(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K,
+    bool transposed_w) {
+  if (transposed_w) {
+    // the simd size must be a multiple of the number of elements per word
+    if constexpr (simd::max_size<T> % 8 == 0) {
+      mxfp4_qmm_t_simd<T>(result, x, w, scales, M, N, K);
+    } else {
+      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
+    }
+  } else {
+    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.ndim() > 1 ? x.shape(-2) : 1;
+  int N = out.shape(-1);
+  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
+  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
+  int batch_size = x.size() / (K * M);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  for (int i = 0; i < batch_size; i++) {
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
+        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case bfloat16:
+      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float16:
+      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float32:
+      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 template <typename T>
 void _bs_qmm_dispatch_typed(
     array& out,
@@ -513,115 +765,198 @@ void _bs_qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_bs_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.shape(-2);
+  int N = out.shape(-1);
+
+  int w_els = w.shape(-1) * w.shape(-2);
+  int g_els = scales.shape(-1) * scales.shape(-2);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
+  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
+
+  for (int i = 0; i < lhs_indices.size(); i++) {
+    int x_idx = lhs_indices_ptr[elem_to_loc(
+        i, lhs_indices.shape(), lhs_indices.strides())];
+    int w_idx = rhs_indices_ptr[elem_to_loc(
+        i, rhs_indices.shape(), rhs_indices.strides())];
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
+        scales_ptr +
+            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case float32:
+      mxfp4_bs_qmm_dispatch_typed<float>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case float16:
+      mxfp4_bs_qmm_dispatch_typed<float16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case bfloat16:
+      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 } // namespace
 
 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 4);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
 
-  std::vector<array> temps;
-  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous(x_pre);
   auto w = ensure_row_contiguous(w_pre);
   auto scales = ensure_row_contiguous(scales_pre);
-  auto biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
+    });
+  }
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 6);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
-  auto& lhs_indices = inputs[4];
-  auto& rhs_indices = inputs[5];
+  auto& lhs_indices = inputs[inputs.size() - 2];
+  auto& rhs_indices = inputs[inputs.size() - 1];
 
-  std::vector<array> temps;
+  auto& encoder = cpu::get_command_encoder(stream());
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &temps](const array& arr) {
+                                          &encoder](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous_last_dims(x_pre);
   auto w = ensure_row_contiguous_last_dims(w_pre);
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
-  auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_input_array(lhs_indices);
   encoder.set_input_array(rhs_indices);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    lhs_indices = array::unsafe_weak_copy(lhs_indices),
-                    rhs_indices = array::unsafe_weak_copy(rhs_indices),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _bs_qmm_dispatch(
-        out,
-        x,
-        w,
-        scales,
-        biases,
-        lhs_indices,
-        rhs_indices,
-        group_size_,
-        bits_,
-        transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _bs_qmm_dispatch(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          group_size_,
+          bits_,
+          transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_bs_qmm_dispatch(
+          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
+    });
+  }
 }
 
 template <typename T, typename U>
@@ -705,16 +1040,14 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::AffineQuantize::eval_cpu(
+void fast::Quantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto ensure_row_contiguous = [s = stream()](const array& arr) {
     if (arr.flags().row_contiguous) {
       return std::make_pair(arr, false);
     } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General, s);
-      return std::make_pair(arr_copy, true);
+      return std::make_pair(contiguous_copy_cpu(arr, s), true);
     }
   };
 
@@ -766,7 +1099,7 @@ void fast::AffineQuantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
     }
   });
 }

mlx/backend/cpu/reduce.cpp

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

mlx/backend/cpu/scan.cpp

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

mlx/backend/cpu/simd/accelerate_simd.h

@@ -234,6 +234,7 @@ Simd<T, N> remainder(Simd<T, N> a, Simd<T, N> b) {
 
 template <typename MaskT, typename T1, typename T2, int N>
 Simd<T1, N> select(Simd<MaskT, N> mask, Simd<T1, N> x, Simd<T2, N> y) {
+  static_assert(std::is_same_v<MaskT, bool>);
   if constexpr (sizeof(T1) == 1) {
     return asd::bitselect(y.value, x.value, asd::convert<char>(mask.value));
   } else if constexpr (sizeof(T1) == 2) {
@@ -251,9 +252,13 @@ Simd<T, N> pow(Simd<T, N> base, Simd<T, N> exp) {
     return asd::pow(base.value, exp.value);
   } else {
     Simd<T, N> res = 1;
-    while (any(exp)) {
-      res = select(exp & 1, res * base, res);
-      base = select(exp, base * base, base);
+    // Raising an integer to a negative power is undefined
+    if (any(exp < 0)) {
+      return 0;
+    }
+    while (any(exp > 0)) {
+      res = select((exp & 1) != 0, res * base, res);
+      base = select(exp > 0, base * base, base);
       exp = exp >> 1;
     }
     return res;

mlx/backend/cpu/softmax.cpp

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

mlx/backend/cpu/sort.cpp

@@ -8,7 +8,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
-
+#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
@@ -333,45 +333,24 @@ void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
+  int axis = axis_;
+  if (axis < 0) {
+    axis += in.ndim();
+  }
+
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);
-  encoder.dispatch(
-      [out = array::unsafe_weak_copy(out), axis_ = axis_]() mutable {
-        switch (out.dtype()) {
-          case bool_:
-            return sort<bool>(out, axis_);
-          case uint8:
-            return sort<uint8_t>(out, axis_);
-          case uint16:
-            return sort<uint16_t>(out, axis_);
-          case uint32:
-            return sort<uint32_t>(out, axis_);
-          case uint64:
-            return sort<uint64_t>(out, axis_);
-          case int8:
-            return sort<int8_t>(out, axis_);
-          case int16:
-            return sort<int16_t>(out, axis_);
-          case int32:
-            return sort<int32_t>(out, axis_);
-          case int64:
-            return sort<int64_t>(out, axis_);
-          case float32:
-            return sort<float>(out, axis_);
-          case float64:
-            return sort<double>(out, axis_);
-          case float16:
-            return sort<float16_t>(out, axis_);
-          case bfloat16:
-            return sort<bfloat16_t>(out, axis_);
-          case complex64:
-            return sort<complex64_t>(out, axis_);
-        }
-      });
+  encoder.dispatch([out = array::unsafe_weak_copy(out), axis]() mutable {
+    dispatch_all_types(out.dtype(), [&](auto type_tag) {
+      sort<MLX_GET_TYPE(type_tag)>(out, axis);
+    });
+  });
 }
 
 void ArgPartition::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -426,8 +405,10 @@ void Partition::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);

mlx/backend/cpu/svd.cpp

@@ -31,7 +31,7 @@ void svd_impl(
 
   // lapack clobbers the input, so we have to make a copy.
   array in(a.shape(), a.dtype(), nullptr, {});
-  copy(
+  copy_cpu(
       a,
       in,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
@@ -81,9 +81,7 @@ void svd_impl(
     // Vᵀ of shape N x N. (M x M in lapack).
     const int ldvt = M;
 
-    auto job_u = (u_ptr) ? "V" : "N";
-    auto job_vt = (u_ptr) ? "V" : "N";
-    static constexpr auto range = "A";
+    auto jobz = (u_ptr) ? "A" : "N";
 
     // Will contain the number of singular values after the call has returned.
     int ns = 0;
@@ -91,30 +89,20 @@ void svd_impl(
 
     // 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) * 12 * K)};
+    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
 
     static const int lwork_query = -1;
 
-    static const int ignored_int = 0;
-    static const T ignored_float = 0;
-
     int info;
 
     // Compute workspace size.
-    gesvdx<T>(
-        /* jobu = */ job_u,
-        /* jobvt = */ job_vt,
-        /* range = */ range,
+    gesdd<T>(
+        /* jobz = */ jobz,
         // 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,
@@ -136,20 +124,13 @@ void svd_impl(
 
     // Loop over matrices.
     for (int i = 0; i < num_matrices; i++) {
-      gesvdx<T>(
-          /* jobu = */ job_u,
-          /* jobvt = */ job_vt,
-          /* range = */ range,
+      gesdd<T>(
+          /* jobz = */ jobz,
           // M and N are swapped since lapack expects column-major.
           /* m = */ &N,
           /* n = */ &M,
           /* a = */ in_ptr + M * N * i,
           /* lda = */ &lda,
-          /* 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,
@@ -167,13 +148,6 @@ 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/cuda/CMakeLists.txt

@@ -6,41 +6,65 @@
 target_sources(
   mlx
   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
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_contiguous.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general.cu
           ${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}/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}/layer_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/logsumexp.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/random.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/all_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/col_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/init_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/segmented_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
           ${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/quantized.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/binary)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/unary)
+
+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)
+else()
+  target_sources(
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_0.cpp)
+endif()
+
 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
 
 # Embed kernel sources in binary for JIT compilation.
@@ -65,6 +89,11 @@ target_include_directories(mlx PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/gen")
 target_compile_options(mlx
                        PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>")
 
+# Enable calling host constexpr functions from device. This is needed because
+# the constexpr version of isnan is host only.
+target_compile_options(
+  mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>")
+
 # CUDA 12.8 emits warning #20280-D for copy kernels which is a false positive.
 # Explicitly pass this flag to suppress the warning, it is safe to set it to
 # true but the warning wouldn't be suppressed.
@@ -78,11 +107,18 @@ endif()
 target_compile_options(
   mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--Wno-deprecated-gpu-targets>")
 
-# Compute capability 7 is required for synchronization between CPU/GPU with
-# managed memory. TODO: Add more architectures for potential performance gain.
-set(MLX_CUDA_ARCHITECTURES
-    "70;80"
-    CACHE STRING "CUDA architectures")
+# Use stronger binaries compression. This feature was introduced in CUDA 12.8
+# and requires drivers released after CUDA 12.4.
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
+  target_compile_options(
+    mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--compress-mode=size>")
+endif()
+
+# Compute capability >= 7.0 is required for synchronization between CPU/GPU with
+# managed memory.
+if(NOT DEFINED MLX_CUDA_ARCHITECTURES)
+  set(MLX_CUDA_ARCHITECTURES "native")
+endif()
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
                                      "${MLX_CUDA_ARCHITECTURES}")
@@ -114,6 +150,27 @@ target_link_libraries(mlx PRIVATE CUDA::cublasLt)
 # Use NVRTC and driver APIs.
 target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 
+# Use the frontend APIs of cuDNN.
+FetchContent_Declare(
+  cudnn
+  GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
+  GIT_TAG v1.14.0
+  GIT_SHALLOW TRUE
+  EXCLUDE_FROM_ALL)
+set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)
+set(CUDNN_FRONTEND_BUILD_SAMPLES OFF)
+set(CUDNN_FRONTEND_BUILD_TESTS OFF)
+set(CUDNN_FRONTEND_BUILD_PYTHON_BINDINGS OFF)
+FetchContent_MakeAvailable(cudnn)
+target_link_libraries(mlx PRIVATE cudnn_frontend)
+# Link with the actual cuDNN libraries.
+include(${cudnn_frontend_SOURCE_DIR}/cmake/cuDNN.cmake)
+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

@@ -2,7 +2,7 @@
 
 #include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/utils.h"
-#include "mlx/backend/cuda/worker.h"
+#include "mlx/utils.h"
 
 #include <cuda_runtime.h>
 #include <fmt/format.h>
@@ -14,14 +14,75 @@ namespace mlx::core {
 
 namespace cu {
 
+constexpr int page_size = 16384;
+
+// Any allocations smaller than this will try to use the small pool
+constexpr int small_block_size = 8;
+
+// The small pool size in bytes. This should be a multiple of the host page
+// size and small_block_size.
+constexpr int small_pool_size = 4 * page_size;
+
+SmallSizePool::SmallSizePool() {
+  auto num_blocks = small_pool_size / small_block_size;
+  buffer_ = new Block[num_blocks];
+
+  next_free_ = buffer_;
+
+  CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
+#if CUDART_VERSION >= 13000
+  cudaMemLocation loc;
+  loc.type = cudaMemLocationTypeDevice;
+  loc.id = 0;
+#else
+  int loc = 0;
+#endif // CUDART_VERSION >= 13000
+  CHECK_CUDA_ERROR(
+      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, loc));
+
+  auto curr = next_free_;
+  for (size_t i = 1; i < num_blocks; ++i) {
+    curr->next = buffer_ + i;
+    curr = curr->next;
+  }
+  curr->next = nullptr;
+}
+
+SmallSizePool::~SmallSizePool() {
+  CHECK_CUDA_ERROR(cudaFree(data_));
+  delete[] buffer_;
+}
+
+CudaBuffer* SmallSizePool::malloc() {
+  if (next_free_ == nullptr) {
+    return nullptr;
+  }
+  Block* b = next_free_;
+  uint64_t i = next_free_ - buffer_;
+  next_free_ = next_free_->next;
+  b->buf.data = static_cast<char*>(data_) + i * small_block_size;
+  b->buf.size = small_block_size;
+  return &b->buf;
+}
+
+void SmallSizePool::free(CudaBuffer* buf) {
+  auto b = reinterpret_cast<Block*>(buf);
+  b->next = next_free_;
+  next_free_ = b;
+}
+
+bool SmallSizePool::in_pool(CudaBuffer* buf) {
+  constexpr int num_blocks = (small_pool_size / small_block_size);
+  auto b = reinterpret_cast<Block*>(buf);
+  int64_t block_num = b - buffer_;
+  return block_num >= 0 && block_num < num_blocks;
+}
+
 CudaAllocator::CudaAllocator()
     : buffer_cache_(
-          getpagesize(),
+          page_size,
           [](CudaBuffer* buf) { return buf->size; },
-          [this](CudaBuffer* buf) {
-            cuda_free(buf->data);
-            delete buf;
-          }) {
+          [this](CudaBuffer* buf) { cuda_free(buf); }) {
   // TODO: Set memory limit for multi-device.
   size_t free, total;
   CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
@@ -31,22 +92,37 @@ CudaAllocator::CudaAllocator()
 
 Buffer CudaAllocator::malloc(size_t size) {
   // Find available buffer from cache.
+  auto orig_size = size;
   std::unique_lock lock(mutex_);
+  if (size <= small_block_size) {
+    size = 8;
+  } else if (size < page_size) {
+    size = next_power_of_2(size);
+  } else {
+    size = page_size * ((size + page_size - 1) / page_size);
+  }
+
   CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
   if (!buf) {
-    // If we have a lot of memory pressure or are over the maximum cache size,
-    // try to reclaim memory from the cache.
-    size_t mem_required = get_active_memory() + get_cache_memory() + size;
-    if (mem_required >= memory_limit_) {
-      buffer_cache_.release_cached_buffers(mem_required - memory_limit_);
+    // If we have a lot of memory pressure try to reclaim memory from the cache.
+    int64_t mem_to_free =
+        get_active_memory() + get_cache_memory() + size - memory_limit_;
+    if (mem_to_free > 0) {
+      buffer_cache_.release_cached_buffers(mem_to_free);
     }
 
+    // Try the scalar pool first
+    if (size <= small_block_size) {
+      buf = scalar_pool_.malloc();
+    }
     lock.unlock();
-    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)));
+    if (!buf) {
+      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)));
+      }
     }
     lock.lock();
   }
@@ -57,7 +133,6 @@ Buffer CudaAllocator::malloc(size_t size) {
   if (get_cache_memory() > max_pool_size_) {
     buffer_cache_.release_cached_buffers(get_cache_memory() - max_pool_size_);
   }
-
   return Buffer{buf};
 }
 
@@ -72,9 +147,7 @@ void CudaAllocator::free(Buffer buffer) {
   if (get_cache_memory() < max_pool_size_) {
     buffer_cache_.recycle_to_cache(buf);
   } else {
-    lock.unlock();
-    cuda_free(buf->data);
-    delete buf;
+    cuda_free(buf);
   }
 }
 
@@ -86,28 +159,14 @@ size_t CudaAllocator::size(Buffer buffer) const {
   return buf->size;
 }
 
-void CudaAllocator::register_this_thread() {
-  std::lock_guard lock(worker_mutex_);
-  allowed_threads_.insert(std::this_thread::get_id());
-}
-
-void CudaAllocator::cuda_free(void* buf) {
-  // If cuda_free() is called from a unregistered thread, reschedule the call to
-  // worker.
-  {
-    std::lock_guard lock(worker_mutex_);
-    if (allowed_threads_.count(std::this_thread::get_id()) == 0) {
-      if (!worker_) {
-        worker_.reset(new Worker);
-      }
-      worker_->add_task([this, buf]() { this->cuda_free(buf); });
-      worker_->end_batch();
-      worker_->commit();
-      return;
-    }
+// This must be called with mutex_ aquired
+void CudaAllocator::cuda_free(CudaBuffer* buf) {
+  if (scalar_pool_.in_pool(buf)) {
+    scalar_pool_.free(buf);
+  } else {
+    cudaFree(buf->data);
+    delete buf;
   }
-
-  cudaFree(buf);
 }
 
 size_t CudaAllocator::get_active_memory() const {

mlx/backend/cuda/allocator.h

@@ -7,13 +7,10 @@
 
 #include <mutex>
 #include <set>
-#include <thread>
 #include <utility>
 
 namespace mlx::core::cu {
 
-class Worker;
-
 using allocator::Buffer;
 
 // Stores cuda-managed unified memory.
@@ -22,21 +19,35 @@ struct CudaBuffer {
   size_t size;
 };
 
+class SmallSizePool {
+ private:
+  union Block {
+    Block* next;
+    CudaBuffer buf;
+  };
+
+  Block* buffer_{nullptr};
+  void* data_{nullptr};
+  Block* next_free_{nullptr};
+
+ public:
+  SmallSizePool();
+  ~SmallSizePool();
+
+  SmallSizePool(const SmallSizePool&) = delete;
+  SmallSizePool& operator=(const SmallSizePool&) = delete;
+
+  CudaBuffer* malloc();
+  void free(CudaBuffer* buf);
+  bool in_pool(CudaBuffer* buf);
+};
+
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
   void free(Buffer buffer) override;
   size_t size(Buffer buffer) const override;
 
-  // Register current thread as safe to free buffers.
-  // In cuda freeing a buffer implicitly synchronizes stream, and for threads
-  // that may be waited by gpu stream (for example cpu stream threads), freeing
-  // buffers there would result in dead lock.
-  void register_this_thread();
-
-  // Call cudaFree in the safe thread.
-  void cuda_free(void* buf);
-
   size_t get_active_memory() const;
   size_t get_peak_memory() const;
   void reset_peak_memory();
@@ -47,19 +58,18 @@ class CudaAllocator : public allocator::Allocator {
   void clear_cache();
 
  private:
+  void cuda_free(CudaBuffer* buf);
+
   CudaAllocator();
   friend CudaAllocator& allocator();
 
-  std::mutex worker_mutex_;
-  std::unique_ptr<Worker> worker_;
-  std::set<std::thread::id> allowed_threads_;
-
   std::mutex mutex_;
   size_t memory_limit_;
   size_t max_pool_size_;
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
+  SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();

mlx/backend/cuda/arange.cu

@@ -0,0 +1,69 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/fp16_math.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+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);
+  }
+}
+
+} // namespace cu
+
+void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Arange::eval_gpu");
+  if (out.size() == 0) {
+    return;
+  }
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& encoder = cu::get_command_encoder(stream());
+  encoder.set_output_array(out);
+
+  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,
+          out.data<OutType>(),
+          out.data_size(),
+          static_cast<CTYPE>(start_),
+          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/arg_reduce.cu

@@ -1,7 +1,8 @@
 // Copyright © 2025 Apple Inc.
+
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
+#include "mlx/backend/cuda/device/fp16_math.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -43,8 +44,11 @@ struct ArgMin {
   }
 
   template <int N>
-  __device__ IndexValPair<T>
-  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
+  __device__ IndexValPair<T> reduce_many(
+      IndexValPair<T> best,
+      const AlignedVector<T, N>& vals,
+      uint32_t offset) {
+#pragma unroll
     for (int i = 0; i < N; i++) {
       if (vals[i] < best.val) {
         best.val = vals[i];
@@ -73,8 +77,11 @@ struct ArgMax {
   }
 
   template <int N>
-  __device__ IndexValPair<T>
-  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
+  __device__ IndexValPair<T> reduce_many(
+      IndexValPair<T> best,
+      const AlignedVector<T, N>& vals,
+      uint32_t offset) {
+#pragma unroll
     for (int i = 0; i < N; i++) {
       if (vals[i] > best.val) {
         best.val = vals[i];
@@ -105,16 +112,15 @@ __global__ void arg_reduce_general(
 
   int64_t in_idx = elem_to_loc(index, shape.data(), in_strides.data(), ndim);
   int64_t out_idx = elem_to_loc(index, shape.data(), out_strides.data(), ndim);
+  in += in_idx;
 
   Op op;
   T init = op.init();
   IndexValPair<T> best{0, init};
 
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
-    T vals[N_READS];
     auto tid = r * BLOCK_DIM + block.thread_index().x;
-    cub::LoadDirectBlocked(
-        tid, strided_iterator(in + in_idx, axis_stride), vals, axis_size, init);
+    auto vals = load_vector<N_READS>(in, tid, axis_size, axis_stride, init);
     best = op.reduce_many(best, vals, tid * N_READS);
   }
 
@@ -151,36 +157,30 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_REAL_TYPES_CHECKED(in.dtype(), "ArgReduce", CTYPE, {
-      using InType = cuda_type_t<CTYPE>;
-      constexpr uint32_t N_READS = 4;
-      MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-        dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
-        dim3 block_dims{BLOCK_DIM, 1, 1};
-        auto kernel = &cu::arg_reduce_general<
-            InType,
-            cu::ArgMax<InType>,
-            BLOCK_DIM,
-            N_READS>;
-        if (reduce_type_ == ArgReduce::ArgMin) {
-          kernel = &cu::arg_reduce_general<
-              InType,
-              cu::ArgMin<InType>,
-              BLOCK_DIM,
-              N_READS>;
-        }
-        kernel<<<num_blocks, block_dims, 0, stream>>>(
-            in.data<InType>(),
-            out.data<uint32_t>(),
-            out.size(),
-            const_param(shape),
-            const_param(in_strides),
-            const_param(out_strides),
-            ndim,
-            axis_stride,
-            axis_size);
-      });
+  dispatch_real_types(in.dtype(), "ArgReduce", [&](auto type_tag) {
+    using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr uint32_t N_READS = 4;
+    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
+      auto kernel =
+          cu::arg_reduce_general<T, cu::ArgMax<T>, block_dim(), N_READS>;
+      if (reduce_type_ == ArgReduce::ArgMin) {
+        kernel = cu::arg_reduce_general<T, cu::ArgMin<T>, block_dim(), N_READS>;
+      }
+      encoder.add_kernel_node(
+          kernel,
+          num_blocks,
+          block_dim(),
+          0,
+          in.data<T>(),
+          out.data<uint32_t>(),
+          out.size(),
+          const_param(shape),
+          const_param(in_strides),
+          const_param(out_strides),
+          ndim,
+          axis_stride,
+          axis_size);
     });
   });
 }

mlx/backend/cuda/binary.cu

@@ -1,317 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/common/binary.h"
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/binary_ops.cuh"
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-#include "mlx/primitives.h"
-
-#include <cooperative_groups.h>
-#include <nvtx3/nvtx3.hpp>
-
-namespace mlx::core {
-
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-template <typename Op, typename In, typename Out, typename IdxT>
-__global__ void binary_ss(const In* a, const In* b, Out* out, IdxT size) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[0], b[0]);
-  }
-}
-
-template <typename Op, typename In, typename Out, typename IdxT>
-__global__ void binary_sv(const In* a, const In* b, Out* out, IdxT size) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[0], b[index]);
-  }
-}
-
-template <typename Op, typename In, typename Out, typename IdxT>
-__global__ void binary_vs(const In* a, const In* b, Out* out, IdxT size) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[0]);
-  }
-}
-
-template <typename Op, typename In, typename Out, typename IdxT>
-__global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[index]);
-  }
-}
-
-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,
-    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) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
-        index, shape.data(), a_strides.data(), b_strides.data());
-    out[index] = Op{}(a[a_idx], b[b_idx]);
-  }
-}
-
-template <typename Op, typename In, typename Out, typename IdxT>
-__global__ void binary_g(
-    const In* a,
-    const In* b,
-    Out* out,
-    IdxT size,
-    const __grid_constant__ Shape shape,
-    const __grid_constant__ Strides a_strides,
-    const __grid_constant__ Strides b_strides,
-    int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_4d(
-        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
-    out[index] = Op{}(a[a_idx], b[b_idx]);
-  }
-}
-
-template <typename Op, typename In, typename Out>
-constexpr bool supports_binary_op() {
-  if (std::is_same_v<Op, Add> || std::is_same_v<Op, Divide> ||
-      std::is_same_v<Op, Maximum> || std::is_same_v<Op, Minimum> ||
-      std::is_same_v<Op, Multiply> || std::is_same_v<Op, Subtract> ||
-      std::is_same_v<Op, Power> || std::is_same_v<Op, Remainder>) {
-    return std::is_same_v<In, Out>;
-  }
-  if (std::is_same_v<Op, Equal> || std::is_same_v<Op, Greater> ||
-      std::is_same_v<Op, GreaterEqual> || std::is_same_v<Op, Less> ||
-      std::is_same_v<Op, LessEqual> || std::is_same_v<Op, NotEqual>) {
-    return std::is_same_v<Out, bool>;
-  }
-  if (std::is_same_v<Op, LogicalAnd> || std::is_same_v<Op, LogicalOr>) {
-    return std::is_same_v<Out, bool> && std::is_same_v<In, bool>;
-  }
-  if (std::is_same_v<Op, NaNEqual>) {
-    return std::is_same_v<Out, bool> && is_inexact_v<In>;
-  }
-  if (std::is_same_v<Op, LogAddExp>) {
-    return std::is_same_v<In, Out> && is_inexact_v<In>;
-  }
-  if (std::is_same_v<Op, ArcTan2>) {
-    return std::is_same_v<In, Out> && is_floating_v<In>;
-  }
-  if (std::is_same_v<Op, BitwiseAnd> || std::is_same_v<Op, BitwiseOr> ||
-      std::is_same_v<Op, BitwiseXor>) {
-    return std::is_same_v<In, Out> && std::is_integral_v<In>;
-  }
-  if (std::is_same_v<Op, LeftShift> || std::is_same_v<Op, RightShift>) {
-    return std::is_same_v<In, Out> && std::is_integral_v<In> &&
-        !std::is_same_v<In, bool>;
-  }
-  return false;
-}
-
-} // namespace cu
-
-template <typename Op>
-void binary_op_gpu_inplace(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs,
-    std::string_view op,
-    const Stream& s) {
-  assert(inputs.size() > 1);
-  const auto& a = inputs[0];
-  const auto& b = inputs[1];
-  auto& out = outputs[0];
-  if (out.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);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(a.dtype(), CTYPE_IN, {
-      MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, {
-        if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
-          using InType = cuda_type_t<CTYPE_IN>;
-          using OutType = cuda_type_t<CTYPE_OUT>;
-
-          auto bopt = get_binary_op_type(a, b);
-          if (bopt == BinaryOpType::General) {
-            auto [shape, strides] = collapse_contiguous_dims(a, b, out);
-            auto& a_strides = strides[0];
-            auto& b_strides = strides[1];
-            bool large = a.data_size() > INT32_MAX ||
-                b.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-            MLX_SWITCH_BOOL(large, LARGE, {
-              using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-              int ndim = shape.size();
-              if (ndim <= 3) {
-                MLX_SWITCH_1_2_3(ndim, NDIM, {
-                  auto kernel =
-                      &cu::binary_g_nd<Op, InType, OutType, IdxT, NDIM>;
-                  auto [num_blocks, block_dims] =
-                      get_launch_args(kernel, out, large);
-                  kernel<<<num_blocks, block_dims, 0, stream>>>(
-                      a.data<InType>(),
-                      b.data<InType>(),
-                      out.data<OutType>(),
-                      out.size(),
-                      const_param<NDIM>(shape),
-                      const_param<NDIM>(a_strides),
-                      const_param<NDIM>(b_strides));
-                });
-              } else {
-                auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
-                auto [num_blocks, block_dims] =
-                    get_launch_args(kernel, out, large);
-                kernel<<<num_blocks, block_dims, 0, stream>>>(
-                    a.data<InType>(),
-                    b.data<InType>(),
-                    out.data<OutType>(),
-                    out.size(),
-                    const_param(shape),
-                    const_param(a_strides),
-                    const_param(b_strides),
-                    ndim);
-              }
-            });
-          } else {
-            MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
-              using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
-              auto kernel = cu::binary_ss<Op, InType, OutType, IdxT>;
-              if (bopt == BinaryOpType::ScalarVector) {
-                kernel = cu::binary_sv<Op, InType, OutType, IdxT>;
-              } else if (bopt == BinaryOpType::VectorScalar) {
-                kernel = cu::binary_vs<Op, InType, OutType, IdxT>;
-              } else if (bopt == BinaryOpType::VectorVector) {
-                kernel = cu::binary_vv<Op, InType, OutType, IdxT>;
-              }
-              auto [num_blocks, block_dims] = get_launch_args(
-                  kernel, out.data_size(), out.shape(), out.strides(), LARGE);
-              kernel<<<num_blocks, block_dims, 0, stream>>>(
-                  a.data<InType>(),
-                  b.data<InType>(),
-                  out.data<OutType>(),
-                  out.data_size());
-            });
-          }
-        } else {
-          throw std::runtime_error(fmt::format(
-              "Can not do binary op {} on inputs of {} with result of {}.",
-              op,
-              dtype_to_string(a.dtype()),
-              dtype_to_string(out.dtype())));
-        }
-      });
-    });
-  });
-}
-
-template <typename Op>
-void binary_op_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs,
-    std::string_view op,
-    const Stream& s) {
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, outputs[0], bopt);
-  set_binary_op_output_data(a, b, outputs[1], bopt);
-  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
-}
-
-template <typename Op>
-void binary_op_gpu(
-    const std::vector<array>& inputs,
-    array& out,
-    std::string_view op,
-    const Stream& s) {
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
-  std::vector<array> outputs{out};
-  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
-}
-
-#define BINARY_GPU(func)                                                 \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) {    \
-    nvtx3::scoped_range r(#func "::eval_gpu");                           \
-    auto& s = out.primitive().stream();                                  \
-    binary_op_gpu<cu::func>(inputs, out, get_primitive_string(this), s); \
-  }
-
-#define BINARY_GPU_MULTI(func)                                               \
-  void func::eval_gpu(                                                       \
-      const std::vector<array>& inputs, std::vector<array>& outputs) {       \
-    nvtx3::scoped_range r(#func "::eval_gpu");                               \
-    auto& s = outputs[0].primitive().stream();                               \
-    binary_op_gpu<cu::func>(inputs, outputs, get_primitive_string(this), 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();
-  auto op = get_primitive_string(this);
-  if (equal_nan_) {
-    binary_op_gpu<cu::NaNEqual>(inputs, out, op, s);
-  } else {
-    binary_op_gpu<cu::Equal>(inputs, out, op, s);
-  }
-}
-
-void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
-  auto& s = out.primitive().stream();
-  auto op = get_primitive_string(this);
-  switch (op_) {
-    case BitwiseBinary::And:
-      binary_op_gpu<cu::BitwiseAnd>(inputs, out, op, s);
-      break;
-    case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, op, s);
-      break;
-    case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, op, s);
-      break;
-    case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, op, s);
-      break;
-    case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, op, s);
-      break;
-  }
-}
-
-} // namespace mlx::core

mlx/backend/cuda/binary/CMakeLists.txt

@@ -0,0 +1,21 @@
+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

@@ -0,0 +1,7 @@
+// 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

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

mlx/backend/cuda/binary/binary.cuh

@@ -0,0 +1,379 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_ss(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (int i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec[i] = Op{}(a[0], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_sv(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[i]);
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec[i] = Op{}(a[0], b_vec[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_vs(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[0]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec[i] = Op{}(a_vec[i], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[i]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    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]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <
+    typename Op,
+    typename In,
+    typename Out,
+    typename IdxT,
+    int NDIM,
+    int N_READS>
+__global__ void binary_g_nd(
+    const In* a,
+    const In* b,
+    Out* out,
+    IdxT size_rest,
+    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;
+  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]);
+  }
+  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>
+__global__ void binary_g(
+    const In* a,
+    const In* b,
+    Out* out,
+    IdxT size_rest,
+    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;
+  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]);
+  }
+  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
+}
+
+template <typename Op, typename In, typename Out>
+constexpr bool supports_binary_op() {
+  if (std::is_same_v<Op, Add> || std::is_same_v<Op, Divide> ||
+      std::is_same_v<Op, Maximum> || std::is_same_v<Op, Minimum> ||
+      std::is_same_v<Op, Multiply> || std::is_same_v<Op, Subtract> ||
+      std::is_same_v<Op, Power> || std::is_same_v<Op, Remainder>) {
+    return std::is_same_v<In, Out>;
+  }
+  if (std::is_same_v<Op, Equal> || std::is_same_v<Op, Greater> ||
+      std::is_same_v<Op, GreaterEqual> || std::is_same_v<Op, Less> ||
+      std::is_same_v<Op, LessEqual> || std::is_same_v<Op, NotEqual>) {
+    return std::is_same_v<Out, bool>;
+  }
+  if (std::is_same_v<Op, LogicalAnd> || std::is_same_v<Op, LogicalOr>) {
+    return std::is_same_v<Out, bool> && std::is_same_v<In, bool>;
+  }
+  if (std::is_same_v<Op, NaNEqual>) {
+    return std::is_same_v<Out, bool> && is_inexact_v<In>;
+  }
+  if (std::is_same_v<Op, LogAddExp>) {
+    return std::is_same_v<In, Out> && is_inexact_v<In>;
+  }
+  if (std::is_same_v<Op, ArcTan2>) {
+    return std::is_same_v<In, Out> && is_floating_v<In>;
+  }
+  if (std::is_same_v<Op, BitwiseAnd> || std::is_same_v<Op, BitwiseOr> ||
+      std::is_same_v<Op, BitwiseXor>) {
+    return std::is_same_v<In, Out> && std::is_integral_v<In>;
+  }
+  if (std::is_same_v<Op, LeftShift> || std::is_same_v<Op, RightShift>) {
+    return std::is_same_v<In, Out> && std::is_integral_v<In> &&
+        !std::is_same_v<In, bool>;
+  }
+  return false;
+}
+
+} // namespace cu
+
+template <typename Op>
+void binary_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    array& out,
+    const char* op,
+    const Stream& s) {
+  assert(inputs.size() > 1);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  if (out.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);
+  dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
+      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
+      if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+        using InType = cuda_type_t<CTYPE_IN>;
+        using OutType = cuda_type_t<CTYPE_OUT>;
+        auto bopt = get_binary_op_type(a, b);
+        if (bopt == BinaryOpType::General) {
+          dispatch_bool(
+              a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
+                  out.data_size() > INT32_MAX,
+              [&](auto large) {
+                using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+                Shape shape;
+                std::vector<Strides> strides;
+                std::tie(shape, strides) = collapse_contiguous_dims(a, b, out);
+                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>;
+                    }
+                    encoder.add_kernel_node(
+                        kernel,
+                        {num_blocks_x, num_blocks_y},
+                        block_dims,
+                        0,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out.data<OutType>(),
+                        rest,
+                        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>;
+                  }
+                  encoder.add_kernel_node(
+                      kernel,
+                      {num_blocks_x, num_blocks_y},
+                      block_dims,
+                      0,
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out.data<OutType>(),
+                      rest,
+                      const_param(shape),
+                      const_param(a_strides),
+                      const_param(b_strides),
+                      ndim);
+                }
+              });
+        } else {
+          dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            constexpr int N_READS = 16 / sizeof(InType);
+            auto kernel = cu::binary_ss<Op, InType, OutType, IdxT, N_READS>;
+            if (bopt == BinaryOpType::ScalarVector) {
+              kernel = cu::binary_sv<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorScalar) {
+              kernel = cu::binary_vs<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorVector) {
+              kernel = cu::binary_vv<Op, InType, OutType, IdxT, N_READS>;
+            }
+            auto [num_blocks, block_dims] = get_launch_args(
+                out.data_size(), out.shape(), out.strides(), large(), N_READS);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                a.data<InType>(),
+                b.data<InType>(),
+                out.data<OutType>(),
+                out.data_size());
+          });
+        }
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do binary op {} on inputs of {} with result of {}.",
+            op,
+            dtype_to_string(a.dtype()),
+            dtype_to_string(out.dtype())));
+      }
+    });
+  });
+}
+
+template <typename Op>
+void binary_op_gpu(
+    const std::vector<array>& inputs,
+    array& out,
+    const char* op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+  binary_op_gpu_inplace<Op>(inputs, out, op, s);
+}
+
+#define BINARY_GPU(func)                                              \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    nvtx3::scoped_range r(#func "::eval_gpu");                        \
+    auto& s = out.primitive().stream();                               \
+    binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
+  }
+
+} // namespace mlx::core

mlx/backend/cuda/binary/bitwise_binary.cu

@@ -0,0 +1,27 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,15 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,409 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b_vec[i]);
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec[i], b[0]);
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec[i], b_vec[i]);
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <
+    typename Op,
+    typename In,
+    typename Out,
+    typename IdxT,
+    int NDIM,
+    int N_READS>
+__global__ void binary_two_g_nd(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size_rest,
+    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;
+  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];
+  }
+  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>
+__global__ void binary_two_g(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size_rest,
+    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;
+  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];
+  }
+  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>
+constexpr bool supports_binary_two_op() {
+  if (std::is_same_v<Op, DivMod>) {
+    return std::is_same_v<In, Out> &&
+        (std::is_integral_v<Out> || is_floating_v<Out>);
+  }
+  return false;
+}
+
+} // namespace cu
+
+template <typename Op>
+void binary_two_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const char* op,
+    const Stream& s) {
+  assert(inputs.size() > 1);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  auto& out_a = outputs[0];
+  auto& out_b = outputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out_a, bopt);
+  set_binary_op_output_data(a, b, out_b, bopt);
+
+  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);
+  encoder.set_output_array(out_b);
+  dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out_a.dtype(), [&](auto out_type_tag) {
+      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
+      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
+      if constexpr (cu::supports_binary_two_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+        using InType = cuda_type_t<CTYPE_IN>;
+        using OutType = cuda_type_t<CTYPE_OUT>;
+
+        auto bopt = get_binary_op_type(a, b);
+        if (bopt == BinaryOpType::General) {
+          dispatch_bool(
+              a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
+                  out_a.data_size() > INT32_MAX,
+              [&](auto large) {
+                using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+                Shape shape;
+                std::vector<Strides> strides;
+                std::tie(shape, strides) =
+                    collapse_contiguous_dims(a, b, out_a);
+                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>;
+                    }
+                    encoder.add_kernel_node(
+                        kernel,
+                        {num_blocks_x, num_blocks_y},
+                        block_dims,
+                        0,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out_a.data<OutType>(),
+                        out_b.data<OutType>(),
+                        rest,
+                        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>;
+                  }
+                  encoder.add_kernel_node(
+                      kernel,
+                      {num_blocks_x, num_blocks_y},
+                      block_dims,
+                      0,
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out_a.data<OutType>(),
+                      out_b.data<OutType>(),
+                      rest,
+                      const_param(shape),
+                      const_param(a_strides),
+                      const_param(b_strides),
+                      ndim);
+                }
+              });
+        } else {
+          dispatch_bool(out_a.data_size() > UINT32_MAX, [&](auto large) {
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            constexpr int N_READS = 16 / sizeof(InType);
+            auto kernel = cu::binary_two_ss<Op, InType, OutType, IdxT, N_READS>;
+            if (bopt == BinaryOpType::ScalarVector) {
+              kernel = cu::binary_two_sv<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorScalar) {
+              kernel = cu::binary_two_vs<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorVector) {
+              kernel = cu::binary_two_vv<Op, InType, OutType, IdxT, N_READS>;
+            }
+            auto [num_blocks, block_dims] = get_launch_args(
+                out_a.data_size(),
+                out_a.shape(),
+                out_a.strides(),
+                large(),
+                N_READS);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                a.data<InType>(),
+                b.data<InType>(),
+                out_a.data<OutType>(),
+                out_b.data<OutType>(),
+                out_a.data_size());
+          });
+        }
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do binary op {} on inputs of {} with result of {}.",
+            op,
+            dtype_to_string(a.dtype()),
+            dtype_to_string(out_a.dtype())));
+      }
+    });
+  });
+}
+
+template <typename Op>
+void binary_two_op_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const char* op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, outputs[0], bopt);
+  set_binary_op_output_data(a, b, outputs[1], bopt);
+  binary_two_op_gpu_inplace<Op>(inputs, outputs, op, s);
+}
+
+void DivMod::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("DivMod::eval_gpu");
+  auto& s = outputs[0].primitive().stream();
+  binary_two_op_gpu<cu::DivMod>(inputs, outputs, name(), s);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/compiled.cpp

@@ -3,6 +3,7 @@
 #include "mlx/backend/common/compiled.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/jit_module.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/graph_utils.h"
 #include "mlx/primitives.h"
 
@@ -52,9 +53,10 @@ struct FusedKernelBuilder {
 
     // Build function signature.
     if (contiguous) {
-      os += "template <typename IdxT = uint32_t>\n";
+      os += "template <typename IdxT = uint32_t, int work_per_thread = 1>\n";
     } else {
-      os += "template <int NDIM, typename IdxT = uint32_t>\n";
+      os +=
+          "template <int NDIM, typename IdxT = uint32_t, int work_per_thread = 1>\n";
     }
     os += fmt::format("__global__ void {}(\n", kernel_name + name);
     for (size_t i = 0; i < params.size(); ++i) {
@@ -66,12 +68,77 @@ struct FusedKernelBuilder {
     }
     os += ") {\n";
 
-    // Index.
+    // Index. For non contiguous kernels we create a separate index
+    // variable per variable otherwise everyone uses `index`.
     os +=
-        "  IdxT index = cg::this_grid().thread_rank();\n"
+        "  IdxT index = cg::this_grid().thread_rank() * work_per_thread;\n"
         "  if (index >= size) {\n"
         "    return;\n"
         "  }\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "  IdxT " + xname + "_idx = 0;\n";
+      }
+      os += "  {\n";
+      os += "    IdxT loc = index;\n";
+      os +=
+          "    #pragma unroll\n"
+          "    for (int i = NDIM - 1; i >= 0; i--) {\n";
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "      " + xname + "_idx += (loc \% shape[i]) * IdxT(" + xname +
+            "_strides[i]);\n";
+      }
+      os +=
+          "      loc /= shape[i];\n"
+          "    }\n"
+          "  }\n";
+    }
+
+    // Vectorized read loop
+    if (contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        const std::string& xname = namer.get_name(x);
+        std::string type = dtype_to_cuda_type(x.dtype());
+        os += fmt::format(
+            "  auto vec_{0} = load_vector<work_per_thread, {1}>({0} + index, 0, size - index, 0);\n",
+            xname,
+            type);
+      }
+    }
+
+    // Create some space for the outputs
+    for (const auto& x : outputs) {
+      const std::string& xname = namer.get_name(x);
+      std::string type = dtype_to_cuda_type(x.dtype());
+      os += fmt::format(
+          "  AlignedVector<{}, work_per_thread> vec_{};\n", type, xname);
+    }
+
+    // Work loop
+    if (!contiguous) {
+      os +=
+          "\n"
+          "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
+    } else {
+      os +=
+          "\n"
+          "  #pragma unroll\n"
+          "  for (int i = 0; i < work_per_thread; i++) {\n";
+    }
 
     // Read inputs.
     for (size_t i = 0; i < inputs.size(); ++i) {
@@ -86,14 +153,11 @@ struct FusedKernelBuilder {
       } else if (is_scalar(x)) {
         value = fmt::format("{}[0]", xname);
       } else if (contiguous) {
-        value = fmt::format("{}[index]", xname);
+        value = fmt::format("vec_{}[i]", xname);
       } else {
-        std::string index = fmt::format(
-            "elem_to_loc_nd<NDIM>(index, shape.data(), {}_strides.data())",
-            xname);
-        value = fmt::format("{}[{}]", xname, index);
+        value = fmt::format("{}[{}_idx]", xname, xname);
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write tape.
@@ -105,21 +169,40 @@ struct FusedKernelBuilder {
         value = fmt::format(
             "static_cast<{}>(tmp_{})", type, namer.get_name(x.inputs()[0]));
       } else {
-        std::ostringstream ss;
-        x.primitive().print(ss);
-        value = ss.str();
+        value = x.primitive().name();
         value += "{}(";
         for (size_t i = 0; i < x.inputs().size() - 1; ++i) {
           value += fmt::format("tmp_{}, ", namer.get_name(x.inputs()[i]));
         }
         value += fmt::format("tmp_{})", namer.get_name(x.inputs().back()));
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write output.
     for (const auto& x : outputs) {
-      os += fmt::format("  {0}[index] = tmp_{0};\n", namer.get_name(x));
+      os += fmt::format("    vec_{0}[i] = tmp_{0};\n", namer.get_name(x));
+    }
+
+    // End of work loop
+    if (!contiguous) {
+      os += "\n";
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += fmt::format("    {0}_idx += {0}_strides[NDIM - 1];\n", xname);
+      }
+    }
+    os += "  }\n";
+
+    // Store the output to global memory
+    for (const auto& x : outputs) {
+      os += fmt::format(
+          "  store_vector({0} + index, 0, vec_{0}, size - index);\n",
+          namer.get_name(x));
     }
 
     os += "}\n";
@@ -145,6 +228,15 @@ void Compiled::eval_gpu(
   nvtx3::scoped_range r("Compiled::eval_gpu");
   auto& s = stream();
 
+  // Determine the work per thread for the vectorized reads/writes. We take it
+  // as 16 over the max itemsize for the outputs. Another heuristic could be
+  // over the max itemsize of all arrays.
+  int max_size = 1;
+  for (const auto& x : outputs) {
+    max_size = (max_size > x.itemsize()) ? max_size : x.itemsize();
+  }
+  int work_per_thread = 16 / max_size;
+
   cu::JitModule& mod = cu::get_jit_module(s.device, lib_name(), [&]() {
     // Build source code.
     cu::FusedKernelBuilder builder{
@@ -157,17 +249,26 @@ void Compiled::eval_gpu(
     builder.build("_strided", false);
     builder.os += "\n} // namespace mlx::core::cu\n";
     // Build kernel names.
-    std::vector<std::string> kernel_names = {
-        fmt::format("mlx::core::cu::{}_contiguous<uint32_t>", lib_name()),
-        fmt::format("mlx::core::cu::{}_contiguous<int64_t>", lib_name()),
-    };
-    for (int i = 1; i <= MAX_NDIM; ++i) {
-      kernel_names.push_back(fmt::format(
-          "mlx::core::cu::{}_strided<{}, uint32_t>", lib_name(), i));
-      kernel_names.push_back(
-          fmt::format("mlx::core::cu::{}_strided<{}, int64_t>", lib_name(), i));
+    std::vector<std::string> kernel_names;
+    kernel_names.push_back(fmt::format(
+        "mlx::core::cu::{}_contiguous<uint32_t, {}>",
+        lib_name(),
+        work_per_thread));
+    kernel_names.push_back(fmt::format(
+        "mlx::core::cu::{}_contiguous<int64_t, {}>",
+        lib_name(),
+        work_per_thread));
+    for (auto wpt : std::array<int, 2>{1, work_per_thread}) {
+      for (int i = 1; i <= MAX_NDIM; ++i) {
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, uint32_t, {}>", lib_name(), i, wpt));
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, int64_t, {}>", lib_name(), i, wpt));
+      }
     }
-    return std::make_pair(std::move(builder.os), std::move(kernel_names));
+
+    return std::make_tuple(
+        false, std::move(builder.os), std::move(kernel_names));
   });
 
   // Collapse contiguous dims to route to a faster kernel if possible. Also
@@ -178,6 +279,7 @@ void Compiled::eval_gpu(
   // Whether to use large index.
   bool large = compiled_use_large_index(inputs, outputs, contiguous);
 
+  cu::KernelArgs args;
   // Put inputs.
   int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
@@ -185,35 +287,42 @@ void Compiled::eval_gpu(
       continue;
     }
     const auto& x = inputs[i];
-    mod.append_arg(x);
+    args.append(x);
     if (!contiguous && !is_scalar(x)) {
-      mod.append_arg(strides_vec[strides_index++]);
+      args.append_ptr(strides_vec[strides_index++].data());
     }
   }
 
   // Put outputs.
   compiled_allocate_outputs(inputs, outputs, is_constant_, contiguous);
   for (auto& x : outputs) {
-    mod.append_arg(x);
+    args.append(x);
   }
 
   // Put shape and size.
   if (!contiguous) {
-    mod.append_arg(shape);
+    args.append_ptr(shape.data());
   }
   if (large) {
-    mod.append_arg<int64_t>(outputs[0].data_size());
+    args.append<int64_t>(outputs[0].data_size());
   } else {
-    mod.append_arg<uint32_t>(outputs[0].data_size());
+    args.append<uint32_t>(outputs[0].data_size());
+  }
+
+  // Choose work per thread
+  if (!contiguous && shape.back() % work_per_thread != 0) {
+    work_per_thread = 1;
   }
 
   // Launch kernel.
   const char* index_type = large ? "int64_t" : "uint32_t";
   std::string kernel_name = fmt::format("mlx::core::cu::{}", lib_name());
   if (contiguous) {
-    kernel_name += fmt::format("_contiguous<{}>", index_type);
+    kernel_name +=
+        fmt::format("_contiguous<{}, {}>", index_type, work_per_thread);
   } else {
-    kernel_name += fmt::format("_strided<{}, {}>", shape.size(), index_type);
+    kernel_name += fmt::format(
+        "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
   }
   auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
@@ -222,9 +331,11 @@ void Compiled::eval_gpu(
   for (const auto& out : outputs) {
     encoder.set_output_array(out);
   }
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    mod.launch_kernel(stream, kernel_name, outputs[0], large);
-  });
+
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] =
+      get_launch_args(outputs[0], large, work_per_thread);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 } // namespace mlx::core

mlx/backend/cuda/conv.cpp

@@ -0,0 +1,418 @@
+// 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/lru_cache.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+
+#include <cassert>
+
+namespace mlx::core {
+
+namespace {
+
+// Alias for better readability.
+#define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
+#define CONV_BACKWARD_INPUT \
+  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DATA_DESCRIPTOR
+#define CONV_BACKWARD_WEIGHT \
+  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
+
+// Custom placeholder representing fallback kernel.
+#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
+
+struct ConvCacheKey {
+  int device_id;
+  cudnnDataType_t cudnn_dtype;
+  std::array<int, MAX_NDIM> input_shape;
+  std::array<int, MAX_NDIM> weight_shape;
+  std::array<int, MAX_NDIM> stride;
+  std::array<int, MAX_NDIM> padding_lo;
+  std::array<int, MAX_NDIM> padding_hi;
+  std::array<int, MAX_NDIM> dilation;
+  int groups;
+  bool flip;
+  uint8_t input_alignment;
+  uint8_t weight_alignment;
+  uint8_t output_alignment;
+};
+
+auto& conv_cache() {
+  static LRUBytesKeyCache<
+      ConvCacheKey,
+      std::pair<
+          cudnnBackendDescriptorType_t,
+          std::optional<cudnn_frontend::ExecutionPlan>>>
+      cache(/* capacity */ 128);
+  return cache;
+}
+
+auto get_conv_op_settings(
+    cudnnBackendDescriptorType_t backend_type,
+    array& x,
+    array& w,
+    array& y,
+    const std::vector<int>& kernel_strides,
+    const std::vector<int>& padding_lo_,
+    const std::vector<int>& padding_hi_,
+    const std::vector<int>& kernel_dilation,
+    const std::vector<int>& input_dilation) {
+  auto padding_lo = convert_vector<int64_t>(padding_lo_);
+  auto padding_hi = convert_vector<int64_t>(padding_hi_);
+
+  if (backend_type == CONV_BACKWARD_INPUT) {
+    for (int i = 0; i < padding_lo.size(); ++i) {
+      int wt_size = 1 + kernel_dilation[i] * (w.shape(1 + i) - 1);
+      padding_lo[i] = wt_size - padding_lo[i] - 1;
+      int in_size = 1 + kernel_strides[i] * (x.shape(1 + i) - 1);
+      int out_size = 1 + input_dilation[i] * (y.shape(1 + i) - 1);
+      padding_hi[i] = out_size - in_size + padding_hi[i];
+    }
+    return std::make_tuple(
+        convert_vector<int64_t>(input_dilation),
+        std::move(padding_lo),
+        std::move(padding_hi),
+        convert_vector<int64_t>(kernel_dilation));
+
+  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
+    padding_hi = padding_lo;
+    return std::make_tuple(
+        convert_vector<int64_t>(kernel_dilation),
+        std::move(padding_lo),
+        std::move(padding_hi),
+        convert_vector<int64_t>(kernel_strides));
+
+  } else {
+    return std::make_tuple(
+        convert_vector<int64_t>(kernel_strides),
+        std::move(padding_lo),
+        std::move(padding_hi),
+        convert_vector<int64_t>(kernel_dilation));
+  }
+}
+
+std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
+    cu::CommandEncoder& encoder,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    array& x,
+    array& w,
+    array& y,
+    const SmallVector<int64_t>& stride,
+    const SmallVector<int64_t>& padding_lo,
+    const SmallVector<int64_t>& padding_hi,
+    const SmallVector<int64_t>& dilation) {
+  try {
+    auto compute_dtype = (dtype == float16 || dtype == bfloat16)
+        ? CUDNN_DATA_FLOAT
+        : dtype_to_cudnn_type(dtype);
+    auto conv_desc = cudnn_frontend::ConvDescBuilder()
+                         .setDataType(compute_dtype)
+                         .setMathMode(CUDNN_CROSS_CORRELATION)
+                         .setNDims(stride.size())
+                         .setStrides(stride.size(), stride.data())
+                         .setPrePadding(padding_lo.size(), padding_lo.data())
+                         .setPostPadding(padding_hi.size(), padding_hi.data())
+                         .setDilation(dilation.size(), dilation.data())
+                         .build();
+
+    auto op = cudnn_frontend::OperationBuilder(backend_type)
+                  .setxDesc(build_cudnn_tensor_nchw('x', x))
+                  .setwDesc(build_cudnn_tensor_nchw('w', w))
+                  .setyDesc(build_cudnn_tensor_nchw('y', y))
+                  .setcDesc(conv_desc)
+                  .build();
+
+    std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
+    return cudnn_frontend::OperationGraphBuilder()
+        .setHandle(encoder.device().cudnn_handle())
+        .setOperationGraph(ops.size(), ops.data())
+        .build();
+  } catch (cudnn_frontend::cudnnException& error) {
+    if (error.getCudnnStatus() != CUDNN_STATUS_BAD_PARAM) {
+      throw;
+    }
+    return std::nullopt;
+  }
+}
+
+// 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.
+std::tuple<array, array, array> prepare_args(
+    cu::CommandEncoder& encoder,
+    cudnnBackendDescriptorType_t backend_type,
+    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);
+  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
+    in = group_transpose(in, groups, -1, 0, -1, s);
+    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.
+    Shape shape(out.shape());
+    std::swap(shape.front(), shape.back());
+    Strides strides(shape.size(), 1);
+    for (int i = shape.size() - 2; i >= 0; --i) {
+      strides[i] = shape[i + 1] * strides[i + 1];
+    }
+    array intermediate(std::move(shape), out.dtype(), nullptr, {});
+    intermediate.copy_shared_buffer(
+        out, std::move(strides), {true, true, false}, out.data_size());
+    out = intermediate;
+  }
+
+  // cuDNN requires contiguous input.
+  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);
+  }
+
+  return {std::move(in), std::move(wt), std::move(out)};
+}
+
+// Get the x/w/y args from the in/wt/out args depending on backend type.
+inline std::tuple<array&, array&, array&> dispatch_args(
+    cudnnBackendDescriptorType_t backend_type,
+    array& in,
+    array& wt,
+    array& out) {
+  switch (backend_type) {
+    case CONV_BACKWARD_INPUT:
+      return {out, wt, in};
+    case CONV_BACKWARD_WEIGHT:
+      return {in, out, wt};
+    default:
+      return {in, wt, out};
+  }
+}
+
+// Register inputs and outputs before actually running conv op. Can only be
+// called once per eval_gpu.
+void register_args(
+    cu::CommandEncoder& encoder,
+    cudnnBackendDescriptorType_t backend_type,
+    array& in,
+    array& wt,
+    array& intermediate_out,
+    array& final_out) {
+  encoder.set_input_array(in);
+  encoder.set_input_array(wt);
+  encoder.set_output_array(final_out);
+
+  if (backend_type == CONV_BACKWARD_WEIGHT) {
+    // Turn |out| into a strided array, which will have C_in and C_out swapped
+    // in vjp and the final |grad_weight| will then be contiguous.
+    Strides strides = intermediate_out.strides();
+    std::swap(strides.front(), strides.back());
+    final_out.copy_shared_buffer(
+        intermediate_out,
+        std::move(strides),
+        {false, false, false},
+        intermediate_out.data_size());
+  }
+}
+
+} // namespace
+
+void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
+  nvtx3::scoped_range r("Convolution::eval_gpu");
+  if (out_.size() == 0) {
+    return;
+  }
+
+  assert(inputs.size() == 2);
+  array in = inputs[0];
+  array wt = inputs[1];
+  array out = out_;
+  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_),
+      groups_,
+      flip_,
+      get_alignment(in),
+      get_alignment(wt),
+      get_alignment(out)};
+  if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
+    auto& [backend_type, plan] = it->second;
+    if (plan) {
+      // Run cached plan.
+      std::tie(in, wt, out) =
+          prepare_args(encoder, backend_type, in, wt, out, groups_, s);
+      register_args(encoder, backend_type, in, wt, out, out_);
+      auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
+      if (!encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+        throw std::runtime_error("[conv] Cached plan failed to execute.");
+      }
+    } else {
+      // Run fallback kernel.
+      gemm_conv(
+          encoder,
+          in,
+          wt,
+          out,
+          kernel_strides_,
+          padding_lo_,
+          kernel_dilation_,
+          input_dilation_,
+          groups_,
+          flip_,
+          s);
+    }
+    return;
+  }
+
+  // There is no reliable way to deduce the proper cuDNN backend for the
+  // convolution, so we make a best guess and then try.
+  SmallVector<cudnnBackendDescriptorType_t, 2> try_backends;
+  if (flip_) {
+    // When weight is flipped, we assume it is backward input convolution.
+    try_backends.push_back(CONV_BACKWARD_INPUT);
+  } else {
+    // Otherwise it could be backward weight convolution or forward convolution,
+    // mathematically there is no difference so we have to use heuristics.
+    // Empirically backward convolutions have large kernel dimensions, and
+    // usually have |in| and |wt| transposed.
+    if (!in.flags().row_contiguous && !wt.flags().row_contiguous &&
+        wt.shape(2) > out.shape(2)) {
+      try_backends = {CONV_BACKWARD_WEIGHT, CONV_FORWARD};
+    } else {
+      try_backends = {CONV_FORWARD, CONV_BACKWARD_WEIGHT};
+    }
+  }
+
+  // Try to build op graph.
+  cudnnBackendDescriptorType_t backend_type;
+  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);
+    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,
+        x,
+        w,
+        y,
+        kernel_strides_,
+        padding_lo_,
+        padding_hi_,
+        kernel_dilation_,
+        input_dilation_);
+    op_graph = build_conv_op_graph(
+        encoder,
+        try_backend,
+        dtype,
+        x,
+        w,
+        y,
+        stride,
+        padding_lo,
+        padding_hi,
+        dilation);
+    if (op_graph) {
+      backend_type = try_backend;
+      in = std::move(in_copy);
+      wt = std::move(wt_copy);
+      out = std::move(out_copy);
+      break;
+    }
+  }
+
+  if (op_graph) {
+    // Setup inputs and outputs.
+    register_args(encoder, backend_type, in, wt, out, out_);
+
+    // 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) {
+      throw std::runtime_error("[conv] Unable to find an execution plan.");
+    }
+    auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
+    if (encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+      conv_cache().emplace(
+          cache_key, std::make_pair(backend_type, std::move(*plan)));
+      return;
+    }
+  }
+
+  // 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));
+}
+
+} // namespace mlx::core

mlx/backend/cuda/conv/conv.h

@@ -0,0 +1,126 @@
+// 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

@@ -0,0 +1,217 @@
+// 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(allocator::malloc(unfolded.nbytes()));
+  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,
+        in.data<DataType>(),
+        unfolded.data<DataType>(),
+        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

@@ -0,0 +1,231 @@
+// 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(allocator::malloc(unfolded.nbytes()));
+  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,
+        in.data<DataType>(),
+        unfolded.data<DataType>(),
+        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

@@ -15,8 +15,8 @@ void copy_gpu_inplace(
     int64_t offset_out,
     CopyType ctype,
     const Stream& s,
-    const std::optional<array>& dynamic_offset_in,
-    const std::optional<array>& dynamic_offset_out) {
+    std::optional<array> dynamic_offset_in,
+    std::optional<array> dynamic_offset_out) {
   if (out.size() == 0) {
     return;
   }
@@ -24,7 +24,6 @@ void copy_gpu_inplace(
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-
   if (ctype == CopyType::Scalar || ctype == CopyType::Vector) {
     copy_contiguous(encoder, ctype, in, out, offset_in, offset_out);
     return;
@@ -45,6 +44,16 @@ 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,
@@ -55,8 +64,8 @@ void copy_gpu_inplace(
             shape_collapsed,
             strides_vec[0],
             strides_vec[1],
-            dynamic_offset_in ? *dynamic_offset_in : array(0, int64),
-            dynamic_offset_out ? *dynamic_offset_out : array(0, int64));
+            *dynamic_offset_in,
+            *dynamic_offset_out);
       } else {
         copy_general(
             encoder,

mlx/backend/cuda/copy/copy.cuh

@@ -10,15 +10,6 @@
 
 namespace mlx::core {
 
-#define MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, ...) \
-  MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE_IN, {               \
-    MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, {           \
-      using InType = cuda_type_t<CTYPE_IN>;                  \
-      using OutType = cuda_type_t<CTYPE_OUT>;                \
-      __VA_ARGS__;                                           \
-    });                                                      \
-  })
-
 void copy_contiguous(
     cu::CommandEncoder& encoder,
     CopyType ctype,

mlx/backend/cuda/copy/copy_contiguous.cu

@@ -10,19 +10,43 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_s(const In* in, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = CastOp<In, Out>{}(in[0]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec[i] = cast_to<Out>(in[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_v(const In* in, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = CastOp<In, Out>{}(in[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[i]);
+    }
+  } else {
+    auto in_vec = load_vector<N_READS>(in, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec[i] = cast_to<Out>(in_vec[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -35,17 +59,24 @@ void copy_contiguous(
     array& out,
     int64_t in_offset,
     int64_t out_offset) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
-        auto kernel = cu::copy_s<InType, OutType, IdxT>;
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+        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, uint32_t>;
+        constexpr int N_READS = 16 / sizeof(InType);
+        auto kernel = cu::copy_s<InType, OutType, IdxT, N_READS>;
         if (ctype == CopyType::Vector) {
-          kernel = cu::copy_v<InType, OutType, IdxT>;
+          kernel = cu::copy_v<InType, OutType, IdxT, N_READS>;
         }
         auto [num_blocks, block_dims] = get_launch_args(
-            kernel, out.data_size(), out.shape(), out.strides(), LARGE);
-        kernel<<<num_blocks, block_dims, 0, stream>>>(
+            out.data_size(), out.shape(), out.strides(), large(), N_READS);
+        encoder.add_kernel_node(
+            kernel,
+            num_blocks,
+            block_dims,
+            0,
             in.data<InType>() + in_offset,
             out.data<OutType>() + out_offset,
             out.data_size());

mlx/backend/cuda/copy/copy_general.cu

@@ -10,37 +10,80 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM>
+template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
 __global__ void copy_gg_nd(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     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) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
-        index, shape.data(), strides_in.data(), strides_out.data());
-    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  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;
+  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]);
+  }
+  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_gg(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides strides_in,
     const __grid_constant__ Strides strides_out,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_4d(
-        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
-    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  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;
+  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]);
+  }
+  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
 } // namespace cu
@@ -55,39 +98,72 @@ void copy_general(
     const Shape& shape,
     const Strides& strides_in,
     const Strides& strides_out) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      const InType* in_ptr = in.data<InType>() + offset_in;
-      OutType* out_ptr = out.data<OutType>() + offset_out;
-      bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-      MLX_SWITCH_BOOL(large, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-        int ndim = shape.size();
-        if (ndim <= 3) {
-          MLX_SWITCH_1_2_3(ndim, NDIM, {
-            auto kernel = cu::copy_gg_nd<InType, OutType, IdxT, NDIM>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-            kernel<<<num_blocks, block_dims, 0, stream>>>(
-                in_ptr,
-                out_ptr,
-                out.size(),
-                const_param<NDIM>(shape),
-                const_param<NDIM>(strides_in),
-                const_param<NDIM>(strides_out));
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
+            using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            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>;
+                }
+                encoder.add_kernel_node(
+                    kernel,
+                    {num_blocks_x, num_blocks_y},
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    rest,
+                    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>;
+              }
+              encoder.add_kernel_node(
+                  kernel,
+                  {num_blocks_x, num_blocks_y},
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  rest,
+                  const_param(shape),
+                  const_param(strides_in),
+                  const_param(strides_out),
+                  ndim);
+            }
           });
-        } else { // ndim >= 4
-          auto kernel = cu::copy_gg<InType, OutType, IdxT>;
-          auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in_ptr,
-              out_ptr,
-              out.size(),
-              const_param(shape),
-              const_param(strides_in),
-              const_param(strides_out),
-              ndim);
-        }
-      });
     });
   });
 }

mlx/backend/cuda/copy/copy_general_dynamic.cu

@@ -41,7 +41,7 @@ __global__ void copy_gg_dynamic(
     const int64_t* offset_out) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_4d(
+    auto [idx_in, idx_out] = elem_to_loc(
         index, shape.data(), strides_in.data(), strides_out.data(), ndim);
     out[idx_out + *offset_out] = CastOp<In, Out>{}(in[idx_in + *offset_in]);
   }
@@ -61,43 +61,56 @@ void copy_general_dynamic(
     const Strides& strides_out,
     const array& dynamic_offset_in,
     const array& dynamic_offset_out) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      const InType* in_ptr = in.data<InType>() + offset_in;
-      OutType* out_ptr = out.data<OutType>() + offset_out;
-      bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-      MLX_SWITCH_BOOL(large, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-        int ndim = shape.size();
-        if (ndim <= 3) {
-          MLX_SWITCH_1_2_3(ndim, NDIM, {
-            auto kernel = cu::copy_gg_dynamic_nd<InType, OutType, IdxT, NDIM>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-            kernel<<<num_blocks, block_dims, 0, stream>>>(
-                in_ptr,
-                out_ptr,
-                out.size(),
-                const_param<NDIM>(shape),
-                const_param<NDIM>(strides_in),
-                const_param<NDIM>(strides_out),
-                dynamic_offset_in.data<int64_t>(),
-                dynamic_offset_out.data<int64_t>());
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
+            using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            int ndim = shape.size();
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                auto [num_blocks, block_dims] = get_launch_args(out, large());
+                encoder.add_kernel_node(
+                    cu::copy_gg_dynamic_nd<
+                        InType,
+                        OutType,
+                        IdxT,
+                        dims_constant()>,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    out.size(),
+                    const_param<dims_constant()>(shape),
+                    const_param<dims_constant()>(strides_in),
+                    const_param<dims_constant()>(strides_out),
+                    dynamic_offset_in.data<int64_t>(),
+                    dynamic_offset_out.data<int64_t>());
+              });
+            } else { // ndim >= 4
+              auto [num_blocks, block_dims] = get_launch_args(out, large());
+              encoder.add_kernel_node(
+                  cu::copy_gg_dynamic<InType, OutType, IdxT>,
+                  num_blocks,
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  out.size(),
+                  const_param(shape),
+                  const_param(strides_in),
+                  const_param(strides_out),
+                  ndim,
+                  dynamic_offset_in.data<int64_t>(),
+                  dynamic_offset_out.data<int64_t>());
+            }
           });
-        } else { // ndim >= 4
-          auto kernel = cu::copy_gg_dynamic<InType, OutType, IdxT>;
-          auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in_ptr,
-              out_ptr,
-              out.size(),
-              const_param(shape),
-              const_param(strides_in),
-              const_param(strides_out),
-              ndim,
-              dynamic_offset_in.data<int64_t>(),
-              dynamic_offset_out.data<int64_t>());
-        }
-      });
     });
   });
 }

mlx/backend/cuda/copy/copy_general_input.cu

@@ -10,33 +10,67 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM>
+template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
 __global__ void copy_g_nd(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
-    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]);
+    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;
   }
+
+  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>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_g(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
-    const __grid_constant__ Strides strides_in,
+    const __grid_constant__ Strides strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    IdxT idx_in = elem_to_loc_4d(index, shape.data(), strides_in.data(), ndim);
-    out[index] = CastOp<In, Out>{}(in[idx_in]);
+  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 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
@@ -50,37 +84,65 @@ void copy_general_input(
     int64_t offset_out,
     const Shape& shape,
     const Strides& strides_in) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      const InType* in_ptr = in.data<InType>() + offset_in;
-      OutType* out_ptr = out.data<OutType>() + offset_out;
-      bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-      MLX_SWITCH_BOOL(large, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-        int ndim = shape.size();
-        if (ndim <= 3) {
-          MLX_SWITCH_1_2_3(ndim, NDIM, {
-            auto kernel = cu::copy_g_nd<InType, OutType, IdxT, NDIM>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-            kernel<<<num_blocks, block_dims, 0, stream>>>(
-                in_ptr,
-                out_ptr,
-                out.size(),
-                const_param<NDIM>(shape),
-                const_param<NDIM>(strides_in));
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
+            using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            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>;
+                }
+                encoder.add_kernel_node(
+                    kernel,
+                    {num_blocks_x, num_blocks_y},
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    rest,
+                    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>;
+              }
+              encoder.add_kernel_node(
+                  kernel,
+                  {num_blocks_x, num_blocks_y},
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  rest,
+                  const_param(shape),
+                  const_param(strides_in),
+                  ndim);
+            }
           });
-        } else { // ndim >= 4
-          auto kernel = cu::copy_g<InType, OutType, IdxT>;
-          auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in_ptr,
-              out_ptr,
-              out.size(),
-              const_param(shape),
-              const_param(strides_in),
-              ndim);
-        }
-      });
     });
   });
 }

mlx/backend/cuda/cudnn_utils.cpp

@@ -0,0 +1,272 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/cudnn_utils.h"
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core {
+
+namespace {
+
+// Create a cudnn tensor descriptor.
+template <typename Vec>
+inline cudnn_frontend::Tensor build_cudnn_tensor(
+    int64_t id,
+    const array& x,
+    const Vec& shape,
+    const Vec& strides) {
+  return cudnn_frontend::TensorBuilder()
+      .setDim(shape.size(), shape.data())
+      .setStrides(strides.size(), strides.data())
+      .setId(id)
+      .setAlignment(get_alignment(x))
+      .setDataType(dtype_to_cudnn_type(x.dtype()))
+      .build();
+}
+
+// In MLX a singleton dim (shape[dim] == 1) can have any stride, but in cuDNN
+// whether a tensor is contiguous is determined with:
+// shape[dim] == shape[dim + 1] * strides[dim + 1]
+// So a contiguous array with singleton dims in MLX may be mistakenly treated
+// as strided in cuDNN, and we work around it by normalizing the strides.
+Strides normalized_strides(const array& x) {
+  if (!x.flags().row_contiguous || x.ndim() < 2) {
+    return x.strides();
+  }
+  Strides strides = x.strides();
+  for (int i = x.ndim() - 2; i >= 0; --i) {
+    if (x.shape(i) == 1) {
+      strides[i] = x.shape(i + 1) * strides[i + 1];
+    }
+  }
+  return strides;
+}
+
+// Return the shape and strides after transposing from NHWC to NCHW.
+auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
+  assert(shape.size() >= 3);
+  shape.insert(shape.begin() + 1, shape.back());
+  shape.erase(shape.end() - 1);
+  strides.insert(strides.begin() + 1, strides.back());
+  strides.erase(strides.end() - 1);
+  return std::make_tuple(std::move(shape), std::move(strides));
+}
+
+inline auto nhwc_to_nchw(const array& x) {
+  return nhwc_to_nchw(
+      convert_vector<int64_t>(x.shape()), normalized_strides(x));
+}
+
+// Return available engines for a |op_graph|.
+cudnn_frontend::EngineConfigList get_cudnn_engine_configs(
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph,
+    bool use_fallback = true) {
+  SmallVector<cudnn_frontend::GeneratorSource, 2> sources;
+  sources.push_back([](auto& op_graph) {
+    auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
+                          .setOperationGraph(op_graph)
+                          .setHeurMode(CUDNN_HEUR_MODE_A)
+                          .build();
+    return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
+  });
+  if (use_fallback) {
+    sources.push_back([&backend_type](auto& op_graph) {
+      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
+                          .setOperationGraph(op_graph)
+                          .setOperation(backend_type)
+                          .build();
+      return fallback.getFallbackList();
+    });
+  }
+
+  auto configs =
+      cudnn_frontend::EngineConfigGenerator(sources.size(), sources.data())
+          .generate_engine_config(op_graph);
+
+  cudnn_frontend::EngineConfigList filtered_configs;
+  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
+    if (cudnn_frontend::hasNumericalNote<
+            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
+      return true;
+    }
+    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
+        dtype == float32 && !env::enable_tf32()) {
+      return true;
+    }
+    return false;
+  });
+  return filtered_configs;
+}
+
+// Take |engine_configs| and |op_graph| and find a working execution plans
+// from them.
+std::optional<cudnn_frontend::ExecutionPlan>
+find_cudnn_plan_from_engine_configs(
+    cudnnHandle_t handle,
+    const cudnn_frontend::EngineConfigList& engine_configs,
+    const cudnn_frontend::OperationGraph& op_graph) {
+  auto op_graph_tag = op_graph.getTag();
+  for (const auto& config : engine_configs) {
+    try {
+      return cudnn_frontend::ExecutionPlanBuilder()
+          .setHandle(handle)
+          .setEngineConfig(config, op_graph_tag)
+          .build();
+    } catch (cudnn_frontend::cudnnException& error) {
+      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
+        throw;
+      }
+    }
+  }
+  return std::nullopt;
+}
+
+// Prepare workspace and args to execute plan.
+template <typename F>
+bool prepare_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs,
+    F&& execute) {
+  int workspace_size = plan.getWorkspaceSize();
+  array workspace(
+      workspace_size > 0 ? allocator::malloc(workspace_size)
+                         : allocator::Buffer(nullptr),
+      {workspace_size},
+      uint8);
+
+  auto args = cudnn_frontend::VariantPackBuilder()
+                  .setWorkspacePointer(workspace.data<void>())
+                  .setDataPointers(num_args, data_ptrs)
+                  .setUids(num_args, uids)
+                  .build();
+
+  auto handle = encoder.device().cudnn_handle();
+  cudnnSetStream(handle, encoder.stream());
+
+  if (!execute(handle, plan.get_raw_desc(), args.get_raw_desc())) {
+    return false;
+  }
+
+  encoder.add_temporary(workspace);
+  return true;
+}
+
+} // namespace
+
+cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  return build_cudnn_tensor(id, x, shape, normalized_strides(x));
+}
+
+cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x) {
+  auto [shape, strides] = nhwc_to_nchw(x);
+  return build_cudnn_tensor(id, x, shape, strides);
+}
+
+cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x) {
+  if (x.ndim() == 0) {
+    SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
+    return build_cudnn_tensor(id, x, scalar_dims, scalar_dims);
+  }
+  if (x.ndim() == 1) {
+    int64_t s = x.shape(0);
+    SmallVector<int64_t, 4> shape = {1, x.shape(0), 1, 1};
+    SmallVector<int64_t, 4> strides = {s, 1, s, s};
+    return build_cudnn_tensor(id, x, shape, strides);
+  }
+  if (x.ndim() == 2) {
+    int64_t s =
+        x.flags().row_contiguous ? x.shape(1) * x.strides(1) : x.strides(0);
+    SmallVector<int64_t, 4> shape = {x.shape(0), x.shape(1), 1, 1};
+    SmallVector<int64_t, 4> strides = {s, x.strides(1), s, s};
+    return build_cudnn_tensor(id, x, shape, strides);
+  }
+  if (x.ndim() == 3 || x.ndim() == 4) {
+    return build_cudnn_tensor_nchw(id, x);
+  }
+  throw std::runtime_error(
+      fmt::format("Unsupported array with {} dims.", x.ndim()));
+}
+
+cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype) {
+  SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
+  return cudnn_frontend::TensorBuilder()
+      .setDim(scalar_dims.size(), scalar_dims.data())
+      .setStrides(scalar_dims.size(), scalar_dims.data())
+      .setId(id)
+      .setAlignment(16)
+      .setDataType(dtype_to_cudnn_type(dtype))
+      .setByValue(true)
+      .build();
+}
+
+std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
+    cudnnHandle_t handle,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph) {
+  auto engine_configs = get_cudnn_engine_configs(backend_type, dtype, op_graph);
+  return find_cudnn_plan_from_engine_configs(handle, engine_configs, op_graph);
+}
+
+bool encode_cudnn_plan_with_capturing(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs) {
+  return prepare_cudnn_plan(
+      encoder,
+      plan,
+      num_args,
+      uids,
+      data_ptrs,
+      [&](auto handle, auto plan, auto args) {
+        auto capture = encoder.capture_context();
+        if (cudnnBackendExecute(handle, plan, args) != CUDNN_STATUS_SUCCESS) {
+          // Discard the captured graph when failed.
+          capture.discard = true;
+          return false;
+        }
+        return true;
+      });
+}
+
+#if CUDNN_VERSION >= 90500
+bool encode_cudnn_plan_with_graph_api(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs) {
+  return prepare_cudnn_plan(
+      encoder,
+      plan,
+      num_args,
+      uids,
+      data_ptrs,
+      [&](auto handle, auto plan, auto args) {
+        if (!graph) {
+          graph = CudaGraph(encoder.device());
+          if (cudnnBackendPopulateCudaGraph(handle, plan, args, graph) !=
+              CUDNN_STATUS_SUCCESS) {
+            return false;
+          }
+        } else {
+          if (cudnnBackendUpdateCudaGraph(handle, plan, args, graph) !=
+              CUDNN_STATUS_SUCCESS) {
+            return false;
+          }
+        }
+        encoder.add_graph_node(graph);
+        return true;
+      });
+}
+#endif
+
+} // namespace mlx::core

mlx/backend/cuda/cudnn_utils.h

@@ -0,0 +1,164 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/cuda/device/config.h"
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/dtype_utils.h"
+
+#include <cudnn_frontend.h>
+#include <cudnn_frontend_find_plan.h>
+#include <fmt/format.h>
+
+#include <algorithm>
+#include <array>
+
+namespace mlx::core {
+
+namespace cu {
+class CommandEncoder;
+}
+
+// Return pointer alignment of |x|'s data.
+inline uint8_t get_alignment(const array& x) {
+  uint8_t alignment = 1;
+  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
+  for (; alignment < 32; alignment *= 2) {
+    if (address % (alignment * 2)) {
+      return alignment;
+    }
+  }
+  return alignment;
+}
+
+// Convert the type of elements in |vec| to |T|.
+template <typename T, typename Vec>
+inline SmallVector<T> convert_vector(const Vec& vec) {
+  return SmallVector<T>(vec.begin(), vec.end());
+}
+
+// Return an array that can be used as map key for |vec| with size <= MAX_NDIM.
+//
+// There are 2 differences from the const_param util from kernel_utils.cuh:
+// 1. The rest of array is filled with 0.
+// 2. This util can be used in .cpp files.
+template <typename T, template <typename U> class Vec>
+inline std::array<T, MAX_NDIM> vector_key(const Vec<T>& vec) {
+  if (vec.size() > MAX_NDIM) {
+    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;
+}
+
+// Helpers used by get_data_ptrs to get pointers.
+inline void* get_data_ptr(const array& arr) {
+  return const_cast<void*>(arr.data<void>());
+}
+
+template <typename T, typename = std::enable_if_t<std::is_scalar_v<T>>>
+inline void* get_data_ptr(T& scalar) {
+  return &scalar;
+}
+
+// Return an array filled with data pointers of args.
+template <typename... Args>
+inline std::array<void*, sizeof...(Args)> get_data_ptrs(Args&... args) {
+  return {get_data_ptr(args)...};
+}
+
+// Map dtype to cudnn data type.
+inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return CUDNN_DATA_INT8;
+    case int32:
+      return CUDNN_DATA_INT32;
+    case uint8:
+      return CUDNN_DATA_UINT8;
+    case float16:
+      return CUDNN_DATA_HALF;
+    case bfloat16:
+      return CUDNN_DATA_BFLOAT16;
+    case float32:
+      return CUDNN_DATA_FLOAT;
+    case float64:
+      return CUDNN_DATA_DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+  }
+}
+
+// Create a tensor descriptor from |x|.
+cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x);
+
+// Create a tensor descriptor from |x|, and transpose from NHWC to NCHW.
+cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x);
+
+// Create a tensor descriptor from |x|, make sure it is 4D, and transpose it
+// from NHWC to NCHW.
+cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x);
+
+// Create a 4D scalar tensor descriptor, which is passed by value.
+cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype);
+
+// Find a working plan for |op_graph|.
+std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
+    cudnnHandle_t handle,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph);
+
+// Encode the plan to command buffer by capturing.
+bool encode_cudnn_plan_with_capturing(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs);
+
+#if CUDNN_VERSION >= 90500
+// Encode the plan to command buffer by using native graph api of cudnn. If the
+// |graph| is empty it will be populated, otherwise it will be updated.
+bool encode_cudnn_plan_with_graph_api(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs);
+#endif
+
+// Helpers to make calls like encode_cudnn_plan(..., {'x', 'y', 'z'}, x, y, z).
+template <typename... Args>
+bool encode_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    std::initializer_list<int64_t> uids,
+    Args&... args) {
+  assert(uids.size() == sizeof...(args));
+  auto data_ptrs = get_data_ptrs(args...);
+  return encode_cudnn_plan_with_capturing(
+      encoder, plan, uids.size(), uids.begin(), data_ptrs.data());
+}
+
+#if CUDNN_VERSION >= 90500
+template <typename... Args>
+bool encode_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    std::initializer_list<int64_t> uids,
+    Args&... args) {
+  assert(uids.size() == sizeof...(args));
+  auto data_ptrs = get_data_ptrs(args...);
+  return encode_cudnn_plan_with_graph_api(
+      encoder, plan, graph, uids.size(), uids.begin(), data_ptrs.data());
+}
+#endif
+
+} // namespace mlx::core