version (#2470)

* revert default cuda install

* revert default cuda install

.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:
@@ -24,8 +18,8 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "15.2.0"
-    resource_class: macos.m1.medium.gen1
+      xcode: "16.2.0"
+    resource_class: m2pro.medium
     steps:
       - checkout
       - run:
@@ -38,7 +32,7 @@ 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 +64,9 @@ 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:
@@ -84,36 +78,36 @@ 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_COMPILE_WARNING_AS_ERROR=ON" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF \
-              CMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py develop
+            uv venv
+            uv pip install cmake
+            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 -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:
@@ -124,60 +118,63 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "15.2.0"
+        default: "16.2.0"
+      macosx_deployment_target:
+        type: string
+        default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.medium.gen1
+    environment:
+      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
+    resource_class: m2pro.medium
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
-            brew install python@3.9
-            brew install openmpi
-            python3.9 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
-            pip install torch
-            pip install tensorflow
-            pip install unittest-xml-reporting
+            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="CMAKE_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
@@ -186,7 +183,7 @@ 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 \
@@ -198,13 +195,60 @@ 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 .
             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:<< 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
+            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: |
+            uv venv
+            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              uv pip install -e ".[dev]" -v
+      - run:
+          name: Run Python tests
+          command: |
+            source .venv/bin/activate
+            LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
+            LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
+      - run:
+          name: 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:
@@ -213,13 +257,18 @@ jobs:
         default: "3.9"
       xcode_version:
         type: string
-        default: "15.2.0"
+        default: "16.2.0"
       build_env:
         type: string
         default: ""
+      macosx_deployment_target:
+        type: string
+        default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.medium.gen1
+    resource_class: m2pro.medium
+    environment:
+      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
     steps:
       - checkout
       - run:
@@ -240,22 +289,30 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
+            env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
             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
+            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: |
+                  source env/bin/activate
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
           condition: << parameters.build_env >>
           steps:
@@ -272,52 +329,100 @@ 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: large
+    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/
 
@@ -329,21 +434,23 @@ 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:
-              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test
+      - cuda_build_and_test:
+          matrix:
+            parameters:
+              image_date: ["2023.11.1", "2025.05.1"]
       - build_documentation 
 
   build_pypi_release:
     when:
       and:
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
         - not: << pipeline.parameters.test_release >>
     jobs:
       - build_release:
@@ -355,8 +462,70 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              xcode_version: ["15.0.0", "15.2.0"]
+              macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["PYPI_RELEASE=1"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
       - build_documentation:
           filters:
             tags:
@@ -364,6 +533,25 @@ 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:
@@ -379,9 +567,14 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
       - 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:
@@ -392,27 +585,140 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              xcode_version: ["15.0.0", "15.2.0"]
-  weekly_build:
+              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"
+      - 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:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
+              macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["DEV_RELEASE=1"]
-  linux_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.linux_release >>
-    jobs:
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
       - build_linux_release:
           matrix:
             parameters:
               python_version: ["3.9", "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"]

.gitignore

@@ -36,6 +36,7 @@ share/python-wheels/
 .installed.cfg
 *.egg
 MANIFEST
+uv.lock
 
 # vim
 *.swp

ACKNOWLEDGMENTS.md

@@ -19,6 +19,7 @@ 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" />

CMakeLists.txt

@@ -34,13 +34,16 @@ option(MLX_BUILD_BENCHMARKS "Build benchmarks for mlx" OFF)
 option(MLX_BUILD_PYTHON_BINDINGS "Build python bindings for mlx" OFF)
 option(MLX_BUILD_METAL "Build metal backend" ON)
 option(MLX_BUILD_CPU "Build cpu backend" ON)
+option(MLX_BUILD_CUDA "Build cuda backend" OFF)
 option(MLX_METAL_DEBUG "Enhance metal debug workflow" OFF)
 option(MLX_ENABLE_X64_MAC "Enable building for x64 macOS" OFF)
 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(
@@ -63,10 +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 -----------------------------
@@ -83,6 +93,10 @@ if(MLX_BUILD_METAL)
   set(QUARTZ_LIB "-framework QuartzCore")
 endif()
 
+if(MLX_BUILD_CUDA)
+  enable_language(CUDA)
+endif()
+
 if(MLX_BUILD_METAL AND NOT METAL_LIB)
   message(STATUS "Metal not found. Unable to build GPU")
   set(MLX_BUILD_METAL OFF)
@@ -212,24 +226,6 @@ else()
   set(MLX_BUILD_ACCELERATE OFF)
 endif()
 
-find_package(MPI)
-if(MPI_FOUND)
-  execute_process(
-    COMMAND zsh "-c" "mpirun --version"
-    OUTPUT_VARIABLE MPI_VERSION
-    ERROR_QUIET)
-  if(${MPI_VERSION} MATCHES ".*Open MPI.*")
-    target_include_directories(mlx PRIVATE ${MPI_INCLUDE_PATH})
-  elseif(MPI_VERSION STREQUAL "")
-    set(MPI_FOUND FALSE)
-    message(
-      WARNING "MPI found but mpirun is not available. Building without MPI.")
-  else()
-    set(MPI_FOUND FALSE)
-    message(WARNING "MPI which is not OpenMPI found. Building without MPI.")
-  endif()
-endif()
-
 message(STATUS "Downloading json")
 FetchContent_Declare(
   json
@@ -244,12 +240,19 @@ target_include_directories(
   mlx PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
              $<INSTALL_INTERFACE:include>)
 
-FetchContent_Declare(
-  fmt
-  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
-  GIT_TAG 10.2.1
-  EXCLUDE_FROM_ALL)
-FetchContent_MakeAvailable(fmt)
+# Do not add mlx_EXPORTS define for shared library.
+set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")
+
+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)

CONTRIBUTING.md

@@ -5,26 +5,26 @@ possible.
 
 ## Pull Requests
 
-1. Fork and submit pull requests to the repo. 
+1. Fork and submit pull requests to the repo.
 2. If you've added code that should be tested, add tests.
 3. If a change is likely to impact efficiency, run some of the benchmarks before
    and after the change. Examples of benchmarks can be found in `benchmarks/python/`.
 4. If you've changed APIs, update the documentation.
-5. Every PR should have passing tests and at least one review. 
+5. Every PR should have passing tests and at least one review.
 6. For code formatting install `pre-commit` using something like `pip install pre-commit` and run `pre-commit install`.
    This should install hooks for running `black` and `clang-format` to ensure
    consistent style for C++ and python code.
- 
+
    You can also run the formatters manually as follows:
- 
-     ```
-     clang-format -i file.cpp
-     ```
- 
-     ```
-     black file.py
-     ```
- 
+
+   ```shell
+   clang-format -i file.cpp
+   ```
+
+   ```shell
+   black file.py
+   ```
+
    or run `pre-commit run --all-files` to check all files in the repo.
 
 ## Issues

MANIFEST.in

@@ -1,4 +1,6 @@
 include CMakeLists.txt
+include mlx.pc.in
 recursive-include mlx/ *
+include cmake/*
 include python/src/*
 include python/mlx/py.typed # support type hinting as in PEP-561

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/irregular_strides.cpp

@@ -1,5 +1,6 @@
 // Copyright © 2023 Apple Inc.
 
+#include <cstring>
 #include <iostream>
 #include <sstream>
 

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/conv_unaligned_bench.py

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

benchmarks/python/gather_mm_bench.py

@@ -0,0 +1,74 @@
+# Copyright © 2025 Apple Inc.
+
+import mlx.core as mx
+from time_utils import time_fn
+
+N = 1024
+D = 1024
+M = 1024
+E = 32
+I = 4
+
+
+def gather_sort(x, indices):
+    N, M = indices.shape
+    indices = indices.flatten()
+    order = mx.argsort(indices)
+    inv_order = mx.argsort(order)
+    return x.flatten(0, -3)[order // M], indices[order], inv_order
+
+
+def scatter_unsort(x, inv_order, shape=None):
+    x = x[inv_order]
+    if shape is not None:
+        x = mx.unflatten(x, 0, shape)
+    return x
+
+
+def gather_mm_simulate(x, w, indices):
+    x, idx, inv_order = gather_sort(x, indices)
+    for i in range(2):
+        y = mx.concatenate([x[i] @ w[j].T for i, j in enumerate(idx.tolist())], axis=0)
+        x = y[:, None]
+    x = scatter_unsort(x, inv_order, indices.shape)
+    return x
+
+
+def time_gather_mm():
+    x = mx.random.normal((N, 1, 1, D)) / 1024**0.5
+    w1 = mx.random.normal((E, M, D)) / 1024**0.5
+    w2 = mx.random.normal((E, D, M)) / 1024**0.5
+    indices = (mx.random.uniform(shape=(N, I)) * E).astype(mx.uint32)
+    sorted_indices = mx.sort(indices.flatten()).reshape(N, I)
+    mx.eval(x, w1, w2, indices, sorted_indices)
+
+    def gather_mm(x, w1, w2, indices, sort):
+        idx = indices
+        inv_order = None
+        if sort:
+            x, idx, inv_order = gather_sort(x, indices)
+        x = mx.gather_mm(x, w1.swapaxes(-1, -2), rhs_indices=idx, sorted_indices=sort)
+        x = mx.gather_mm(x, w2.swapaxes(-1, -2), rhs_indices=idx, sorted_indices=sort)
+        if sort:
+            x = scatter_unsort(x, inv_order, indices.shape)
+        return x
+
+    time_fn(gather_mm, x, w1, w2, indices, False)
+    time_fn(gather_mm, x, w1, w2, sorted_indices, False)
+    time_fn(gather_mm, x, w1, w2, indices, True)
+
+    x = mx.random.normal((N * I, D)) / 1024**0.5
+    w1 = mx.random.normal((M, D)) / 1024**0.5
+    w2 = mx.random.normal((D, M)) / 1024**0.5
+    mx.eval(x, w1, w2)
+
+    def equivalent_matmul(x, w1, w2):
+        x = x @ w1.T
+        x = x @ w2.T
+        return x
+
+    time_fn(equivalent_matmul, x, w1, w2)
+
+
+if __name__ == "__main__":
+    time_gather_mm()

benchmarks/python/gather_qmm_bench.py

@@ -0,0 +1,84 @@
+# Copyright © 2025 Apple Inc.
+
+import mlx.core as mx
+from time_utils import time_fn
+
+N = 1024
+D = 1024
+M = 1024
+E = 32
+I = 4
+
+
+def gather_sort(x, indices):
+    N, M = indices.shape
+    indices = indices.flatten()
+    order = mx.argsort(indices)
+    inv_order = mx.argsort(order)
+    return x.flatten(0, -3)[order // M], indices[order], inv_order
+
+
+def scatter_unsort(x, inv_order, shape=None):
+    x = x[inv_order]
+    if shape is not None:
+        x = mx.unflatten(x, 0, shape)
+    return x
+
+
+def gather_mm_simulate(x, w, indices):
+    x, idx, inv_order = gather_sort(x, indices)
+    for i in range(2):
+        y = mx.concatenate(
+            [
+                mx.quantized_matmul(x[i], w[0][j], w[1][j], w[2][j], transpose=True)
+                for i, j in enumerate(idx.tolist())
+            ],
+            axis=0,
+        )
+        x = y[:, None]
+    x = scatter_unsort(x, inv_order, indices.shape)
+    return x
+
+
+def time_gather_qmm():
+    x = mx.random.normal((N, 1, 1, D)) / 1024**0.5
+    w1 = mx.random.normal((E, M, D)) / 1024**0.5
+    w2 = mx.random.normal((E, D, M)) / 1024**0.5
+    w1 = mx.quantize(w1)
+    w2 = mx.quantize(w2)
+    indices = (mx.random.uniform(shape=(N, I)) * E).astype(mx.uint32)
+    sorted_indices = mx.sort(indices.flatten()).reshape(N, I)
+    mx.eval(x, w1, w2, indices, sorted_indices)
+
+    def gather_mm(x, w1, w2, indices, sort):
+        idx = indices
+        inv_order = None
+        if sort:
+            x, idx, inv_order = gather_sort(x, indices)
+        x = mx.gather_qmm(x, *w1, transpose=True, rhs_indices=idx, sorted_indices=sort)
+        x = mx.gather_qmm(x, *w2, transpose=True, rhs_indices=idx, sorted_indices=sort)
+        if sort:
+            x = scatter_unsort(x, inv_order, indices.shape)
+        return x
+
+    time_fn(gather_mm, x, w1, w2, indices, False)
+    time_fn(gather_mm, x, w1, w2, sorted_indices, False)
+    time_fn(gather_mm, x, w1, w2, indices, True)
+
+    x = mx.random.normal((N * I, D)) / 1024**0.5
+    w1 = mx.random.normal((M, D)) / 1024**0.5
+    w2 = mx.random.normal((D, M)) / 1024**0.5
+    w1 = mx.quantize(w1)
+    w2 = mx.quantize(w2)
+    mx.eval(x, w1, w2)
+
+    def equivalent_matmul(x, w1, w2):
+        x = mx.quantized_matmul(x, *w1, transpose=True)
+        x = mx.quantized_matmul(x, *w2, transpose=True)
+        return x
+
+    time_fn(equivalent_matmul, x, w1, w2)
+
+
+if __name__ == "__main__":
+    time_gather_qmm()

benchmarks/python/layer_norm_bench.py

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

benchmarks/python/sdpa_bench.py

@@ -28,11 +28,34 @@ def bench(f, *args):
     return (e - s) * 1e-9
 
 
-def mlx_sdpa_fused_inner(q, k, v, scale):
-    return mx.fast.scaled_dot_product_attention(q, k, v, scale=scale, mask=None)
+def prepare_inputs(B, qL, kL, D, qH, kH, mask, transpose, dtype):
+    np_dtype = getattr(np, dtype)
+
+    shape_q = (B, qL, qH, D) if transpose else (B, qH, qL, D)
+    shape_kv = (B, kL, kH, D) if transpose else (B, kH, kL, D)
+
+    scale = 1.0 / math.sqrt(D)
+
+    q_np = np.random.normal(0.0, 1.0, shape_q).astype(np_dtype)
+    k_np = np.random.normal(0.0, scale, shape_kv).astype(np_dtype)
+    v_np = np.random.normal(0.0, scale, shape_kv).astype(np_dtype)
+
+    q_mx = mx.array(q_np)
+    k_mx = mx.array(k_np)
+    v_mx = mx.array(v_np)
+
+    if mask is not None:
+        if mask == "additive":
+            mask_np = np.random.normal(0.0, 1.0, (B, qH, qL, kL)).astype(np_dtype)
+            mask = mx.array(mask_np)
+        elif mask == "bool":
+            mask_np = np.random.uniform(0.0, 1.0, (B, qH, qL, kL)) < 0.5
+            mask = mx.array(mask_np)
+
+    return q_mx, k_mx, v_mx, scale, mask
 
 
-def mlx_sdpa_unfused_inner(q, k, v, scale, f32softmax=False):
+def mlx_ref_attn(q, k, v, scale=1.0, mask=None):
     q_dtype = q.dtype
     q = q * mx.array(scale, q_dtype)
     n_q_heads = q.shape[-3]
@@ -41,6 +64,7 @@ def mlx_sdpa_unfused_inner(q, k, v, scale, f32softmax=False):
 
     B = q.shape[0]
     L = q.shape[2]
+    kL = k.shape[2]
 
     if n_repeats > 1:
         q = mx.reshape(q, [B, n_kv_heads, n_repeats, L, -1])
@@ -48,10 +72,27 @@ def mlx_sdpa_unfused_inner(q, k, v, scale, f32softmax=False):
         v = mx.expand_dims(v, 2)
 
     scores = q @ mx.swapaxes(k, -1, -2)
-    if f32softmax:
-        scores = mx.softmax(scores.astype(mx.float32), axis=-1).astype(q_dtype)
-    else:
-        scores = mx.softmax(scores, axis=-1)
+
+    if mask is not None:
+
+        if mask == "causal":
+            q_offset = max(0, kL - L)
+            q_indices = mx.arange(q_offset, q_offset + L)
+            k_indices = mx.arange(kL)
+            mask = q_indices[:, None] >= k_indices[None]
+
+        if n_repeats > 1 and mask.ndim >= 3:
+            if mask.shape[-3] == 1:
+                mask = mx.expand_dims(mask, -3)
+            else:
+                mask = mx.unflatten(mask, -3, (n_kv_heads, n_repeats))
+
+        if mask.dtype == mx.bool_:
+            scores = mx.where(mask, scores, -np.float32(np.inf))
+        else:
+            scores += mask
+
+    scores = mx.softmax(scores, axis=-1, precise=True)
 
     out = scores @ v
     if n_repeats > 1:
@@ -60,74 +101,55 @@ def mlx_sdpa_unfused_inner(q, k, v, scale, f32softmax=False):
     return out
 
 
-def mlx_spda_unfused(q, k, v, scale, transpose):
-    q_out = q
+def mlx_fused_attn(q, k, v, scale, mask):
+    return mx.fast.scaled_dot_product_attention(q, k, v, scale=scale, mask=mask)
+
+
+def do_attention(f, q, k, v, scale, mask=None, transpose=False):
     if transpose:
-        k = mx.transpose(k, (0, 2, 1, 3))
-        v = mx.transpose(v, (0, 2, 1, 3))
+        q_t = mx.transpose(q, (0, 2, 1, 3))
+        k_t = mx.transpose(k, (0, 2, 1, 3))
+        v_t = mx.transpose(v, (0, 2, 1, 3))
+        o_t = f(q_t, k_t, v_t, scale=scale, mask=mask)
+        return mx.transpose(o_t, (0, 2, 1, 3))
+    else:
+        return f(q, k, v, scale=scale, mask=mask)
+
+
+def do_attention_bench(f, q, k, v, scale, mask=None, transpose=False):
+    q_out = q
 
     for i in range(N_iter_func):
-        if transpose:
-            q_out = mx.transpose(q_out, (0, 2, 1, 3))
-        q_out = mlx_sdpa_unfused_inner(q_out, k, v, scale)
-        if transpose:
-            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+        q_out = do_attention(f, q_out, k, v, scale, mask=mask, transpose=transpose)
 
     mx.eval(q_out)
     return q_out
 
 
-def mlx_spda_fused(q, k, v, scale, transpose):
-    q_out = q
-    if transpose:
-        k = mx.transpose(k, (0, 2, 1, 3))
-        v = mx.transpose(v, (0, 2, 1, 3))
-
-    for i in range(N_iter_func):
-        if transpose:
-            q_out = mx.transpose(q_out, (0, 2, 1, 3))
-        q_out = mlx_sdpa_fused_inner(q_out, k, v, scale)
-        if transpose:
-            q_out = mx.transpose(q_out, (0, 2, 1, 3))
-
-    mx.eval(q_out)
-    return q_out
-
-
-def bench_shape(B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, np_dtype, transpose=True):
-    shape_q = (
-        (B, qsl, n_q_heads, head_dim) if transpose else (B, n_q_heads, qsl, head_dim)
-    )
-    shape_kv = (
-        (B, ksl, n_kv_heads, head_dim) if transpose else (B, n_kv_heads, ksl, head_dim)
+def bench_shape(
+    B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, dtype, transpose=True, mask_in=None
+):
+    q_mx, k_mx, v_mx, scale, mask = prepare_inputs(
+        B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, mask_in, transpose, dtype
     )
 
-    q_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_q).astype(np_dtype)
-    k_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_kv).astype(np_dtype)
-    v_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_kv).astype(np_dtype)
+    time_mlx_unfused = bench(
+        do_attention_bench, mlx_ref_attn, q_mx, k_mx, v_mx, scale, mask, transpose
+    )
+    time_mlx_fused = bench(
+        do_attention_bench, mlx_fused_attn, q_mx, k_mx, v_mx, scale, mask, transpose
+    )
 
-    scale = math.sqrt(1.0 / head_dim)
+    o_mlx_fused = do_attention(mlx_ref_attn, q_mx, k_mx, v_mx, scale, mask, transpose)
+    o_mlx_unfused = do_attention(
+        mlx_fused_attn, q_mx, k_mx, v_mx, scale, mask, transpose
+    )
 
-    q_mx = mx.array(q_np)
-    k_mx = mx.array(k_np)
-    v_mx = mx.array(v_np)
+    atol = 1e-5 if dtype == "float32" else 2e-4
 
-    time_mlx_unfused = bench(mlx_spda_unfused, q_mx, k_mx, v_mx, scale, transpose)
-    time_mlx_fused = bench(mlx_spda_fused, q_mx, k_mx, v_mx, scale, transpose)
-
-    if transpose:
-        q_mx = mx.transpose(q_mx, (0, 2, 1, 3))
-        k_mx = mx.transpose(k_mx, (0, 2, 1, 3))
-        v_mx = mx.transpose(v_mx, (0, 2, 1, 3))
-
-    o_mlx_fused = mlx_sdpa_fused_inner(q_mx, k_mx, v_mx, scale)
-    o_mlx_unfused = mlx_sdpa_unfused_inner(q_mx, k_mx, v_mx, scale, f32softmax=True)
-
-    atol = 1e-5 if np_dtype == np.float32 else 1e-4
-
-    if not mx.allclose(o_mlx_fused, o_mlx_unfused, atol=atol):
+    if not mx.allclose(o_mlx_fused, o_mlx_unfused, atol=atol, rtol=atol):
         print(
-            f"Failed at (B: {B}, qsl: {qsl}, ksl: {ksl}, head_dim: {head_dim}, n_qh: {n_q_heads}, n_kvh: {n_kv_heads}) [tpose = {transpose}] with max(|a - b|) = {mx.max(mx.abs(o_mlx_unfused - o_mlx_fused)):3.2e}"
+            f"Failed at (B: {B}, qsl: {qsl}, ksl: {ksl}, head_dim: {head_dim}, n_qh: {n_q_heads}, n_kvh: {n_kv_heads}, mask: {mask_in}) [tpose = {transpose}] with max(|a - b|) = {mx.max(mx.abs(o_mlx_unfused - o_mlx_fused)):3.2e}"
         )
 
     return time_mlx_fused, time_mlx_unfused
@@ -151,39 +173,51 @@ if __name__ == "__main__":
           (  1,   128,   128,       64,   32,    32),
           (  1,   256,   256,       64,   32,    32),
           (  1,   512,   512,       64,   32,    32),
-          (  1,  1024,  1024,       64,   32,    32),
-          (  1,  2048,  2048,       64,   32,    32),
-          (  1,  4096,  4096,       64,   32,    32),
+          (  1,  1024,  1024,       64,   32,     8),
+          (  1,  2048,  2048,       64,   32,     8),
+          (  1,  4096,  4096,       64,   32,     8),
     )
 
     shapes_80 = (
         # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
-          (  1,  1024,  1024,       80,   32,    32),
-          (  1,  2048,  2048,       80,   32,    32),
-          (  1,  4096,  4096,       80,   32,    32),
+          (  1,  1024,  1024,       80,   32,     8),
+          (  1,  2048,  2048,       80,   32,     8),
+          (  1,  4096,  4096,       80,   32,     8),
     )
 
     shapes_128 = (
         # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
-          (  1,  1024,  1024,      128,   32,    32),
-          (  1,  2048,  2048,      128,   32,    32),
-          (  1,  4096,  4096,      128,   32,    32),
+          (  1,  1024,  1024,      128,   32,     8),
+          (  1,  2048,  2048,      128,   32,     8),
+          (  1,  4096,  4096,      128,   32,     8),
     )
     # fmt: on
 
     shapes = shapes_64 + shapes_80 + shapes_128
 
-    print("  B,   qsl,   ksl, hdim, n_qh, n_kvh, tpose,   dtype, t_unfs, t_fuse, diff%")
+    masks = [None, "bool", "causal"]
+
+    print(
+        "  B,   qsl,   ksl, hdim, n_qh, n_kvh, t,   dtype,     mask, t_unfs, t_fuse, diff%"
+    )
 
     for dtype in dtypes:
         for transpose in transposes:
             for B, qsl, ksl, head_dim, n_q_heads, n_kv_heads in shapes:
-                np_dtype = getattr(np, dtype)
-                time_mlx_fused, time_mlx_unfused = bench_shape(
-                    B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, np_dtype, transpose
-                )
-                diff = time_mlx_unfused / time_mlx_fused - 1.0
-                t_str = 1 if transpose else 0
-                print(
-                    f"{B:3d}, {qsl:5d}, {ksl:5d}, {head_dim:4d}, {n_q_heads:4d}, {n_kv_heads:5d}, {t_str:5d}, {dtype}, {time_mlx_unfused: 2.3f}, {time_mlx_fused: 2.3f}, {100. * diff:+5.2f}%"
-                )
+                for mask_in in masks:
+                    time_mlx_fused, time_mlx_unfused = bench_shape(
+                        B,
+                        qsl,
+                        ksl,
+                        head_dim,
+                        n_q_heads,
+                        n_kv_heads,
+                        dtype,
+                        transpose,
+                        mask_in,
+                    )
+                    diff = time_mlx_unfused / time_mlx_fused - 1.0
+                    t_str = 1 if transpose else 0
+                    print(
+                        f"{B:3d}, {qsl:5d}, {ksl:5d}, {head_dim:4d}, {n_q_heads:4d}, {n_kv_heads:5d}, {t_str:1d}, {dtype}, {str(mask_in):>8}, {time_mlx_unfused: 2.3f}, {time_mlx_fused: 2.3f}, {100. * diff:+5.2f}%"
+                    )

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/extension.cmake

@@ -11,13 +11,14 @@ include(CMakeParseArguments)
 # Args: TARGET: Custom target to be added for the metal library TITLE: Name of
 # the .metallib OUTPUT_DIRECTORY: Where to place ${TITLE}.metallib SOURCES: List
 # of source files INCLUDE_DIRS: List of include dirs DEPS: List of dependency
-# files (like headers)
+# files (like headers) DEBUG: Boolean, if true, enables debug compile options
+# for this specific library. If not provided, uses global MLX_METAL_DEBUG.
 #
 # clang format on
 
 macro(mlx_build_metallib)
   # Parse args
-  set(oneValueArgs TARGET TITLE OUTPUT_DIRECTORY)
+  set(oneValueArgs TARGET TITLE OUTPUT_DIRECTORY DEBUG)
   set(multiValueArgs SOURCES INCLUDE_DIRS DEPS)
   cmake_parse_arguments(MTLLIB "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
@@ -26,6 +27,10 @@ macro(mlx_build_metallib)
 
   # Collect compile options
   set(MTLLIB_COMPILE_OPTIONS -Wall -Wextra -fno-fast-math -Wno-c++17-extensions)
+  if(MLX_METAL_DEBUG OR MTLLIB_DEBUG)
+    set(MTLLIB_COMPILE_OPTIONS ${MTLLIB_COMPILE_OPTIONS} -gline-tables-only
+                               -frecord-sources)
+  endif()
 
   # Prepare metallib build command
   add_custom_command(

docs/Doxyfile

@@ -13,7 +13,7 @@ EXCLUDE_PATTERNS       = */private/*
 CREATE_SUBDIRS         = NO
 FULL_PATH_NAMES        = YES
 RECURSIVE              = YES
-GENERATE_HTML          = YES
+GENERATE_HTML          = NO
 GENERATE_LATEX         = NO
 GENERATE_XML           = YES
 XML_PROGRAMLISTING     = YES

docs/src/conf.py

@@ -10,7 +10,7 @@ import mlx.core as mx
 # -- Project information -----------------------------------------------------
 
 project = "MLX"
-copyright = "2023, MLX Contributors"
+copyright = "2023, Apple"
 author = "MLX Contributors"
 version = ".".join(mx.__version__.split(".")[:3])
 release = version

docs/src/dev/custom_metal_kernels.rst

@@ -8,23 +8,26 @@ MLX supports writing custom Metal kernels through the Python and C++ APIs.
 Simple Example
 --------------
 
+.. currentmodule:: mlx.core
+
 Let's write a custom kernel that computes ``exp`` elementwise:
 
 .. code-block:: python
 
-  def exp_elementwise(a: mx.array):
-      source = """
-          uint elem = thread_position_in_grid.x;
-          T tmp = inp[elem];
-          out[elem] = metal::exp(tmp);
-      """
+  source = """
+      uint elem = thread_position_in_grid.x;
+      T tmp = inp[elem];
+      out[elem] = metal::exp(tmp);
+  """
 
-      kernel = mx.fast.metal_kernel(
-          name="myexp",
-          input_names=["inp"],
-          output_names=["out"],
-          source=source,
-      )
+  kernel = mx.fast.metal_kernel(
+      name="myexp",
+      input_names=["inp"],
+      output_names=["out"],
+      source=source,
+  )
+
+  def exp_elementwise(a: mx.array):
       outputs = kernel(
           inputs=[a],
           template=[("T", mx.float32)],
@@ -39,8 +42,13 @@ Let's write a custom kernel that computes ``exp`` elementwise:
   b = exp_elementwise(a)
   assert mx.allclose(b, mx.exp(a))
 
+Every time you make a kernel, a new Metal library is created and possibly
+JIT compiled. To reduce the overhead from that, build the kernel once with
+:func:`fast.metal_kernel` and then use it many times.
+
 .. note::
-    We are only required to pass the body of the Metal kernel in ``source``.
+   Only pass the body of the Metal kernel in ``source``. The function
+   signature is generated automatically.
 
 The full function signature will be generated using:
 
@@ -78,44 +86,51 @@ Putting this all together, the generated function signature for ``myexp`` is as
 
   template [[host_name("custom_kernel_myexp_float")]] [[kernel]] decltype(custom_kernel_myexp_float<float>) custom_kernel_myexp_float<float>;
 
-Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads <https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_ function.
-This means we will launch ``mx.prod(grid)`` threads, subdivided into ``threadgroup`` size threadgroups.
-For optimal performance, each thread group dimension should be less than or equal to the corresponding grid dimension.
+Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads
+<https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_
+function. This means we will launch ``mx.prod(grid)`` threads, subdivided into
+``threadgroup`` size threadgroups.  For optimal performance, each thread group
+dimension should be less than or equal to the corresponding grid dimension.
 
-Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
+Passing ``verbose=True`` to :func:`ast.metal_kernel.__call__` will print the
+generated code for debugging purposes.
 
 Using Shape/Strides
 -------------------
 
-``mx.fast.metal_kernel`` supports an argument ``ensure_row_contiguous`` which is ``True`` by default.
-This will copy the ``mx.array`` inputs if needed before the kernel is launched to ensure that the memory layout is row contiguous.
-Generally this makes writing the kernel easier, since we don't have to worry about gaps or the ordering of the dims
-when indexing.
+:func:`fast.metal_kernel` supports an argument ``ensure_row_contiguous`` which
+is ``True`` by default. This will copy the array inputs if needed
+before the kernel is launched to ensure that the memory layout is row
+contiguous.  Generally this makes writing the kernel easier, since we don't
+have to worry about gaps or the ordering of the dims when indexing.
 
-If we want to avoid this copy, ``metal_kernel`` automatically passes ``a_shape``, ``a_strides`` and ``a_ndim`` for each
-input array ``a`` if any are present in ``source``.
-We can then use MLX's built in indexing utils to fetch the right elements for each thread.
+If we want to avoid this copy, :func:`fast.metal_kernel` automatically passes
+``a_shape``, ``a_strides`` and ``a_ndim`` for each input array ``a`` if any are
+present in ``source``. We can then use MLX's built in indexing utils to fetch
+the right elements for each thread.
 
-Let's convert ``myexp`` above to support arbitrarily strided arrays without relying on a copy from ``ensure_row_contiguous``:
+Let's convert ``myexp`` above to support arbitrarily strided arrays without
+relying on a copy from ``ensure_row_contiguous``:
 
 .. code-block:: python
+   
+  source = """
+      uint elem = thread_position_in_grid.x;
+      // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
+      uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
+      T tmp = inp[loc];
+      // Output arrays are always row contiguous
+      out[elem] = metal::exp(tmp);
+  """
+
+  kernel = mx.fast.metal_kernel(
+      name="myexp_strided",
+      input_names=["inp"],
+      output_names=["out"],
+      source=source
+  )
 
   def exp_elementwise(a: mx.array):
-      source = """
-          uint elem = thread_position_in_grid.x;
-          // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
-          uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
-          T tmp = inp[loc];
-          // Output arrays are always row contiguous
-          out[elem] = metal::exp(tmp);
-      """
-
-      kernel = mx.fast.metal_kernel(
-          name="myexp_strided",
-          input_names=["inp"],
-          output_names=["out"],
-          source=source
-      )
       outputs = kernel(
           inputs=[a],
           template=[("T", mx.float32)],
@@ -142,137 +157,139 @@ We'll start with the following MLX implementation using standard ops:
 
 .. code-block:: python
 
-    def grid_sample_ref(x, grid):
-        N, H_in, W_in, _ = x.shape
-        ix = ((grid[..., 0] + 1) * W_in - 1) / 2
-        iy = ((grid[..., 1] + 1) * H_in - 1) / 2
+  def grid_sample_ref(x, grid):
+      N, H_in, W_in, _ = x.shape
+      ix = ((grid[..., 0] + 1) * W_in - 1) / 2
+      iy = ((grid[..., 1] + 1) * H_in - 1) / 2
 
-        ix_nw = mx.floor(ix).astype(mx.int32)
-        iy_nw = mx.floor(iy).astype(mx.int32)
+      ix_nw = mx.floor(ix).astype(mx.int32)
+      iy_nw = mx.floor(iy).astype(mx.int32)
 
-        ix_ne = ix_nw + 1
-        iy_ne = iy_nw
+      ix_ne = ix_nw + 1
+      iy_ne = iy_nw
 
-        ix_sw = ix_nw
-        iy_sw = iy_nw + 1
+      ix_sw = ix_nw
+      iy_sw = iy_nw + 1
 
-        ix_se = ix_nw + 1
-        iy_se = iy_nw + 1
+      ix_se = ix_nw + 1
+      iy_se = iy_nw + 1
 
-        nw = (ix_se - ix)    * (iy_se - iy)
-        ne = (ix    - ix_sw) * (iy_sw - iy)
-        sw = (ix_ne - ix)    * (iy    - iy_ne)
-        se = (ix    - ix_nw) * (iy    - iy_nw)
+      nw = (ix_se - ix)    * (iy_se - iy)
+      ne = (ix    - ix_sw) * (iy_sw - iy)
+      sw = (ix_ne - ix)    * (iy    - iy_ne)
+      se = (ix    - ix_nw) * (iy    - iy_nw)
 
-        I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
-        I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
-        I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
-        I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
+      I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
+      I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
+      I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
+      I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
 
-        mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
-        mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
-        mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
-        mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
+      mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
+      mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
+      mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
+      mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
 
-        I_nw *= mask_nw[..., None]
-        I_ne *= mask_ne[..., None]
-        I_sw *= mask_sw[..., None]
-        I_se *= mask_se[..., None]
+      I_nw *= mask_nw[..., None]
+      I_ne *= mask_ne[..., None]
+      I_sw *= mask_sw[..., None]
+      I_se *= mask_se[..., None]
 
-        output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
+      output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
 
-        return output
+      return output
 
-Now let's use ``mx.custom_function`` together with ``mx.fast.metal_kernel``
+Now let's use :func:`custom_function` together with :func:`fast.metal_kernel`
 to write a fast GPU kernel for both the forward and backward passes.
 
 First we'll implement the forward pass as a fused kernel:
 
 .. code-block:: python
 
-    @mx.custom_function
-    def grid_sample(x, grid):
+  source = """
+      uint elem = thread_position_in_grid.x;
+      int H = x_shape[1];
+      int W = x_shape[2];
+      int C = x_shape[3];
+      int gH = grid_shape[1];
+      int gW = grid_shape[2];
 
-        assert x.ndim == 4, "`x` must be 4D."
-        assert grid.ndim == 4, "`grid` must be 4D."
+      int w_stride = C;
+      int h_stride = W * w_stride;
+      int b_stride = H * h_stride;
 
-        B, _, _, C = x.shape
-        _, gN, gM, D = grid.shape
-        out_shape = (B, gN, gM, C)
+      uint grid_idx = elem / C * 2;
+      float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+      float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
 
-        assert D == 2, "Last dim of `grid` must be size 2."
+      int ix_nw = floor(ix);
+      int iy_nw = floor(iy);
 
-        source = """
-            uint elem = thread_position_in_grid.x;
-            int H = x_shape[1];
-            int W = x_shape[2];
-            int C = x_shape[3];
-            int gH = grid_shape[1];
-            int gW = grid_shape[2];
+      int ix_ne = ix_nw + 1;
+      int iy_ne = iy_nw;
 
-            int w_stride = C;
-            int h_stride = W * w_stride;
-            int b_stride = H * h_stride;
+      int ix_sw = ix_nw;
+      int iy_sw = iy_nw + 1;
 
-            uint grid_idx = elem / C * 2;
-            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
-            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+      int ix_se = ix_nw + 1;
+      int iy_se = iy_nw + 1;
 
-            int ix_nw = floor(ix);
-            int iy_nw = floor(iy);
+      T nw = (ix_se - ix)    * (iy_se - iy);
+      T ne = (ix    - ix_sw) * (iy_sw - iy);
+      T sw = (ix_ne - ix)    * (iy    - iy_ne);
+      T se = (ix    - ix_nw) * (iy    - iy_nw);
 
-            int ix_ne = ix_nw + 1;
-            int iy_ne = iy_nw;
+      int batch_idx = elem / C / gH / gW * b_stride;
+      int channel_idx = elem % C;
+      int base_idx = batch_idx + channel_idx;
 
-            int ix_sw = ix_nw;
-            int iy_sw = iy_nw + 1;
+      T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
+      T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
+      T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
+      T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
 
-            int ix_se = ix_nw + 1;
-            int iy_se = iy_nw + 1;
+      I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
+      I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
+      I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
+      I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
 
-            T nw = (ix_se - ix)    * (iy_se - iy);
-            T ne = (ix    - ix_sw) * (iy_sw - iy);
-            T sw = (ix_ne - ix)    * (iy    - iy_ne);
-            T se = (ix    - ix_nw) * (iy    - iy_nw);
+      out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
+  """
 
-            int batch_idx = elem / C / gH / gW * b_stride;
-            int channel_idx = elem % C;
-            int base_idx = batch_idx + channel_idx;
+  kernel = mx.fast.metal_kernel(
+      name="grid_sample",
+      input_names=["x", "grid"],
+      output_names=["out"],
+      source=source,
+  )
 
-            T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
-            T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
-            T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
-            T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
+  @mx.custom_function
+  def grid_sample(x, grid):
 
-            I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
-            I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
-            I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
-            I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
+      assert x.ndim == 4, "`x` must be 4D."
+      assert grid.ndim == 4, "`grid` must be 4D."
 
-            out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
-        """
-        kernel = mx.fast.metal_kernel(
-            name="grid_sample",
-            input_names=["x", "grid"],
-            output_names=["out"],
-            source=source,
-        )
-        outputs = kernel(
-            inputs=[x, grid],
-            template=[("T", x.dtype)],
-            output_shapes=[out_shape],
-            output_dtypes=[x.dtype],
-            grid=(np.prod(out_shape), 1, 1),
-            threadgroup=(256, 1, 1),
-        )
-        return outputs[0]
+      B, _, _, C = x.shape
+      _, gN, gM, D = grid.shape
+      out_shape = (B, gN, gM, C)
+
+      assert D == 2, "Last dim of `grid` must be size 2."
+
+      outputs = kernel(
+          inputs=[x, grid],
+          template=[("T", x.dtype)],
+          output_shapes=[out_shape],
+          output_dtypes=[x.dtype],
+          grid=(np.prod(out_shape), 1, 1),
+          threadgroup=(256, 1, 1),
+      )
+      return outputs[0]
 
 For a reasonably sized input such as:
 
 .. code-block:: python
 
-    x.shape = (8, 1024, 1024, 64)
-    grid.shape = (8, 256, 256, 2)
+  x.shape = (8, 1024, 1024, 64)
+  grid.shape = (8, 256, 256, 2)
 
 On an M1 Max, we see a big performance improvement:
 
@@ -281,11 +298,11 @@ On an M1 Max, we see a big performance improvement:
 Grid Sample VJP
 ---------------
 
-Since we decorated ``grid_sample`` with ``mx.custom_function``, we can now define
-its custom vjp transform so MLX can differentiate it.
+Since we decorated ``grid_sample`` with :func:`custom_function`, we can now
+define its custom vjp transform so MLX can differentiate it.
 
 The backwards pass requires atomically updating ``x_grad``/``grid_grad`` and so
-requires a few extra ``mx.fast.metal_kernel`` features:
+requires a few extra :func:`fast.metal_kernel` features:
 
 * ``init_value=0``
     Initialize all of the kernel's outputs to this value before it runs. This allows us to update only part of the output arrays with the kernel.
@@ -299,128 +316,129 @@ We can then implement the backwards pass as follows:
 
 .. code-block:: python
 
-    @grid_sample.vjp
-    def grid_sample_vjp(primals, cotangent, _):
-        x, grid = primals
-        B, _, _, C = x.shape
-        _, gN, gM, D = grid.shape
+  source = """
+      uint elem = thread_position_in_grid.x;
+      int H = x_shape[1];
+      int W = x_shape[2];
+      int C = x_shape[3];
+      // Pad C to the nearest larger simdgroup size multiple
+      int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
 
-        assert D == 2, "Last dim of `grid` must be size 2."
+      int gH = grid_shape[1];
+      int gW = grid_shape[2];
 
-        source = """
-            uint elem = thread_position_in_grid.x;
-            int H = x_shape[1];
-            int W = x_shape[2];
-            int C = x_shape[3];
-            // Pad C to the nearest larger simdgroup size multiple
-            int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
+      int w_stride = C;
+      int h_stride = W * w_stride;
+      int b_stride = H * h_stride;
 
-            int gH = grid_shape[1];
-            int gW = grid_shape[2];
+      uint grid_idx = elem / C_padded * 2;
+      float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+      float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
 
-            int w_stride = C;
-            int h_stride = W * w_stride;
-            int b_stride = H * h_stride;
+      int ix_nw = floor(ix);
+      int iy_nw = floor(iy);
 
-            uint grid_idx = elem / C_padded * 2;
-            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
-            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+      int ix_ne = ix_nw + 1;
+      int iy_ne = iy_nw;
 
-            int ix_nw = floor(ix);
-            int iy_nw = floor(iy);
+      int ix_sw = ix_nw;
+      int iy_sw = iy_nw + 1;
 
-            int ix_ne = ix_nw + 1;
-            int iy_ne = iy_nw;
+      int ix_se = ix_nw + 1;
+      int iy_se = iy_nw + 1;
 
-            int ix_sw = ix_nw;
-            int iy_sw = iy_nw + 1;
+      T nw = (ix_se - ix)    * (iy_se - iy);
+      T ne = (ix    - ix_sw) * (iy_sw - iy);
+      T sw = (ix_ne - ix)    * (iy    - iy_ne);
+      T se = (ix    - ix_nw) * (iy    - iy_nw);
 
-            int ix_se = ix_nw + 1;
-            int iy_se = iy_nw + 1;
+      int batch_idx = elem / C_padded / gH / gW * b_stride;
+      int channel_idx = elem % C_padded;
+      int base_idx = batch_idx + channel_idx;
 
-            T nw = (ix_se - ix)    * (iy_se - iy);
-            T ne = (ix    - ix_sw) * (iy_sw - iy);
-            T sw = (ix_ne - ix)    * (iy    - iy_ne);
-            T se = (ix    - ix_nw) * (iy    - iy_nw);
+      T gix = T(0);
+      T giy = T(0);
+      if (channel_idx < C) {
+          int cot_index = elem / C_padded * C + channel_idx;
+          T cot = cotangent[cot_index];
+          if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
+              int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
 
-            int batch_idx = elem / C_padded / gH / gW * b_stride;
-            int channel_idx = elem % C_padded;
-            int base_idx = batch_idx + channel_idx;
+              T I_nw = x[offset];
+              gix -= I_nw * (iy_se - iy) * cot;
+              giy -= I_nw * (ix_se - ix) * cot;
+          }
+          if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
+              int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
 
-            T gix = T(0);
-            T giy = T(0);
-            if (channel_idx < C) {
-                int cot_index = elem / C_padded * C + channel_idx;
-                T cot = cotangent[cot_index];
-                if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
-                    int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
+              T I_ne = x[offset];
+              gix += I_ne * (iy_sw - iy) * cot;
+              giy -= I_ne * (ix - ix_sw) * cot;
+          }
+          if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
+              int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
 
-                    T I_nw = x[offset];
-                    gix -= I_nw * (iy_se - iy) * cot;
-                    giy -= I_nw * (ix_se - ix) * cot;
-                }
-                if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
-                    int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
+              T I_sw = x[offset];
+              gix -= I_sw * (iy - iy_ne) * cot;
+              giy += I_sw * (ix_ne - ix) * cot;
+          }
+          if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
+              int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
 
-                    T I_ne = x[offset];
-                    gix += I_ne * (iy_sw - iy) * cot;
-                    giy -= I_ne * (ix - ix_sw) * cot;
-                }
-                if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
-                    int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
+              T I_se = x[offset];
+              gix += I_se * (iy - iy_nw) * cot;
+              giy += I_se * (ix - ix_nw) * cot;
+          }
+      }
 
-                    T I_sw = x[offset];
-                    gix -= I_sw * (iy - iy_ne) * cot;
-                    giy += I_sw * (ix_ne - ix) * cot;
-                }
-                if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
-                    int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
+      T gix_mult = W / 2;
+      T giy_mult = H / 2;
 
-                    T I_se = x[offset];
-                    gix += I_se * (iy - iy_nw) * cot;
-                    giy += I_se * (ix - ix_nw) * cot;
-                }
-            }
+      // Reduce across each simdgroup first.
+      // This is much faster than relying purely on atomics.
+      gix = simd_sum(gix);
+      giy = simd_sum(giy);
 
-            T gix_mult = W / 2;
-            T giy_mult = H / 2;
+      if (thread_index_in_simdgroup == 0) {
+          atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
+          atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
+      }
+  """
+  kernel = mx.fast.metal_kernel(
+      name="grid_sample_grad",
+      input_names=["x", "grid", "cotangent"],
+      output_names=["x_grad", "grid_grad"],
+      source=source,
+      atomic_outputs=True,
+  )
 
-            // Reduce across each simdgroup first.
-            // This is much faster than relying purely on atomics.
-            gix = simd_sum(gix);
-            giy = simd_sum(giy);
+  @grid_sample.vjp
+  def grid_sample_vjp(primals, cotangent, _):
+      x, grid = primals
+      B, _, _, C = x.shape
+      _, gN, gM, D = grid.shape
 
-            if (thread_index_in_simdgroup == 0) {
-                atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
-                atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
-            }
-        """
-        kernel = mx.fast.metal_kernel(
-            name="grid_sample_grad",
-            input_names=["x", "grid", "cotangent"],
-            output_names=["x_grad", "grid_grad"],
-            source=source,
-            atomic_outputs=True,
-        )
-        # pad the output channels to simd group size
-        # so that our `simd_sum`s don't overlap.
-        simdgroup_size = 32
-        C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
-        grid_size = B * gN * gM * C_padded
-        outputs = kernel(
-            inputs=[x, grid, cotangent],
-            template=[("T", x.dtype)],
-            output_shapes=[x.shape, grid.shape],
-            output_dtypes=[x.dtype, x.dtype],
-            grid=(grid_size, 1, 1),
-            threadgroup=(256, 1, 1),
-            init_value=0,
-        )
-        return outputs[0], outputs[1]
+      assert D == 2, "Last dim of `grid` must be size 2."
+
+      # pad the output channels to simd group size
+      # so that our `simd_sum`s don't overlap.
+      simdgroup_size = 32
+      C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
+      grid_size = B * gN * gM * C_padded
+      outputs = kernel(
+          inputs=[x, grid, cotangent],
+          template=[("T", x.dtype)],
+          output_shapes=[x.shape, grid.shape],
+          output_dtypes=[x.dtype, x.dtype],
+          grid=(grid_size, 1, 1),
+          threadgroup=(256, 1, 1),
+          init_value=0,
+      )
+      return outputs[0], outputs[1]
 
 There's an even larger speed up for the vjp:
 

docs/src/dev/extensions.rst

@@ -93,9 +93,9 @@ Primitives
 ^^^^^^^^^^^
 
 A :class:`Primitive` is part of the computation graph of an :class:`array`. It
-defines how to create outputs arrays given a input arrays. Further, a
+defines how to create output arrays given input arrays. Further, a
 :class:`Primitive` has methods to run on the CPU or GPU and for function
-transformations such as ``vjp`` and ``jvp``.  Lets go back to our example to be
+transformations such as ``vjp`` and ``jvp``.  Let's go back to our example to be
 more concrete:
 
 .. code-block:: C++
@@ -128,7 +128,7 @@ more concrete:
         /** The vector-Jacobian product. */
         std::vector<array> vjp(
             const std::vector<array>& primals,
-            const array& cotan,
+            const std::vector<array>& cotangents,
             const std::vector<int>& argnums,
             const std::vector<array>& outputs) override;
 
@@ -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 **/
@@ -247,9 +247,7 @@ point-wise. This is captured in the templated function :meth:`axpby_impl`.
       float alpha_,
       float beta_,
       mx::Stream stream) {
-    // Allocate the output with `malloc_or_wait` which synchronously allocates
-    // memory, potentially waiting if the system is under memory pressure
-    out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
+    out.set_data(mx::allocator::malloc(out.nbytes()));
 
     // Get the CPU command encoder and register input and output arrays
     auto& encoder = mx::cpu::get_command_encoder(stream);
@@ -393,17 +391,17 @@ below.
         auto& d = metal::device(s.device);
 
         // Allocate output memory
-        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+        out.set_data(allocator::malloc(out.nbytes()));
 
         // Resolve name of kernel
-        std::ostringstream kname;
-        kname << "axpby_" << "general_" << type_to_name(out);
+        std::stream kname;
+        kname = "axpby_general_" + type_to_name(out);
 
-        // Make sure the metal library is available
-        d.register_library("mlx_ext");
+        // Load the metal library
+        auto lib = d.get_library("mlx_ext", current_binary_dir());
 
         // Make a kernel from this metal library
-        auto kernel = d.get_kernel(kname.str(), "mlx_ext");
+        auto kernel = d.get_kernel(kname, lib);
 
         // Prepare to encode kernel
         auto& compute_encoder = d.get_command_encoder(s.index);
@@ -471,7 +469,7 @@ one we just defined:
             const std::vector<array>& tangents,
             const std::vector<int>& argnums) {
         // Forward mode diff that pushes along the tangents
-        // The jvp transform on the primitive can built with ops
+        // The jvp transform on the primitive can be built with ops
         // that are scheduled on the same stream as the primitive
 
         // If argnums = {0}, we only push along x in which case the
@@ -483,7 +481,7 @@ one we just defined:
             auto scale_arr = array(scale, tangents[0].dtype());
             return {multiply(scale_arr, tangents[0], stream())};
         }
-        // If, argnums = {0, 1}, we take contributions from both
+        // If argnums = {0, 1}, we take contributions from both
         // which gives us jvp = tangent_x * alpha + tangent_y * beta
         else {
             return {axpby(tangents[0], tangents[1], alpha_, beta_, stream())};
@@ -737,7 +735,7 @@ Let's look at a simple script and its results:
 
     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 is correct: {mx.all(c == 6.0).item()}")
 
 Output:
 
@@ -745,7 +743,7 @@ Output:
 
     c shape: [3, 4]
     c dtype: float32
-    c correctness: True
+    c is correct: True
 
 Results
 ^^^^^^^

docs/src/index.rst

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

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 -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/array.rst

@@ -19,6 +19,8 @@ Array
     array.ndim
     array.shape
     array.size
+    array.real
+    array.imag
     array.abs
     array.all
     array.any
@@ -38,6 +40,7 @@ Array
     array.log10
     array.log1p
     array.log2
+    array.logcumsumexp
     array.logsumexp
     array.max
     array.mean

docs/src/python/fft.rst

@@ -20,3 +20,5 @@ FFT
   irfft2
   rfftn
   irfftn
+  fftshift
+  ifftshift

docs/src/python/linalg.rst

@@ -16,9 +16,12 @@ Linear Algebra
     cross
     qr
     svd
+    eigvals
+    eig
     eigvalsh
     eigh
     lu
     lu_factor
+    pinv
     solve
     solve_triangular

docs/src/python/memory_management.rst

@@ -0,0 +1,16 @@
+Memory Management
+=================
+
+.. currentmodule:: mlx.core
+
+.. autosummary::
+  :toctree: _autosummary
+
+  get_active_memory
+  get_peak_memory
+  reset_peak_memory
+  get_cache_memory
+  set_memory_limit
+  set_cache_limit
+  set_wired_limit
+  clear_cache

docs/src/python/metal.rst

@@ -8,13 +8,5 @@ Metal
 
   is_available
   device_info
-  get_active_memory
-  get_peak_memory
-  reset_peak_memory
-  get_cache_memory
-  set_memory_limit
-  set_cache_limit
-  set_wired_limit
-  clear_cache
   start_capture
   stop_capture

docs/src/python/ops.rst

@@ -36,10 +36,12 @@ Operations
    bitwise_or
    bitwise_xor
    block_masked_mm
+   broadcast_arrays
    broadcast_to
    ceil
    clip
    concatenate
+   contiguous
    conj
    conjugate
    convolve
@@ -101,6 +103,7 @@ Operations
    log10
    log1p
    logaddexp
+   logcumsumexp
    logical_not
    logical_and
    logical_or

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

@@ -18,3 +18,5 @@ Common Optimizers
    AdamW
    Adamax
    Lion
+   MultiOptimizer
+   Muon

docs/src/python/transforms.rst

@@ -9,6 +9,7 @@ Transforms
   :toctree: _autosummary
 
    eval
+   async_eval
    compile
    custom_function
    disable_compile

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

@@ -107,6 +107,16 @@ same array:
   >>> a
   array([1, 2, 0], dtype=int32)
 
+
+Note, unlike NumPy, updates to the same location are nondeterministic:
+
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> a[[0, 0]] = mx.array([4, 5])
+
+The first element of ``a`` could be ``4`` or ``5``.
+
 Transformations of functions which use in-place updates are allowed and work as
 expected. For example:
 

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
 ///////////////////////////////////////////////////////////////////////////////
@@ -72,9 +86,7 @@ void axpby_impl(
     float alpha_,
     float beta_,
     mx::Stream stream) {
-  // Allocate the output with `malloc_or_wait` which synchronously allocates
-  // memory, potentially waiting if the system is under memory pressure
-  out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(mx::allocator::malloc(out.nbytes()));
 
   // Get the CPU command encoder and register input and output arrays
   auto& encoder = mx::cpu::get_command_encoder(stream);
@@ -160,25 +172,24 @@ void Axpby::eval_gpu(
   // Allocate output memory with strides based on specialization
   if (contiguous_kernel) {
     out.set_data(
-        mx::allocator::malloc_or_wait(x.data_size() * out.itemsize()),
+        mx::allocator::malloc(x.data_size() * out.itemsize()),
         x.data_size(),
         x.strides(),
         x.flags());
   } else {
-    out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
+    out.set_data(mx::allocator::malloc(out.nbytes()));
   }
 
   // Resolve name of kernel (corresponds to axpby.metal)
-  std::ostringstream kname;
-  kname << "axpby_";
-  kname << (contiguous_kernel ? "contiguous_" : "general_");
-  kname << type_to_name(out);
+  std::string kname = "axpby_";
+  kname += (contiguous_kernel ? "contiguous_" : "general_");
+  kname += type_to_name(out);
 
-  // Make sure the metal library is available
-  d.register_library("mlx_ext");
+  // Load the metal library
+  auto lib = d.get_library("mlx_ext", current_binary_dir());
 
   // Make a kernel from this metal library
-  auto kernel = d.get_kernel(kname.str(), "mlx_ext");
+  auto kernel = d.get_kernel(kname, lib);
 
   // Prepare to encode kernel
   auto& compute_encoder = d.get_command_encoder(s.index);

examples/extensions/axpby/axpby.h

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

examples/extensions/requirements.txt

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

examples/extensions/test.py

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

mlx/CMakeLists.txt

@@ -5,6 +5,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/compile.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/dtype.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/dtype_utils.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/export.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/einsum.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/fast.cpp
@@ -20,7 +21,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/metal.h)
 
 # Define MLX_VERSION only in the version.cpp file.
-add_library(mlx_version STATIC ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp)
+add_library(mlx_version OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp)
 target_compile_definitions(mlx_version PRIVATE MLX_VERSION="${MLX_VERSION}")
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:mlx_version>)
 
@@ -48,5 +49,19 @@ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
 if(MLX_BUILD_METAL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/metal)
 else()
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_metal)
+  target_sources(mlx
+                 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/no_metal.cpp)
+endif()
+
+if(MLX_BUILD_CUDA)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda)
+else()
+  target_sources(mlx
+                 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda/no_cuda.cpp)
+endif()
+
+if(MLX_BUILD_METAL OR MLX_BUILD_CUDA)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/gpu)
+else()
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_gpu)
 endif()

mlx/allocator.cpp

@@ -4,12 +4,11 @@
 #include <sstream>
 
 #include "mlx/allocator.h"
-#include "mlx/scheduler.h"
 
 namespace mlx::core::allocator {
 
 Buffer malloc(size_t size) {
-  auto buffer = allocator().malloc(size, /* allow_swap */ true);
+  auto buffer = allocator().malloc(size);
   if (size && !buffer.ptr()) {
     std::ostringstream msg;
     msg << "[malloc] Unable to allocate " << size << " bytes.";
@@ -22,45 +21,4 @@ void free(Buffer buffer) {
   allocator().free(buffer);
 }
 
-Buffer CommonAllocator::malloc(size_t size, bool) {
-  void* ptr = std::malloc(size + sizeof(size_t));
-  if (ptr != nullptr) {
-    *static_cast<size_t*>(ptr) = size;
-  }
-  return Buffer{ptr};
-}
-
-void CommonAllocator::free(Buffer buffer) {
-  std::free(buffer.ptr());
-}
-
-size_t CommonAllocator::size(Buffer buffer) const {
-  if (buffer.ptr() == nullptr) {
-    return 0;
-  }
-  return *static_cast<size_t*>(buffer.ptr());
-}
-
-Buffer malloc_or_wait(size_t size) {
-  auto buffer = allocator().malloc(size);
-
-  while (size && !buffer.ptr() && scheduler::n_active_tasks() > 0) {
-    scheduler::wait_for_one();
-    buffer = allocator().malloc(size);
-  }
-
-  // Try swapping if needed
-  if (size && !buffer.ptr()) {
-    buffer = allocator().malloc(size, /* allow_swap = */ true);
-  }
-
-  if (size && !buffer.ptr()) {
-    std::ostringstream msg;
-    msg << "[malloc_or_wait] Unable to allocate " << size << " bytes.";
-    throw std::runtime_error(msg.str());
-  }
-
-  return buffer;
-}
-
 } // namespace mlx::core::allocator

mlx/allocator.h

@@ -32,14 +32,10 @@ Buffer malloc(size_t size);
 
 void free(Buffer buffer);
 
-// Wait for running tasks to finish and free up memory
-// if allocation fails
-Buffer malloc_or_wait(size_t size);
-
 class Allocator {
   /** Abstract base class for a memory allocator. */
  public:
-  virtual Buffer malloc(size_t size, bool allow_swap = false) = 0;
+  virtual Buffer malloc(size_t size) = 0;
   virtual void free(Buffer buffer) = 0;
   virtual size_t size(Buffer buffer) const = 0;
 
@@ -53,16 +49,4 @@ class Allocator {
 
 Allocator& allocator();
 
-class CommonAllocator : public Allocator {
-  /** A general CPU allocator. */
- public:
-  virtual Buffer malloc(size_t size, bool allow_swap = false) override;
-  virtual void free(Buffer buffer) override;
-  virtual size_t size(Buffer buffer) const override;
-
- private:
-  CommonAllocator() = default;
-  friend Allocator& allocator();
-};
-
 } // namespace mlx::core::allocator

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
@@ -224,6 +225,10 @@ class array {
     // Not copyable
     Data(const Data& d) = delete;
     Data& operator=(const Data& d) = delete;
+    Data(Data&& o) : buffer(o.buffer), d(o.d) {
+      o.buffer = allocator::Buffer(nullptr);
+      o.d = [](allocator::Buffer) {};
+    }
     ~Data() {
       d(buffer);
     }
@@ -339,11 +344,11 @@ class array {
     return allocator::allocator().size(buffer());
   }
 
-  // Return a copy of the shared pointer
-  // to the array::Data struct
-  std::shared_ptr<Data> data_shared_ptr() const {
+  // Return the shared pointer to the array::Data struct
+  const std::shared_ptr<Data>& data_shared_ptr() const {
     return array_desc_->data;
   }
+
   // Return a raw pointer to the arrays data
   template <typename T>
   T* data() {
@@ -356,7 +361,7 @@ class array {
   }
 
   enum Status {
-    // The ouptut of a computation which has not been scheduled.
+    // The output of a computation which has not been scheduled.
     // For example, the status of `x` in `auto x = a + b`.
     unscheduled,
 

mlx/backend/common/CMakeLists.txt

@@ -1,6 +1,7 @@
 target_sources(
   mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/broadcasting.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/common.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp

mlx/backend/common/binary.h

@@ -44,14 +44,14 @@ inline void set_binary_op_output_data(
   switch (bopt) {
     case BinaryOpType::ScalarScalar:
       out.set_data(
-          allocator::malloc_or_wait(out.itemsize()), 1, a.strides(), a.flags());
+          allocator::malloc(out.itemsize()), 1, a.strides(), a.flags());
       break;
     case BinaryOpType::ScalarVector:
       if (b_donatable) {
         out.copy_shared_buffer(b);
       } else {
         out.set_data(
-            allocator::malloc_or_wait(b.data_size() * out.itemsize()),
+            allocator::malloc(b.data_size() * out.itemsize()),
             b.data_size(),
             b.strides(),
             b.flags());
@@ -62,7 +62,7 @@ inline void set_binary_op_output_data(
         out.copy_shared_buffer(a);
       } else {
         out.set_data(
-            allocator::malloc_or_wait(a.data_size() * out.itemsize()),
+            allocator::malloc(a.data_size() * out.itemsize()),
             a.data_size(),
             a.strides(),
             a.flags());
@@ -75,7 +75,7 @@ inline void set_binary_op_output_data(
         out.copy_shared_buffer(b);
       } else {
         out.set_data(
-            allocator::malloc_or_wait(a.data_size() * out.itemsize()),
+            allocator::malloc(a.data_size() * out.itemsize()),
             a.data_size(),
             a.strides(),
             a.flags());
@@ -88,7 +88,7 @@ inline void set_binary_op_output_data(
           b_donatable && b.flags().row_contiguous && b.size() == out.size()) {
         out.copy_shared_buffer(b);
       } else {
-        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+        out.set_data(allocator::malloc(out.nbytes()));
       }
       break;
   }

mlx/backend/common/broadcasting.cpp

@@ -0,0 +1,24 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+void broadcast(const array& in, array& out) {
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+  Strides strides(out.ndim(), 0);
+  int diff = out.ndim() - in.ndim();
+  for (int i = in.ndim() - 1; i >= 0; --i) {
+    strides[i + diff] = (in.shape()[i] == 1) ? 0 : in.strides()[i];
+  }
+  auto flags = in.flags();
+  if (out.size() > in.size()) {
+    flags.row_contiguous = flags.col_contiguous = false;
+  }
+  out.copy_shared_buffer(in, strides, flags, in.data_size());
+}
+
+} // namespace mlx::core

mlx/backend/common/broadcasting.h

@@ -0,0 +1,11 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+
+namespace mlx::core {
+
+void broadcast(const array& in, array& out);
+
+} // namespace mlx::core

mlx/backend/common/buffer_cache.h

@@ -0,0 +1,157 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <cassert>
+#include <functional>
+#include <map>
+
+namespace mlx::core {
+
+template <typename T>
+class BufferCache {
+ public:
+  BufferCache(
+      size_t page_size,
+      std::function<size_t(T*)> get_size,
+      std::function<void(T*)> free)
+      : page_size_(page_size),
+        get_size_(std::move(get_size)),
+        free_(std::move(free)) {}
+
+  ~BufferCache() {
+    clear();
+  }
+
+  BufferCache(const BufferCache&) = delete;
+  BufferCache& operator=(const BufferCache&) = delete;
+
+  T* reuse_from_cache(size_t size) {
+    // Find the closest buffer in pool.
+    auto it = buffer_pool_.lower_bound(size);
+    if (it == buffer_pool_.end() ||
+        it->first >= std::min(2 * size, size + 2 * page_size_)) {
+      return nullptr;
+    }
+
+    // Collect from the cache.
+    T* buf = it->second->buf;
+    pool_size_ -= it->first;
+
+    // Remove from record.
+    remove_from_list(it->second);
+    buffer_pool_.erase(it);
+    return buf;
+  }
+
+  void recycle_to_cache(T* buf) {
+    assert(buf);
+    // Add to cache.
+    BufferHolder* bh = new BufferHolder(buf);
+    add_at_head(bh);
+    size_t size = get_size_(buf);
+    pool_size_ += size;
+    buffer_pool_.emplace(size, bh);
+  }
+
+  int release_cached_buffers(size_t min_bytes_to_free) {
+    if (min_bytes_to_free >= 0.9 * pool_size_) {
+      return clear();
+    } else {
+      int n_release = 0;
+      size_t total_bytes_freed = 0;
+
+      while (tail_ && (total_bytes_freed < min_bytes_to_free)) {
+        // Release buffer.
+        size_t size = get_size_(tail_->buf);
+        total_bytes_freed += size;
+        free_(tail_->buf);
+        n_release++;
+
+        // Remove from record.
+        auto its = buffer_pool_.equal_range(size);
+        auto it = std::find_if(its.first, its.second, [this](const auto& el) {
+          return el.second == tail_;
+        });
+        assert(it != buffer_pool_.end());
+        buffer_pool_.erase(it);
+        remove_from_list(tail_);
+      }
+
+      pool_size_ -= total_bytes_freed;
+      return n_release;
+    }
+  }
+
+  int clear() {
+    int n_release = 0;
+    for (auto& [size, holder] : buffer_pool_) {
+      free_(holder->buf);
+      n_release++;
+      delete holder;
+    }
+    buffer_pool_.clear();
+    pool_size_ = 0;
+    head_ = nullptr;
+    tail_ = nullptr;
+    return n_release;
+  }
+
+  size_t cache_size() const {
+    return pool_size_;
+  }
+
+  size_t page_size() const {
+    return page_size_;
+  }
+
+ private:
+  struct BufferHolder {
+   public:
+    explicit BufferHolder(T* buf_) : buf(buf_) {}
+
+    BufferHolder* prev{nullptr};
+    BufferHolder* next{nullptr};
+    T* buf;
+  };
+
+  void add_at_head(BufferHolder* to_add) {
+    if (!head_) {
+      head_ = to_add;
+      tail_ = to_add;
+    } else {
+      head_->prev = to_add;
+      to_add->next = head_;
+      head_ = to_add;
+    }
+  }
+
+  void remove_from_list(BufferHolder* to_remove) {
+    if (to_remove->prev && to_remove->next) { // if middle
+      to_remove->prev->next = to_remove->next;
+      to_remove->next->prev = to_remove->prev;
+    } else if (to_remove->prev && to_remove == tail_) { // if tail
+      tail_ = to_remove->prev;
+      tail_->next = nullptr;
+    } else if (to_remove == head_ && to_remove->next) { // if head
+      head_ = to_remove->next;
+      head_->prev = nullptr;
+    } else if (to_remove == head_ && to_remove == tail_) { // if only element
+      head_ = nullptr;
+      tail_ = nullptr;
+    }
+
+    delete to_remove;
+  }
+
+  std::multimap<size_t, BufferHolder*> buffer_pool_;
+  BufferHolder* head_{nullptr};
+  BufferHolder* tail_{nullptr};
+  size_t pool_size_{0};
+
+  const size_t page_size_;
+  std::function<size_t(T*)> get_size_;
+  std::function<void(T*)> free_;
+};
+
+} // namespace mlx::core

mlx/backend/common/common.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2024 Apple Inc.
 #include <cassert>
 
+#include "mlx/backend/common/broadcasting.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/primitives.h"
 
@@ -42,23 +43,6 @@ void AsStrided::eval(const std::vector<array>& inputs, array& out) {
   return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
 }
 
-void broadcast(const array& in, array& out) {
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-  Strides strides(out.ndim(), 0);
-  int diff = out.ndim() - in.ndim();
-  for (int i = in.ndim() - 1; i >= 0; --i) {
-    strides[i + diff] = (in.shape()[i] == 1) ? 0 : in.strides()[i];
-  }
-  auto flags = in.flags();
-  if (out.size() > in.size()) {
-    flags.row_contiguous = flags.col_contiguous = false;
-  }
-  out.copy_shared_buffer(in, strides, flags, in.data_size());
-}
-
 void Broadcast::eval(const std::vector<array>& inputs, array& out) {
   broadcast(inputs[0], out);
 }
@@ -103,7 +87,7 @@ void ExpandDims::eval(const std::vector<array>& inputs, array& out) {
 
 void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   double numel = 1;
   for (auto ax : axes_) {

mlx/backend/common/compiled.cpp

@@ -1,8 +1,7 @@
 // Copyright © 2023-2024 Apple Inc.
 
 #include "mlx/backend/common/compiled.h"
-#include "mlx/graph_utils.h"
-#include "mlx/primitives.h"
+#include "mlx/backend/common/utils.h"
 #include "mlx/utils.h"
 
 namespace mlx::core {
@@ -15,6 +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:
@@ -51,6 +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_:
@@ -79,55 +82,6 @@ std::string get_type_string(Dtype d) {
   }
 }
 
-std::string build_lib_name(
-    const std::vector<array>& inputs,
-    const std::vector<array>& outputs,
-    const std::vector<array>& tape,
-    const std::unordered_set<uintptr_t>& constant_ids) {
-  NodeNamer namer;
-  std::ostringstream os;
-  std::ostringstream constant_hasher;
-
-  // Fill the input names. This is not really necessary, I just like having A,
-  // B, C, ... as the inputs.
-  for (auto& x : inputs) {
-    namer.get_name(x);
-  }
-
-  // The primitives describing the tape. For unary and binary primitives this
-  // must be enough to describe the full computation.
-  for (auto& a : tape) {
-    // name and type of output
-    os << namer.get_name(a) << kindof(a.dtype()) << a.itemsize();
-    // computation performed
-    a.primitive().print(os);
-    // name of inputs to the function
-    for (auto& inp : a.inputs()) {
-      os << namer.get_name(inp);
-    }
-  }
-  os << "_";
-
-  for (auto& x : inputs) {
-    if (constant_ids.find(x.id()) != constant_ids.end()) {
-      os << "C";
-      print_constant(constant_hasher, x);
-    } else {
-      os << (is_scalar(x) ? "S" : "V");
-    }
-  }
-  os << "_";
-  for (auto& x : inputs) {
-    if (constant_ids.find(x.id()) != constant_ids.end()) {
-      continue;
-    }
-    os << kindof(x.dtype()) << x.itemsize();
-  }
-  os << "_" << std::hash<std::string>{}(constant_hasher.str());
-
-  return os.str();
-}
-
 bool compiled_check_contiguity(
     const std::vector<array>& inputs,
     const Shape& shape) {
@@ -159,8 +113,7 @@ bool compiled_check_contiguity(
 void compiled_allocate_outputs(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::vector<array>& inputs_,
-    const std::unordered_set<uintptr_t>& constant_ids_,
+    const std::function<bool(size_t)>& is_constant,
     bool contiguous) {
   if (contiguous) {
     int o = 0;
@@ -175,8 +128,7 @@ void compiled_allocate_outputs(
       // - Donatable
       // - Not a constant
       if (in.itemsize() == outputs[o].itemsize() && !is_scalar(in) &&
-          in.is_donatable() &&
-          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+          in.is_donatable() && is_constant(i)) {
         outputs[o++].copy_shared_buffer(in);
       }
       // Get representative input flags to properly set non-donated outputs
@@ -188,7 +140,7 @@ void compiled_allocate_outputs(
     }
     for (; o < outputs.size(); ++o) {
       outputs[o].set_data(
-          allocator::malloc_or_wait(data_size * outputs[o].itemsize()),
+          allocator::malloc(data_size * outputs[o].itemsize()),
           data_size,
           strides,
           flags);
@@ -204,16 +156,86 @@ void compiled_allocate_outputs(
       // - Not a constant
       if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
           in.itemsize() == outputs[o].itemsize() && in.is_donatable() &&
-          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+          is_constant(i)) {
         outputs[o].copy_shared_buffer(
             in, outputs[o].strides(), in.flags(), in.data_size());
         o++;
       }
     }
     for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(allocator::malloc_or_wait(outputs[o].nbytes()));
+      outputs[o].set_data(allocator::malloc(outputs[o].nbytes()));
     }
   }
 }
 
+std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(
+    const std::vector<array>& inputs,
+    const array& out,
+    const std::function<bool(size_t)>& is_constant) {
+  const Shape& shape = out.shape();
+  bool contiguous = compiled_check_contiguity(inputs, shape);
+  if (contiguous) {
+    return {true, shape, {}};
+  }
+
+  std::vector<Strides> strides_vec{out.strides()};
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    // Skip constants.
+    if (is_constant(i)) {
+      continue;
+    }
+
+    // Skip scalar inputs.
+    const auto& x = inputs[i];
+    if (is_scalar(x)) {
+      continue;
+    }
+
+    // Broadcast the inputs to the output shape.
+    Strides xstrides;
+    size_t j = 0;
+    for (; j < shape.size() - x.ndim(); ++j) {
+      if (shape[j] == 1) {
+        xstrides.push_back(out.strides()[j]);
+      } else {
+        xstrides.push_back(0);
+      }
+    }
+    for (size_t i = 0; i < x.ndim(); ++i, ++j) {
+      if (x.shape(i) == 1) {
+        if (shape[j] == 1) {
+          xstrides.push_back(out.strides()[j]);
+        } else {
+          xstrides.push_back(0);
+        }
+      } else {
+        xstrides.push_back(x.strides()[i]);
+      }
+    }
+    strides_vec.push_back(std::move(xstrides));
+  }
+
+  auto tup = collapse_contiguous_dims(shape, strides_vec, INT32_MAX);
+  return {false, std::move(std::get<0>(tup)), std::move(std::get<1>(tup))};
+}
+
+bool compiled_use_large_index(
+    const std::vector<array>& inputs,
+    const std::vector<array>& outputs,
+    bool contiguous) {
+  if (contiguous) {
+    size_t max_size = 0;
+    for (const auto& in : inputs) {
+      max_size = std::max(max_size, in.data_size());
+    }
+    return max_size > UINT32_MAX;
+  } else {
+    size_t max_size = 0;
+    for (const auto& o : outputs) {
+      max_size = std::max(max_size, o.size());
+    }
+    return max_size > UINT32_MAX;
+  }
+}
+
 } // namespace mlx::core

mlx/backend/common/compiled.h

@@ -1,9 +1,8 @@
 // Copyright © 2023-2024 Apple Inc.
 #pragma once
 
+#include <functional>
 #include <iomanip>
-#include <sstream>
-#include <unordered_set>
 
 #include "mlx/array.h"
 #include "mlx/primitives.h"
@@ -14,19 +13,17 @@ inline bool is_static_cast(const Primitive& p) {
   return (typeid(p) == typeid(Broadcast) || typeid(p) == typeid(AsType));
 }
 
-std::string build_lib_name(
-    const std::vector<array>& inputs,
-    const std::vector<array>& outputs,
-    const std::vector<array>& tape,
-    const std::unordered_set<uintptr_t>& constant_ids);
-
 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>
@@ -60,8 +57,19 @@ bool compiled_check_contiguity(
 void compiled_allocate_outputs(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::vector<array>& inputs_,
-    const std::unordered_set<uintptr_t>& constant_ids_,
+    const std::function<bool(size_t)>& is_constant,
+    bool contiguous);
+
+// Collapse contiguous dims ignoring scalars and constants.
+std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(
+    const std::vector<array>& inputs,
+    const array& out,
+    const std::function<bool(size_t)>& is_constant);
+
+// Return whether the kernel should use large index.
+bool compiled_use_large_index(
+    const std::vector<array>& inputs,
+    const std::vector<array>& outputs,
     bool contiguous);
 
 } // namespace mlx::core

mlx/backend/common/copy.h

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/array.h"
+#include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
@@ -26,19 +26,19 @@ inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
   if (ctype == CopyType::Vector) {
     // If the input is donateable, we are doing a vector copy and the types
     // have the same size, then the input buffer can hold the output.
-    if (in.is_donatable() && in.itemsize() == out.itemsize()) {
+    if (is_donatable(in, out)) {
       out.copy_shared_buffer(in);
       return true;
     } else {
       out.set_data(
-          allocator::malloc_or_wait(in.data_size() * out.itemsize()),
+          allocator::malloc(in.data_size() * out.itemsize()),
           in.data_size(),
           in.strides(),
           in.flags());
       return false;
     }
   } else {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    out.set_data(allocator::malloc(out.nbytes()));
     return false;
   }
 }

mlx/backend/common/hadamard.h

@@ -99,7 +99,11 @@ inline std::pair<int, int> decompose_hadamard(int n) {
           "[hadamard] Only supports n = m*2^k where m in (1, 12, 20, 28).");
     }
   }
+  if (n > (1 << 26)) {
+    throw std::invalid_argument(
+        "[hadamard] Only supports n = m*2^k where k <= 26");
+  }
   return {n, m};
 }
 
-} // namespace mlx::core
+} // namespace mlx::core

mlx/backend/common/load.cpp

@@ -28,7 +28,7 @@ void swap_endianness(uint8_t* data_bytes, size_t N) {
 namespace mlx::core {
 
 void Load::eval_cpu(const std::vector<array>& inputs, array& out) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
   auto read_task = [out_ptr = out.data<char>(),
                     size = out.size(),
                     itemsize = out.itemsize(),

mlx/backend/common/matmul.h

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

mlx/backend/common/reduce.cpp

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

mlx/backend/common/reduce.h

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

mlx/backend/common/ternary.h

@@ -48,12 +48,12 @@ inline void set_ternary_op_output_data(
   switch (topt) {
     case TernaryOpType::ScalarScalarScalar:
       out.set_data(
-          allocator::malloc_or_wait(out.itemsize()), 1, b.strides(), b.flags());
+          allocator::malloc(out.itemsize()), 1, b.strides(), b.flags());
       break;
     case TernaryOpType::VectorVectorVector:
       if (!(maybe_donate(a) || maybe_donate(b) || maybe_donate(c))) {
         out.set_data(
-            allocator::malloc_or_wait(out.itemsize() * b.data_size()),
+            allocator::malloc(out.itemsize() * b.data_size()),
             b.data_size(),
             b.strides(),
             b.flags());
@@ -64,7 +64,7 @@ inline void set_ternary_op_output_data(
       if (!((a.flags().row_contiguous && maybe_donate(a)) ||
             (b.flags().row_contiguous && maybe_donate(b)) ||
             (c.flags().row_contiguous && maybe_donate(c)))) {
-        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+        out.set_data(allocator::malloc(out.nbytes()));
       }
       break;
   }

mlx/backend/common/unary.h

@@ -0,0 +1,26 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+inline void set_unary_output_data(const array& in, array& out) {
+  if (in.flags().contiguous) {
+    if (is_donatable(in, out)) {
+      out.copy_shared_buffer(in);
+    } else {
+      out.set_data(
+          allocator::malloc(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+    }
+  } else {
+    out.set_data(allocator::malloc(out.nbytes()));
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/utils.cpp

@@ -1,9 +1,22 @@
 // Copyright © 2023-2024 Apple Inc.
 
+#include <dlfcn.h>
+
 #include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
+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(
     const Shape& shape,
     const std::vector<Strides>& strides,
@@ -101,4 +114,145 @@ std::pair<Shape, Strides> collapse_contiguous_dims(
   return collapse_contiguous_dims(a.shape(), a.strides(), size_cap);
 }
 
+Dims get_block_dims_common(int dim0, int dim1, int dim2, int pow2 /* = 10 */) {
+  int pows[3] = {0, 0, 0};
+  int sum = 0;
+  while (true) {
+    int presum = sum;
+    // Check all the pows
+    if (dim0 >= (1 << (pows[0] + 1))) {
+      pows[0]++;
+      sum++;
+    }
+    if (sum == 10) {
+      break;
+    }
+    if (dim1 >= (1 << (pows[1] + 1))) {
+      pows[1]++;
+      sum++;
+    }
+    if (sum == 10) {
+      break;
+    }
+    if (dim2 >= (1 << (pows[2] + 1))) {
+      pows[2]++;
+      sum++;
+    }
+    if (sum == presum || sum == pow2) {
+      break;
+    }
+  }
+  return std::make_tuple(1ul << pows[0], 1ul << pows[1], 1ul << pows[2]);
+}
+
+Dims get_2d_grid_dims_common(const Shape& shape, const Strides& strides) {
+  // Dims with strides of 0 are ignored as they
+  // correspond to broadcasted dimensions
+  size_t grid_x = 1;
+  size_t grid_y = 1;
+  for (int i = 0; i < shape.size(); ++i) {
+    if (strides[i] == 0) {
+      continue;
+    }
+    if (grid_x * shape[i] < UINT32_MAX) {
+      grid_x *= shape[i];
+    } else {
+      grid_y *= shape[i];
+    }
+  }
+  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);
+  }
+  return std::make_tuple(
+      static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
+}
+
+Dims get_2d_grid_dims_common(
+    const Shape& shape,
+    const Strides& strides,
+    size_t divisor) {
+  // Compute the 2d grid dimensions such that the total size of the grid is
+  // divided by divisor.
+  size_t grid_x = 1;
+  size_t grid_y = 1;
+  for (int i = 0; i < shape.size(); ++i) {
+    if (strides[i] == 0) {
+      continue;
+    }
+
+    // No need to add this shape we can just remove it from the divisor.
+    if (divisor % shape[i] == 0) {
+      divisor /= shape[i];
+      continue;
+    }
+
+    if (grid_x * shape[i] < UINT32_MAX) {
+      grid_x *= shape[i];
+    } else {
+      grid_y *= shape[i];
+    }
+
+    if (divisor > 1) {
+      if (grid_x % divisor == 0) {
+        grid_x /= divisor;
+        divisor = 1;
+      } else if (grid_y % divisor == 0) {
+        grid_y /= divisor;
+        divisor = 1;
+      }
+    }
+  }
+  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX) {
+    throw std::runtime_error("Unable to safely factor shape.");
+  }
+  if (grid_y > grid_x) {
+    std::swap(grid_x, grid_y);
+  }
+  if (divisor > 1) {
+    grid_x = ((grid_x + divisor - 1) / divisor) * divisor;
+  }
+  return std::make_tuple(
+      static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
+}
+
+std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
+  auto [bx, by, bz] = get_block_dims_common(dim0, dim1, dim2);
+  auto gx = (dim0 + bx - 1) / bx;
+  auto gy = (dim1 + by - 1) / by;
+  auto gz = (dim2 + bz - 1) / bz;
+
+  return std::make_pair(
+      std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
+}
+
+array swapaxes_in_eval(const array& x, int axis1, int axis2) {
+  int ndim = x.ndim();
+  if (axis1 < 0) {
+    axis1 += ndim;
+  }
+  if (axis2 < 0) {
+    axis2 += ndim;
+  }
+
+  auto shape = x.shape();
+  std::swap(shape[axis1], shape[axis2]);
+  auto strides = x.strides();
+  std::swap(strides[axis1], strides[axis2]);
+
+  auto [data_size, row_contiguous, col_contiguous] =
+      check_contiguity(shape, strides);
+  bool contiguous = data_size == x.data_size();
+
+  array out(std::move(shape), x.dtype(), nullptr, {});
+  out.copy_shared_buffer(
+      x,
+      std::move(strides),
+      {contiguous, row_contiguous, col_contiguous},
+      x.data_size());
+  return out;
+}
+
 } // namespace mlx::core

mlx/backend/common/utils.h

@@ -2,12 +2,17 @@
 
 #pragma once
 
+#include <filesystem>
+#include <tuple>
 #include <vector>
 
 #include "mlx/array.h"
 
 namespace mlx::core {
 
+// 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) {
   int64_t loc = 0;
@@ -70,6 +75,31 @@ std::pair<Shape, Strides> collapse_contiguous_dims(
     const array& a,
     int64_t size_cap = std::numeric_limits<int32_t>::max());
 
+// Compute the thread block dimensions which fit the given
+// input dimensions.
+// - The thread block dimensions will be powers of two
+// - The thread block size will be less than 2^pow2
+using Dims = std::tuple<uint32_t, uint32_t, uint32_t>;
+Dims get_block_dims_common(int dim0, int dim1, int dim2, int pow2 = 10);
+
+// Computes a 2D grid where each element is < UINT_MAX
+// Assumes:
+// - overall size (product of non-broadcasted dimensions) is < UINT_MAX^2
+// - shape and strides correspond to a contiguous (no holes) but
+//   possibly broadcasted array
+Dims get_2d_grid_dims_common(const Shape& shape, const Strides& strides);
+
+// Same as above but we do an implicit division with divisor.
+// Basically, equivalent to factorizing
+//    Prod(s \forall s in shape if strides[s] > 0) / divisor.
+Dims get_2d_grid_dims_common(
+    const Shape& shape,
+    const Strides& strides,
+    size_t divisor);
+
+// Get both the block and a grid of blocks that covers dim0, dim1 and dim2.
+std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2);
+
 struct ContiguousIterator {
   inline void step() {
     int dims = shape_.size();
@@ -165,4 +195,14 @@ void shared_buffer_reshape(
     const array& in,
     const Strides& out_strides,
     array& out);
+
+// Like the swapaxes op but safe to call in eval_gpu.
+array swapaxes_in_eval(const array& x, int axis1, int axis2);
+
+template <typename T>
+inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
+  vec.erase(std::next(vec.begin(), index));
+  return vec;
+}
+
 } // namespace mlx::core

mlx/backend/cpu/CMakeLists.txt

@@ -40,11 +40,13 @@ add_dependencies(mlx cpu_compiled_preamble)
 
 target_sources(
   mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/available.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eig.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/eigh.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/encoder.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
@@ -58,6 +60,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/logsumexp.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
@@ -73,8 +76,8 @@ target_sources(
 if(MLX_BUILD_ACCELERATE)
   target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/bnns.cpp)
 else()
-  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/no_fp16.cpp
-                             ${CMAKE_CURRENT_SOURCE_DIR}/gemms/no_bf16.cpp)
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/simd_fp16.cpp
+                             ${CMAKE_CURRENT_SOURCE_DIR}/gemms/simd_bf16.cpp)
 endif()
 
 if(IOS)

mlx/backend/cpu/arg_reduce.cpp

@@ -14,10 +14,8 @@ template <typename InT, typename OpT>
 void arg_reduce(const array& in, array& out, const OpT& op, int axis) {
   auto axis_size = in.shape()[axis];
   auto axis_stride = in.strides()[axis];
-  Strides strides = in.strides();
-  Shape shape = in.shape();
-  strides.erase(strides.begin() + axis);
-  shape.erase(shape.begin() + axis);
+  Strides strides = remove_index(in.strides(), axis);
+  Shape shape = remove_index(in.shape(), axis);
   auto in_ptr = in.data<InT>();
   auto out_ptr = out.data<uint32_t>();
 
@@ -68,7 +66,7 @@ void arg_reduce_dispatch(
 void ArgReduce::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(in);
   encoder.set_output_array(out);

mlx/backend/cpu/available.cpp

@@ -0,0 +1,11 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cpu/available.h"
+
+namespace mlx::core::cpu {
+
+bool is_available() {
+  return true;
+}
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/available.h

@@ -0,0 +1,9 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+namespace mlx::core::cpu {
+
+bool is_available();
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/binary.cpp

@@ -172,9 +172,12 @@ void binary_float(
       case bfloat16:
         binary_op<bfloat16_t, Op>(a, b, out, bopt);
         break;
+      case complex64:
+        binary_op<complex64_t, Op>(a, b, out, bopt);
+        break;
       default:
         throw std::runtime_error(
-            "[binary_float] Only supports non-complex floating point types.");
+            "[binary_float] Only supports floating point types.");
     }
   });
 }

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

@@ -40,7 +40,10 @@ struct CompilerCache {
   std::shared_mutex mtx;
 };
 
-static CompilerCache cache{};
+static CompilerCache& cache() {
+  static CompilerCache cache_;
+  return cache_;
+};
 
 // GPU compile is always available if the GPU is available and since we are in
 // this file CPU compile is also available.
@@ -56,14 +59,16 @@ void* compile(
     const std::string& kernel_name,
     const std::function<std::string(void)>& source_builder) {
   {
-    std::shared_lock lock(cache.mtx);
-    if (auto it = cache.kernels.find(kernel_name); it != cache.kernels.end()) {
+    std::shared_lock lock(cache().mtx);
+    if (auto it = cache().kernels.find(kernel_name);
+        it != cache().kernels.end()) {
       return it->second;
     }
   }
 
-  std::unique_lock lock(cache.mtx);
-  if (auto it = cache.kernels.find(kernel_name); it != cache.kernels.end()) {
+  std::unique_lock lock(cache().mtx);
+  if (auto it = cache().kernels.find(kernel_name);
+      it != cache().kernels.end()) {
     return it->second;
   }
   std::string source_code = source_builder();
@@ -120,10 +125,10 @@ void* compile(
   }
 
   // load library
-  cache.libs.emplace_back(shared_lib_path);
+  cache().libs.emplace_back(shared_lib_path);
 
   // Load function
-  void* fun = dlsym(cache.libs.back().lib, kernel_name.c_str());
+  void* fun = dlsym(cache().libs.back().lib, kernel_name.c_str());
   if (!fun) {
     std::ostringstream msg;
     msg << "[Compile::eval_cpu] Failed to load compiled function "
@@ -131,7 +136,7 @@ void* compile(
         << dlerror();
     throw std::runtime_error(msg.str());
   }
-  cache.kernels.insert({kernel_name, fun});
+  cache().kernels.insert({kernel_name, fun});
   return fun;
 }
 
@@ -141,18 +146,9 @@ inline void build_kernel(
     const std::vector<array>& inputs,
     const std::vector<array>& outputs,
     const std::vector<array>& tape,
-    const std::unordered_set<uintptr_t>& constant_ids,
+    const std::function<bool(size_t)>& is_constant,
     bool contiguous,
     int ndim) {
-  // All outputs should have the exact same shape and will be row contiguous
-  auto output_shape = outputs[0].shape();
-  auto output_strides = outputs[0].strides();
-
-  // Constants are scalars that are captured by value and cannot change
-  auto is_constant = [&constant_ids](const array& x) {
-    return constant_ids.find(x.id()) != constant_ids.end();
-  };
-
   NodeNamer namer;
 
 #ifdef _MSC_VER
@@ -165,14 +161,15 @@ inline void build_kernel(
 
   // Add the input arguments
   int cnt = 0;
-  for (auto& x : inputs) {
-    auto& xname = namer.get_name(x);
-
+  for (size_t i = 0; i < inputs.size(); ++i) {
     // Skip constants from the input list
-    if (is_constant(x)) {
+    if (is_constant(i)) {
       continue;
     }
 
+    const auto& x = inputs[i];
+    auto& xname = namer.get_name(x);
+
     auto tstr = get_type_string(x.dtype());
     os << "  " << tstr << "* " << xname << " = (" << tstr << "*)args[" << cnt++
        << "];" << std::endl;
@@ -206,10 +203,11 @@ inline void build_kernel(
   }
 
   // Read the inputs in tmps
-  for (auto& x : inputs) {
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    const auto& x = inputs[i];
     auto& xname = namer.get_name(x);
 
-    if (is_constant(x)) {
+    if (is_constant(i)) {
       os << "  " << get_type_string(x.dtype()) << " tmp_" << xname << " = ";
       print_constant(os, x);
       os << ";" << std::endl;
@@ -233,7 +231,7 @@ 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]) << ", ";
@@ -259,8 +257,9 @@ inline void build_kernel(
   } else {
     for (int d = ndim - 1; d >= 0; --d) {
       // Update pointers
-      for (auto& x : inputs) {
-        if (is_constant(x) || is_scalar(x)) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        if (is_constant(i) || is_scalar(x)) {
           continue;
         }
         auto& xname = namer.get_name(x);
@@ -282,65 +281,45 @@ inline void build_kernel(
 void Compiled::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
-  if (kernel_lib_.empty()) {
-    kernel_lib_ = build_lib_name(inputs_, outputs_, tape_, constant_ids_);
-  }
-
-  // Figure out which kernel we are using
-  auto& shape = outputs[0].shape();
-  auto contiguous = compiled_check_contiguity(inputs, shape);
   auto& encoder = cpu::get_command_encoder(stream());
 
-  // Handle all broadcasting and collect function input arguments
+  // Collapse contiguous dims to route to a faster kernel if possible. Also
+  // handle all broadcasting.
+  auto [contiguous, shape, strides] =
+      compiled_collapse_contiguous_dims(inputs, outputs[0], is_constant_);
+
+  // Force allocating shape/strides on heap so we can take their data() first
+  // and then std::move them.
+  // TODO: Refactor code to avoid heap allocation.
+  shape.grow();
+  for (auto& s : strides) {
+    s.grow();
+  }
+
+  // Collect function input arguments.
   std::vector<void*> args;
-  std::vector<std::vector<size_t>> strides;
-  for (int i = 0; i < inputs.size(); i++) {
-    // Skip constants.
-    if (constant_ids_.find(inputs_[i].id()) != constant_ids_.end()) {
+  int strides_index = 1;
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    if (is_constant_(i)) {
       continue;
     }
-    auto& x = inputs[i];
+    const auto& x = inputs[i];
     encoder.set_input_array(x);
     args.push_back((void*)x.data<void>());
-
-    if (contiguous || is_scalar(x)) {
-      continue;
+    if (!contiguous && !is_scalar(x)) {
+      args.push_back(strides[strides_index++].data());
     }
-
-    // Broadcast the input to the output shape.
-    std::vector<size_t> xstrides;
-    int j = 0;
-    for (; j < shape.size() - x.ndim(); j++) {
-      if (shape[j] == 1) {
-        xstrides.push_back(outputs[0].strides()[j]);
-      } else {
-        xstrides.push_back(0);
-      }
-    }
-    for (int i = 0; i < x.ndim(); i++, j++) {
-      if (x.shape(i) == 1) {
-        if (shape[j] == 1) {
-          xstrides.push_back(outputs[0].strides()[j]);
-        } else {
-          xstrides.push_back(0);
-        }
-      } else {
-        xstrides.push_back(x.strides()[i]);
-      }
-    }
-    strides.push_back(std::move(xstrides));
-    args.push_back(strides.back().data());
   }
 
   // Get the kernel name from the lib
   int ndim = shape.size();
   auto kernel_name = kernel_lib_ + (contiguous ? "_contiguous" : "_strided_");
   if (!contiguous) {
-    kernel_name += std::to_string(shape.size());
+    kernel_name += std::to_string(ndim);
   }
 
   // Get the function
-  auto fn_ptr = compile(kernel_name, [&]() {
+  auto fn_ptr = compile(kernel_name, [&, contiguous = contiguous]() {
     std::ostringstream kernel;
     kernel << get_kernel_preamble() << std::endl;
     kernel << "extern \"C\"  {" << std::endl;
@@ -350,7 +329,7 @@ void Compiled::eval_cpu(
         inputs_,
         outputs_,
         tape_,
-        constant_ids_,
+        is_constant_,
         contiguous,
         ndim);
     // Close extern "C"
@@ -358,26 +337,22 @@ void Compiled::eval_cpu(
     return kernel.str();
   });
 
-  compiled_allocate_outputs(
-      inputs, outputs, inputs_, constant_ids_, contiguous);
+  compiled_allocate_outputs(inputs, outputs, is_constant_, contiguous);
 
   for (auto& x : outputs) {
     args.push_back(x.data<void>());
     encoder.set_output_array(x);
   }
-  Shape out_shape;
   if (!contiguous) {
-    out_shape = outputs[0].shape();
-    args.push_back((void*)out_shape.data());
+    args.push_back((void*)shape.data());
   } else {
     args.push_back((void*)outputs[0].data_size());
   }
   auto fun = (void (*)(void**))fn_ptr;
-  encoder.dispatch(
-      [fun,
-       args = std::move(args),
-       strides = std::move(strides),
-       out_shape = std::move(out_shape)]() mutable { fun(args.data()); });
+  encoder.dispatch([fun,
+                    args = std::move(args),
+                    strides = std::move(strides),
+                    shape = std::move(shape)]() mutable { fun(args.data()); });
 }
 
 } // namespace mlx::core

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};
   }
 };
 
@@ -30,12 +28,11 @@ void AllReduce::eval_cpu(
       if (in.is_donatable()) {
         out.copy_shared_buffer(in);
       } else {
-        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+        out.set_data(allocator::malloc(out.nbytes()));
       }
       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;
     }
@@ -46,8 +43,15 @@ void AllReduce::eval_cpu(
     case Sum:
       distributed::detail::all_sum(group(), in, outputs[0], stream());
       break;
+    case Max:
+      distributed::detail::all_max(group(), in, outputs[0], stream());
+      break;
+    case Min:
+      distributed::detail::all_min(group(), in, outputs[0], stream());
+      break;
     default:
-      throw std::runtime_error("Only all reduce sum is supported for now");
+      throw std::runtime_error(
+          "Only all reduce sum, min and max are supported for now");
   }
 }
 
@@ -58,7 +62,7 @@ void AllGather::eval_cpu(
   assert(outputs.size() == 1);
 
   auto [in, copied] = ensure_row_contiguous(inputs[0], stream());
-  outputs[0].set_data(allocator::malloc_or_wait(outputs[0].nbytes()));
+  outputs[0].set_data(allocator::malloc(outputs[0].nbytes()));
   distributed::detail::all_gather(group(), in, outputs[0], stream());
   if (copied) {
     auto& enc = cpu::get_command_encoder(stream());
@@ -87,7 +91,7 @@ void Recv::eval_cpu(
   assert(inputs.size() == 0);
   assert(outputs.size() == 1);
 
-  outputs[0].set_data(allocator::malloc_or_wait(outputs[0].nbytes()));
+  outputs[0].set_data(allocator::malloc(outputs[0].nbytes()));
   distributed::detail::recv(group(), outputs[0], src_, stream());
 }
 

mlx/backend/cpu/eig.cpp

@@ -0,0 +1,174 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/allocator.h"
+#include "mlx/array.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/lapack.h"
+#include "mlx/linalg.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename T>
+void eig_impl(
+    array& a,
+    array& vectors,
+    array& values,
+    bool compute_eigenvectors,
+    Stream stream) {
+  using OT = std::complex<T>;
+  auto a_ptr = a.data<T>();
+  auto eig_ptr = values.data<OT>();
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(values);
+  OT* vec_ptr = nullptr;
+  if (compute_eigenvectors) {
+    encoder.set_output_array(vectors);
+    vec_ptr = vectors.data<OT>();
+  }
+  encoder.dispatch([a_ptr,
+                    vec_ptr,
+                    eig_ptr,
+                    compute_eigenvectors,
+                    N = vectors.shape(-1),
+                    size = vectors.size()]() mutable {
+    // Work query
+    char jobr = 'N';
+    char jobl = compute_eigenvectors ? 'V' : 'N';
+    int n_vecs_r = 1;
+    int n_vecs_l = compute_eigenvectors ? N : 1;
+    int lwork = -1;
+    int info;
+    {
+      T work;
+      int iwork;
+      geev<T>(
+          &jobl,
+          &jobr,
+          &N,
+          nullptr,
+          &N,
+          nullptr,
+          nullptr,
+          nullptr,
+          &n_vecs_l,
+          nullptr,
+          &n_vecs_r,
+          &work,
+          &lwork,
+          &info);
+      lwork = static_cast<int>(work);
+    }
+
+    auto eig_tmp_data = array::Data{allocator::malloc(sizeof(T) * N * 2)};
+    auto vec_tmp_data =
+        array::Data{allocator::malloc(vec_ptr ? sizeof(T) * N * N * 2 : 0)};
+    auto eig_tmp = static_cast<T*>(eig_tmp_data.buffer.raw_ptr());
+    auto vec_tmp = static_cast<T*>(vec_tmp_data.buffer.raw_ptr());
+    auto work_buf = array::Data{allocator::malloc(sizeof(T) * lwork)};
+    for (size_t i = 0; i < size / (N * N); ++i) {
+      geev<T>(
+          &jobl,
+          &jobr,
+          &N,
+          a_ptr,
+          &N,
+          eig_tmp,
+          eig_tmp + N,
+          vec_tmp,
+          &n_vecs_l,
+          nullptr,
+          &n_vecs_r,
+          static_cast<T*>(work_buf.buffer.raw_ptr()),
+          &lwork,
+          &info);
+      for (int i = 0; i < N; ++i) {
+        eig_ptr[i] = {eig_tmp[i], eig_tmp[N + i]};
+      }
+      if (vec_ptr) {
+        for (int i = 0; i < N; ++i) {
+          if (eig_ptr[i].imag() != 0) {
+            // This vector and the next are a pair
+            for (int j = 0; j < N; ++j) {
+              vec_ptr[i * N + j] = {
+                  vec_tmp[i * N + j], -vec_tmp[(i + 1) * N + j]};
+              vec_ptr[(i + 1) * N + j] = {
+                  vec_tmp[i * N + j], vec_tmp[(i + 1) * N + j]};
+            }
+            i += 1;
+          } else {
+            for (int j = 0; j < N; ++j) {
+              vec_ptr[i * N + j] = {vec_tmp[i * N + j], 0};
+            }
+          }
+        }
+        vec_ptr += N * N;
+      }
+      a_ptr += N * N;
+      eig_ptr += N;
+      if (info != 0) {
+        std::stringstream msg;
+        msg << "[Eig::eval_cpu] Eigenvalue decomposition failed with error code "
+            << info;
+        throw std::runtime_error(msg.str());
+      }
+    }
+  });
+  encoder.add_temporary(a);
+}
+
+} // namespace
+
+void Eig::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  const auto& a = inputs[0];
+  auto& values = outputs[0];
+
+  auto vectors = compute_eigenvectors_
+      ? outputs[1]
+      : array(a.shape(), complex64, nullptr, {});
+
+  auto a_copy = array(a.shape(), a.dtype(), nullptr, {});
+  copy_cpu(
+      a,
+      a_copy,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream());
+
+  values.set_data(allocator::malloc(values.nbytes()));
+
+  if (compute_eigenvectors_) {
+    // Set the strides and flags so the eigenvectors
+    // are in the columns of the output
+    auto flags = vectors.flags();
+    auto strides = vectors.strides();
+    auto ndim = a.ndim();
+    std::swap(strides[ndim - 1], strides[ndim - 2]);
+
+    if (a.size() > 1) {
+      flags.row_contiguous = false;
+      if (ndim > 2) {
+        flags.col_contiguous = false;
+      } else {
+        flags.col_contiguous = true;
+      }
+    }
+    vectors.set_data(
+        allocator::malloc(vectors.nbytes()), vectors.size(), strides, flags);
+  }
+  switch (a.dtype()) {
+    case float32:
+      eig_impl<float>(a_copy, vectors, values, compute_eigenvectors_, stream());
+      break;
+    default:
+      throw std::runtime_error("[Eig::eval_cpu] only supports float32.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/eigh.cpp

@@ -12,6 +12,133 @@ namespace mlx::core {
 
 namespace {
 
+template <typename T, class Enable = void>
+struct EighWork {};
+
+template <typename T>
+struct EighWork<
+    T,
+    typename std::enable_if<std::is_floating_point<T>::value>::type> {
+  using R = T;
+
+  char jobz;
+  char uplo;
+  int N;
+  int lwork;
+  int liwork;
+  int info;
+  std::vector<array::Data> buffers;
+
+  EighWork(char jobz_, char uplo_, int N_)
+      : jobz(jobz_), uplo(uplo_), N(N_), lwork(-1), liwork(-1) {
+    T work;
+    int iwork;
+    syevd<T>(
+        &jobz,
+        &uplo,
+        &N,
+        nullptr,
+        &N,
+        nullptr,
+        &work,
+        &lwork,
+        &iwork,
+        &liwork,
+        &info);
+    lwork = static_cast<int>(work);
+    liwork = iwork;
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+    buffers.emplace_back(allocator::malloc(sizeof(int) * liwork));
+  }
+
+  void run(T* vectors, T* values) {
+    syevd<T>(
+        &jobz,
+        &uplo,
+        &N,
+        vectors,
+        &N,
+        values,
+        static_cast<T*>(buffers[0].buffer.raw_ptr()),
+        &lwork,
+        static_cast<int*>(buffers[1].buffer.raw_ptr()),
+        &liwork,
+        &info);
+  }
+};
+
+template <>
+struct EighWork<std::complex<float>> {
+  using T = std::complex<float>;
+  using R = float;
+
+  char jobz;
+  char uplo;
+  int N;
+  int lwork;
+  int lrwork;
+  int liwork;
+  int info;
+  std::vector<array::Data> buffers;
+
+  EighWork(char jobz_, char uplo_, int N_)
+      : jobz(jobz_), uplo(uplo_), N(N_), lwork(-1), lrwork(-1), liwork(-1) {
+    T work;
+    R rwork;
+    int iwork;
+    heevd<T>(
+        &jobz,
+        &uplo,
+        &N,
+        nullptr,
+        &N,
+        nullptr,
+        &work,
+        &lwork,
+        &rwork,
+        &lrwork,
+        &iwork,
+        &liwork,
+        &info);
+    lwork = static_cast<int>(work.real());
+    lrwork = static_cast<int>(rwork);
+    liwork = iwork;
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+    buffers.emplace_back(allocator::malloc(sizeof(R) * lrwork));
+    buffers.emplace_back(allocator::malloc(sizeof(int) * liwork));
+  }
+
+  void run(T* vectors, R* values) {
+    heevd<T>(
+        &jobz,
+        &uplo,
+        &N,
+        vectors,
+        &N,
+        values,
+        static_cast<T*>(buffers[0].buffer.raw_ptr()),
+        &lwork,
+        static_cast<R*>(buffers[1].buffer.raw_ptr()),
+        &lrwork,
+        static_cast<int*>(buffers[2].buffer.raw_ptr()),
+        &liwork,
+        &info);
+    if (jobz == 'V') {
+      // We have pre-transposed the vectors but we also must conjugate them
+      // when they are complex.
+      //
+      // We could vectorize this but it is so fast in comparison to heevd that
+      // it doesn't really matter.
+      for (int i = 0; i < N; i++) {
+        for (int j = 0; j < N; j++) {
+          *vectors = std::conj(*vectors);
+          vectors++;
+        }
+      }
+    }
+  }
+};
+
 template <typename T>
 void eigh_impl(
     array& vectors,
@@ -19,8 +146,10 @@ void eigh_impl(
     const std::string& uplo,
     bool compute_eigenvectors,
     Stream stream) {
+  using R = typename EighWork<T>::R;
+
   auto vec_ptr = vectors.data<T>();
-  auto eig_ptr = values.data<T>();
+  auto eig_ptr = values.data<R>();
   char jobz = compute_eigenvectors ? 'V' : 'N';
 
   auto& encoder = cpu::get_command_encoder(stream);
@@ -33,50 +162,17 @@ void eigh_impl(
                     N = vectors.shape(-1),
                     size = vectors.size()]() mutable {
     // Work query
-    int lwork = -1;
-    int liwork = -1;
-    int info;
-    {
-      T work;
-      int iwork;
-      syevd<T>(
-          &jobz,
-          &uplo,
-          &N,
-          nullptr,
-          &N,
-          nullptr,
-          &work,
-          &lwork,
-          &iwork,
-          &liwork,
-          &info);
-      lwork = static_cast<int>(work);
-      liwork = iwork;
-    }
+    EighWork<T> work(jobz, uplo, N);
 
-    auto work_buf = array::Data{allocator::malloc_or_wait(sizeof(T) * lwork)};
-    auto iwork_buf =
-        array::Data{allocator::malloc_or_wait(sizeof(int) * liwork)};
+    // Work loop
     for (size_t i = 0; i < size / (N * N); ++i) {
-      syevd<T>(
-          &jobz,
-          &uplo,
-          &N,
-          vec_ptr,
-          &N,
-          eig_ptr,
-          static_cast<T*>(work_buf.buffer.raw_ptr()),
-          &lwork,
-          static_cast<int*>(iwork_buf.buffer.raw_ptr()),
-          &liwork,
-          &info);
+      work.run(vec_ptr, eig_ptr);
       vec_ptr += N * N;
       eig_ptr += N;
-      if (info != 0) {
+      if (work.info != 0) {
         std::stringstream msg;
         msg << "[Eigh::eval_cpu] Eigenvalue decomposition failed with error code "
-            << info;
+            << work.info;
         throw std::runtime_error(msg.str());
       }
     }
@@ -98,9 +194,9 @@ void Eigh::eval_cpu(
       ? outputs[1]
       : array(a.shape(), a.dtype(), nullptr, {});
 
-  values.set_data(allocator::malloc_or_wait(values.nbytes()));
+  values.set_data(allocator::malloc(values.nbytes()));
 
-  copy(
+  copy_cpu(
       a,
       vectors,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
@@ -132,6 +228,10 @@ void Eigh::eval_cpu(
       eigh_impl<double>(
           vectors, values, uplo_, compute_eigenvectors_, stream());
       break;
+    case complex64:
+      eigh_impl<std::complex<float>>(
+          vectors, values, uplo_, compute_eigenvectors_, stream());
+      break;
     default:
       throw std::runtime_error(
           "[Eigh::eval_cpu] only supports float32 or float64.");

mlx/backend/cpu/encoder.h

@@ -9,6 +9,9 @@
 
 namespace mlx::core::cpu {
 
+// Number of dispatches per scheduler task
+constexpr int DISPATCHES_PER_TASK = 10;
+
 struct CommandEncoder {
   CommandEncoder(Stream stream) : stream_(stream) {}
 
@@ -39,13 +42,24 @@ struct CommandEncoder {
 
   template <class F, class... Args>
   void dispatch(F&& f, Args&&... args) {
+    num_ops_ = (num_ops_ + 1) % DISPATCHES_PER_TASK;
     auto task = std::bind(std::forward<F>(f), std::forward<Args>(args)...);
-    scheduler::enqueue(stream_, std::move(task));
+    if (num_ops_ == 0) {
+      scheduler::notify_new_task(stream_);
+      auto task_wrap = [s = stream_, task = std::move(task)]() mutable {
+        task();
+        scheduler::notify_task_completion(s);
+      };
+      scheduler::enqueue(stream_, std::move(task_wrap));
+    } else {
+      scheduler::enqueue(stream_, std::move(task));
+    }
   }
 
  private:
   Stream stream_;
   std::vector<array> temporaries_;
+  int num_ops_{0};
 };
 
 CommandEncoder& get_command_encoder(Stream stream);

mlx/backend/cpu/eval.cpp

@@ -33,12 +33,8 @@ void eval(array& arr) {
     buffers.erase(it);
   }
   auto& encoder = cpu::get_command_encoder(s);
-  scheduler::notify_new_task(s);
-  encoder.dispatch([s,
-                    buffers = std::move(buffers),
-                    temps = std::move(encoder.temporaries())]() {
-    scheduler::notify_task_completion(s);
-  });
+  encoder.dispatch([buffers = std::move(buffers),
+                    temps = std::move(encoder.temporaries())]() {});
 }
 
 } // namespace mlx::core::cpu

mlx/backend/cpu/fft.cpp

@@ -22,7 +22,7 @@ void FFT::eval_cpu(const std::vector<array>& inputs, array& out) {
     s *= out.itemsize();
   }
 
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   std::vector<size_t> shape;
   if (out.dtype() == float32) {

mlx/backend/cpu/gemms/no_bf16.cpp

@@ -1,27 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cpu/gemm.h"
-
-namespace mlx::core {
-
-template <>
-void matmul<bfloat16_t>(
-    const bfloat16_t*,
-    const bfloat16_t*,
-    bfloat16_t*,
-    bool,
-    bool,
-    size_t,
-    size_t,
-    size_t,
-    float,
-    float,
-    size_t,
-    const Shape&,
-    const Strides&,
-    const Shape&,
-    const Strides&) {
-  throw std::runtime_error("[Matmul::eval_cpu] bfloat16 not supported.");
-}
-
-} // namespace mlx::core

mlx/backend/cpu/gemms/no_fp16.cpp

@@ -1,27 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cpu/gemm.h"
-
-namespace mlx::core {
-
-template <>
-void matmul<float16_t>(
-    const float16_t*,
-    const float16_t*,
-    float16_t*,
-    bool,
-    bool,
-    size_t,
-    size_t,
-    size_t,
-    float,
-    float,
-    size_t,
-    const Shape&,
-    const Strides&,
-    const Shape&,
-    const Strides&) {
-  throw std::runtime_error("[Matmul::eval_cpu] float16 not supported.");
-}
-
-} // namespace mlx::core

mlx/backend/cpu/gemms/simd_bf16.cpp

@@ -0,0 +1,45 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/gemm.h"
+#include "mlx/backend/cpu/gemms/simd_gemm.h"
+
+namespace mlx::core {
+
+template <>
+void matmul<bfloat16_t>(
+    const bfloat16_t* a,
+    const bfloat16_t* b,
+    bfloat16_t* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
+  for (int i = 0; i < batch_size; ++i) {
+    simd_gemm<bfloat16_t, float>(
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
+        out + M * N * i,
+        a_transposed,
+        b_transposed,
+        M,
+        N,
+        K,
+        alpha,
+        beta);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/gemms/simd_fp16.cpp

@@ -0,0 +1,45 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/gemm.h"
+#include "mlx/backend/cpu/gemms/simd_gemm.h"
+
+namespace mlx::core {
+
+template <>
+void matmul<float16_t>(
+    const float16_t* a,
+    const float16_t* b,
+    float16_t* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
+  for (int i = 0; i < batch_size; ++i) {
+    simd_gemm<float16_t, float>(
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
+        out + M * N * i,
+        a_transposed,
+        b_transposed,
+        M,
+        N,
+        K,
+        alpha,
+        beta);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/gemms/simd_gemm.h

@@ -0,0 +1,139 @@
+// Copyright © 2025 Apple Inc.
+#pragma once
+
+#include "mlx/backend/cpu/simd/simd.h"
+
+namespace mlx::core {
+
+inline int ceildiv(int a, int b) {
+  return (a + b - 1) / b;
+}
+
+template <int block_size, typename T, typename AccT>
+void load_block(
+    const T* in,
+    AccT* out,
+    int M,
+    int N,
+    int i,
+    int j,
+    bool transpose) {
+  if (transpose) {
+    for (int ii = 0; ii < block_size && i * block_size + ii < M; ++ii) {
+      for (int jj = 0; jj < block_size && j * block_size + jj < N; ++jj) {
+        out[jj * block_size + ii] =
+            in[(i * block_size + ii) * N + j * block_size + jj];
+      }
+    }
+  } else {
+    for (int ii = 0; ii < block_size && i * block_size + ii < M; ++ii) {
+      for (int jj = 0; jj < block_size && j * block_size + jj < N; ++jj) {
+        out[ii * block_size + jj] =
+            in[(i * block_size + ii) * N + j * block_size + jj];
+      }
+    }
+  }
+}
+
+template <typename T, typename AccT>
+void simd_gemm(
+    const T* a,
+    const T* b,
+    T* c,
+    bool a_trans,
+    bool b_trans,
+    int M,
+    int N,
+    int K,
+    float alpha,
+    float beta) {
+  constexpr int block_size = 16;
+  constexpr int simd_size = simd::max_size<AccT>;
+  static_assert(
+      (block_size % simd_size) == 0,
+      "Block size must be divisible by SIMD size");
+
+  int last_k_block_size = K - block_size * (K / block_size);
+  int last_k_simd_block = (last_k_block_size / simd_size) * simd_size;
+  for (int i = 0; i < ceildiv(M, block_size); i++) {
+    for (int j = 0; j < ceildiv(N, block_size); j++) {
+      AccT c_block[block_size * block_size] = {0.0};
+      AccT a_block[block_size * block_size];
+      AccT b_block[block_size * block_size];
+
+      int k = 0;
+      for (; k < K / block_size; k++) {
+        // Load a and b blocks
+        if (a_trans) {
+          load_block<block_size>(a, a_block, K, M, k, i, true);
+        } else {
+          load_block<block_size>(a, a_block, M, K, i, k, false);
+        }
+        if (b_trans) {
+          load_block<block_size>(b, b_block, N, K, j, k, false);
+        } else {
+          load_block<block_size>(b, b_block, K, N, k, j, true);
+        }
+
+        // Multiply and accumulate
+        for (int ii = 0; ii < block_size && i * block_size + ii < M; ++ii) {
+          for (int jj = 0; jj < block_size && j * block_size + jj < N; ++jj) {
+            for (int kk = 0; kk < block_size; kk += simd_size) {
+              auto av =
+                  simd::load<AccT, simd_size>(a_block + ii * block_size + kk);
+              auto bv =
+                  simd::load<AccT, simd_size>(b_block + jj * block_size + kk);
+              c_block[ii * block_size + jj] += simd::sum(av * bv);
+            }
+          }
+        }
+      }
+      if (last_k_block_size) {
+        // Load a and b blocks
+        if (a_trans) {
+          load_block<block_size>(a, a_block, K, M, k, i, true);
+        } else {
+          load_block<block_size>(a, a_block, M, K, i, k, false);
+        }
+        if (b_trans) {
+          load_block<block_size>(b, b_block, N, K, j, k, false);
+        } else {
+          load_block<block_size>(b, b_block, K, N, k, j, true);
+        }
+
+        // Multiply and accumulate
+        for (int ii = 0; ii < block_size && i * block_size + ii < M; ++ii) {
+          for (int jj = 0; jj < block_size && j * block_size + jj < N; ++jj) {
+            int kk = 0;
+            for (; kk < last_k_simd_block; kk += simd_size) {
+              auto av =
+                  simd::load<AccT, simd_size>(a_block + ii * block_size + kk);
+              auto bv =
+                  simd::load<AccT, simd_size>(b_block + jj * block_size + kk);
+              c_block[ii * block_size + jj] += simd::sum(av * bv);
+            }
+            for (; kk < last_k_block_size; ++kk) {
+              c_block[ii * block_size + jj] +=
+                  a_block[ii * block_size + kk] * b_block[jj * block_size + kk];
+            }
+          }
+        }
+      }
+
+      // Store
+      for (int ii = 0; ii < block_size && i * block_size + ii < M; ++ii) {
+        for (int jj = 0; jj < block_size && j * block_size + jj < N; ++jj) {
+          auto c_idx = (i * block_size + ii) * N + j * block_size + jj;
+          if (beta != 0) {
+            c[c_idx] = static_cast<T>(
+                alpha * c_block[ii * block_size + jj] + beta * c[c_idx]);
+          } else {
+            c[c_idx] = static_cast<T>(alpha * c_block[ii * block_size + jj]);
+          }
+        }
+      }
+    }
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/hadamard.cpp

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

mlx/backend/cpu/indexing.cpp

@@ -197,7 +197,7 @@ void dispatch_gather(
 }
 
 void Gather::eval_cpu(const std::vector<array>& inputs, array& out) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& src = inputs[0];
   std::vector<array> inds;
@@ -257,15 +257,11 @@ void gather_axis(
     const array& ind,
     array& out,
     const int axis) {
-  auto strides = ind.strides();
-  strides.erase(strides.begin() + axis);
-  auto shape = ind.shape();
-  shape.erase(shape.begin() + axis);
-  ContiguousIterator ind_it(shape, strides, src.ndim() - 1);
-
-  strides = src.strides();
-  strides.erase(strides.begin() + axis);
-  ContiguousIterator src_it(shape, strides, src.ndim() - 1);
+  auto shape = remove_index(ind.shape(), axis);
+  ContiguousIterator ind_it(
+      shape, remove_index(ind.strides(), axis), src.ndim() - 1);
+  ContiguousIterator src_it(
+      shape, remove_index(src.strides(), axis), src.ndim() - 1);
 
   auto ind_ptr = ind.data<IdxT>();
   auto src_ptr = src.data<T>();
@@ -354,7 +350,7 @@ void dispatch_gather_axis(
 }
 
 void GatherAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& src = inputs[0];
   auto& inds = inputs[1];
@@ -521,7 +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;
@@ -585,15 +581,11 @@ void Scatter::eval_cpu(const std::vector<array>& inputs, array& out) {
 
 template <typename T, typename IdxT, typename OpT>
 void scatter_axis(array& out, const array idx, const array& upd, int axis) {
-  auto strides = idx.strides();
-  strides.erase(strides.begin() + axis);
-  auto shape = idx.shape();
-  shape.erase(shape.begin() + axis);
-  ContiguousIterator idx_it(shape, strides, upd.ndim() - 1);
-
-  strides = upd.strides();
-  strides.erase(strides.begin() + axis);
-  ContiguousIterator upd_it(shape, strides, upd.ndim() - 1);
+  auto shape = remove_index(idx.shape(), axis);
+  ContiguousIterator idx_it(
+      shape, remove_index(idx.strides(), axis), upd.ndim() - 1);
+  ContiguousIterator upd_it(
+      shape, remove_index(upd.strides(), axis), upd.ndim() - 1);
 
   auto idx_ptr = idx.data<IdxT>();
   auto upd_ptr = upd.data<T>();
@@ -694,7 +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

@@ -11,7 +11,7 @@ namespace mlx::core {
 template <typename T>
 void general_inv(T* inv, int N) {
   int info;
-  auto ipiv = array::Data{allocator::malloc_or_wait(sizeof(int) * N)};
+  auto ipiv = array::Data{allocator::malloc(sizeof(int) * N)};
   // Compute LU factorization.
   getrf<T>(
       /* m = */ &N,
@@ -49,7 +49,7 @@ void general_inv(T* inv, int N) {
   }
 
   const int lwork = workspace_size;
-  auto scratch = array::Data{allocator::malloc_or_wait(sizeof(T) * lwork)};
+  auto scratch = array::Data{allocator::malloc(sizeof(T) * lwork)};
 
   // Compute inverse.
   getri<T>(
@@ -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

@@ -2,14 +2,14 @@
 
 #pragma once
 
-// Required for Visual Studio.
-// https://github.com/OpenMathLib/OpenBLAS/blob/develop/docs/install.md
-#ifdef _MSC_VER
 #include <complex>
 #define LAPACK_COMPLEX_CUSTOM
 #define lapack_complex_float std::complex<float>
 #define lapack_complex_double std::complex<double>
-#endif
+#define lapack_complex_float_real(z) ((z).real())
+#define lapack_complex_float_imag(z) ((z).imag())
+#define lapack_complex_double_real(z) ((z).real())
+#define lapack_complex_double_imag(z) ((z).imag())
 
 #ifdef MLX_USE_ACCELERATE
 #include <Accelerate/Accelerate.h>
@@ -32,7 +32,7 @@
 
 #endif
 
-#define INSTANTIATE_LAPACK_TYPES(FUNC)                       \
+#define INSTANTIATE_LAPACK_REAL(FUNC)                        \
   template <typename T, typename... Args>                    \
   void FUNC(Args... args) {                                  \
     if constexpr (std::is_same_v<T, float>) {                \
@@ -42,11 +42,24 @@
     }                                                        \
   }
 
-INSTANTIATE_LAPACK_TYPES(geqrf)
-INSTANTIATE_LAPACK_TYPES(orgqr)
-INSTANTIATE_LAPACK_TYPES(syevd)
-INSTANTIATE_LAPACK_TYPES(potrf)
-INSTANTIATE_LAPACK_TYPES(gesvdx)
-INSTANTIATE_LAPACK_TYPES(getrf)
-INSTANTIATE_LAPACK_TYPES(getri)
-INSTANTIATE_LAPACK_TYPES(trtri)
+INSTANTIATE_LAPACK_REAL(geqrf)
+INSTANTIATE_LAPACK_REAL(orgqr)
+INSTANTIATE_LAPACK_REAL(syevd)
+INSTANTIATE_LAPACK_REAL(geev)
+INSTANTIATE_LAPACK_REAL(potrf)
+INSTANTIATE_LAPACK_REAL(gesvdx)
+INSTANTIATE_LAPACK_REAL(getrf)
+INSTANTIATE_LAPACK_REAL(getri)
+INSTANTIATE_LAPACK_REAL(trtri)
+
+#define INSTANTIATE_LAPACK_COMPLEX(FUNC)                            \
+  template <typename T, typename... Args>                           \
+  void FUNC(Args... args) {                                         \
+    if constexpr (std::is_same_v<T, std::complex<float>>) {         \
+      MLX_LAPACK_FUNC(c##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, std::complex<double>>) { \
+      MLX_LAPACK_FUNC(z##FUNC)(std::forward<Args>(args)...);        \
+    }                                                               \
+  }
+
+INSTANTIATE_LAPACK_COMPLEX(heevd)

mlx/backend/cpu/logsumexp.cpp

@@ -0,0 +1,139 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <cassert>
+#include <cmath>
+
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/simd/simd.h"
+#include "mlx/primitives.h"
+#include "mlx/types/limits.h"
+
+namespace mlx::core {
+
+namespace {
+
+using namespace mlx::core::simd;
+
+template <typename T, typename AccT>
+void logsumexp(const array& in, array& out, Stream stream) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+
+  const T* in_ptr = in.data<T>();
+  T* out_ptr = out.data<T>();
+
+  int M = in.shape().back();
+  int L = in.data_size() / M;
+
+  encoder.dispatch([in_ptr, out_ptr, M, L]() mutable {
+    constexpr int N = std::min(max_size<AccT>, max_size<T>);
+
+    const T* current_in_ptr;
+
+    for (int i = 0; i < L; i++, in_ptr += M, out_ptr += 1) {
+      // Find the maximum
+      current_in_ptr = in_ptr;
+      Simd<AccT, N> vmaximum(-numeric_limits<AccT>::infinity());
+      size_t s = M;
+      while (s >= N) {
+        Simd<AccT, N> vals = load<T, N>(current_in_ptr);
+        vmaximum = maximum(vals, vmaximum);
+        current_in_ptr += N;
+        s -= N;
+      }
+
+      AccT maximum = max(vmaximum);
+      while (s-- > 0) {
+        maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
+        current_in_ptr++;
+      }
+
+      // Compute the normalizer and the exponentials
+      Simd<AccT, N> vnormalizer(0.0);
+      current_in_ptr = in_ptr;
+      s = M;
+      while (s >= N) {
+        Simd<AccT, N> vexp = load<T, N>(current_in_ptr);
+        vexp = exp(vexp - maximum);
+        vnormalizer = vnormalizer + vexp;
+        current_in_ptr += N;
+        s -= N;
+      }
+      AccT normalizer = sum(vnormalizer);
+      while (s-- > 0) {
+        AccT _exp = std::exp(*current_in_ptr - maximum);
+        normalizer += _exp;
+        current_in_ptr++;
+      }
+      // Normalize
+      *out_ptr = std::isinf(maximum)
+          ? static_cast<T>(maximum)
+          : static_cast<T>(std::log(normalizer) + maximum);
+    }
+  });
+}
+
+} // namespace
+
+void LogSumExp::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+
+  // Make sure that the last dimension is contiguous
+  auto s = stream();
+  auto& encoder = cpu::get_command_encoder(s);
+  auto ensure_contiguous = [&s, &encoder](const array& x) {
+    if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
+      return x;
+    } else {
+      array x_copy = contiguous_copy_cpu(x, s);
+      encoder.add_temporary(x_copy);
+      return x_copy;
+    }
+  };
+
+  auto in = ensure_contiguous(inputs[0]);
+  if (in.flags().row_contiguous) {
+    out.set_data(allocator::malloc(out.nbytes()));
+  } else {
+    auto n = in.shape(-1);
+    auto flags = in.flags();
+    auto strides = in.strides();
+    for (auto& s : strides) {
+      s /= n;
+    }
+    bool col_contig = strides[0] == 1;
+    for (int i = 1; col_contig && i < strides.size(); ++i) {
+      col_contig &=
+          (out.shape(i) == 1 || strides[i - 1] == out.shape(i) * strides[i]);
+    }
+    flags.col_contiguous = col_contig;
+    out.set_data(
+        allocator::malloc(in.nbytes() / n),
+        in.data_size() / n,
+        std::move(strides),
+        flags);
+  }
+
+  switch (in.dtype()) {
+    case float32:
+      logsumexp<float, float>(in, out, stream());
+      break;
+    case float16:
+      logsumexp<float16_t, float>(in, out, stream());
+      break;
+    case bfloat16:
+      logsumexp<bfloat16_t, float>(in, out, stream());
+      break;
+    case float64:
+      logsumexp<double, double>(in, out, stream());
+      break;
+    default:
+      throw std::runtime_error(
+          "[logsumexp] only supports floating point types");
+      break;
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cpu/luf.cpp

@@ -30,9 +30,8 @@ void luf_impl(
   auto strides = lu.strides();
   strides[ndim - 1] = M;
   strides[ndim - 2] = 1;
-  lu.set_data(
-      allocator::malloc_or_wait(lu.nbytes()), lu.nbytes(), strides, flags);
-  copy_inplace(
+  lu.set_data(allocator::malloc(lu.nbytes()), lu.nbytes(), strides, flags);
+  copy_cpu_inplace(
       a,
       lu,
       a.shape(),
@@ -44,8 +43,8 @@ void luf_impl(
       stream);
 
   auto a_ptr = lu.data<T>();
-  pivots.set_data(allocator::malloc_or_wait(pivots.nbytes()));
-  row_indices.set_data(allocator::malloc_or_wait(row_indices.nbytes()));
+  pivots.set_data(allocator::malloc(pivots.nbytes()));
+  row_indices.set_data(allocator::malloc(row_indices.nbytes()));
   auto pivots_ptr = pivots.data<uint32_t>();
   auto row_indices_ptr = row_indices.data<uint32_t>();
   size_t num_matrices = a.size() / (M * N);

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) {
@@ -59,7 +112,7 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     throw std::runtime_error(
         "[BlockMaskedMM::eval] Currently only supports float32.");
   }
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
@@ -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);
         }
       };
@@ -318,7 +370,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     throw std::runtime_error(
         "[GatherMM::eval] Currently only supports float32.");
   }
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
@@ -333,7 +385,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       int64_t stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -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());
     }
@@ -115,7 +115,7 @@ void matmul_general(
 }
 
 void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
   if (inputs[0].shape(-1) == 0) {
     auto& encoder = cpu::get_command_encoder(stream());
     encoder.set_output_array(out);
@@ -132,14 +132,20 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     throw std::runtime_error(
         "[AddMM::eval_cpu] Currently only supports float32.");
   }
+  if (out.size() == 0) {
+    out.set_data(allocator::malloc(out.nbytes()));
+    return;
+  }
 
   // Fill output with C
   auto& c = inputs[2];
   CopyType ctype = c.data_size() == 1
       ? CopyType::Scalar
       : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-  copy(c, out, ctype, stream());
-
+  copy_cpu(c, out, ctype, stream());
+  if (inputs[0].shape(-1) == 0) {
+    return;
+  }
   matmul_general(inputs[0], inputs[1], out, stream(), alpha_, beta_);
 }
 

mlx/backend/cpu/primitives.cpp

@@ -21,8 +21,8 @@ namespace mlx::core {
 void reshape(const array& in, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    copy_inplace(in, out, CopyType::General, out.primitive().stream());
+    out.set_data(allocator::malloc(out.nbytes()));
+    copy_cpu_inplace(in, out, CopyType::General, out.primitive().stream());
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -39,7 +39,7 @@ static std::pair<array, bool> compute_dynamic_offset(
   if (donate) {
     offset.copy_shared_buffer(indices);
   } else {
-    offset.set_data(allocator::malloc_or_wait(offset.itemsize()));
+    offset.set_data(allocator::malloc(offset.itemsize()));
   }
 
   auto& encoder = cpu::get_command_encoder(stream);
@@ -124,7 +124,7 @@ void Transpose::eval_cpu(const std::vector<array>& inputs, array& out) {
 
 void Arange::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 0);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
   switch (out.dtype()) {
     case bool_:
       throw std::runtime_error("Bool type unsupported for arange.");
@@ -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) {
@@ -186,7 +186,7 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
   std::partial_sum(sizes.cbegin(), sizes.cend(), sizes.begin());
 
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto strides = out.strides();
   auto flags = out.flags();
@@ -198,18 +198,20 @@ 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());
   }
 }
 
 void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  if (in.flags().row_contiguous ||
-      (allow_col_major_ && in.flags().col_contiguous)) {
+  constexpr size_t extra_bytes = 16384;
+  if (in.buffer_size() <= out.nbytes() + extra_bytes &&
+      (in.flags().row_contiguous ||
+       (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());
   }
 }
 
@@ -233,7 +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) {
@@ -249,7 +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;
@@ -264,7 +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) {
@@ -276,7 +278,7 @@ void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   size_t elems_per_key = out.size() / num_keys;
   size_t bytes_per_key = out.itemsize() * elems_per_key;
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto kptr = inputs[0].data<uint32_t>();
   auto cptr = out.data<char>();
@@ -335,10 +337,10 @@ void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
     return;
   }
   auto& in = inputs[0];
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  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(),
@@ -370,11 +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(),
@@ -410,14 +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(),
@@ -450,13 +452,13 @@ void View::eval_cpu(const std::vector<array>& inputs, array& out) {
   } else {
     auto tmp = array(
         in.shape(), in.dtype() == bool_ ? uint8 : in.dtype(), nullptr, {});
-    tmp.set_data(allocator::malloc_or_wait(tmp.nbytes()));
+    tmp.set_data(allocator::malloc(tmp.nbytes()));
     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

@@ -25,12 +25,11 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
   auto strides = in.strides();
   strides[in.ndim() - 2] = 1;
   strides[in.ndim() - 1] = M;
-  in.set_data(
-      allocator::malloc_or_wait(in.nbytes()), in.nbytes(), strides, flags);
-  copy_inplace(a, in, CopyType::GeneralGeneral, stream);
+  in.set_data(allocator::malloc(in.nbytes()), in.nbytes(), strides, flags);
+  copy_cpu_inplace(a, in, CopyType::GeneralGeneral, stream);
   auto& encoder = cpu::get_command_encoder(stream);
-  q.set_data(allocator::malloc_or_wait(q.nbytes()));
-  r.set_data(allocator::malloc_or_wait(r.nbytes()));
+  q.set_data(allocator::malloc(q.nbytes()));
+  r.set_data(allocator::malloc(r.nbytes()));
 
   auto in_ptr = in.data<T>();
   auto r_ptr = r.data<T>();
@@ -41,8 +40,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
   encoder.set_output_array(r);
   encoder.dispatch([in_ptr, q_ptr, r_ptr, M, N, lda, num_matrices]() {
     int num_reflectors = std::min(M, N);
-    auto tau =
-        allocator::malloc_or_wait(sizeof(T) * num_matrices * num_reflectors);
+    auto tau = allocator::malloc(sizeof(T) * num_matrices * num_reflectors);
 
     T optimal_work;
     int lwork = -1;
@@ -53,7 +51,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
 
     // Update workspace size
     lwork = optimal_work;
-    auto work = allocator::malloc_or_wait(sizeof(T) * lwork);
+    auto work = allocator::malloc(sizeof(T) * lwork);
 
     // Loop over matrices
     for (int i = 0; i < num_matrices; ++i) {
@@ -96,7 +94,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
         &lwork,
         &info);
     lwork = optimal_work;
-    work = allocator::malloc_or_wait(sizeof(T) * lwork);
+    work = allocator::malloc(sizeof(T) * lwork);
 
     // Loop over matrices
     for (int i = 0; i < num_matrices; ++i) {

mlx/backend/cpu/quantized.cpp

@@ -13,9 +13,18 @@ namespace mlx::core {
 
 namespace {
 
+inline constexpr short get_pack_factor(int bits, int wsize = 8) {
+  return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
+}
+
+inline constexpr short get_bytes_per_pack(int bits, int wsize = 8) {
+  auto power_of_2_bits = (bits & (bits - 1)) == 0;
+  return power_of_2_bits ? (wsize / 8) : (bits == 5 ? 5 : 3);
+}
+
 template <typename T, int bits>
 void extract_bits(const uint8_t* w_in, T* w_out) {
-  assert(bits == 3 || bits == 6);
+  static_assert(bits == 3 || bits == 5 || bits == 6);
   if (bits == 3) {
     w_out[0] = static_cast<T>(w_in[0] & 0x7);
     w_out[1] = static_cast<T>((w_in[0] & 0x38) >> 3);
@@ -25,6 +34,16 @@ void extract_bits(const uint8_t* w_in, T* w_out) {
     w_out[5] = static_cast<T>(((w_in[1] & 0x80) >> 7) + ((w_in[2] & 0x3) << 1));
     w_out[6] = static_cast<T>((w_in[2] & 0x1c) >> 2);
     w_out[7] = static_cast<T>((w_in[2] & 0xe0) >> 5);
+  } else if (bits == 5) {
+    w_out[0] = static_cast<T>(w_in[0] & 0x1f);
+    w_out[1] = static_cast<T>(((w_in[0] & 0xe0) >> 5) + ((w_in[1] & 0x3) << 3));
+    w_out[2] = static_cast<T>((w_in[1] & 0x7c) >> 2);
+    w_out[3] = static_cast<T>(((w_in[1] & 0x80) >> 7) + ((w_in[2] & 0xf) << 1));
+    w_out[4] = static_cast<T>(((w_in[2] & 0xf0) >> 4) + ((w_in[3] & 0x1) << 4));
+    w_out[5] = static_cast<T>((w_in[3] & 0x3e) >> 1);
+    w_out[6] = static_cast<T>(((w_in[3] & 0xc0) >> 6) + ((w_in[4] & 0x7) << 2));
+    w_out[7] = static_cast<T>((w_in[4] & 0xf8) >> 3);
+
   } else if (bits == 6) {
     w_out[0] = static_cast<T>(w_in[0] & 0x3f);
     w_out[1] =
@@ -46,8 +65,8 @@ void _qmm(
     int N,
     int K) {
   constexpr int bitmask = (1 << bits) - 1;
-  constexpr int pack_factor = bits == 3 ? 8 : bits == 6 ? 4 : 8 / bits;
-  constexpr int bytes_per_pack = (bits == 3 || bits == 6) ? 3 : 1;
+  constexpr int pack_factor = get_pack_factor(bits, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(bits);
   constexpr int packs_in_group = group_size / pack_factor;
 
   for (int m = 0; m < M; m++) {
@@ -65,7 +84,7 @@ void _qmm(
         T scale = *scales_local++;
         T bias = *biases_local++;
         for (int ng = 0; ng < packs_in_group; ng++) {
-          if (bits == 3 || bits == 6) {
+          if constexpr (bits == 3 || bits == 5 || bits == 6) {
             T wl[pack_factor];
             extract_bits<T, bits>(w_local, wl);
 #pragma clang loop unroll(full)
@@ -104,8 +123,9 @@ void _qmm_t(
     int N,
     int K) {
   constexpr int bitmask = (1 << bits) - 1;
-  constexpr int pack_factor = bits == 3 ? 8 : bits == 6 ? 4 : 8 / bits;
-  constexpr int bytes_per_pack = (bits == 3 || bits == 6) ? 3 : 1;
+
+  constexpr int pack_factor = get_pack_factor(bits, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(bits);
   constexpr int packs_in_group = group_size / pack_factor;
 
   for (int m = 0; m < M; m++) {
@@ -121,7 +141,7 @@ void _qmm_t(
         T bias = *biases_local++;
 
         for (int kw = 0; kw < packs_in_group; kw++) {
-          if (bits == 3 || bits == 6) {
+          if constexpr (bits == 3 || bits == 5 || bits == 6) {
             T wl[pack_factor];
             extract_bits<T, bits>(w_local, wl);
 #pragma clang loop unroll(full)
@@ -304,6 +324,10 @@ void _qmm_dispatch_typed(
       _qmm_dispatch_group<T, 4>(
           result, x, w, scales, biases, M, N, K, group_size, transposed_w);
       break;
+    case 5:
+      _qmm_dispatch_group<T, 5>(
+          result, x, w, scales, biases, M, N, K, group_size, transposed_w);
+      break;
     case 6:
       _qmm_dispatch_group<T, 6>(
           result, x, w, scales, biases, M, N, K, group_size, transposed_w);
@@ -505,7 +529,7 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
       return 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);
       return temps.back();
     }
   };
@@ -515,7 +539,7 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto scales = ensure_row_contiguous(scales_pre);
   auto biases = ensure_row_contiguous(biases_pre);
 
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.add_temporaries(std::move(temps));
@@ -555,7 +579,7 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return 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);
       return temps.back();
     }
   };
@@ -565,7 +589,7 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
   auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.add_temporaries(std::move(temps));
@@ -613,9 +637,8 @@ void quantize(
   float eps = 1e-7;
 
   bool power_of_2_bits = is_power_of_2(bits);
-  int el_per_int = bits == 3 ? 8 : bits == 6 ? 4 : 32 / bits;
-  // For 3/6 bits we read 3 uint8s at a time instead of 1 uint32
-  int bytes_per_pack = power_of_2_bits ? 1 : 3;
+  int el_per_int = get_pack_factor(bits, 32);
+  int bytes_per_pack = get_bytes_per_pack(bits);
   int int_per_group = group_size * bytes_per_pack / el_per_int;
   size_t n_groups = w_size / group_size;
 
@@ -640,15 +663,21 @@ void quantize(
     }
     size_t out_idx = i * int_per_group;
     for (int j = 0; j < int_per_group / bytes_per_pack; ++j) {
-      uint32_t out_el = 0;
+      uint64_t out_el = 0;
       for (int k = 0; k < el_per_int; ++k) {
         float w_el = w[w_idx + j * el_per_int + k];
         w_el = std::rint((w_el - bias) / scale);
         w_el = std::min(std::max(w_el, 0.0f), n_bins);
-        out_el |= static_cast<uint32_t>(w_el) << (k * bits);
+        out_el |= static_cast<uint64_t>(w_el) << (k * bits);
       }
       if (power_of_2_bits) {
         out[out_idx + j] = out_el;
+      } else if (bits == 5) {
+        out[out_idx + bytes_per_pack * j] = out_el & 0xff;
+        out[out_idx + bytes_per_pack * j + 1] = (out_el & 0xff00) >> 8;
+        out[out_idx + bytes_per_pack * j + 2] = (out_el & 0xff0000) >> 16;
+        out[out_idx + bytes_per_pack * j + 3] = (out_el & 0xff000000) >> 24;
+        out[out_idx + bytes_per_pack * j + 4] = (out_el & 0xff00000000) >> 32;
       } else {
         out[out_idx + bytes_per_pack * j] = out_el & 0xff;
         out[out_idx + bytes_per_pack * j + 1] = (out_el & 0xff00) >> 8;
@@ -683,20 +712,18 @@ void fast::AffineQuantize::eval_cpu(
     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);
     }
   };
 
   auto [w, copied] = ensure_row_contiguous(inputs[0]);
   auto& out = outputs[0];
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& scales = outputs[1];
   auto& biases = outputs[2];
-  scales.set_data(allocator::malloc_or_wait(scales.nbytes()));
-  biases.set_data(allocator::malloc_or_wait(biases.nbytes()));
+  scales.set_data(allocator::malloc(scales.nbytes()));
+  biases.set_data(allocator::malloc(biases.nbytes()));
   auto& encoder = cpu::get_command_encoder(stream());
   if (copied) {
     encoder.add_temporary(w);

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);
   };
 };
@@ -433,7 +449,7 @@ void reduce_dispatch_min_max(
 void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(in);
   encoder.set_output_array(out);
@@ -527,10 +543,10 @@ 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

@@ -3,6 +3,7 @@
 #include <cassert>
 
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/binary_ops.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
@@ -226,6 +227,16 @@ void scan_dispatch(
       scan_op<T, U>(in, out, axis, reverse, inclusive, op, init);
       break;
     }
+    case Scan::LogAddExp: {
+      auto op = [](U a, T b) {
+        return detail::LogAddExp{}(a, static_cast<U>(b));
+      };
+      auto init = (issubdtype(in.dtype(), floating))
+          ? static_cast<U>(-std::numeric_limits<float>::infinity())
+          : std::numeric_limits<U>::min();
+      scan_op<T, U>(in, out, axis, reverse, inclusive, op, init);
+      break;
+    }
   }
 }
 
@@ -239,12 +250,10 @@ 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_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   encoder.set_input_array(in);
   encoder.set_output_array(out);
@@ -319,7 +328,8 @@ void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
             reduce_type_, in, out, axis_, reverse_, inclusive_);
         break;
       case complex64:
-        throw std::runtime_error("Scan ops do not support complex types yet");
+        scan_dispatch<complex64_t, complex64_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
         break;
     }
   });

mlx/backend/cpu/simd/accelerate_fp16_simd.h

@@ -17,7 +17,7 @@ struct ScalarT<float16_t, N> {
 #endif
 
 template <>
-static constexpr int max_size<float16_t> = N;
+inline constexpr int max_size<float16_t> = N;
 
 #define SIMD_FP16_DEFAULT_UNARY(op)                    \
   template <>                                          \

mlx/backend/cpu/simd/accelerate_simd.h

@@ -83,25 +83,25 @@ struct Simd {
 // Values chosen based on benchmarks on M3 Max
 // TODO: consider choosing these more optimally
 template <>
-static constexpr int max_size<int8_t> = 16;
+inline constexpr int max_size<int8_t> = 16;
 template <>
-static constexpr int max_size<int16_t> = 16;
+inline constexpr int max_size<int16_t> = 16;
 template <>
-static constexpr int max_size<int> = 8;
+inline constexpr int max_size<int> = 8;
 template <>
-static constexpr int max_size<int64_t> = 4;
+inline constexpr int max_size<int64_t> = 4;
 template <>
-static constexpr int max_size<uint8_t> = 16;
+inline constexpr int max_size<uint8_t> = 16;
 template <>
-static constexpr int max_size<uint16_t> = 16;
+inline constexpr int max_size<uint16_t> = 16;
 template <>
-static constexpr int max_size<uint32_t> = 8;
+inline constexpr int max_size<uint32_t> = 8;
 template <>
-static constexpr int max_size<uint64_t> = 4;
+inline constexpr int max_size<uint64_t> = 4;
 template <>
-static constexpr int max_size<float> = 8;
+inline constexpr int max_size<float> = 8;
 template <>
-static constexpr int max_size<double> = 4;
+inline constexpr int max_size<double> = 4;
 
 #define SIMD_DEFAULT_UNARY(name, op) \
   template <typename T, int N>       \

mlx/backend/cpu/simd/base_simd.h

@@ -87,14 +87,45 @@ DEFAULT_UNARY(cosh, std::cosh)
 DEFAULT_UNARY(expm1, std::expm1)
 DEFAULT_UNARY(floor, std::floor)
 DEFAULT_UNARY(log, std::log)
-DEFAULT_UNARY(log2, std::log2)
 DEFAULT_UNARY(log10, std::log10)
-DEFAULT_UNARY(log1p, std::log1p)
 DEFAULT_UNARY(sinh, std::sinh)
 DEFAULT_UNARY(sqrt, std::sqrt)
 DEFAULT_UNARY(tan, std::tan)
 DEFAULT_UNARY(tanh, std::tanh)
 
+template <typename T>
+Simd<T, 1> log1p(Simd<T, 1> in) {
+  if constexpr (is_complex<T>) {
+    auto x = in.value.real();
+    auto y = in.value.imag();
+    auto zabs = std::abs(in.value);
+    auto theta = std::atan2(y, x + 1);
+    if (zabs < 0.5) {
+      auto r = x * (2 + x) + y * y;
+      if (r == 0) { // handle underflow
+        return Simd<T, 1>{T{x, theta}};
+      }
+      return Simd<T, 1>{T{((typeof(x))(0.5)) * std::log1p(r), theta}};
+    } else {
+      auto z0 = std::hypot(x + 1, y);
+      return Simd<T, 1>{T{std::log(z0), theta}};
+    }
+  } else {
+    return Simd<T, 1>{std::log1p(in.value)};
+  }
+}
+
+template <typename T>
+Simd<T, 1> log2(Simd<T, 1> in) {
+  if constexpr (is_complex<T>) {
+    auto out = std::log(in.value);
+    auto scale = decltype(out.real())(M_LN2);
+    return Simd<T, 1>{T{out.real() / scale, out.imag() / scale}};
+  } else {
+    return Simd<T, 1>{std::log2(in.value)};
+  }
+}
+
 template <typename T>
 Simd<T, 1> operator~(Simd<T, 1> in) {
   return ~in.value;

mlx/backend/cpu/softmax.cpp

@@ -119,25 +119,19 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   // Make sure that the last dimension is contiguous
   auto set_output = [s = stream(), &out](const array& x) {
-    bool no_copy = x.strides()[x.ndim() - 1] == 1;
-    if (x.ndim() > 1) {
-      auto s = x.strides()[x.ndim() - 2];
-      no_copy &= (s == 0 || s == x.shape().back());
-    }
-    if (no_copy) {
+    if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
       if (x.is_donatable()) {
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            allocator::malloc_or_wait(x.data_size() * x.itemsize()),
+            allocator::malloc(x.data_size() * x.itemsize()),
             x.data_size(),
             x.strides(),
             x.flags());
       }
       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;
     }
@@ -146,18 +140,6 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto in = set_output(inputs[0]);
 
   switch (in.dtype()) {
-    case bool_:
-    case uint8:
-    case uint16:
-    case uint32:
-    case uint64:
-    case int8:
-    case int16:
-    case int32:
-    case int64:
-      throw std::runtime_error(
-          "Softmax is defined only for floating point types");
-      break;
     case float32:
       softmax<float, float>(in, out, stream());
       break;
@@ -178,9 +160,9 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
     case float64:
       softmax<double, double>(in, out, stream());
       break;
-    case complex64:
-      throw std::invalid_argument(
-          "[Softmax] Not yet implemented for complex64");
+    default:
+      throw std::runtime_error(
+          "[softmax] Only defined for floating point types.");
       break;
   }
 }

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 {
@@ -288,7 +288,7 @@ void ArgSort::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Allocate output
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(in);
@@ -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) {
@@ -379,7 +358,7 @@ void ArgPartition::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Allocate output
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  out.set_data(allocator::malloc(out.nbytes()));
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(in);
@@ -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);