.circleci/config.yml

@@ -7,6 +7,15 @@ 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:
@@ -15,8 +24,8 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "16.2.0"
-    resource_class: m2pro.medium
+      xcode: "15.2.0"
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
@@ -29,7 +38,7 @@ jobs:
             pip install --upgrade pip
             pip install --upgrade cmake
             pip install -r docs/requirements.txt
-            pip install . -v
+            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install . -v
       - when:
           condition:
             not: << parameters.upload-docs >>
@@ -61,9 +70,9 @@ jobs:
                  git push -f origin gh-pages
 
   linux_build_and_test:
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
+    docker:
+      - image: cimg/python:3.9
+
     steps:
       - checkout
       - run:
@@ -75,17 +84,20 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            sudo apt-get upgrade -y
             pip install --upgrade cmake
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
+            pip install nanobind==2.2.0
+            pip install numpy
+            sudo apt-get update
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
       - run:
           name: Install Python package
           command: |
-            pip install -e ".[dev]"
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              python3 setup.py build_ext --inplace
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              python3 setup.py develop
       - run:
           name: Generate package stubs
           command: |
@@ -95,10 +107,7 @@ jobs:
       - run:
           name: Run Python tests
           command: |
-            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
+            python3 -m unittest discover python/tests -v
       - run:
           name: Build CPP only
           command: |
@@ -113,15 +122,10 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "16.2.0"
-      macosx_deployment_target:
-        type: string
-        default: ""
+        default: "15.2.0"
     macos:
       xcode: << parameters.xcode_version >>
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m2pro.medium
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
@@ -133,7 +137,7 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install nanobind==2.4.0
+            pip install nanobind==2.2.0
             pip install numpy
             pip install torch
             pip install tensorflow
@@ -142,8 +146,7 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              pip install -e . -v
+            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
@@ -157,8 +160,6 @@ jobs:
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
             mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build example extension
           command: |
@@ -193,34 +194,13 @@ jobs:
           name: Run Python tests with JIT
           command: |
             source env/bin/activate
-            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
+            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
+              CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
               pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               python -m xmlrunner discover -v python/tests -o test-results/gpu_jit
 
-  cuda_build_and_test:
-    machine:
-      image: linux-cuda-12:2023.11.1
-      resource_class: gpu.nvidia.small.gen2
-    steps:
-      - checkout
-      - run:
-          name: Install Python package
-          command: |
-            sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            python3 -m venv env
-            source env/bin/activate
-            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              pip install -e ".[dev]"
-      - run:
-          name: Run Python tests
-          command: |
-            source env/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
-
   build_release:
     parameters:
       python_version:
@@ -228,18 +208,13 @@ jobs:
         default: "3.9"
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "15.2.0"
       build_env:
         type: string
         default: ""
-      macosx_deployment_target:
-        type: string
-        default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: m2pro.medium
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
@@ -251,7 +226,7 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install nanobind==2.4.0
+            pip install nanobind==2.2.0
             pip install --upgrade setuptools
             pip install numpy
             pip install twine
@@ -260,7 +235,8 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
+            DEV_RELEASE=1 \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               pip install . -v
       - run:
           name: Generate package stubs
@@ -272,17 +248,9 @@ jobs:
           name: Build Python package
           command: |
             source env/bin/activate
-            << 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
+            << parameters.build_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
+              python -m build -w
       - when:
           condition: << parameters.build_env >>
           steps:
@@ -299,99 +267,52 @@ jobs:
       python_version:
         type: string
         default: "3.9"
-      build_env:
+      extra_env:
         type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
+        default: "DEV_RELEASE=1"
+    docker:
+      - image: ubuntu:20.04
     steps:
       - checkout
       - run:
           name: Build wheel
           command: |
             PYTHON=python<< parameters.python_version >>
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            sudo apt-get upgrade -y
-            TZ=Etc/UTC sudo apt-get -y install tzdata
-            sudo apt-get install -y apt-utils
-            sudo apt-get install -y software-properties-common
-            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
-            sudo apt-get install -y build-essential git
+            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
             $PYTHON -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
+            pip install nanobind==2.2.0
+            pip install --upgrade setuptools
+            pip install numpy
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.build_env >> pip install ".[dev]" -v
+            << parameters.extra_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              pip install . -v
             pip install typing_extensions
             python setup.py generate_stubs 
-            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: linux-cuda-12:default
-      resource_class: gpu.nvidia.small.gen2
-    steps:
-      - checkout
+            << parameters.extra_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              python -m build --wheel
+            auditwheel show dist/*
+            auditwheel repair dist/* --plat manylinux_2_31_x86_64
       - run:
-          name: Build wheel
+          name: Upload package
           command: |
-            sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install zip
-            python -m venv env
             source env/bin/activate
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            << 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: |
-                  source env/bin/activate
-                  twine upload wheelhouse/*.whl
+            twine upload wheelhouse/*
       - store_artifacts:
           path: wheelhouse/
 
@@ -403,19 +324,22 @@ workflows:
             pattern: "^(?!pull/)[-\\w]+$"
             value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
+        - not: << pipeline.parameters.weekly_build >>
+        - not: << pipeline.parameters.test_release >>
     jobs:
       - mac_build_and_test:
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test
-      - cuda_build_and_test 
       - 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:
           filters:
@@ -426,70 +350,8 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
+              xcode_version: ["15.0.0", "15.2.0"]
               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:
@@ -497,25 +359,6 @@ 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:
@@ -531,11 +374,9 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test:
           requires: [ hold ]
-      - cuda_build_and_test:
-          requires: [ hold ]
   nightly_build:
     when:
       and:
@@ -546,56 +387,27 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["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"
+              xcode_version: ["15.0.0", "15.2.0"]
+  weekly_build:
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.weekly_build >>
+    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"]
+              build_env: ["DEV_RELEASE=1"]
+  linux_test_release:
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.linux_release >>
+    jobs:
       - build_linux_release:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-      - build_cuda_release
+              extra_env: ["PYPI_RELEASE=1"]

.gitignore

@@ -36,7 +36,6 @@ share/python-wheels/
 .installed.cfg
 *.egg
 MANIFEST
-uv.lock
 
 # vim
 *.swp
@@ -77,9 +76,6 @@ build/
 *.out
 *.app
 
-# Debug symbols
-*.pdb
-
 # VSCode 
 .vscode/
 .DS_Store

.pre-commit-config.yaml

@@ -1,16 +1,15 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v19.1.7
+    rev: v18.1.8
     hooks:
     -   id: clang-format
 # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 25.1.0
+    rev: 24.8.0
     hooks:
     -   id: black
-    
 -   repo: https://github.com/pycqa/isort
-    rev: 6.0.0
+    rev: 5.13.2
     hooks:
     -   id: isort
         args:

ACKNOWLEDGMENTS.md

@@ -7,7 +7,7 @@ with a short description of your contribution(s) below. For example:
 
 MLX was developed with contributions from the following individuals:
 
-- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops. Added `clip_grad_norm` along with `tree_reduce`. Added `cross`. Added `orthogonal` initializer.
+- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops. Added `clip_grad_norm` along with `tree_reduce`. Added `cross`.
 - Juarez Bochi: Fixed bug in cross attention.
 - Justin Deschenaux: Sine, Cosine, arange, randint, truncated normal, bernoulli, lion optimizer, Dropout2d, linear and logistic regression python example.
 - Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream`, safetensors support, `einsum`, and `einsum_path`.

CMakeLists.txt

@@ -1,24 +1,6 @@
-cmake_minimum_required(VERSION 3.25)
+cmake_minimum_required(VERSION 3.24)
 
-if(NOT MLX_VERSION)
-  file(STRINGS "mlx/version.h" _mlx_h_version REGEX "^#define MLX_VERSION_.*$")
-  string(REGEX MATCH "#define MLX_VERSION_MAJOR ([0-9]+)" _ "${_mlx_h_version}")
-  set(_major ${CMAKE_MATCH_1})
-  string(REGEX MATCH "#define MLX_VERSION_MINOR ([0-9]+)" _ "${_mlx_h_version}")
-  set(_minor ${CMAKE_MATCH_1})
-  string(REGEX MATCH "#define MLX_VERSION_PATCH ([0-9]+)" _ "${_mlx_h_version}")
-  set(_patch ${CMAKE_MATCH_1})
-  set(MLX_PROJECT_VERSION "${_major}.${_minor}.${_patch}")
-  set(MLX_VERSION ${MLX_PROJECT_VERSION})
-else()
-  string(REGEX REPLACE "^([0-9]+\.[0-9]+\.[0-9]+).*" "\\1" MLX_PROJECT_VERSION
-                       ${MLX_VERSION})
-endif()
-
-project(
-  mlx
-  LANGUAGES C CXX
-  VERSION ${MLX_PROJECT_VERSION})
+project(mlx LANGUAGES C CXX)
 
 # ----------------------------- Setup -----------------------------
 set(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
@@ -34,21 +16,26 @@ 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(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
 
+if(NOT MLX_VERSION)
+  set(MLX_VERSION 0.20.0)
+endif()
+
 # --------------------- Processor tests -------------------------
+
 message(
   STATUS
     "Building MLX for ${CMAKE_SYSTEM_PROCESSOR} processor on ${CMAKE_SYSTEM_NAME}"
 )
 
+set(MLX_BUILD_ARM OFF)
+
 if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
   if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
     if(NOT MLX_ENABLE_X64_MAC)
@@ -64,8 +51,14 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
       message(WARNING "Building for x86_64 arch is not officially supported.")
     endif()
   endif()
+
 else()
   set(MLX_BUILD_METAL OFF)
+  message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
+endif()
+
+if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm64")
+  set(MLX_BUILD_ARM ON)
 endif()
 
 # ----------------------------- Lib -----------------------------
@@ -82,10 +75,6 @@ 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)
@@ -100,26 +89,25 @@ elseif(MLX_BUILD_METAL)
   # Throw an error if xcrun not found
   execute_process(
     COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
-    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
+    OUTPUT_VARIABLE MACOS_VERSION COMMAND_ERROR_IS_FATAL ANY)
 
-  if(${MACOS_SDK_VERSION} LESS 14.0)
+  if(${MACOS_VERSION} LESS 14.0)
     message(
       FATAL_ERROR
         "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON")
   endif()
-  message(STATUS "Building with macOS SDK version ${MACOS_SDK_VERSION}")
+  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
 
   set(METAL_CPP_URL
-      https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18.zip)
-
-  if(NOT CMAKE_OSX_DEPLOYMENT_TARGET STREQUAL "")
-    set(XCRUN_FLAGS "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
-  endif()
+      https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18-beta.zip
+  )
+  # Get the metal version
   execute_process(
     COMMAND
       zsh "-c"
-      "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal ${XCRUN_FLAGS} -E -x metal -P - | tail -1 | tr -d '\n'"
+      "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal -E -x metal -P - | tail -1 | tr -d '\n'"
     OUTPUT_VARIABLE MLX_METAL_VERSION COMMAND_ERROR_IS_FATAL ANY)
+
   FetchContent_Declare(metal_cpp URL ${METAL_CPP_URL})
 
   FetchContent_MakeAvailable(metal_cpp)
@@ -127,58 +115,20 @@ elseif(MLX_BUILD_METAL)
     mlx PUBLIC $<BUILD_INTERFACE:${metal_cpp_SOURCE_DIR}>
                $<INSTALL_INTERFACE:include/metal_cpp>)
   target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
-endif()
 
-if(WIN32)
-  if(MSVC)
-    # GGUF does not build with MSVC.
-    set(MLX_BUILD_GGUF OFF)
-    # There is no prebuilt OpenBLAS distribution for MSVC.
-    set(MLX_BUILD_BLAS_FROM_SOURCE ON)
-  endif()
-  # Windows implementation of dlfcn.h APIs.
-  FetchContent_Declare(
-    dlfcn-win32
-    GIT_REPOSITORY https://github.com/dlfcn-win32/dlfcn-win32.git
-    GIT_TAG v1.4.1
-    EXCLUDE_FROM_ALL)
-  block()
-  set(BUILD_SHARED_LIBS OFF)
-  FetchContent_MakeAvailable(dlfcn-win32)
-  endblock()
-  target_include_directories(mlx PRIVATE "${dlfcn-win32_SOURCE_DIR}/src")
-  target_link_libraries(mlx PRIVATE dl)
+  add_compile_definitions("MLX_METAL_VERSION=${MLX_METAL_VERSION}")
 endif()
 
 if(MLX_BUILD_CPU)
   find_library(ACCELERATE_LIBRARY Accelerate)
-  if(ACCELERATE_LIBRARY)
+  if(MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
     message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
     set(MLX_BUILD_ACCELERATE ON)
+    target_link_libraries(mlx PUBLIC ${ACCELERATE_LIBRARY})
+    add_compile_definitions(ACCELERATE_NEW_LAPACK)
   else()
     message(STATUS "Accelerate or arm neon not found, using default backend.")
     set(MLX_BUILD_ACCELERATE OFF)
-  endif()
-
-  if(MLX_BUILD_ACCELERATE)
-    target_link_libraries(mlx PUBLIC ${ACCELERATE_LIBRARY})
-    add_compile_definitions(MLX_USE_ACCELERATE)
-    add_compile_definitions(ACCELERATE_NEW_LAPACK)
-  elseif(MLX_BUILD_BLAS_FROM_SOURCE)
-    # Download and build OpenBLAS from source code.
-    FetchContent_Declare(
-      openblas
-      GIT_REPOSITORY https://github.com/OpenMathLib/OpenBLAS.git
-      GIT_TAG v0.3.28
-      EXCLUDE_FROM_ALL)
-    set(BUILD_STATIC_LIBS ON) # link statically
-    set(NOFORTRAN ON) # msvc has no fortran compiler
-    FetchContent_MakeAvailable(openblas)
-    target_link_libraries(mlx PRIVATE openblas)
-    target_include_directories(
-      mlx PRIVATE "${openblas_SOURCE_DIR}/lapack-netlib/LAPACKE/include"
-                  "${CMAKE_BINARY_DIR}/generated" "${CMAKE_BINARY_DIR}")
-  else()
     if(${CMAKE_HOST_APPLE})
       # The blas shipped in macOS SDK is not supported, search homebrew for
       # openblas instead.
@@ -196,7 +146,7 @@ if(MLX_BUILD_CPU)
     message(STATUS "Lapack lib " ${LAPACK_LIBRARIES})
     message(STATUS "Lapack include " ${LAPACK_INCLUDE_DIRS})
     target_include_directories(mlx PRIVATE ${LAPACK_INCLUDE_DIRS})
-    target_link_libraries(mlx PRIVATE ${LAPACK_LIBRARIES})
+    target_link_libraries(mlx PUBLIC ${LAPACK_LIBRARIES})
     # List blas after lapack otherwise we may accidentally incldue an old
     # version of lapack.h from the include dirs of blas.
     find_package(BLAS REQUIRED)
@@ -209,19 +159,29 @@ if(MLX_BUILD_CPU)
     message(STATUS "Blas lib " ${BLAS_LIBRARIES})
     message(STATUS "Blas include " ${BLAS_INCLUDE_DIRS})
     target_include_directories(mlx PRIVATE ${BLAS_INCLUDE_DIRS})
-    target_link_libraries(mlx PRIVATE ${BLAS_LIBRARIES})
+    target_link_libraries(mlx PUBLIC ${BLAS_LIBRARIES})
   endif()
 else()
   set(MLX_BUILD_ACCELERATE OFF)
 endif()
 
-message(STATUS "Downloading json")
-FetchContent_Declare(
-  json
-  URL https://github.com/nlohmann/json/releases/download/v3.11.3/json.tar.xz)
-FetchContent_MakeAvailable(json)
-target_include_directories(
-  mlx PRIVATE $<BUILD_INTERFACE:${json_SOURCE_DIR}/single_include/nlohmann>)
+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()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
 
@@ -229,9 +189,6 @@ target_include_directories(
   mlx PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
              $<INSTALL_INTERFACE:include>)
 
-# Do not add mlx_EXPORTS define for shared library.
-set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")
-
 FetchContent_Declare(
   fmt
   GIT_REPOSITORY https://github.com/fmtlib/fmt.git
@@ -249,7 +206,8 @@ if(MLX_BUILD_PYTHON_BINDINGS)
   execute_process(
     COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
     OUTPUT_STRIP_TRAILING_WHITESPACE
-    OUTPUT_VARIABLE nanobind_ROOT)
+    OUTPUT_VARIABLE NB_DIR)
+  list(APPEND CMAKE_PREFIX_PATH "${NB_DIR}")
   find_package(nanobind CONFIG REQUIRED)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/python/src)
 endif()

CONTRIBUTING.md

@@ -17,13 +17,13 @@ possible.
  
    You can also run the formatters manually as follows:
  
-   ```shell
-   clang-format -i file.cpp
-   ```
+     ```
+     clang-format -i file.cpp
+     ```
  
-   ```shell
-   black file.py
-   ```
+     ```
+     black file.py
+     ```
  
    or run `pre-commit run --all-files` to check all files in the repo.
 

MANIFEST.in

@@ -1,6 +1,4 @@
 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

benchmarks/cpp/autograd.cpp

@@ -5,35 +5,35 @@
 #include "mlx/mlx.h"
 #include "time_utils.h"
 
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 void time_value_and_grad() {
-  auto x = mx::ones({200, 1000});
-  mx::eval(x);
-  auto fn = [](mx::array x) {
+  auto x = ones({200, 1000});
+  eval(x);
+  auto fn = [](array x) {
     for (int i = 0; i < 20; ++i) {
-      x = mx::log(mx::exp(x));
+      x = log(exp(x));
     }
-    return mx::sum(x);
+    return sum(x);
   };
 
-  auto grad_fn = mx::grad(fn);
+  auto grad_fn = grad(fn);
   auto independent_value_and_grad = [&]() {
     auto value = fn(x);
     auto dfdx = grad_fn(x);
-    return std::vector<mx::array>{value, dfdx};
+    return std::vector<array>{value, dfdx};
   };
   TIME(independent_value_and_grad);
 
-  auto value_and_grad_fn = mx::value_and_grad(fn);
+  auto value_and_grad_fn = value_and_grad(fn);
   auto combined_value_and_grad = [&]() {
     auto [value, dfdx] = value_and_grad_fn(x);
-    return std::vector<mx::array>{value, dfdx};
+    return std::vector<array>{value, dfdx};
   };
   TIME(combined_value_and_grad);
 }
 
 int main() {
-  std::cout << "Benchmarks for " << mx::default_device() << std::endl;
+  std::cout << "Benchmarks for " << default_device() << std::endl;
   time_value_and_grad();
 }

benchmarks/cpp/compare_devices.cpp

@@ -4,21 +4,21 @@
 #include "mlx/mlx.h"
 #include "time_utils.h"
 
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 void time_add_op() {
   std::vector<int> sizes(1, 1);
   for (int i = 0; i < 9; ++i) {
     sizes.push_back(10 * sizes.back());
   }
-  set_default_device(mx::Device::cpu);
+  set_default_device(Device::cpu);
   for (auto size : sizes) {
-    auto a = mx::random::uniform({size});
-    auto b = mx::random::uniform({size});
-    mx::eval(a, b);
+    auto a = random::uniform({size});
+    auto b = random::uniform({size});
+    eval(a, b);
     std::cout << "Size " << size << std::endl;
-    TIMEM("cpu", mx::add, a, b, mx::Device::cpu);
-    TIMEM("gpu", mx::add, a, b, mx::Device::gpu);
+    TIMEM("cpu", add, a, b, Device::cpu);
+    TIMEM("gpu", add, a, b, Device::gpu);
   }
 }
 

benchmarks/cpp/irregular_strides.cpp

@@ -1,111 +1,110 @@
 // Copyright © 2023 Apple Inc.
 
-#include <cstring>
 #include <iostream>
 #include <sstream>
 
 #include "mlx/mlx.h"
 #include "time_utils.h"
 
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 void time_irregular_binary_ops_1D() {
-  auto device = mx::default_device();
+  auto device = default_device();
   int size = 1000000;
   int step = 2;
-  auto a = mx::random::uniform({size});
-  auto b = mx::random::uniform({size});
-  mx::eval(a, b);
+  auto a = random::uniform({size});
+  auto b = random::uniform({size});
+  eval(a, b);
   a = slice(a, {0}, {size}, {step});
   b = slice(b, {0}, {size}, {step});
-  TIMEM("1D strided", mx::add, a, b, device);
+  TIMEM("1D strided", add, a, b, device);
 }
 
 void time_irregular_binary_ops_2D() {
-  auto device = mx::default_device();
+  auto device = default_device();
   int size = 2048;
-  auto a = mx::random::uniform({size, size});
-  auto b = mx::random::uniform({size, size});
-  mx::eval(a, b);
-  TIMEM("2D regular", mx::add, a, b, device);
+  auto a = random::uniform({size, size});
+  auto b = random::uniform({size, size});
+  eval(a, b);
+  TIMEM("2D regular", add, a, b, device);
 
-  b = mx::transpose(b);
-  mx::eval(b);
-  TIMEM("2D mx::transpose", mx::add, a, b, device);
+  b = transpose(b);
+  eval(b);
+  TIMEM("2D transpose", add, a, b, device);
 
-  b = mx::random::uniform({size});
-  mx::eval(b);
-  TIMEM("2D broadcast dim 0", mx::add, a, b, device);
+  b = random::uniform({size});
+  eval(b);
+  TIMEM("2D broadcast dim 0", add, a, b, device);
 
-  b = mx::reshape(b, {size, 1});
-  mx::eval(b);
-  TIMEM("2D broadcast dim 1", mx::add, a, b, device);
+  b = reshape(b, {size, 1});
+  eval(b);
+  TIMEM("2D broadcast dim 1", add, a, b, device);
 }
 
 void time_irregular_binary_ops_3D() {
-  auto device = mx::default_device();
+  auto device = default_device();
   int d0 = 32;
   int d1 = 512;
   int d2 = 512;
-  auto a = mx::random::uniform({d0, d1, d2});
-  auto b = mx::random::uniform({d0, d1, d2});
-  TIMEM("3D regular", mx::add, a, b, device);
+  auto a = random::uniform({d0, d1, d2});
+  auto b = random::uniform({d0, d1, d2});
+  TIMEM("3D regular", add, a, b, device);
 
-  b = mx::transpose(b, {0, 2, 1});
-  TIMEM("3D mx::transpose", mx::add, a, b, device);
+  b = transpose(b, {0, 2, 1});
+  TIMEM("3D transpose", add, a, b, device);
 
-  b = mx::random::uniform({d1, d2});
-  TIMEM("3D broadcast dim 0", mx::add, a, b, device);
+  b = random::uniform({d1, d2});
+  TIMEM("3D broadcast dim 0", add, a, b, device);
 
-  b = mx::random::uniform({d0, 1, d2});
-  TIMEM("3D broadcast dim 1", mx::add, a, b, device);
+  b = random::uniform({d0, 1, d2});
+  TIMEM("3D broadcast dim 1", add, a, b, device);
 
-  b = mx::random::uniform({d0, d1, 1});
-  TIMEM("3D broadcast dim 2", mx::add, a, b, device);
+  b = random::uniform({d0, d1, 1});
+  TIMEM("3D broadcast dim 2", add, a, b, device);
 
-  b = mx::random::uniform({d2});
-  TIMEM("3D broadcast dims 0, 1", mx::add, a, b, device);
+  b = random::uniform({d2});
+  TIMEM("3D broadcast dims 0, 1", add, a, b, device);
 
-  b = mx::random::uniform({d1, 1});
-  TIMEM("3D broadcast dims 0, 2", mx::add, a, b, device);
+  b = random::uniform({d1, 1});
+  TIMEM("3D broadcast dims 0, 2", add, a, b, device);
 
-  b = mx::random::uniform({d0, 1, 1});
-  TIMEM("3D broadcast dims 1, 2", mx::add, a, b, device);
+  b = random::uniform({d0, 1, 1});
+  TIMEM("3D broadcast dims 1, 2", add, a, b, device);
 }
 
 void time_irregular_binary_ops_4D() {
-  auto device = mx::default_device();
+  auto device = default_device();
   std::vector<int> shape = {8, 8, 512, 512};
-  auto a = mx::random::uniform(shape);
-  auto b = mx::random::uniform(shape);
+  auto a = random::uniform(shape);
+  auto b = random::uniform(shape);
 
-  TIMEM("4D regular", mx::add, a, b, device);
+  TIMEM("4D regular", add, a, b, device);
 
-  b = mx::transpose(b, {0, 1, 3, 2});
-  TIMEM("4D mx::transpose", mx::add, a, b, device);
+  b = transpose(b, {0, 1, 3, 2});
+  TIMEM("4D transpose", add, a, b, device);
 
   std::string om = "4D broadcast dims ";
   for (int i = 0; i < shape.size(); ++i) {
     shape[i] = 1;
-    b = mx::random::uniform(shape);
+    b = random::uniform(shape);
     std::ostringstream msg;
     msg << om << i;
-    TIMEM(msg.str(), mx::add, a, b, device);
+    TIMEM(msg.str(), add, a, b, device);
 
     for (int j = i + 1; j < shape.size(); ++j) {
       shape[j] = 1;
       std::ostringstream msg;
       msg << om << i << ", " << j;
-      b = mx::random::uniform(shape);
-      TIMEM(msg.str(), mx::add, a, b, device);
+      b = random::uniform(shape);
+      TIMEM(msg.str(), add, a, b, device);
       shape[j] = a.shape(j);
 
       for (int k = j + 1; k < shape.size(); ++k) {
         shape[k] = 1;
         std::ostringstream msg;
         msg << om << i << ", " << j << ", " << k;
-        b = mx::random::uniform(shape);
-        TIMEM(msg.str(), mx::add, a, b, device);
+        b = random::uniform(shape);
+        TIMEM(msg.str(), add, a, b, device);
         shape[k] = a.shape(k);
       }
     }
@@ -114,83 +113,83 @@ void time_irregular_binary_ops_4D() {
 }
 
 void time_irregular_reshape() {
-  auto device = mx::default_device();
+  auto device = default_device();
   std::vector<int> shape;
-  auto reshape_fn = [&shape, device](const mx::array& a) {
-    return mx::reshape(a, shape, device);
+  auto reshape_fn = [&shape, device](const array& a) {
+    return reshape(a, shape, device);
   };
 
   int size = 64;
   int d = 2 * size;
 
-  auto a = mx::random::uniform({d, d, d});
+  auto a = random::uniform({d, d, d});
 
   shape = {8 * size, size, size};
   TIMEM("3D contiguous", reshape_fn, a);
 
-  a = mx::transpose(a);
+  a = transpose(a);
   shape = {8 * size, size, size};
-  TIMEM("3D mx::transpose", reshape_fn, a);
+  TIMEM("3D transpose", reshape_fn, a);
 
-  a = mx::transpose(a, {1, 2, 0});
+  a = transpose(a, {1, 2, 0});
   shape = {8 * size, size, size};
-  TIMEM("3D mx::transpose dims 1 2", reshape_fn, a);
+  TIMEM("3D transpose dims 1 2", reshape_fn, a);
 
-  a = mx::broadcast_to(mx::random::uniform({d, d}), {d, d, d});
+  a = broadcast_to(random::uniform({d, d}), {d, d, d});
   TIMEM("3D broadcast dim 0", reshape_fn, a);
 
-  a = mx::broadcast_to(mx::random::uniform({d, 1, d}), {d, d, d});
+  a = broadcast_to(random::uniform({d, 1, d}), {d, d, d});
   TIMEM("3D broadcast dim 1", reshape_fn, a);
 
-  a = mx::broadcast_to(mx::random::uniform({d, d, 1}), {d, d, d});
+  a = broadcast_to(random::uniform({d, d, 1}), {d, d, d});
   TIMEM("3D broadcast dim 2", reshape_fn, a);
 
-  a = mx::broadcast_to(mx::random::uniform({d}), {d, d, d});
+  a = broadcast_to(random::uniform({d}), {d, d, d});
   TIMEM("3D broadcast dims 0, 1", reshape_fn, a);
 
-  a = mx::broadcast_to(mx::random::uniform({d, 1}), {d, d, d});
+  a = broadcast_to(random::uniform({d, 1}), {d, d, d});
   TIMEM("3D broadcast dims 0, 2", reshape_fn, a);
 
-  a = mx::broadcast_to(mx::random::uniform({d, 1, 1}), {d, d, d});
+  a = broadcast_to(random::uniform({d, 1, 1}), {d, d, d});
   TIMEM("3D broadcast dims 1, 2", reshape_fn, a);
 
-  a = mx::broadcast_to(mx::random::uniform({1, 1, 1}), {d, d, d});
+  a = broadcast_to(random::uniform({1, 1, 1}), {d, d, d});
   TIMEM("3D broadcast dims 1, 2, 3", reshape_fn, a);
 }
 
 void time_irregular_astype_1D() {
-  auto device = mx::default_device();
+  auto device = default_device();
   int size = 1000000;
   int step = 2;
-  auto a = mx::random::uniform({size});
+  auto a = random::uniform({size});
   a = slice(a, {0}, {size}, {step});
-  TIMEM("1D strided", mx::astype, a, mx::int32, device);
+  TIMEM("1D strided", astype, a, int32, device);
 }
 
 void time_irregular_astype_2D() {
-  auto device = mx::default_device();
+  auto device = default_device();
   int size = 2048;
   std::vector<int> shape = {size, size};
 
-  auto a = mx::random::uniform(shape);
-  TIMEM("2D regular", mx::astype, a, mx::int32, device);
+  auto a = random::uniform(shape);
+  TIMEM("2D regular", astype, a, int32, device);
 
-  a = mx::transpose(a);
-  TIMEM("2D mx::transpose", mx::astype, a, mx::int32, device);
+  a = transpose(a);
+  TIMEM("2D transpose", astype, a, int32, device);
 
-  a = mx::broadcast_to(mx::random::uniform({size}), shape);
-  TIMEM("2D broadcast dim 0", mx::astype, a, mx::int32, device);
+  a = broadcast_to(random::uniform({size}), shape);
+  TIMEM("2D broadcast dim 0", astype, a, int32, device);
 
-  a = mx::broadcast_to(mx::random::uniform({size, 1}), shape);
-  TIMEM("2D broadcast dim 1", mx::astype, a, mx::int32, device);
+  a = broadcast_to(random::uniform({size, 1}), shape);
+  TIMEM("2D broadcast dim 1", astype, a, int32, device);
 }
 
 int main(int argc, char** argv) {
   if (argc > 1) {
     bool use_gpu = !strcmp(argv[1], "gpu");
-    set_default_device(use_gpu ? mx::Device::gpu : mx::Device::cpu);
+    set_default_device(use_gpu ? Device::gpu : Device::cpu);
   }
-  std::cout << "Benchmarks for " << mx::default_device() << std::endl;
+  std::cout << "Benchmarks for " << default_device() << std::endl;
   time_irregular_binary_ops_1D();
   time_irregular_binary_ops_2D();
   time_irregular_binary_ops_3D();

benchmarks/cpp/single_ops.cpp

@@ -3,20 +3,20 @@
 #include "mlx/mlx.h"
 #include "time_utils.h"
 
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 void time_creation_ops() {
   int M = 2000;
   int N = 500;
   auto shape = {M, N};
-  auto full_fp32 = [&]() { return mx::full(shape, 3.3f); };
+  auto full_fp32 = [&]() { return full(shape, 3.3f); };
   TIME(full_fp32);
-  auto zeros_fp32 = [&]() { return mx::zeros(shape, mx::float32); };
+  auto zeros_fp32 = [&]() { return zeros(shape, float32); };
   TIME(zeros_fp32);
-  auto ones_fp32 = [&]() { return mx::ones(shape, mx::float32); };
+  auto ones_fp32 = [&]() { return ones(shape, float32); };
   TIME(ones_fp32);
 
-  auto arange_fp32 = [&]() { return mx::arange(0.0, 10.0, 1e-4); };
+  auto arange_fp32 = [&]() { return arange(0.0, 10.0, 1e-4); };
   TIME(arange_fp32);
 }
 
@@ -24,212 +24,194 @@ void time_type_conversions() {
   int M = 2000;
   int N = 500;
   auto shape = {M, N};
-  auto device = mx::default_device();
+  auto device = default_device();
 
-  auto a = mx::zeros(shape, mx::float32);
-  mx::eval(a);
-  TIMEM("mx::float32 to mx::int32", mx::astype, a, mx::int32, device);
-  TIMEM("mx::float32 to mx::uint32", mx::astype, a, mx::uint32, device);
+  auto a = zeros(shape, float32);
+  eval(a);
+  TIMEM("float32 to int32", astype, a, int32, device);
+  TIMEM("float32 to uint32", astype, a, uint32, device);
 
-  a = mx::zeros(shape, mx::int32);
-  mx::eval(a);
-  TIMEM("mx::int32 to mx::float32", mx::astype, a, mx::float32, device);
+  a = zeros(shape, int32);
+  eval(a);
+  TIMEM("int32 to float32", astype, a, float32, device);
 
-  a = mx::zeros(shape, mx::bool_);
-  mx::eval(a);
-  TIMEM("bool to mx::float32", mx::astype, a, mx::float32, device);
-  TIMEM("bool to mx::int32", mx::astype, a, mx::int32, device);
-  TIMEM("bool to mx::uint32", mx::astype, a, mx::uint32, device);
+  a = zeros(shape, bool_);
+  eval(a);
+  TIMEM("bool to float32", astype, a, float32, device);
+  TIMEM("bool to int32", astype, a, int32, device);
+  TIMEM("bool to uint32", astype, a, uint32, device);
 }
 
 void time_random_generation() {
   int M = 2000;
   int N = 500;
 
-  auto uniform = [&]() { return mx::random::uniform({M, N}, mx::float32); };
+  auto uniform = [&]() { return random::uniform({M, N}, float32); };
   TIME(uniform);
-  auto normal = [&]() { return mx::random::normal({M, N}, mx::float32); };
+  auto normal = [&]() { return random::normal({M, N}, float32); };
   TIME(normal);
 }
 
 void time_unary_ops() {
   int M = 2000;
   int N = 500;
-  auto device = mx::default_device();
+  auto device = default_device();
 
-  auto a = mx::random::normal({M, N});
-  mx::eval(a);
+  auto a = random::normal({M, N});
+  eval(a);
   TIME(mlx::core::abs, a, device);
-  TIME(mx::negative, a, device);
-  TIME(mx::sign, a, device);
-  TIME(mx::square, a, device);
+  TIME(negative, a, device);
+  TIME(sign, a, device);
+  TIME(square, a, device);
   TIME(mlx::core::sqrt, a, device);
-  TIME(mx::rsqrt, a, device);
+  TIME(rsqrt, a, device);
   TIME(mlx::core::exp, a, device);
 
-  a = mx::random::uniform({M, N});
+  a = random::uniform({M, N});
   TIME(mlx::core::log, a, device);
 }
 
 void time_binary_ops() {
   int M = 1000, N = 100, K = 10;
-  auto condition = mx::random::randint(0, 2, {M, N, K});
-  auto a = mx::random::uniform({M, N, K});
-  auto b = mx::random::uniform({M, N, K});
-  auto device = mx::default_device();
-  mx::eval(a, b);
+  auto condition = random::randint(0, 2, {M, N, K});
+  auto a = random::uniform({M, N, K});
+  auto b = random::uniform({M, N, K});
+  auto device = default_device();
+  eval(a, b);
 
-  TIME(mx::add, a, b, device);
-  TIME(mx::subtract, a, b, device);
-  TIME(mx::multiply, a, b, device);
-  TIME(mx::divide, a, b, device);
-  TIME(mx::maximum, a, b, device);
-  TIME(mx::minimum, a, b, device);
-  TIME(mx::where, condition, a, b, device);
+  TIME(add, a, b, device);
+  TIME(subtract, a, b, device);
+  TIME(multiply, a, b, device);
+  TIME(divide, a, b, device);
+  TIME(maximum, a, b, device);
+  TIME(minimum, a, b, device);
+  TIME(where, condition, a, b, device);
 
-  condition = mx::array({true});
-  b = mx::random::uniform({1});
-  mx::eval(b);
-  TIMEM("scalar", mx::add, a, b, device);
-  TIMEM("vector-scalar", mx::subtract, a, b, device);
-  TIMEM("scalar-vector", mx::subtract, b, a, device);
-  TIMEM("scalar", mx::multiply, a, b, device);
-  TIMEM("vector-scalar", mx::divide, a, b, device);
-  TIMEM("scalar-vector", mx::divide, b, a, device);
-  TIMEM("scalar-vector", mx::where, condition, a, b, device);
+  condition = array({true});
+  b = random::uniform({1});
+  eval(b);
+  TIMEM("scalar", add, a, b, device);
+  TIMEM("vector-scalar", subtract, a, b, device);
+  TIMEM("scalar-vector", subtract, b, a, device);
+  TIMEM("scalar", multiply, a, b, device);
+  TIMEM("vector-scalar", divide, a, b, device);
+  TIMEM("scalar-vector", divide, b, a, device);
+  TIMEM("scalar-vector", where, condition, a, b, device);
 
-  condition = mx::broadcast_to(mx::array({true}), {1000, 100});
-  a = mx::broadcast_to(mx::random::uniform({1}), {1000, 100});
-  b = mx::broadcast_to(mx::random::uniform({1}), {1000, 100});
-  mx::eval(a, b);
-  TIMEM("scalar-scalar broadcast", mx::add, a, b, device);
-  TIMEM("scalar-scalar broadcast", mx::subtract, a, b, device);
-  TIMEM("scalar-scalar broadcast", mx::multiply, a, b, device);
-  TIMEM("scalar-scalar broadcast", mx::divide, a, b, device);
-  TIMEM("scalar-scalar broadcast", mx::where, condition, a, b, device);
+  condition = broadcast_to(array({true}), {1000, 100});
+  a = broadcast_to(random::uniform({1}), {1000, 100});
+  b = broadcast_to(random::uniform({1}), {1000, 100});
+  eval(a, b);
+  TIMEM("scalar-scalar broadcast", add, a, b, device);
+  TIMEM("scalar-scalar broadcast", subtract, a, b, device);
+  TIMEM("scalar-scalar broadcast", multiply, a, b, device);
+  TIMEM("scalar-scalar broadcast", divide, a, b, device);
+  TIMEM("scalar-scalar broadcast", where, condition, a, b, device);
 }
 
 void time_strided_ops() {
   int M = 50, N = 50, O = 50, P = 50;
-  auto a = mx::random::uniform({M, N, O, P});
-  auto b = mx::random::uniform({M, N, O, P});
-  auto device = mx::default_device();
-  mx::eval(a, b);
-  TIMEM("non-strided", mx::add, a, b, device);
-  a = mx::transpose(a, {1, 0, 2, 3});
-  b = mx::transpose(b, {3, 2, 0, 1});
-  mx::eval(a, b);
-  TIMEM("strided", mx::add, a, b, device);
+  auto a = random::uniform({M, N, O, P});
+  auto b = random::uniform({M, N, O, P});
+  auto device = default_device();
+  eval(a, b);
+  TIMEM("non-strided", add, a, b, device);
+  a = transpose(a, {1, 0, 2, 3});
+  b = transpose(b, {3, 2, 0, 1});
+  eval(a, b);
+  TIMEM("strided", add, a, b, device);
 }
 
 void time_comparisons() {
   int M = 1000, N = 100, K = 10;
-  auto a = mx::random::uniform({M, N, K});
-  auto b = mx::random::uniform({M, N, K});
-  auto device = mx::default_device();
-  mx::eval(a, b);
-  TIME(mx::equal, a, b, device);
-  TIME(mx::greater, a, b, device);
-  TIME(mx::greater_equal, a, b, device);
-  TIME(mx::less, a, b, device);
-  TIME(mx::less_equal, a, b, device);
+  auto a = random::uniform({M, N, K});
+  auto b = random::uniform({M, N, K});
+  auto device = default_device();
+  eval(a, b);
+  TIME(equal, a, b, device);
+  TIME(greater, a, b, device);
+  TIME(greater_equal, a, b, device);
+  TIME(less, a, b, device);
+  TIME(less_equal, a, b, device);
 }
 
 void time_matvec() {
   int M = 2000, N = 200;
-  auto a = mx::random::uniform({M, N});
-  auto b = mx::random::uniform({N});
-  auto c = mx::random::uniform({M});
-  mx::eval(a, b, c);
-  auto matvec = [&]() { return mx::matmul(a, b); };
+  auto a = random::uniform({M, N});
+  auto b = random::uniform({N});
+  auto c = random::uniform({M});
+  eval(a, b, c);
+  auto matvec = [&]() { return matmul(a, b); };
   TIME(matvec);
 
-  auto matvec_transpose = [&]() { return mx::matmul(mx::transpose(a), c); };
+  auto matvec_transpose = [&]() { return matmul(transpose(a), c); };
   TIME(matvec_transpose);
 }
 
 void time_matmul() {
   int M = 1000, N = 1000, K = 1000;
-  auto a = mx::random::uniform({M, K});
-  auto b = mx::random::uniform({K, N});
-  auto device = mx::default_device();
-  mx::eval(a, b);
-  TIME(mx::matmul, a, b, device);
+  auto a = random::uniform({M, K});
+  auto b = random::uniform({K, N});
+  auto device = default_device();
+  eval(a, b);
+  TIME(matmul, a, b, device);
 
-  auto transpose_matmul = [&]() { return mx::matmul(mx::transpose(a), b); };
+  auto transpose_matmul = [&]() { return matmul(transpose(a), b); };
   TIME(transpose_matmul);
 }
 
 void time_reductions() {
-  auto a = mx::random::normal({10000, 1000});
-  mx::eval(a);
-  auto sum_all = [&a]() { return mx::sum(a, false); };
+  auto a = random::normal({10000, 1000});
+  eval(a);
+  auto sum_all = [&a]() { return sum(a, false); };
   TIME(sum_all);
 
-  auto sum_along_0 = [&a]() { return mx::sum(a, 0, false); };
+  auto sum_along_0 = [&a]() { return sum(a, 0, false); };
   TIME(sum_along_0);
 
-  auto sum_along_1 = [&a]() { return mx::sum(a, 1, false); };
+  auto sum_along_1 = [&a]() { return sum(a, 1, false); };
   TIME(sum_along_1);
 
-  auto prod_all = [&a]() { return mx::prod(a, false); };
+  auto prod_all = [&a]() { return prod(a, false); };
   TIME(prod_all);
 
-  auto all_true = [&a]() { return mx::all(a, false); };
+  auto all_true = [&a]() { return all(a, false); };
   TIME(all_true);
 
-  auto all_along_0 = [&a]() { return mx::all(a, 0, false); };
+  auto all_along_0 = [&a]() { return all(a, 0, false); };
   TIME(all_along_0);
 
-  auto all_along_1 = [&a]() { return mx::all(a, 1, false); };
+  auto all_along_1 = [&a]() { return all(a, 1, false); };
   TIME(all_along_1);
 
-  auto any_true = [&a]() { return mx::any(a, false); };
+  auto any_true = [&a]() { return any(a, false); };
   TIME(any_true);
 
-  auto argmin_along_0 = [&a]() { return mx::argmin(a, 0, false); };
+  auto argmin_along_0 = [&a]() { return argmin(a, 0, false); };
   TIME(argmin_along_0);
 
-  auto argmin_along_1 = [&a]() { return mx::argmin(a, 1, false); };
+  auto argmin_along_1 = [&a]() { return 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() {
-  auto a = mx::random::normal({1000, 768});
-  mx::eval(a);
-  auto indices = mx::random::randint(0, 1000, {256});
-  mx::eval(indices);
+  auto a = random::normal({1000, 768});
+  eval(a);
+  auto indices = random::randint(0, 1000, {256});
+  eval(indices);
 
-  auto embedding_lookup = [&a, &indices]() { return mx::take(a, indices, 0); };
+  auto embedding_lookup = [&a, &indices]() { return take(a, indices, 0); };
   TIME(embedding_lookup);
 
-  indices = mx::random::randint(0, 768 * 1000, {256 * 768});
-  mx::eval(indices);
+  indices = random::randint(0, 768 * 1000, {256 * 768});
+  eval(indices);
 
-  auto single_element_lookup = [&a, &indices]() {
-    return mx::take(a, indices);
-  };
+  auto single_element_lookup = [&a, &indices]() { return take(a, indices); };
   TIME(single_element_lookup);
 
-  indices = mx::random::randint(0, 1000, {256});
-  auto updates = mx::random::normal({256, 1, 768});
-  mx::eval(indices, updates);
+  indices = random::randint(0, 1000, {256});
+  auto updates = random::normal({256, 1, 768});
+  eval(indices, updates);
 
   auto embedding_update = [&a, &indices, &updates]() {
     return scatter(a, indices, updates, 0);
@@ -241,10 +223,10 @@ void time_gather_scatter() {
   };
   TIME(embedding_add);
 
-  a = mx::reshape(a, {-1});
-  indices = mx::random::randint(0, 768 * 1000, {768 * 256});
-  updates = mx::random::normal({256 * 768, 1});
-  mx::eval(a, indices, updates);
+  a = reshape(a, {-1});
+  indices = random::randint(0, 768 * 1000, {768 * 256});
+  updates = random::normal({256 * 768, 1});
+  eval(a, indices, updates);
 
   auto single_element_update = [&a, &indices, &updates]() {
     return scatter(a, indices, updates, 0);
@@ -258,21 +240,21 @@ void time_gather_scatter() {
 }
 
 void time_divmod() {
-  auto a = mx::random::normal({1000});
-  auto b = mx::random::normal({1000});
-  mx::eval({a, b});
+  auto a = random::normal({1000});
+  auto b = random::normal({1000});
+  eval({a, b});
 
-  auto divmod_fused = [&a, &b]() { return mx::divmod(a, b); };
+  auto divmod_fused = [&a, &b]() { return divmod(a, b); };
   TIME(divmod_fused);
 
   auto divmod_separate = [&a, &b]() {
-    return std::vector<mx::array>{mx::floor_divide(a, b), mx::remainder(a, b)};
+    return std::vector<array>{floor_divide(a, b), remainder(a, b)};
   };
   TIME(divmod_separate);
 }
 
 int main() {
-  std::cout << "Benchmarks for " << mx::default_device() << std::endl;
+  std::cout << "Benchmarks for " << default_device() << std::endl;
   time_creation_ops();
   time_type_conversions();
   time_unary_ops();

benchmarks/python/comparative/bench_torch.py

@@ -5,7 +5,6 @@ import os
 import time
 
 import torch
-import torch.cuda
 import torch.mps
 
 
@@ -45,10 +44,8 @@ def bench(f, *args):
 
 
 def sync_if_needed(x):
-    if x.device == torch.device("mps"):
+    if x.device != torch.device("cpu"):
         torch.mps.synchronize()
-    elif x.device == torch.device("cuda"):
-        torch.cuda.synchronize()
 
 
 @torch.no_grad()
@@ -102,14 +99,6 @@ 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 = []
@@ -351,11 +340,7 @@ if __name__ == "__main__":
         args.axis.pop(0)
 
     torch.set_num_threads(1)
-    device = "mps"
-    if torch.cuda.is_available():
-        device = "cuda"
-    if args.cpu:
-        device = "cpu"
+    device = "cpu" if args.cpu else "mps"
 
     types = args.dtype
     if not types:
@@ -475,8 +460,5 @@ 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

@@ -1,107 +0,0 @@
-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_bench.py

@@ -1,6 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.
 
 import argparse
+from time import time
 
 import mlx.core as mx
 import torch

benchmarks/python/gather_mm_bench.py

@@ -1,74 +0,0 @@
-# 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

@@ -1,84 +0,0 @@
-# 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

@@ -1,7 +1,5 @@
 # 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
@@ -12,71 +10,32 @@ def layer_norm(x, w, b, eps):
     x = x.astype(mx.float32)
     mu = mx.mean(x, -1, keepdims=True)
     v = mx.var(x, -1, keepdims=True)
-    y = (x - mu) * mx.rsqrt(v + eps)
-    if w is not None:
-        y = y * w
-    if b is not None:
-        y = y + b
-    return y
+    return (x - mu) * mx.rsqrt(v + eps) * w + b
 
 
-def time_layer_norm(N, dt):
-    L = 1024
+def time_layer_norm():
     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, 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)
+    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)
     mx.eval(x, w, b, y)
 
-    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):
+    def layer_norm_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_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, 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_grad_x_loop(g, x):
-        gx = x
-        for _ in range(32):
-            gx = g(gx, y)
-        return gx
-
-    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)
+    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)
 
 
 if __name__ == "__main__":
-    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)
+    time_layer_norm()

benchmarks/python/rms_norm_bench.py

@@ -9,10 +9,7 @@ def rms_norm(x, w, eps):
     ot = x.dtype
     x = x.astype(mx.float32)
     n = mx.rsqrt(x.square().mean(-1, keepdims=True) + eps)
-    y = (x * n).astype(ot)
-    if w is not None:
-        y = y * w
-    return y
+    return (x * n).astype(ot) * w
 
 
 def time_rms_norm():
@@ -37,27 +34,6 @@ def time_rms_norm():
     time_fn(rms_norm_loop, mx.compile(g1), x, w)
     time_fn(rms_norm_loop, mx.compile(g2), x, w)
 
-    f1 = lambda x, y: (rms_norm(x, None, 1e-5) * y).sum()
-    f2 = lambda x, y: (mx.fast.rms_norm(x, 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)
-    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
-    mx.eval(x, w, y)
-
-    def rms_norm_loop(g, x):
-        gx = x
-        for _ in range(32):
-            gx = g(gx, y)
-        return gx
-
-    time_fn(rms_norm_loop, g1, x)
-    time_fn(rms_norm_loop, g2, x)
-    time_fn(rms_norm_loop, mx.compile(g1), x)
-    time_fn(rms_norm_loop, mx.compile(g2), x)
-
 
 if __name__ == "__main__":
     time_rms_norm()

benchmarks/python/sdpa_bench.py

@@ -1,223 +1,62 @@
-# Copyright © 2024 Apple Inc.
-
 import argparse
 import math
-import os
-import subprocess
-import time
 
 import mlx.core as mx
-import numpy as np
+from time_utils import time_fn
 
-device_name = subprocess.check_output(["sysctl", "-n", "machdep.cpu.brand_string"])
-device_name = device_name.decode("utf-8").strip("\n")
-
-N_warmup = 5
-N_iter_bench = 40
-N_iter_func = 8
+MAX_SEQ = 300
+START_SEQ = 100
+SEQ_INCREMENT = 50
 
 
-def bench(f, *args):
-    for i in range(N_warmup):
-        f(*args)
+def time_self_attention_primitives():
+    mx.random.seed(3)
+    B = 2
+    H = 38
+    D = 64
+    for R in range(START_SEQ, MAX_SEQ, SEQ_INCREMENT):
+        q = mx.random.uniform(shape=(B, H, R, D))
+        k = mx.random.uniform(shape=(B, H, R, D))
+        v = mx.random.uniform(shape=(B, H, R, D))
+        scale = 1.0 / math.sqrt(float(D))
+        mx.eval(q, k, v)
 
-    s = time.perf_counter_ns()
-    for i in range(N_iter_bench):
-        f(*args)
-    e = time.perf_counter_ns()
-    return (e - s) * 1e-9
+        def sdpa_primitives(qs, ks, vs, alpha):
+            s = (alpha * qs) @ ks.transpose(0, 1, 3, 2)
+            p = mx.softmax(s.astype(mx.float32), axis=-1).astype(s.dtype)
+            o = p @ vs
+            return o
+
+        time_fn(sdpa_primitives, q, k, v, scale)
 
 
-def prepare_inputs(B, qL, kL, D, qH, kH, mask, transpose, dtype):
-    np_dtype = getattr(np, dtype)
+def time_self_attention_sdpa():
+    mx.random.seed(3)
+    B = 2
+    H = 38
+    D = 64
+    for R in range(START_SEQ, MAX_SEQ, SEQ_INCREMENT):
+        q = mx.random.uniform(shape=(B, H, R, D))
+        k = mx.random.uniform(shape=(B, H, R, D))
+        v = mx.random.uniform(shape=(B, H, R, D))
+        scale = 1.0 / math.sqrt(float(D))
+        mx.eval(q, k, v)
 
-    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)
+        def sdpa_fused(qs, ks, vs, alpha):
+            o = mx.fast.scaled_dot_product_attention(qs, ks, vs, scale=alpha)
+            return o
 
-    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_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]
-    n_kv_heads = k.shape[-3]
-    n_repeats = n_q_heads // n_kv_heads
-
-    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])
-        k = mx.expand_dims(k, 2)
-        v = mx.expand_dims(v, 2)
-
-    scores = q @ mx.swapaxes(k, -1, -2)
-
-    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:
-        out = mx.reshape(out, [B, n_q_heads, L, -1])
-
-    return out
-
-
-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:
-        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):
-        q_out = do_attention(f, q_out, k, v, scale, mask=mask, transpose=transpose)
-
-    mx.eval(q_out)
-    return q_out
-
-
-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
-    )
-
-    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
-    )
-
-    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
-    )
-
-    atol = 1e-5 if dtype == "float32" else 2e-4
-
-    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}, 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
-
-
-def get_gflop_count(B, M, N, K):
-    return float(2.0 * N_iter_bench * N_iter_func * B * M * N * K) / float(1024.0**3)
+        time_fn(sdpa_fused, q, k, v, scale)
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Run gemm benchmarks")
+    parser = argparse.ArgumentParser("MLX benchmarks.")
+    parser.add_argument("--gpu", action="store_true", help="Use the Metal back-end.")
+    args = parser.parse_args()
+    if args.gpu:
+        mx.set_default_device(mx.gpu)
+    else:
+        mx.set_default_device(mx.cpu)
 
-    dtypes = ("float16", "float32")[:1]
-    transposes = (False,)
-
-    # fmt: off
-    shapes_64 = (
-        # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
-          (  1,    32,    32,       64,   32,    32),
-          (  1,    64,    64,       64,   32,    32),
-          (  1,   128,   128,       64,   32,    32),
-          (  1,   256,   256,       64,   32,    32),
-          (  1,   512,   512,       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,     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,     8),
-          (  1,  2048,  2048,      128,   32,     8),
-          (  1,  4096,  4096,      128,   32,     8),
-    )
-    # fmt: on
-
-    shapes = shapes_64 + shapes_80 + shapes_128
-
-    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:
-                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}%"
-                    )
+    time_self_attention_sdpa()
+    time_self_attention_primitives()

benchmarks/python/sdpa_vector_bench.py

@@ -4,92 +4,46 @@ import math
 import mlx.core as mx
 from time_utils import time_fn
 
-L = 16384
+L = 1024
 H = 32
-H_k = H // 4
+H_k = 32 // 4
 D = 128
-V = 128
-dtype = mx.float16
-loops = 10
 
 
-def upproject(x, w):
-    if w is None:
-        return x
-    else:
-        return x @ w.T
+def attention(q, k, v):
+    B, Hq, L, D = q.shape
+    _, Hk, S, _ = k.shape
+    q = q.reshape(B, Hk, Hq // Hk, L, D)
+    k = k[:, :, None, :, :]
+    v = v[:, :, None, :, :]
+    s = q @ k.transpose(0, 1, 2, 4, 3)
+    p = mx.softmax(s.astype(mx.float32), axis=-1).astype(s.dtype)
+    o = p @ v
+    return o.reshape(B, Hq, L, D)
 
 
-def attention(q, k, v, mask=None, w=None):
-    def _sdpa(q, k, v):
-        B, Hq, L, D = q.shape
-        _, Hk, S, _ = k.shape
-        _, _, _, V = v.shape
-        q = q.reshape(B, Hk, Hq // Hk, L, D)
-        k = k[:, :, None, :, :]
-        v = v[:, :, None, :, :]
-        s = q @ k.transpose(0, 1, 2, 4, 3)
-        if mask is not None:
-            m = mx.broadcast_to(mask, (B, Hq, L, S)).reshape(B, Hk, Hq // Hk, L, S)
-            s = mx.where(m, s, mx.finfo(s.dtype).min)
-        p = mx.softmax(s.astype(mx.float32), axis=-1).astype(s.dtype)
-        o = p @ v
-        return o.reshape(B, Hq, L, V)
-
-    for i in range(loops):
-        q = _sdpa(q, k, v)
-        q = upproject(q, w)
-    return q
-
-
-def sdpa(q, k, v, mask=None, w=None):
-    for i in range(loops):
-        q = mx.fast.scaled_dot_product_attention(q, k, v, scale=1.0, mask=mask)
-        q = upproject(q, w)
-    return q
+def sdpa(q, k, v):
+    return mx.fast.scaled_dot_product_attention(q, k, v, scale=1.0)
 
 
 def time_self_attention_primitives():
     mx.random.seed(3)
-    q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
-    k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    v = mx.random.uniform(shape=(1, H_k, L, V)).astype(dtype)
-    w = mx.random.uniform(shape=(D, V)).astype(dtype) if V != D else None
-    mx.eval(q, k, v, w)
-    time_fn(attention, q, k, v, w=w)
+    q = mx.random.uniform(shape=(1, H, 1, D))
+    k = mx.random.uniform(shape=(1, H_k, L, D))
+    v = mx.random.uniform(shape=(1, H_k, L, D))
+    mx.eval(q, k, v)
+    time_fn(attention, q, k, v)
 
 
 def time_self_attention_sdpa():
     mx.random.seed(3)
-    q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
-    k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    v = mx.random.uniform(shape=(1, H_k, L, V)).astype(dtype)
-    w = mx.random.uniform(shape=(D, V)).astype(dtype) if V != D else None
-    mx.eval(q, k, v, w)
-    time_fn(sdpa, q, k, v, w=w)
-
-
-def time_self_attention_sdpa_with_mask():
-    mx.random.seed(3)
-    q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
-    k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
-    v = mx.random.uniform(shape=(1, H_k, L, V)).astype(dtype)
-    w = mx.random.uniform(shape=(D, V)).astype(dtype) if V != D else None
-    mask = mx.full((L,), True)
-    mask[L // 2 :] = False
-    mx.eval(q, k, v, mask, w)
-
-    def sdpa_mask(*args):
-        return sdpa(*args, mask=mask, w=w)
-
-    def attention_mask(*args):
-        return attention(*args, mask=mask, w=w)
-
-    time_fn(attention_mask, q, k, v)
-    time_fn(sdpa_mask, q, k, v)
+    q = mx.random.uniform(shape=(1, H, 1, D))
+    k = mx.random.uniform(shape=(1, H_k, L, D))
+    v = mx.random.uniform(shape=(1, H_k, L, D))
+    mx.eval(q, k, v)
+    time_fn(sdpa, q, k, v)
 
 
 if __name__ == "__main__":
     time_self_attention_sdpa()
     time_self_attention_primitives()
-    time_self_attention_sdpa_with_mask()

benchmarks/python/single_ops.py

@@ -51,20 +51,6 @@ 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)
@@ -122,8 +108,6 @@ if __name__ == "__main__":
 
     time_add()
     time_matmul()
-    time_min()
-    time_max()
     time_maximum()
     time_exp()
     time_negative()

benchmarks/python/synchronize_bench.py

@@ -1,55 +0,0 @@
-import time
-
-import mlx.core as mx
-
-rank = mx.distributed.init().rank()
-
-
-def timeit(fn, a):
-
-    # warmup
-    for _ in range(5):
-        mx.eval(fn(a))
-
-    its = 10
-    tic = time.perf_counter()
-    for _ in range(its):
-        mx.eval(fn(a))
-    toc = time.perf_counter()
-    ms = 1000 * (toc - tic) / its
-    return ms
-
-
-def all_reduce_benchmark():
-    a = mx.ones((5, 5), mx.int32)
-
-    its_per_eval = 100
-
-    def fn(x):
-        for _ in range(its_per_eval):
-            x = mx.distributed.all_sum(x)
-            x = x - 1
-        return x
-
-    ms = timeit(fn, a) / its_per_eval
-    if rank == 0:
-        print(f"All Reduce: time per iteration {ms:.6f} (ms)")
-
-
-def all_gather_benchmark():
-    a = mx.ones((5, 5), mx.int32)
-    its_per_eval = 100
-
-    def fn(x):
-        for _ in range(its_per_eval):
-            x = mx.distributed.all_gather(x)[0]
-        return x
-
-    ms = timeit(fn, a) / its_per_eval
-    if rank == 0:
-        print(f"All gather: time per iteration {ms:.6f} (ms)")
-
-
-if __name__ == "__main__":
-    all_reduce_benchmark()
-    all_gather_benchmark()

cmake/extension.cmake

@@ -1,7 +1,5 @@
 include(CMakeParseArguments)
 
-# clang format off
-#
 # ##############################################################################
 # Build metal library
 #
@@ -11,14 +9,11 @@ include(CMakeParseArguments)
 # Args: TARGET: Custom target to be added for the metal library TITLE: Name of
 # the .metallib OUTPUT_DIRECTORY: Where to place ${TITLE}.metallib SOURCES: List
 # of source files INCLUDE_DIRS: List of include dirs DEPS: List of dependency
-# files (like headers) DEBUG: Boolean, if true, enables debug compile options
-# for this specific library. If not provided, uses global MLX_METAL_DEBUG.
+# files (like headers)
 #
-# clang format on
-
 macro(mlx_build_metallib)
   # Parse args
-  set(oneValueArgs TARGET TITLE OUTPUT_DIRECTORY DEBUG)
+  set(oneValueArgs TARGET TITLE OUTPUT_DIRECTORY)
   set(multiValueArgs SOURCES INCLUDE_DIRS DEPS)
   cmake_parse_arguments(MTLLIB "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
@@ -26,11 +21,7 @@ macro(mlx_build_metallib)
   set(MTLLIB_BUILD_TARGET "${MTLLIB_OUTPUT_DIRECTORY}/${MTLLIB_TITLE}.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()
+  set(MTLLIB_COMPILE_OPTIONS -Wall -Wextra -fno-fast-math)
 
   # 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          = NO
+GENERATE_HTML          = YES
 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, Apple"
+copyright = "2023, MLX Contributors"
 author = "MLX Contributors"
 version = ".".join(mx.__version__.split(".")[:3])
 release = version

docs/src/dev/custom_metal_kernels.rst

@@ -8,26 +8,23 @@ 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
 
-  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,
-  )
-
   def exp_elementwise(a: mx.array):
+      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,
+      )
       outputs = kernel(
           inputs=[a],
           template=[("T", mx.float32)],
@@ -42,13 +39,8 @@ 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::
-   Only pass the body of the Metal kernel in ``source``. The function
-   signature is generated automatically.
+    We are only required to pass the body of the Metal kernel in ``source``.
 
 The full function signature will be generated using:
 
@@ -86,51 +78,44 @@ 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 :func:`ast.metal_kernel.__call__` will print the
-generated code for debugging purposes.
+Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
 
 Using Shape/Strides
 -------------------
 
-: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.
+``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.
 
-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.
+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.
 
-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)],
@@ -157,139 +142,137 @@ 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 :func:`custom_function` together with :func:`fast.metal_kernel`
+Now let's use ``mx.custom_function`` together with ``mx.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
 
-  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];
+    @mx.custom_function
+    def grid_sample(x, grid):
 
-      int w_stride = C;
-      int h_stride = W * w_stride;
-      int b_stride = H * h_stride;
+        assert x.ndim == 4, "`x` must be 4D."
+        assert grid.ndim == 4, "`grid` must be 4D."
 
-      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;
+        B, _, _, C = x.shape
+        _, gN, gM, D = grid.shape
+        out_shape = (B, gN, gM, C)
 
-      int ix_nw = floor(ix);
-      int iy_nw = floor(iy);
+        assert D == 2, "Last dim of `grid` must be size 2."
 
-      int ix_ne = ix_nw + 1;
-      int iy_ne = iy_nw;
+        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_sw = ix_nw;
-      int iy_sw = iy_nw + 1;
+            int w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
 
-      int ix_se = ix_nw + 1;
-      int iy_se = 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;
 
-      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_nw = floor(ix);
+            int iy_nw = floor(iy);
 
-      int batch_idx = elem / C / gH / gW * b_stride;
-      int channel_idx = elem % C;
-      int base_idx = batch_idx + channel_idx;
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
 
-      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_sw = ix_nw;
+            int iy_sw = 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;
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
 
-      out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
-  """
+            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);
 
-  kernel = mx.fast.metal_kernel(
-      name="grid_sample",
-      input_names=["x", "grid"],
-      output_names=["out"],
-      source=source,
-  )
+            int batch_idx = elem / C / gH / gW * b_stride;
+            int channel_idx = elem % C;
+            int base_idx = batch_idx + channel_idx;
 
-  @mx.custom_function
-  def grid_sample(x, grid):
+            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];
 
-      assert x.ndim == 4, "`x` must be 4D."
-      assert grid.ndim == 4, "`grid` must be 4D."
+            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;
 
-      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]
+            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]
 
 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:
 
@@ -298,11 +281,11 @@ On an M1 Max, we see a big performance improvement:
 Grid Sample VJP
 ---------------
 
-Since we decorated ``grid_sample`` with :func:`custom_function`, we can now
-define its custom vjp transform so MLX can differentiate it.
+Since we decorated ``grid_sample`` with ``mx.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 :func:`fast.metal_kernel` features:
+requires a few extra ``mx.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.
@@ -316,129 +299,128 @@ We can then implement the backwards pass as follows:
 
 .. code-block:: python
 
-  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;
+    @grid_sample.vjp
+    def grid_sample_vjp(primals, cotangent, _):
+        x, grid = primals
+        B, _, _, C = x.shape
+        _, gN, gM, D = grid.shape
 
-      int gH = grid_shape[1];
-      int gW = grid_shape[2];
+        assert D == 2, "Last dim of `grid` must be size 2."
 
-      int w_stride = C;
-      int h_stride = W * w_stride;
-      int b_stride = H * h_stride;
+        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;
 
-      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 gH = grid_shape[1];
+            int gW = grid_shape[2];
 
-      int ix_nw = floor(ix);
-      int iy_nw = floor(iy);
+            int w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
 
-      int ix_ne = ix_nw + 1;
-      int iy_ne = iy_nw;
+            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_sw = ix_nw;
-      int iy_sw = iy_nw + 1;
+            int ix_nw = floor(ix);
+            int iy_nw = floor(iy);
 
-      int ix_se = ix_nw + 1;
-      int iy_se = iy_nw + 1;
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
 
-      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_sw = ix_nw;
+            int iy_sw = 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;
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
 
-      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 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 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);
+            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_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 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_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_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_se = x[offset];
-              gix += I_se * (iy - iy_nw) * cot;
-              giy += I_se * (ix - ix_nw) * cot;
-          }
-      }
+                    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 gix_mult = W / 2;
-      T giy_mult = H / 2;
+                    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);
 
-      // Reduce across each simdgroup first.
-      // This is much faster than relying purely on atomics.
-      gix = simd_sum(gix);
-      giy = simd_sum(giy);
+                    T I_se = x[offset];
+                    gix += I_se * (iy - iy_nw) * cot;
+                    giy += I_se * (ix - ix_nw) * cot;
+                }
+            }
 
-      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,
-  )
+            T gix_mult = W / 2;
+            T giy_mult = H / 2;
 
-  @grid_sample.vjp
-  def grid_sample_vjp(primals, cotangent, _):
-      x, grid = primals
-      B, _, _, C = x.shape
-      _, gN, gM, D = grid.shape
+            // Reduce across each simdgroup first.
+            // This is much faster than relying purely on atomics.
+            gix = simd_sum(gix);
+            giy = simd_sum(giy);
 
-      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]
+            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]
 
 There's an even larger speed up for the vjp:
 

docs/src/dev/extensions.rst

@@ -22,12 +22,12 @@ You can do that in MLX directly:
 This function performs that operation while leaving the implementation and
 function transformations to MLX.
 
-However, you may want to customize the underlying implementation, perhaps to
-make it faster. In this tutorial we will go through adding custom extensions.
-It will cover:
+However you may need to customize the underlying implementation, perhaps to
+make it faster or for custom differentiation. In this tutorial we will go
+through adding custom extensions. It will cover:
 
 * The structure of the MLX library.
-* Implementing a CPU operation.
+* Implementing a CPU operation that redirects to Accelerate_ when appropriate.
 * Implementing a GPU operation using metal.
 * Adding the ``vjp`` and ``jvp`` function transformation.
 * Building a custom extension and binding it to python.
@@ -45,7 +45,7 @@ Operations
 Operations are the front-end functions that operate on arrays. They are defined
 in the C++ API (:ref:`cpp_ops`), and the Python API (:ref:`ops`) binds them.
 
-We would like an operation :meth:`axpby` that takes in two arrays, ``x`` and
+We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and
 ``y``, and two scalars, ``alpha`` and ``beta``. This is how to define it in
 C++:
 
@@ -55,7 +55,7 @@ C++:
     *  Scale and sum two vectors element-wise
     *  z = alpha * x + beta * y
     *
-    *  Use NumPy-style broadcasting between x and y
+    *  Follow numpy style broadcasting between x and y
     *  Inputs are upcasted to floats if needed
     **/
     array axpby(
@@ -66,7 +66,7 @@ C++:
         StreamOrDevice s = {} // Stream on which to schedule the operation
     );
 
-The simplest way to implement this is with existing operations:
+The simplest way to this operation is in terms of existing operations:
 
 .. code-block:: C++
 
@@ -93,9 +93,9 @@ Primitives
 ^^^^^^^^^^^
 
 A :class:`Primitive` is part of the computation graph of an :class:`array`. It
-defines how to create output arrays given input arrays. Further, a
+defines how to create outputs arrays given a input arrays. Further, a
 :class:`Primitive` has methods to run on the CPU or GPU and for function
-transformations such as ``vjp`` and ``jvp``.  Let's go back to our example to be
+transformations such as ``vjp`` and ``jvp``.  Lets 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 std::vector<array>& cotangents,
+            const array& cotan,
             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.
         */
-        std::pair<std::vector<array>, std::vector<int>> vmap(
+        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) override;
 
-        /** The name of primitive. */
-        const char* name() const override {
-          return "Axpby";
+        /** Print the primitive. */
+        void print(std::ostream& os) override {
+            os << "Axpby";
         }
 
         /** Equivalence check **/
@@ -153,6 +153,9 @@ more concrete:
       private:
         float alpha_;
         float beta_;
+
+        /** Fall back implementation for evaluation on CPU */
+        void eval(const std::vector<array>& inputs, array& out);
     };
 
 The :class:`Axpby` class derives from the base :class:`Primitive` class. The
@@ -185,7 +188,7 @@ Let's reimplement our operation now in terms of our :class:`Axpby` primitive.
         auto promoted_dtype = promote_types(x.dtype(), y.dtype());
 
         // Upcast to float32 for non-floating point inputs x and y
-        auto out_dtype = issubdtype(promoted_dtype, float32)
+        auto out_dtype = is_floating_point(promoted_dtype)
             ? promoted_dtype
             : promote_types(promoted_dtype, float32);
 
@@ -231,57 +234,49 @@ the execution of the computation graph, and calls :meth:`Axpby::eval_cpu` or
 Implementing the CPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Let's start by implementing :meth:`Axpby::eval_cpu`.
+Let's start by implementing a naive and generic version of
+:meth:`Axpby::eval_cpu`. We declared this as a private member function of
+:class:`Axpby` earlier called :meth:`Axpby::eval`.
 
-The method will go over each element of the output array, find the
+Our naive method will go over each element of the output array, find the
 corresponding input elements of ``x`` and ``y`` and perform the operation
 point-wise. This is captured in the templated function :meth:`axpby_impl`.
 
 .. code-block:: C++
 
-  template <typename T>
-  void axpby_impl(
-      const mx::array& x,
-      const mx::array& y,
-      mx::array& out,
-      float alpha_,
-      float beta_,
-      mx::Stream stream) {
-    out.set_data(mx::allocator::malloc(out.nbytes()));
+    template <typename T>
+    void axpby_impl(
+            const array& x,
+            const array& y,
+            array& out,
+            float alpha_,
+            float beta_) {
+        // We only allocate memory when we are ready to fill the output
+        // malloc_or_wait synchronously allocates available memory
+        // There may be a wait executed here if the allocation is requested
+        // under memory-pressured conditions
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
-    // Get the CPU command encoder and register input and output arrays
-    auto& encoder = mx::cpu::get_command_encoder(stream);
-    encoder.set_input_array(x);
-    encoder.set_input_array(y);
-    encoder.set_output_array(out);
+        // Collect input and output data pointers
+        const T* x_ptr = x.data<T>();
+        const T* y_ptr = y.data<T>();
+        T* out_ptr = out.data<T>();
 
-    // Launch the CPU kernel
-    encoder.dispatch([x_ptr = x.data<T>(),
-                      y_ptr = y.data<T>(),
-                      out_ptr = out.data<T>(),
-                      size = out.size(),
-                      shape = out.shape(),
-                      x_strides = x.strides(),
-                      y_strides = y.strides(),
-                      alpha_,
-                      beta_]() {
+        // Cast alpha and beta to the relevant types
+        T alpha = static_cast<T>(alpha_);
+        T beta = static_cast<T>(beta_);
 
-      // Cast alpha and beta to the relevant types
-      T alpha = static_cast<T>(alpha_);
-      T beta = static_cast<T>(beta_);
+        // Do the element-wise operation for each output
+        for (size_t out_idx = 0; out_idx < out.size(); out_idx++) {
+            // Map linear indices to offsets in x and y
+            auto x_offset = elem_to_loc(out_idx, x.shape(), x.strides());
+            auto y_offset = elem_to_loc(out_idx, y.shape(), y.strides());
 
-      // Do the element-wise operation for each output
-      for (size_t out_idx = 0; out_idx < size; out_idx++) {
-        // Map linear indices to offsets in x and y
-        auto x_offset = mx::elem_to_loc(out_idx, shape, x_strides);
-        auto y_offset = mx::elem_to_loc(out_idx, shape, y_strides);
-
-        // We allocate the output to be contiguous and regularly strided
-        // (defaults to row major) and hence it doesn't need additional mapping
-        out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
-      }
-    });
-  }
+            // We allocate the output to be contiguous and regularly strided
+            // (defaults to row major) and hence it doesn't need additional mapping
+            out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
+        }
+    }
 
 Our implementation should work for all incoming floating point arrays.
 Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
@@ -289,32 +284,112 @@ Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
 
 .. code-block:: C++
 
-    void Axpby::eval_cpu(
-        const std::vector<mx::array>& inputs,
-        std::vector<mx::array>& outputs) {
-      auto& x = inputs[0];
-      auto& y = inputs[1];
-      auto& out = outputs[0];
+    /** Fall back implementation for evaluation on CPU */
+    void Axpby::eval(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
+        auto& x = inputs[0];
+        auto& y = inputs[1];
+        auto& out = outputs[0];
 
-      // Dispatch to the correct dtype
-      if (out.dtype() == mx::float32) {
-        return axpby_impl<float>(x, y, out, alpha_, beta_, stream());
-      } else if (out.dtype() == mx::float16) {
-        return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_, stream());
-      } else if (out.dtype() == mx::bfloat16) {
-        return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_, stream());
-      } else if (out.dtype() == mx::complex64) {
-        return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_, stream());
-      } else {
-        throw std::runtime_error(
-            "Axpby is only supported for floating point types.");
-      }
+        // Dispatch to the correct dtype
+        if (out.dtype() == float32) {
+            return axpby_impl<float>(x, y, out, alpha_, beta_);
+        } else if (out.dtype() == float16) {
+            return axpby_impl<float16_t>(x, y, out, alpha_, beta_);
+        } else if (out.dtype() == bfloat16) {
+            return axpby_impl<bfloat16_t>(x, y, out, alpha_, beta_);
+        } else if (out.dtype() == complex64) {
+            return axpby_impl<complex64_t>(x, y, out, alpha_, beta_);
+        } else {
+            throw std::runtime_error(
+                "[Axpby] Only supports floating point types.");
+        }
+    }
+
+This is good as a fallback implementation. We can use the ``axpby`` routine
+provided by the Accelerate_ framework for a faster implementation in certain
+cases:
+
+#.  Accelerate does not provide implementations of ``axpby`` for half precision
+    floats. We can only use it for ``float32`` types.
+#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all
+    elements have fixed strides between them. We only direct to Accelerate
+    if both ``x`` and ``y`` are row contiguous or column contiguous.
+#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` in-place.
+    MLX expects to write the output to a new array. We must copy the elements
+    of ``y`` into the output and use that as an input to ``axpby``.
+
+Let's write an implementation that uses Accelerate in the right conditions.
+It allocates data for the output, copies ``y`` into it, and then calls the
+:func:`catlas_saxpby` from accelerate.
+
+.. code-block:: C++
+
+    template <typename T>
+    void axpby_impl_accelerate(
+            const array& x,
+            const array& y,
+            array& out,
+            float alpha_,
+            float beta_) {
+        // Accelerate library provides catlas_saxpby which does
+        // Y = (alpha * X) + (beta * Y) in place
+        // To use it, we first copy the data in y over to the output array
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+        // We then copy over the elements using the contiguous vector specialization
+        copy_inplace(y, out, CopyType::Vector);
+
+        // Get x and y pointers for catlas_saxpby
+        const T* x_ptr = x.data<T>();
+        T* y_ptr = out.data<T>();
+
+        T alpha = static_cast<T>(alpha_);
+        T beta = static_cast<T>(beta_);
+
+        // Call the inplace accelerate operator
+        catlas_saxpby(
+            /* N = */ out.size(),
+            /* ALPHA = */ alpha,
+            /* X = */ x_ptr,
+            /* INCX = */ 1,
+            /* BETA = */ beta,
+            /* Y = */ y_ptr,
+            /* INCY = */ 1);
+    }
+
+For inputs that do not fit the criteria for accelerate, we fall back to
+:meth:`Axpby::eval`. With this in mind, let's finish our
+:meth:`Axpby::eval_cpu`.
+
+.. code-block:: C++
+
+    /** Evaluate primitive on CPU using accelerate specializations */
+    void Axpby::eval_cpu(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
+        assert(inputs.size() == 2);
+        auto& x = inputs[0];
+        auto& y = inputs[1];
+        auto& out = outputs[0];
+
+        // Accelerate specialization for contiguous single precision float arrays
+        if (out.dtype() == float32 &&
+            ((x.flags().row_contiguous && y.flags().row_contiguous) ||
+            (x.flags().col_contiguous && y.flags().col_contiguous))) {
+            axpby_impl_accelerate<float>(x, y, out, alpha_, beta_);
+            return;
+        }
+
+        // Fall back to common back-end if specializations are not available
+        eval(inputs, outputs);
     }
 
 Just this much is enough to run the operation :meth:`axpby` on a CPU stream! If
 you do not plan on running the operation on the GPU or using transforms on
 computation graphs that contain :class:`Axpby`, you can stop implementing the
-primitive here.
+primitive here and enjoy the speed-ups you get from the Accelerate library.
 
 Implementing the GPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -345,8 +420,8 @@ element in the output.
             constant const float& alpha [[buffer(3)]],
             constant const float& beta [[buffer(4)]],
             constant const int* shape [[buffer(5)]],
-            constant const int64_t* x_strides [[buffer(6)]],
-            constant const int64_t* y_strides [[buffer(7)]],
+            constant const size_t* x_strides [[buffer(6)]],
+            constant const size_t* y_strides [[buffer(7)]],
             constant const int& ndim [[buffer(8)]],
             uint index [[thread_position_in_grid]]) {
         // Convert linear indices to offsets in array
@@ -363,10 +438,24 @@ each instantiation a unique host name so we can identify it.
 
 .. code-block:: C++
 
-    instantiate_kernel("axpby_general_float32", axpby_general, float)
-    instantiate_kernel("axpby_general_float16", axpby_general, float16_t)
-    instantiate_kernel("axpby_general_bfloat16", axpby_general, bfloat16_t)
-    instantiate_kernel("axpby_general_complex64", axpby_general, complex64_t)
+    #define instantiate_axpby(type_name, type)              \
+        template [[host_name("axpby_general_" #type_name)]] \
+        [[kernel]] void axpby_general<type>(                \
+            device const type* x [[buffer(0)]],             \
+            device const type* y [[buffer(1)]],             \
+            device type* out [[buffer(2)]],                 \
+            constant const float& alpha [[buffer(3)]],      \
+            constant const float& beta [[buffer(4)]],       \
+            constant const int* shape [[buffer(5)]],        \
+            constant const size_t* x_strides [[buffer(6)]], \
+            constant const size_t* y_strides [[buffer(7)]], \
+            constant const int& ndim [[buffer(8)]],         \
+            uint index [[thread_position_in_grid]]);
+
+    instantiate_axpby(float32, float);
+    instantiate_axpby(float16, half);
+    instantiate_axpby(bfloat16, bfloat16_t);
+    instantiate_axpby(complex64, complex64_t);
 
 The logic to determine the kernel, set the inputs, resolve the grid dimensions,
 and dispatch to the GPU are contained in :meth:`Axpby::eval_gpu` as shown
@@ -391,21 +480,21 @@ below.
         auto& d = metal::device(s.device);
 
         // Allocate output memory
-        out.set_data(allocator::malloc(out.nbytes()));
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
         // Resolve name of kernel
         std::ostringstream kname;
         kname << "axpby_" << "general_" << type_to_name(out);
 
-        // Load the metal library
-        auto lib = d.get_library("mlx_ext");
+        // Make sure the metal library is available
+        d.register_library("mlx_ext");
 
         // Make a kernel from this metal library
-        auto kernel = d.get_kernel(kname.str(), lib);
+        auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
         // Prepare to encode kernel
         auto& compute_encoder = d.get_command_encoder(s.index);
-        compute_encoder.set_compute_pipeline_state(kernel);
+        compute_encoder->setComputePipelineState(kernel);
 
         // Kernel parameters are registered with buffer indices corresponding to
         // those in the kernel declaration at axpby.metal
@@ -420,14 +509,14 @@ below.
         compute_encoder.set_output_array(out, 2);
 
         // Encode alpha and beta
-        compute_encoder.set_bytes(alpha_, 3);
-        compute_encoder.set_bytes(beta_, 4);
+        compute_encoder->setBytes(&alpha_, sizeof(float), 3);
+        compute_encoder->setBytes(&beta_, sizeof(float), 4);
 
         // Encode shape, strides and ndim
-        compute_encoder.set_vector_bytes(x.shape(), 5);
-        compute_encoder.set_vector_bytes(x.strides(), 6);
-        compute_encoder.set_bytes(y.strides(), 7);
-        compute_encoder.set_bytes(ndim, 8);
+        compute_encoder->setBytes(x.shape().data(), ndim * sizeof(int), 5);
+        compute_encoder->setBytes(x.strides().data(), ndim * sizeof(size_t), 6);
+        compute_encoder->setBytes(y.strides().data(), ndim * sizeof(size_t), 7);
+        compute_encoder->setBytes(&ndim, sizeof(int), 8);
 
         // We launch 1 thread for each input and make sure that the number of
         // threads in any given threadgroup is not higher than the max allowed
@@ -441,7 +530,7 @@ below.
 
         // Launch the grid with the given number of threads divided among
         // the given threadgroups
-        compute_encoder.dispatch_threads(grid_dims, group_dims);
+        compute_encoder.dispatchThreads(grid_dims, group_dims);
     }
 
 We can now call the :meth:`axpby` operation on both the CPU and the GPU!
@@ -469,7 +558,7 @@ 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 be built with ops
+        // The jvp transform on the primitive can 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
@@ -481,7 +570,7 @@ 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())};
@@ -735,7 +824,7 @@ 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 is correct: {mx.all(c == 6.0).item()}")
+    print(f"c correct: {mx.all(c == 6.0).item()}")
 
 Output:
 
@@ -743,13 +832,13 @@ Output:
 
     c shape: [3, 4]
     c dtype: float32
-    c is correct: True
+    c correctness: True
 
 Results
 ^^^^^^^
 
 Let's run a quick benchmark and see how our new ``axpby`` operation compares
-with the naive :meth:`simple_axpby` we first defined.
+with the naive :meth:`simple_axpby` we first defined on the CPU.
 
 .. code-block:: python
 
@@ -757,11 +846,13 @@ with the naive :meth:`simple_axpby` we first defined.
     from mlx_sample_extensions import axpby
     import time
 
+    mx.set_default_device(mx.cpu)
+
     def simple_axpby(x: mx.array, y: mx.array, alpha: float, beta: float) -> mx.array:
         return alpha * x + beta * y
 
-    M = 4096
-    N = 4096
+    M = 256
+    N = 512
 
     x = mx.random.normal((M, N))
     y = mx.random.normal((M, N))
@@ -772,24 +863,24 @@ with the naive :meth:`simple_axpby` we first defined.
 
     def bench(f):
         # Warm up
-        for i in range(5):
+        for i in range(100):
             z = f(x, y, alpha, beta)
             mx.eval(z)
 
         # Timed run
         s = time.time()
-        for i in range(100):
+        for i in range(5000):
             z = f(x, y, alpha, beta)
             mx.eval(z)
         e = time.time()
-        return 1000 * (e - s) / 100
+        return e - s
 
     simple_time = bench(simple_axpby)
     custom_time = bench(axpby)
 
-    print(f"Simple axpby: {simple_time:.3f} ms | Custom axpby: {custom_time:.3f} ms")
+    print(f"Simple axpby: {simple_time:.3f} s | Custom axpby: {custom_time:.3f} s")
 
-The results are ``Simple axpby: 1.559 ms | Custom axpby: 0.774 ms``. We see
+The results are ``Simple axpby: 0.114 s | Custom axpby: 0.109 s``. We see
 modest improvements right away!
 
 This operation is now good to be used to build other operations, in

docs/src/dev/mlx_in_cpp.rst

@@ -1,121 +0,0 @@
-.. _mlx_in_cpp:
-
-Using MLX in C++
-================
-
-You can use MLX in a C++ project with CMake.
-
-.. note::
-
-  This guide is based one the following `example using MLX in C++ 
-  <https://github.com/ml-explore/mlx/tree/main/examples/cmake_project>`_
-
-First install MLX:
-
-.. code-block:: bash
-
-  pip install -U mlx
-
-You can also install the MLX Python package from source or just the C++
-library. For more information see the :ref:`documentation on installing MLX
-<build_and_install>`.
-
-Next make an example program in ``example.cpp``: 
-
-.. code-block:: C++
-
-  #include <iostream>
-
-  #include "mlx/mlx.h"
-
-  namespace mx = mlx::core;
-
-  int main() {
-    auto x = mx::array({1, 2, 3});
-    auto y = mx::array({1, 2, 3});
-    std::cout << x + y << std::endl;
-    return 0;
-  }
-
-The next step is to setup a CMake file in ``CMakeLists.txt``:
-
-.. code-block:: cmake
-
-  cmake_minimum_required(VERSION 3.27)
-
-  project(example LANGUAGES CXX)
-
-  set(CMAKE_CXX_STANDARD 17)
-  set(CMAKE_CXX_STANDARD_REQUIRED ON)
-
-
-Depending on how you installed MLX, you may need to tell CMake where to
-find it. 
-
-If you installed MLX with Python, then add the following to the CMake file:
-
-.. code-block:: cmake
-
-  find_package(
-    Python 3.9
-    COMPONENTS Interpreter Development.Module
-    REQUIRED)
-  execute_process(
-    COMMAND "${Python_EXECUTABLE}" -m mlx --cmake-dir
-    OUTPUT_STRIP_TRAILING_WHITESPACE
-    OUTPUT_VARIABLE MLX_ROOT)
-
-If you installed the MLX C++ package to a system path, then CMake should be
-able to find it. If you installed it to a non-standard location or CMake can't
-find MLX then set ``MLX_ROOT`` to the location where MLX is installed:
-
-.. code-block:: cmake
-
-  set(MLX_ROOT "/path/to/mlx/")
-
-Next, instruct CMake to find MLX:
-
-.. code-block:: cmake
-
-  find_package(MLX CONFIG REQUIRED)
-
-Finally, add the ``example.cpp`` program as an executable and link MLX.
-
-.. code-block:: cmake
-
-  add_executable(example example.cpp)
-  target_link_libraries(example PRIVATE mlx)
-
-You can build the example with:
-
-.. code-block:: bash
-
-  cmake -B build -DCMAKE_BUILD_TYPE=Release
-  cmake --build build
-
-And run it with:
-
-.. code-block:: bash
-
-  ./build/example
-
-Note ``find_package(MLX CONFIG REQUIRED)`` sets the following variables:
-
-.. list-table:: Package Variables
-   :widths: 20 20 
-   :header-rows: 1
-
-   * - Variable 
-     - Description 
-   * - MLX_FOUND
-     - ``True`` if MLX is found
-   * - MLX_INCLUDE_DIRS
-     - Include directory
-   * - MLX_LIBRARIES
-     - Libraries to link against
-   * - MLX_CXX_FLAGS
-     - Additional compiler flags
-   * - MLX_BUILD_ACCELERATE
-     - ``True`` if MLX was built with Accelerate 
-   * - MLX_BUILD_METAL
-     - ``True`` if MLX was built with Metal

docs/src/index.rst

@@ -45,7 +45,6 @@ are the CPU and GPU.
    usage/numpy
    usage/distributed
    usage/using_streams
-   usage/export
 
 .. toctree::
    :caption: Examples
@@ -62,7 +61,6 @@ are the CPU and GPU.
    python/array
    python/data_types
    python/devices_and_streams
-   python/export
    python/ops
    python/random
    python/transforms
@@ -70,7 +68,6 @@ are the CPU and GPU.
    python/fft
    python/linalg
    python/metal
-   python/memory_management
    python/nn
    python/optimizers
    python/distributed
@@ -89,4 +86,3 @@ are the CPU and GPU.
    dev/extensions
    dev/metal_debugger
    dev/custom_metal_kernels
-   dev/mlx_in_cpp

docs/src/install.rst

@@ -1,5 +1,3 @@
-.. _build_and_install:
-
 Build and Install
 =================
 
@@ -23,24 +21,13 @@ 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 has a CUDA backend which you can use on any Linux platform with CUDA 12
-and SM 7.0 (Volta) and up. To install MLX with CUDA support, run:
+MLX is also available on conda-forge. To install MLX with conda do:
 
 .. code-block:: shell
 
-    pip install "mlx[cuda]"
+   conda install conda-forge::mlx
 
-CPU-only (Linux)
-^^^^^^^^^^^^^^^^
-
-For a CPU-only version of MLX that runs on Linux use:
-
-.. code-block:: shell
-
-    pip install "mlx[cpu]"
 
 Troubleshooting
 ^^^^^^^^^^^^^^^
@@ -66,7 +53,7 @@ Build Requirements
 ^^^^^^^^^^^^^^^^^^
 
 - A C++ compiler with C++17 support (e.g. Clang >= 5.0)
-- `cmake <https://cmake.org/>`_ -- version 3.25 or later, and ``make``
+- `cmake <https://cmake.org/>`_ -- version 3.24 or later, and ``make``
 - Xcode >= 15.0 and macOS SDK >= 14.0
 
 .. note::
@@ -76,8 +63,6 @@ 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>`_:
 
@@ -89,20 +74,20 @@ Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-  pip install .
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
 
 For developing, install the package with development dependencies, and use an
 editable install:
 
 .. code-block:: shell
 
-  pip install -e ".[dev]"
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
 
 Once the development dependencies are installed, you can build faster with:
 
 .. code-block:: shell
 
- python setup.py build_ext --inplace
+ CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
 
 Run the tests with:
 
@@ -120,8 +105,6 @@ 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
@@ -200,7 +183,6 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
-
 Binary Size Minimization
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -227,51 +209,7 @@ Metal library by run-time compiling kernels the first time they are used in MLX
 on a given machine. Note run-time compilation incurs a cold-start cost which can
 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
-
+Metal kernel cache persists accross reboots.
 
 Troubleshooting
 ^^^^^^^^^^^^^^^

docs/src/python/array.rst

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

docs/src/python/data_types.rst

@@ -51,20 +51,11 @@ The default floating point type is ``float32`` and the default integer type is
    * - ``float32``
      - 4 
      - 32-bit float
-   * - ``float64``
-     - 4 
-     - 64-bit double
    * - ``complex64``
      - 8 
      - 64-bit complex float
 
 
-.. note::
-
-    Arrays with type ``float64`` only work with CPU operations. Using
-    ``float64`` arrays on the GPU will result in an exception.
-
-
 Data type are aranged in a hierarchy. See the :obj:`DtypeCategory` object
 documentation for more information. Use :func:`issubdtype` to determine if one
 ``dtype`` (or category) is a subtype of another category.
@@ -75,4 +66,3 @@ documentation for more information. Use :func:`issubdtype` to determine if one
    Dtype
    DtypeCategory
    issubdtype
-   finfo

docs/src/python/export.rst

@@ -1,14 +0,0 @@
-.. _export:
-
-Export Functions
-================
-
-.. currentmodule:: mlx.core
-
-.. autosummary::
-  :toctree: _autosummary
-
-   export_function
-   import_function
-   exporter
-   export_to_dot

docs/src/python/fast.rst

@@ -12,4 +12,5 @@ Fast
   layer_norm
   rope
   scaled_dot_product_attention
+  affine_quantize
   metal_kernel

docs/src/python/fft.rst

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

docs/src/python/linalg.rst

@@ -16,12 +16,5 @@ Linear Algebra
     cross
     qr
     svd
-    eigvals
-    eig
     eigvalsh
     eigh
-    lu
-    lu_factor
-    pinv
-    solve
-    solve_triangular

docs/src/python/memory_management.rst

@@ -1,16 +0,0 @@
-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

@@ -8,5 +8,13 @@ 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/nn.rst

@@ -174,7 +174,6 @@ In detail:
 
    value_and_grad
    quantize
-   average_gradients
 
 .. toctree::
 

docs/src/python/nn/layers.rst

@@ -12,7 +12,6 @@ Layers
    ALiBi
    AvgPool1d
    AvgPool2d
-   AvgPool3d
    BatchNorm
    CELU
    Conv1d
@@ -42,7 +41,6 @@ Layers
    LSTM
    MaxPool1d
    MaxPool2d
-   MaxPool3d
    Mish
    MultiHeadAttention
    PReLU

docs/src/python/ops.rst

@@ -32,16 +32,13 @@ Operations
    atleast_2d
    atleast_3d
    bitwise_and
-   bitwise_invert
    bitwise_or
    bitwise_xor
    block_masked_mm
-   broadcast_arrays
    broadcast_to
    ceil
    clip
    concatenate
-   contiguous
    conj
    conjugate
    convolve
@@ -92,7 +89,6 @@ Operations
    isneginf
    isposinf
    issubdtype
-   kron
    left_shift
    less
    less_equal
@@ -103,7 +99,6 @@ Operations
    log10
    log1p
    logaddexp
-   logcumsumexp
    logical_not
    logical_and
    logical_or
@@ -149,8 +144,6 @@ Operations
    sign
    sin
    sinh
-   slice
-   slice_update
    softmax
    sort
    split
@@ -175,7 +168,6 @@ Operations
    tri
    tril
    triu
-   unflatten
    var
    view
    where

docs/src/python/optimizers/common_optimizers.rst

@@ -18,4 +18,3 @@ Common Optimizers
    AdamW
    Adamax
    Lion
-   MultiOptimizer

docs/src/python/transforms.rst

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

docs/src/usage/compile.rst

@@ -421,77 +421,3 @@ the most opportunity to optimize the computation graph:
   # Compiling the outer function is good to do as it will likely
   # be faster even though the inner functions are compiled
   fun = mx.compile(outer)
-
-
-
-.. _shapeless_compile:
-
-Shapeless Compilation
----------------------
-
-When the shape of an input to a compiled function changes, the function is
-recompiled. You can compile a function once and run it on inputs with
-variable shapes by specifying ``shapeless=True`` to :func:`compile`. In this
-case changes to the shapes of the inputs do not cause the function to be
-recompiled.
-
-.. code-block:: python
-
-  def fun(x, y):
-      return mx.abs(x + y)
-
-  compiled_fun = mx.compile(fun, shapeless=True)
-
-  x = mx.array(1.0)
-  y = mx.array(-2.0)
-
-  # Firt call compiles the function
-  print(compiled_fun(x, y))
-
-  # Second call with different shapes
-  # does not recompile the function
-  x = mx.array([1.0, -6.0])
-  y = mx.array([-2.0, 3.0])
-  print(compiled_fun(x, y))
-
-
-Use shapeless compilations carefully. Since compilation is not triggered when
-shapes change, any graphs which are conditional on the input shapes will not
-work as expected. Shape-dependent computations are common and sometimes subtle
-to detect. For example:
-
-.. code-block:: python
-
-  def fun(x):
-      return x.reshape(x.shape[0] * x.shape[1], -1)
-
-  compiled_fun = mx.compile(fun, shapeless=True)
-
-  x = mx.random.uniform(shape=(2, 3, 4))
-
-  out = compiled_fun(x)
-
-  x = mx.random.uniform(shape=(5, 5, 3))
-
-  # Error, can't reshape (5, 5, 3) to (6, -1)
-  out = compiled_fun(x)
-
-The second call to the ``compiled_fun`` fails because of the call to
-:func:`reshape` which uses the static shape of ``x`` in the first call. We can
-fix this by using :func:`flatten` to avoid hardcoding the shape of ``x``:
-
-.. code-block:: python
-
-  def fun(x):
-      return x.flatten(0, 1)
-
-  compiled_fun = mx.compile(fun, shapeless=True)
-
-  x = mx.random.uniform(shape=(2, 3, 4))
-
-  out = compiled_fun(x)
-
-  x = mx.random.uniform(shape=(5, 5, 3))
-
-  # Ok
-  out = compiled_fun(x)

docs/src/usage/distributed.rst

@@ -5,27 +5,21 @@ Distributed Communication
 
 .. currentmodule:: mlx.core.distributed
 
-MLX supports distributed communication operations that allow the computational cost
-of training or inference to be shared across many physical machines. At the
-moment we support two different communication backends:
-
-* `MPI <https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ a
-  full-featured and mature distributed communications library
-* A **ring** backend of our own that uses native TCP sockets and should be
-  faster for thunderbolt connections.
-
-The list of all currently supported operations and their documentation can be
-seen in the :ref:`API docs<distributed>`.
+MLX utilizes `MPI <https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ to
+provide distributed communication operations that allow the computational cost
+of training or inference to be shared across many physical machines. You can
+see a list of the supported operations in the :ref:`API docs<distributed>`.
 
 .. note::
-   Some operations may not be supported or not as fast as they should be.
+   A lot of operations may not be supported or not as fast as they should be.
    We are adding more and tuning the ones we have as we are figuring out the
    best way to do distributed computing on Macs using MLX.
 
 Getting Started
 ---------------
 
-A distributed program in MLX is as simple as:
+MLX already comes with the ability to "talk" to MPI if it is installed on the
+machine. The minimal distributed program in MLX is as simple as:
 
 .. code:: python
 
@@ -36,79 +30,74 @@ A distributed program in MLX is as simple as:
     print(world.rank(), x)
 
 The program above sums the array ``mx.ones(10)`` across all
-distributed processes. However, when this script is run with ``python`` only
-one process is launched and no distributed communication takes place. Namely,
-all operations in ``mx.distributed`` are noops when the distributed group has a
-size of one. This property allows us to avoid code that checks if we are in a
-distributed setting similar to the one below:
+distributed processes. If simply run with ``python``, however, only one
+process is launched and no distributed communication takes place.
 
-.. code:: python
-
-    import mlx.core as mx
-
-    x = ...
-    world = mx.distributed.init()
-    # No need for the check we can simply do x = mx.distributed.all_sum(x)
-    if world.size() > 1:
-        x = mx.distributed.all_sum(x)
-
-Running Distributed Programs
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-MLX provides ``mlx.launch`` a helper script to launch distributed programs.
-Continuing with our initial example we can run it on localhost with 4 processes using
+To launch the program in distributed mode we need to use ``mpirun`` or
+``mpiexec`` depending on the MPI installation. The simplest possible way is the
+following:
 
 .. code:: shell
 
-    $ mlx.launch -n 4 my_script.py
-    3 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
-    2 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
-    1 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
-    0 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    $ mpirun -np 2 python test.py
+    1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+    0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
 
-We can also run it on some remote hosts by providing their IPs (provided that
-the script exists on all hosts and they are reachable by ssh)
+The above launches two processes on the same (local) machine and we can see
+both standard output streams. The processes send the array of 1s to each other
+and compute the sum which is printed. Launching with ``mpirun -np 4 ...`` would
+print 4 etc.
+
+Installing MPI
+---------------
+
+MPI can be installed with Homebrew, using the Anaconda package manager or
+compiled from source. Most of our testing is done using ``openmpi`` installed
+with the Anaconda package manager as follows:
 
 .. code:: shell
 
-    $ mlx.launch --hosts ip1,ip2,ip3,ip4 my_script.py
-    3 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
-    2 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
-    1 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
-    0 array([4, 4, 4, ..., 4, 4, 4], dtype=float32)
+    $ conda install openmpi
 
-Consult the dedicated :doc:`usage guide<launching_distributed>` for more
-information on using ``mlx.launch``.
+Installing with Homebrew may require specifying the location of ``libmpi.dyld``
+so that MLX can find it and load it at runtime. This can simply be achieved by
+passing the ``DYLD_LIBRARY_PATH`` environment variable to ``mpirun``.
 
-Selecting Backend
-^^^^^^^^^^^^^^^^^
+.. code:: shell
 
-You can select the backend you want to use when calling :func:`init` by passing
-one of ``{'any', 'ring', 'mpi'}``. When passing ``any``, MLX will try to
-initialize the ``ring`` backend and if it fails the ``mpi`` backend. If they
-both fail then a singleton group is created.
+    $ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python test.py
+
+Setting up Remote Hosts
+-----------------------
+
+MPI can automatically connect to remote hosts and set up the communication over
+the network if the remote hosts can be accessed via ssh. A good checklist to
+debug connectivity issues is the following:
+
+* ``ssh hostname`` works from all machines to all machines without asking for
+  password or host confirmation
+* ``mpirun`` is accessible on all machines. You can call ``mpirun`` using its
+  full path to force all machines to use a specific path.
+* Ensure that the ``hostname`` used by MPI is the one that you have configured
+  in the ``.ssh/config`` files on all machines.
 
 .. note::
-   After a distributed backend is successfully initialized :func:`init` will
-   return **the same backend** if called without arguments or with backend set to
-   ``any``.
+  For an example hostname ``foo.bar.com`` MPI can use only ``foo`` as
+  the hostname passed to ssh if the current hostname matches ``*.bar.com``.
 
-The following examples aim to clarify the backend initialization logic in MLX:
+An easy way to pass the host names to MPI is using a host file. A host file
+looks like the following, where ``host1`` and ``host2`` should be the fully
+qualified domain names or IPs for these hosts.
 
-.. code:: python
+.. code::
 
-    # Case 1: Initialize MPI regardless if it was possible to initialize the ring backend
-    world = mx.distributed.init(backend="mpi")
-    world2 = mx.distributed.init()  # subsequent calls return the MPI backend!
+    host1 slots=1
+    host2 slots=1
 
-    # Case 2: Initialize any backend
-    world = mx.distributed.init(backend="any")  # equivalent to no arguments
-    world2 = mx.distributed.init()  # same as above
-
-    # Case 3: Initialize both backends at the same time
-    world_mpi = mx.distributed.init(backend="mpi")
-    world_ring = mx.distributed.init(backend="ring")
-    world_any = mx.distributed.init()  # same as MPI because it was initialized first!
+When using MLX, it is very likely that you want to use 1 slot per host, ie one
+process per host.  The hostfile also needs to contain the current
+host if you want to run on the local host. Passing the host file to
+``mpirun`` is simply done using the ``--hostfile`` command line argument.
 
 Training Example
 ----------------
@@ -152,13 +141,12 @@ everything else remaining the same.
     from mlx.utils import tree_map
 
     def all_reduce_grads(grads):
-        N = mx.distributed.init().size()
+        N = mx.distributed.init()
         if N == 1:
             return grads
         return tree_map(
-            lambda x: mx.distributed.all_sum(x) / N,
-            grads
-        )
+                lambda x: mx.distributed.all_sum(x) / N,
+                grads)
 
     def step(model, x, y):
         loss, grads = loss_grad_fn(model, x, y)
@@ -166,179 +154,13 @@ everything else remaining the same.
         optimizer.update(model, grads)
         return loss
 
-Utilizing ``nn.average_gradients``
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+Tuning All Reduce
+-----------------
 
-Although the code example above works correctly; it performs one communication
-per gradient. It is significantly more efficient to aggregate several gradients
-together and perform fewer communication steps.
+We are working on improving the performance of all reduce on MLX but for now
+the two main things one can do to extract the most out of distributed training with MLX are:
 
-This is the purpose of :func:`mlx.nn.average_gradients`. The final code looks
-almost identical to the example above:
-
-.. code:: python
-
-    model = ...
-    optimizer = ...
-    dataset = ...
-
-    def step(model, x, y):
-        loss, grads = loss_grad_fn(model, x, y)
-        grads = mlx.nn.average_gradients(grads) # <---- This line was added
-        optimizer.update(model, grads)
-        return loss
-
-    for x, y in dataset:
-        loss = step(model, x, y)
-        mx.eval(loss, model.parameters())
-
-
-Getting Started with MPI
-------------------------
-
-MLX already comes with the ability to "talk" to MPI if it is installed on the
-machine. Launching distributed MLX programs that use MPI can be done with
-``mpirun`` as expected. However, in the following examples we will be using
-``mlx.launch --backend mpi`` which takes care of some nuisances such as setting
-absolute paths for the ``mpirun`` executable and the ``libmpi.dyld`` shared
-library.
-
-The simplest possible usage is the following which, assuming the minimal
-example in the beginning of this page, should result in:
-
-.. code:: shell
-
-    $ mlx.launch --backend mpi -n 2 test.py
-    1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
-    0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
-
-The above launches two processes on the same (local) machine and we can see
-both standard output streams. The processes send the array of 1s to each other
-and compute the sum which is printed. Launching with ``mlx.launch -n 4 ...`` would
-print 4 etc.
-
-Installing MPI
-^^^^^^^^^^^^^^
-
-MPI can be installed with Homebrew, using the Anaconda package manager or
-compiled from source. Most of our testing is done using ``openmpi`` installed
-with the Anaconda package manager as follows:
-
-.. code:: shell
-
-    $ conda install conda-forge::openmpi
-
-Installing with Homebrew may require specifying the location of ``libmpi.dyld``
-so that MLX can find it and load it at runtime. This can simply be achieved by
-passing the ``DYLD_LIBRARY_PATH`` environment variable to ``mpirun`` and it is
-done automatically by ``mlx.launch``.
-
-.. code:: shell
-
-    $ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python test.py
-    $ # or simply
-    $ mlx.launch -n 2 test.py
-
-Setting up Remote Hosts
-^^^^^^^^^^^^^^^^^^^^^^^
-
-MPI can automatically connect to remote hosts and set up the communication over
-the network if the remote hosts can be accessed via ssh. A good checklist to
-debug connectivity issues is the following:
-
-* ``ssh hostname`` works from all machines to all machines without asking for
-  password or host confirmation
-* ``mpirun`` is accessible on all machines.
-* Ensure that the ``hostname`` used by MPI is the one that you have configured
-  in the ``.ssh/config`` files on all machines.
-
-Tuning MPI All Reduce
-^^^^^^^^^^^^^^^^^^^^^
-
-.. note::
-
-    For faster all reduce consider using the ring backend either with Thunderbolt
-    connections or over Ethernet.
-
-Configure MPI to use N tcp connections between each host to improve bandwidth
-by passing ``--mca btl_tcp_links N``.
-
-Force MPI to use the most performant network interface by setting ``--mca
-btl_tcp_if_include <iface>`` where ``<iface>`` should be the interface you want
-to use.
-
-Getting Started with Ring
--------------------------
-
-The ring backend does not depend on any third party library so it is always
-available. It uses TCP sockets so the nodes need to be reachable via a network.
-As the name suggests the nodes are connected in a ring which means that rank 1
-can only communicate with rank 0 and rank 2, rank 2 only with rank 1 and rank 3
-and so on and so forth. As a result :func:`send` and :func:`recv` with
-arbitrary sender and receiver is not supported in the ring backend.
-
-Defining a Ring
-^^^^^^^^^^^^^^^
-
-The easiest way to define and use a ring is via a JSON hostfile and the
-``mlx.launch`` :doc:`helper script <launching_distributed>`. For each node one
-defines a hostname to ssh into to run commands on this node and one or more IPs
-that this node will listen to for connections.
-
-For example the hostfile below defines a 4 node ring. ``hostname1`` will be
-rank 0, ``hostname2`` rank 1 etc.
-
-.. code:: json
-
-    [
-        {"ssh": "hostname1", "ips": ["123.123.123.1"]},
-        {"ssh": "hostname2", "ips": ["123.123.123.2"]},
-        {"ssh": "hostname3", "ips": ["123.123.123.3"]},
-        {"ssh": "hostname4", "ips": ["123.123.123.4"]}
-    ]
-
-Running ``mlx.launch --hostfile ring-4.json my_script.py`` will ssh into each
-node, run the script which will listen for connections in each of the provided
-IPs. Specifically, ``hostname1`` will connect to ``123.123.123.2`` and accept a
-connection from ``123.123.123.4`` and so on and so forth.
-
-Thunderbolt Ring
-^^^^^^^^^^^^^^^^
-
-Although the ring backend can have benefits over MPI even for Ethernet, its
-main purpose is to use Thunderbolt rings for higher bandwidth communication.
-Setting up such thunderbolt rings can be done manually, but is a relatively
-tedious process. To simplify this, we provide the utility ``mlx.distributed_config``.
-
-To use ``mlx.distributed_config`` your computers need to be accessible by ssh via
-Ethernet or Wi-Fi. Subsequently, connect them via thunderbolt cables and then call the
-utility as follows:
-
-.. code:: shell
-
-   mlx.distributed_config --verbose --hosts host1,host2,host3,host4
-
-By default the script will attempt to discover the thunderbolt ring and provide
-you with the commands to configure each node as well as the ``hostfile.json``
-to use with ``mlx.launch``. If password-less ``sudo`` is available on the nodes
-then ``--auto-setup`` can be used to configure them automatically.
-
-To validate your connection without configuring anything
-``mlx.distributed_config`` can also plot the ring using DOT format.
-
-.. code:: shell
-
-   mlx.distributed_config --verbose --hosts host1,host2,host3,host4 --dot >ring.dot
-   dot -Tpng ring.dot >ring.png
-   open ring.png
-
-If you want to go through the process manually, the steps are as follows:
-
-* Disable the thunderbolt bridge interface
-* For the cable connecting rank ``i`` to rank ``i + 1`` find the interfaces
-  corresponding to that cable in nodes ``i`` and ``i + 1``.
-* Set up a unique subnetwork connecting the two nodes for the corresponding
-  interfaces. For instance if the cable corresponds to ``en2`` on node ``i``
-  and ``en2`` also on node ``i + 1`` then we may assign IPs ``192.168.0.1`` and
-  ``192.168.0.2`` respectively to the two nodes. For more details you can see
-  the commands prepared by the utility script.
+1. Perform a few large reductions instead of many small ones to improve
+   bandwidth and latency
+2. Pass ``--mca btl_tcp_links 4`` to ``mpirun`` to configure it to use 4 tcp
+   connections between each host to improve bandwidth

docs/src/usage/export.rst

@@ -1,288 +0,0 @@
-.. _export_usage:
-
-Exporting Functions
-===================
-
-.. currentmodule:: mlx.core
-
-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++). 
-
-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 
--------------------
-
-Let's start with a simple example:
- 
-.. code-block:: python
-
-  def fun(x, y):
-    return x + y
-
-  x = mx.array(1.0)
-  y = mx.array(1.0)
-  mx.export_function("add.mlxfn", fun, x, y)
-
-To export a function, provide sample input arrays that the function
-can be called with. The data doesn't matter, but the shapes and types of the
-arrays do. In the above example we exported ``fun`` with two ``float32``
-scalar arrays. We can then import the function and run it:
-
-.. code-block:: python
-
-  add_fun = mx.import_function("add.mlxfn")
-
-  out, = add_fun(mx.array(1.0), mx.array(2.0))
-  # Prints: array(3, dtype=float32)
-  print(out)
-
-  out, = add_fun(mx.array(1.0), mx.array(3.0))
-  # Prints: array(4, dtype=float32)
-  print(out)
-
-  # Raises an exception
-  add_fun(mx.array(1), mx.array(3.0))
-
-  # Raises an exception
-  add_fun(mx.array([1.0, 2.0]), mx.array(3.0))
-
-Notice the third and fourth calls to ``add_fun`` raise exceptions because the
-shapes and types of the inputs are different than the shapes and types of the
-example inputs we exported the function with.
-
-Also notice that even though the original ``fun`` returns a single output
-array, the imported function always returns a tuple of one or more arrays.
-
-The inputs to :func:`export_function` and to an imported function can be
-specified as variable positional arguments or as a tuple of arrays:
-
-.. code-block:: python
-
-  def fun(x, y):
-    return x + y
-
-  x = mx.array(1.0)
-  y = mx.array(1.0)
-   
-  # Both arguments to fun are positional
-  mx.export_function("add.mlxfn", fun, x, y)
-
-  # Same as above
-  mx.export_function("add.mlxfn", fun, (x, y))
-
-  imported_fun = mx.import_function("add.mlxfn")
-
-  # Ok
-  out, = imported_fun(x, y)
-
-  # Also ok
-  out, = imported_fun((x, y))
-
-You can pass example inputs to functions as positional or keyword arguments. If
-you use keyword arguments to export the function, then you have to use the same
-keyword arguments when calling the imported function.
-
-.. code-block:: python
-
-  def fun(x, y):
-    return x + y
-
-  # One argument to fun is positional, the other is a kwarg
-  mx.export_function("add.mlxfn", fun, x, y=y)
-
-  imported_fun = mx.import_function("add.mlxfn")
-
-  # Ok
-  out, = imported_fun(x, y=y)
-
-  # Also ok
-  out, = imported_fun((x,), {"y": y})
-
-  # Raises since the keyword argument is missing
-  out, = imported_fun(x, y)
-
-  # Raises since the keyword argument has the wrong key
-  out, = imported_fun(x, z=y)
-
-
-Exporting Modules
------------------
-
-An :obj:`mlx.nn.Module` can be exported with or without the parameters included
-in the exported function. Here's an example:
-
-.. code-block:: python
-
-   model = nn.Linear(4, 4)
-   mx.eval(model.parameters())
-
-   def call(x):
-      return model(x)
-
-   mx.export_function("model.mlxfn", call, mx.zeros(4))
-
-In the above example, the :obj:`mlx.nn.Linear` module is exported. Its
-parameters are also saved to the ``model.mlxfn`` file.
-
-.. note::
-
-   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.
-
-If you only want to export the ``Module.__call__`` function without the
-parameters, pass them as inputs to the ``call`` wrapper:
-
-.. code-block:: python
-
-   model = nn.Linear(4, 4)
-   mx.eval(model.parameters())
-
-   def call(x, **params):
-     # 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()))
-   mx.export_function("model.mlxfn", call, (mx.zeros(4),), params)
-
-
-Shapeless Exports
------------------
-
-Just like :func:`compile`, functions can also be exported for dynamically shaped
-inputs. Pass ``shapeless=True`` to :func:`export_function` or :func:`exporter`
-to export a function which can be used for inputs with variable shapes:
-
-.. code-block:: python
-
-  mx.export_function("fun.mlxfn", mx.abs, mx.array(0.0), shapeless=True)
-  imported_abs = mx.import_function("fun.mlxfn")
-
-  # Ok
-  out, = imported_abs(mx.array(-1.0))
-  
-  # Also ok 
-  out, = imported_abs(mx.array([-1.0, -2.0]))
-
-With ``shapeless=False`` (which is the default), the second call to
-``imported_abs`` would raise an exception with a shape mismatch.
-
-Shapeless exporting works the same as shapeless compilation and should be
-used carefully. See the :ref:`documentation on shapeless compilation
-<shapeless_compile>` for more information.
-
-Exporting Multiple Traces
--------------------------
-
-In some cases, functions build different computation graphs for different
-input arguments. A simple way to manage this is to export to a new file with
-each set of inputs. This is a fine option in many cases. But it can be
-suboptimal if the exported functions have a large amount of duplicate constant
-data (for example the parameters of a :obj:`mlx.nn.Module`).
-
-The export API in MLX lets you export multiple traces of the same function to
-a single file by creating an exporting context manager with :func:`exporter`:
-
-.. code-block:: python
-
-  def fun(x, y=None):
-      constant = mx.array(3.0)
-      if y is not None:
-        x += y 
-      return x + constant
-
-  with mx.exporter("fun.mlxfn", fun) as exporter:
-      exporter(mx.array(1.0))
-      exporter(mx.array(1.0), y=mx.array(0.0))
-
-  imported_function = mx.import_function("fun.mlxfn")
-
-  # Call the function with y=None
-  out, = imported_function(mx.array(1.0))
-  print(out)
-
-  # Call the function with y specified
-  out, = imported_function(mx.array(1.0), y=mx.array(1.0))
-  print(out)
-
-In the above example the function constant data, (i.e. ``constant``), is only
-saved once. 
-
-Transformations with Imported Functions
----------------------------------------
-
-Function transformations like :func:`grad`, :func:`vmap`, and :func:`compile` work
-on imported functions just like regular Python functions:
-
-.. code-block:: python
-
-  def fun(x):
-      return mx.sin(x)
-
-  x = mx.array(0.0)
-  mx.export_function("sine.mlxfn", fun, x)
-
-  imported_fun = mx.import_function("sine.mlxfn")
-
-  # Take the derivative of the imported function
-  dfdx = mx.grad(lambda x: imported_fun(x)[0])
-  # Prints: array(1, dtype=float32)
-  print(dfdx(x))
-
-  # Compile the imported function 
-  mx.compile(imported_fun)
-  # Prints: array(0, dtype=float32)
-  print(compiled_fun(x)[0])
-
-
-Importing Functions in C++
---------------------------
-
-Importing and running functions in C++ is basically the same as importing and
-running them in Python. First, follow the :ref:`instructions <mlx_in_cpp>` to
-setup a simple C++ project that uses MLX as a library.
-
-Next, export a simple function from Python:
-
-.. code-block:: python
-
-  def fun(x, y):
-      return mx.exp(x + y)
-
-  x = mx.array(1.0)
-  y = mx.array(1.0)
-  mx.export_function("fun.mlxfn", fun, x, y)
-
-
-Import and run the function in C++ with only a few lines of code:
-
-.. code-block:: c++
-
-  auto fun = mx::import_function("fun.mlxfn");
-
-  auto inputs = {mx::array(1.0), mx::array(1.0)};
-  auto outputs = fun(inputs);
-
-  // Prints: array(2, dtype=float32)
-  std::cout << outputs[0] << std::endl;
-
-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++.
-
-More Examples
--------------
-
-Here are a few more complete examples exporting more complex functions from
-Python and importing and running them in C++:
-
-* `Inference and training a multi-layer perceptron <https://github.com/ml-explore/mlx/tree/main/examples/export>`_

docs/src/usage/function_transforms.rst

@@ -184,8 +184,8 @@ Let's time these two different versions:
   print(timeit.timeit(lambda: mx.eval(naive_add(xs, ys)), number=100))
   print(timeit.timeit(lambda: mx.eval(vmap_add(xs, ys)), number=100))
 
-On an M1 Max the naive version takes in total ``5.639`` seconds whereas the
-vectorized version takes only ``0.024`` seconds, more than 200 times faster.
+On an M1 Max the naive version takes in total ``0.390`` seconds whereas the
+vectorized version takes only ``0.025`` seconds, more than ten times faster.
 
 Of course, this operation is quite contrived. A better approach is to simply do
 ``xs + ys.T``, but for more complex functions :func:`vmap` can be quite handy.

docs/src/usage/indexing.rst

@@ -107,16 +107,6 @@ 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:
 

docs/src/usage/launching_distributed.rst

@@ -1,105 +0,0 @@
-:orphan:
-
-.. _usage_launch_distributed:
-
-Launching Distributed Programs
-==============================
-
-.. currentmodule:: mlx.core.distributed
-
-Installing the MLX python package provides a helper script ``mlx.launch`` that
-can be used to run python scripts distributed on several nodes. It allows
-launching using either the MPI backend or the ring backend. See the
-:doc:`distributed docs <distributed>` for the different backends.
-
-Usage
------
-
-The minimal usage example of ``mlx.launch`` is simply
-
-.. code:: shell
-
-    mlx.launch --hosts ip1,ip2 my_script.py
-
-or for testing on localhost
-
-.. code:: shell
-
-    mlx.launch -n 2 my_script.py
-
-The ``mlx.launch`` command connects to the provided host and launches the input
-script on each host. It monitors each of the launched processes and terminates
-the rest if one of them fails unexpectedly or if ``mlx.launch`` is terminated.
-It also takes care of forwarding the output of each remote process to stdout
-and stderr respectively.
-
-Providing Hosts
-^^^^^^^^^^^^^^^^
-
-Hosts can be provided as command line arguments, like above, but the way that
-allows to fully define a list of hosts is via a JSON hostfile. The hostfile has
-a very simple schema. It is simply a list of objects that define each host via
-a hostname to ssh to and a list of IPs to utilize for the communication.
-
-.. code:: json
-
-    [
-        {"ssh": "hostname1", "ips": ["123.123.1.1", "123.123.2.1"]},
-        {"ssh": "hostname2", "ips": ["123.123.1.2", "123.123.2.2"]}
-    ]
-
-You can use ``mlx.distributed_config --over ethernet`` to create a hostfile
-with IPs corresponding to the ``en0`` interface.
-
-Setting up Remote Hosts
-^^^^^^^^^^^^^^^^^^^^^^^^
-
-In order to be able to launch the script on each host we need to be able to
-connect via ssh. Moreover the input script and python binary need to be on each
-host and on the same path. A good checklist to debug errors is the following:
-
-* ``ssh hostname`` works without asking for password or host confirmation
-* the python binary is available on all hosts at the same path. You can use
-  ``mlx.launch --print-python`` to see what that path is.
-* the script you want to run is available on all hosts at the same path
-
-.. _mpi_specifics:
-
-MPI Specifics
--------------
-
-One can use MPI by passing ``--backend mpi`` to ``mlx.launch``. In that case,
-``mlx.launch`` is a thin wrapper over ``mpirun``. Moreover,
-
-* The IPs in the hostfile are ignored
-* The ssh connectivity requirement is stronger as every node needs to be able
-  to connect to every other node
-* ``mpirun`` needs to be available on every node at the same path
-
-Finally, one can pass arguments to ``mpirun`` using ``--mpi-arg``. For instance
-to choose a specific interface for the byte-transfer-layer of MPI we can call
-``mlx.launch`` as follows:
-
-.. code:: shell
-
-    mlx.launch --backend mpi --mpi-arg '--mca btl_tcp_if_include en0' --hostfile hosts.json my_script.py
-
-
-.. _ring_specifics:
-
-Ring Specifics
---------------
-
-The ring backend, which is also the default backend, can be explicitly selected
-with the argument ``--backend ring``. The ring backend has some specific
-requirements and arguments that are different to MPI:
-
-* The argument ``--hosts`` only accepts IPs and not hostnames. If we need to
-  ssh to a hostname that does not correspond to the IP we want to bind to we
-  have to provide a hostfile.
-* ``--starting-port`` defines the port to bind to on the remote hosts.
-  Specifically rank 0 for the first IP will use this port and each subsequent
-  IP or rank will add 1 to this port.
-* ``--connections-per-ip`` allows us to increase the number of connections
-  between neighboring nodes. This corresponds to ``--mca btl_tcp_links 2`` for
-  ``mpirun``.

docs/src/usage/numpy.rst

@@ -21,13 +21,11 @@ Let's convert an array to NumPy and back.
 
 .. note::
 
-    Since NumPy does not support ``bfloat16`` arrays, you will need to convert
-    to ``float16`` or ``float32`` first: ``np.array(a.astype(mx.float32))``.
-    Otherwise, you will receive an error like: ``Item size 2 for PEP 3118
-    buffer format string does not match the dtype V item size 0.``
+    Since NumPy does not support ``bfloat16`` arrays, you will need to convert to ``float16`` or ``float32`` first:
+    ``np.array(a.astype(mx.float32))``.
+    Otherwise, you will receive an error like: ``Item size 2 for PEP 3118 buffer format string does not match the dtype V item size 0.``
 
-By default, NumPy copies data to a new array. This can be prevented by creating
-an array view:
+By default, NumPy copies data to a new array. This can be prevented by creating an array view:
 
 .. code-block:: python
 
@@ -37,16 +35,10 @@ an array view:
   a_view[0] = 1
   print(a[0].item()) # 1
 
-.. note::
+A NumPy array view is a normal NumPy array, except that it does not own its memory.
+This means writing to the view is reflected in the original array.
 
-    NumPy arrays with type ``float64`` will be default converted to MLX arrays
-    with type ``float32``.
-
-A NumPy array view is a normal NumPy array, except that it does not own its
-memory. This means writing to the view is reflected in the original array.
-
-While this is quite powerful to prevent copying arrays, it should be noted that
-external changes to the memory of arrays cannot be reflected in gradients.
+While this is quite powerful to prevent copying arrays, it should be noted that external changes to the memory of arrays cannot be reflected in gradients.
 
 Let's demonstrate this in an example:
 
@@ -64,12 +56,11 @@ Let's demonstrate this in an example:
 
 
 The function ``f`` indirectly modifies the array ``x`` through a memory view.
-However, this modification is not reflected in the gradient, as seen in the
-last line outputting ``1.0``, representing the gradient of the sum operation
-alone.  The squaring of ``x`` occurs externally to MLX, meaning that no
-gradient is incorporated.  It's important to note that a similar issue arises
-during array conversion and copying.  For instance, a function defined as
-``mx.array(np.array(x)**2).sum()`` would also result in an incorrect gradient,
+However, this modification is not reflected in the gradient, as seen in the last line outputting ``1.0``,
+representing the gradient of the sum operation alone.
+The squaring of ``x`` occurs externally to MLX, meaning that no gradient is incorporated.
+It's important to note that a similar issue arises during array conversion and copying.
+For instance, a function defined as ``mx.array(np.array(x)**2).sum()`` would also result in an incorrect gradient,
 even though no in-place operations on MLX memory are executed.
 
 PyTorch
@@ -80,8 +71,7 @@ PyTorch
    PyTorch Support for :obj:`memoryview` is experimental and can break for
    multi-dimensional arrays. Casting to NumPy first is advised for now.
 
-PyTorch supports the buffer protocol, but it requires an explicit
-:obj:`memoryview`.
+PyTorch supports the buffer protocol, but it requires an explicit :obj:`memoryview`.
 
 .. code-block:: python
 
@@ -92,8 +82,7 @@ PyTorch supports the buffer protocol, but it requires an explicit
   b = torch.tensor(memoryview(a))
   c = mx.array(b.numpy())
 
-Conversion from PyTorch tensors back to arrays must be done via intermediate
-NumPy arrays with ``numpy()``.
+Conversion from PyTorch tensors back to arrays must be done via intermediate NumPy arrays with ``numpy()``.
 
 JAX
 ---
@@ -111,8 +100,7 @@ JAX fully supports the buffer protocol.
 TensorFlow
 ----------
 
-TensorFlow supports the buffer protocol, but it requires an explicit
-:obj:`memoryview`.
+TensorFlow supports the buffer protocol, but it requires an explicit :obj:`memoryview`.
 
 .. code-block:: python
 

examples/cmake_project/CMakeLists.txt

@@ -1,22 +0,0 @@
-cmake_minimum_required(VERSION 3.27)
-
-project(example LANGUAGES CXX)
-
-set(CMAKE_CXX_STANDARD 17)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
-
-# Comment the following two commands only the MLX C++ library is installed and
-# set(MLX_ROOT "/path/to/mlx") directly if needed.
-find_package(
-  Python 3.9
-  COMPONENTS Interpreter Development.Module
-  REQUIRED)
-execute_process(
-  COMMAND "${Python_EXECUTABLE}" -m mlx --cmake-dir
-  OUTPUT_STRIP_TRAILING_WHITESPACE
-  OUTPUT_VARIABLE MLX_ROOT)
-
-find_package(MLX CONFIG REQUIRED)
-
-add_executable(example example.cpp)
-target_link_libraries(example PRIVATE mlx)

examples/cmake_project/README.md

@@ -1,26 +0,0 @@
-## Build and Run 
-
-Install MLX with Python:
-
-```bash
-pip install mlx>=0.22
-```
-
-Build the C++ example:
-
-```bash
-cmake -B build -DCMAKE_BUILD_TYPE=Release
-cmake --build build
-```
-
-Run the C++ example:
-
-```
-./build/example
-```
-
-which should output:
-
-```
-array([2, 4, 6], dtype=int32)
-```

examples/cmake_project/example.cpp

@@ -1,14 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include <iostream>
-
-#include "mlx/mlx.h"
-
-namespace mx = mlx::core;
-
-int main() {
-  auto x = mx::array({1, 2, 3});
-  auto y = mx::array({1, 2, 3});
-  std::cout << x + y << std::endl;
-  return 0;
-}

examples/cpp/distributed.cpp

@@ -4,19 +4,19 @@
 
 #include "mlx/mlx.h"
 
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 int main() {
-  if (!mx::distributed::is_available()) {
+  if (!distributed::is_available()) {
     std::cout << "No communication backend found" << std::endl;
     return 1;
   }
 
-  auto global_group = mx::distributed::init();
+  auto global_group = distributed::init();
   std::cout << global_group.rank() << " / " << global_group.size() << std::endl;
 
-  mx::array x = mx::ones({10});
-  mx::array out = mx::distributed::all_sum(x, global_group);
+  array x = ones({10});
+  array out = distributed::all_sum(x, global_group);
 
   std::cout << out << std::endl;
 }

examples/cpp/linear_regression.cpp

@@ -10,7 +10,7 @@
 /**
  * An example of linear regression with MLX.
  */
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 int main() {
   int num_features = 100;
@@ -19,35 +19,35 @@ int main() {
   float learning_rate = 0.01;
 
   // True parameters
-  auto w_star = mx::random::normal({num_features});
+  auto w_star = random::normal({num_features});
 
   // The input examples (design matrix)
-  auto X = mx::random::normal({num_examples, num_features});
+  auto X = random::normal({num_examples, num_features});
 
   // Noisy labels
-  auto eps = 1e-2 * mx::random::normal({num_examples});
-  auto y = mx::matmul(X, w_star) + eps;
+  auto eps = 1e-2 * random::normal({num_examples});
+  auto y = matmul(X, w_star) + eps;
 
   // Initialize random parameters
-  mx::array w = 1e-2 * mx::random::normal({num_features});
+  array w = 1e-2 * random::normal({num_features});
 
-  auto loss_fn = [&](mx::array w) {
-    auto yhat = mx::matmul(X, w);
-    return (0.5f / num_examples) * mx::sum(mx::square(yhat - y));
+  auto loss_fn = [&](array w) {
+    auto yhat = matmul(X, w);
+    return (0.5f / num_examples) * sum(square(yhat - y));
   };
 
-  auto grad_fn = mx::grad(loss_fn);
+  auto grad_fn = grad(loss_fn);
 
   auto tic = timer::time();
   for (int it = 0; it < num_iters; ++it) {
-    auto grads = grad_fn(w);
-    w = w - learning_rate * grads;
-    mx::eval(w);
+    auto grad = grad_fn(w);
+    w = w - learning_rate * grad;
+    eval(w);
   }
   auto toc = timer::time();
 
   auto loss = loss_fn(w);
-  auto error_norm = std::sqrt(mx::sum(mx::square(w - w_star)).item<float>());
+  auto error_norm = std::sqrt(sum(square(w - w_star)).item<float>());
   auto throughput = num_iters / timer::seconds(toc - tic);
   std::cout << "Loss " << loss << ", |w - w*| = " << error_norm
             << ", Throughput " << throughput << " (it/s)." << std::endl;

examples/cpp/logistic_regression.cpp

@@ -10,7 +10,7 @@
 /**
  * An example of logistic regression with MLX.
  */
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 int main() {
   int num_features = 100;
@@ -19,35 +19,35 @@ int main() {
   float learning_rate = 0.1;
 
   // True parameters
-  auto w_star = mx::random::normal({num_features});
+  auto w_star = random::normal({num_features});
 
   // The input examples
-  auto X = mx::random::normal({num_examples, num_features});
+  auto X = random::normal({num_examples, num_features});
 
   // Labels
-  auto y = mx::matmul(X, w_star) > 0;
+  auto y = matmul(X, w_star) > 0;
 
   // Initialize random parameters
-  mx::array w = 1e-2 * mx::random::normal({num_features});
+  array w = 1e-2 * random::normal({num_features});
 
-  auto loss_fn = [&](mx::array w) {
-    auto logits = mx::matmul(X, w);
+  auto loss_fn = [&](array w) {
+    auto logits = matmul(X, w);
     auto scale = (1.0f / num_examples);
-    return scale * mx::sum(mx::logaddexp(mx::array(0.0f), logits) - y * logits);
+    return scale * sum(logaddexp(array(0.0f), logits) - y * logits);
   };
 
-  auto grad_fn = mx::grad(loss_fn);
+  auto grad_fn = grad(loss_fn);
 
   auto tic = timer::time();
   for (int it = 0; it < num_iters; ++it) {
-    auto grads = grad_fn(w);
-    w = w - learning_rate * grads;
-    mx::eval(w);
+    auto grad = grad_fn(w);
+    w = w - learning_rate * grad;
+    eval(w);
   }
   auto toc = timer::time();
 
   auto loss = loss_fn(w);
-  auto acc = mx::sum((mx::matmul(X, w) > 0) == y) / num_examples;
+  auto acc = sum((matmul(X, w) > 0) == y) / num_examples;
   auto throughput = num_iters / timer::seconds(toc - tic);
   std::cout << "Loss " << loss << ", Accuracy, " << acc << ", Throughput "
             << throughput << " (it/s)." << std::endl;

examples/cpp/metal_capture.cpp

@@ -5,27 +5,27 @@
 
 #include "mlx/mlx.h"
 
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 int main() {
   // To use Metal debugging and profiling:
   // 1. Build with the MLX_METAL_DEBUG CMake option (i.e. -DMLX_METAL_DEBUG=ON).
   // 2. Run with MTL_CAPTURE_ENABLED=1.
-  mx::metal::start_capture("mlx_trace.gputrace");
+  metal::start_capture("mlx_trace.gputrace");
 
   // Start at index two because the default GPU and CPU streams have indices
   // zero and one, respectively. This naming matches the label assigned to each
   // stream's command queue.
-  auto s2 = new_stream(mx::Device::gpu);
-  auto s3 = new_stream(mx::Device::gpu);
+  auto s2 = new_stream(Device::gpu);
+  auto s3 = new_stream(Device::gpu);
 
-  auto a = mx::arange(1.f, 10.f, 1.f, mx::float32, s2);
-  auto b = mx::arange(1.f, 10.f, 1.f, mx::float32, s3);
-  auto x = mx::add(a, a, s2);
-  auto y = mx::add(b, b, s3);
+  auto a = arange(1.f, 10.f, 1.f, float32, s2);
+  auto b = arange(1.f, 10.f, 1.f, float32, s3);
+  auto x = add(a, a, s2);
+  auto y = add(b, b, s3);
 
   // The multiply will happen on the default stream.
-  std::cout << mx::multiply(x, y) << std::endl;
+  std::cout << multiply(x, y) << std::endl;
 
-  mx::metal::stop_capture();
+  metal::stop_capture();
 }

examples/cpp/tutorial.cpp

@@ -5,11 +5,11 @@
 
 #include "mlx/mlx.h"
 
-namespace mx = mlx::core;
+using namespace mlx::core;
 
 void array_basics() {
   // Make a scalar array:
-  mx::array x(1.0);
+  array x(1.0);
 
   // Get the value out of it:
   auto s = x.item<float>();
@@ -29,31 +29,31 @@ void array_basics() {
 
   // The datatype should be float32:
   auto dtype = x.dtype();
-  assert(dtype == mx::float32);
+  assert(dtype == float32);
 
   // Specify the dtype when constructing the array:
-  x = mx::array(1, mx::int32);
-  assert(x.dtype() == mx::int32);
+  x = array(1, int32);
+  assert(x.dtype() == int32);
   x.item<int>(); // OK
   // x.item<float>();  // Undefined!
 
   // Make a multidimensional array:
-  x = mx::array({1.0f, 2.0f, 3.0f, 4.0f}, {2, 2});
+  x = array({1.0f, 2.0f, 3.0f, 4.0f}, {2, 2});
   // mlx is row-major by default so the first row of this array
   // is [1.0, 2.0] and the second row is [3.0, 4.0]
 
   // Make an array of shape {2, 2} filled with ones:
-  auto y = mx::ones({2, 2});
+  auto y = ones({2, 2});
 
   // Pointwise add x and y:
-  auto z = mx::add(x, y);
+  auto z = add(x, y);
 
   // Same thing:
   z = x + y;
 
   // mlx is lazy by default. At this point `z` only
   // has a shape and a type but no actual data:
-  assert(z.dtype() == mx::float32);
+  assert(z.dtype() == float32);
   assert(z.shape(0) == 2);
   assert(z.shape(1) == 2);
 
@@ -63,33 +63,33 @@ void array_basics() {
   // and inputs. When `eval` is called on an array (or arrays), the array and
   // all of its dependencies are recursively evaluated to produce the result.
   // Once an array is evaluated, it has data and is detached from its inputs.
-  mx::eval(z);
+  eval(z);
 
-  // Of course the array can still be an input to other operations. You can
-  // even call eval on the array again, this will just be a no-op:
-  mx::eval(z); // no-op
+  // Of course the array can still be an input to other operations. You can even
+  // call eval on the array again, this will just be a no-op:
+  eval(z); // no-op
 
   // Some functions or methods on arrays implicitly evaluate them. For example
   // accessing a value in an array or printing the array implicitly evaluate it:
-  z = mx::ones({1});
+  z = ones({1});
   z.item<float>(); // implicit evaluation
 
-  z = mx::ones({2, 2});
+  z = ones({2, 2});
   std::cout << z << std::endl; // implicit evaluation
 }
 
 void automatic_differentiation() {
-  auto fn = [](mx::array x) { return mx::square(x); };
+  auto fn = [](array x) { return square(x); };
 
   // Computing the derivative function of a function
-  auto grad_fn = mx::grad(fn);
+  auto grad_fn = grad(fn);
   // Call grad_fn on the input to get the derivative
-  auto x = mx::array(1.5);
+  auto x = array(1.5);
   auto dfdx = grad_fn(x);
   // dfdx is 2 * x
 
   // Get the second derivative by composing grad with grad
-  auto d2fdx2 = mx::grad(mx::grad(fn))(x);
+  auto d2fdx2 = grad(grad(fn))(x);
   // d2fdx2 is 2
 }
 

examples/export/CMakeLists.txt

@@ -1,22 +0,0 @@
-cmake_minimum_required(VERSION 3.27)
-
-project(import_mlx LANGUAGES CXX)
-
-set(CMAKE_CXX_STANDARD 17)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
-
-find_package(
-  Python 3.9
-  COMPONENTS Interpreter Development.Module
-  REQUIRED)
-execute_process(
-  COMMAND "${Python_EXECUTABLE}" -m mlx --cmake-dir
-  OUTPUT_STRIP_TRAILING_WHITESPACE
-  OUTPUT_VARIABLE MLX_ROOT)
-find_package(MLX CONFIG REQUIRED)
-
-add_executable(eval_mlp eval_mlp.cpp)
-target_link_libraries(eval_mlp PRIVATE mlx)
-
-add_executable(train_mlp train_mlp.cpp)
-target_link_libraries(train_mlp PRIVATE mlx)

examples/export/README.md

@@ -1,49 +0,0 @@
-## Setup
-
-Install MLX:
-
-```bash
-pip install mlx>=0.22
-```
-
-Build the C++ examples:
-
-```bash
-cmake -B build -DCMAKE_BUILD_TYPE=Release
-cmake --build build
-```
-
-## Run
-
-### Eval MLP
-
-Run the Python script to export the eval function:
-
-```bash
-python eval_mlp.py
-```
-
-Then run the C++ program to import and run the function:
-
-```
-./build/eval_mlp
-```
-
-The Python and C++ programs should output the same result.
-
-### Train MLP
-
-Run the Python script to export the model initialization and training
-functions:
-
-```bash
-python train_mlp.py
-```
-
-Then run the C++ program to import and run the functions:
-
-```
-./build/train_mlp
-```
-
-The Python and C++ programs should output the same results.

examples/export/eval_mlp.cpp

@@ -1,25 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include <mlx/mlx.h>
-#include <iostream>
-
-namespace mx = mlx::core;
-
-int main() {
-  int batch_size = 8;
-  int input_dim = 32;
-
-  // Make the input
-  mx::random::seed(42);
-  auto example_x = mx::random::uniform({batch_size, input_dim});
-
-  // Import the function
-  auto forward = mx::import_function("eval_mlp.mlxfn");
-
-  // Call the imported function
-  auto out = forward({example_x})[0];
-
-  std::cout << out << std::endl;
-
-  return 0;
-}

examples/export/eval_mlp.py

@@ -1,52 +0,0 @@
-# Copyright © 2024 Apple Inc.
-
-import mlx.core as mx
-import mlx.nn as nn
-import mlx.utils
-
-
-class MLP(nn.Module):
-    """A simple MLP."""
-
-    def __init__(
-        self, num_layers: int, input_dim: int, hidden_dim: int, output_dim: int
-    ):
-        super().__init__()
-        layer_sizes = [input_dim] + [hidden_dim] * num_layers + [output_dim]
-        self.layers = [
-            nn.Linear(idim, odim)
-            for idim, odim in zip(layer_sizes[:-1], layer_sizes[1:])
-        ]
-
-    def __call__(self, x):
-        for l in self.layers[:-1]:
-            x = nn.relu(l(x))
-        return self.layers[-1](x)
-
-
-if __name__ == "__main__":
-
-    batch_size = 8
-    input_dim = 32
-    output_dim = 10
-
-    # Load the model
-    mx.random.seed(0)  # Seed for params
-    model = MLP(num_layers=5, input_dim=input_dim, hidden_dim=64, output_dim=output_dim)
-    mx.eval(model)
-
-    # Note, the model parameters are saved in the export function
-    def forward(x):
-        return model(x)
-
-    mx.random.seed(42)  # Seed for input
-    example_x = mx.random.uniform(shape=(batch_size, input_dim))
-
-    mx.export_function("eval_mlp.mlxfn", forward, example_x)
-
-    # Import in Python
-    imported_forward = mx.import_function("eval_mlp.mlxfn")
-    expected = forward(example_x)
-    (out,) = imported_forward(example_x)
-    assert mx.allclose(expected, out)
-    print(out)

examples/export/train_mlp.cpp

@@ -1,35 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include <mlx/mlx.h>
-#include <iostream>
-
-namespace mx = mlx::core;
-
-int main() {
-  int batch_size = 8;
-  int input_dim = 32;
-  int output_dim = 10;
-
-  auto state = mx::import_function("init_mlp.mlxfn")({});
-
-  // Make the input
-  mx::random::seed(42);
-  auto example_X = mx::random::normal({batch_size, input_dim});
-  auto example_y = mx::random::randint(0, output_dim, {batch_size});
-
-  // Import the function
-  auto step = mx::import_function("train_mlp.mlxfn");
-
-  // Call the imported function
-  for (int it = 0; it < 100; ++it) {
-    state.insert(state.end(), {example_X, example_y});
-    state = step(state);
-    eval(state);
-    auto loss = state.back();
-    state.pop_back();
-    if (it % 10 == 0) {
-      std::cout << "Loss " << loss.item<float>() << std::endl;
-    }
-  }
-  return 0;
-}

examples/export/train_mlp.py

@@ -1,76 +0,0 @@
-# Copyright © 2024 Apple Inc.
-
-import mlx.core as mx
-import mlx.nn as nn
-import mlx.optimizers as optim
-import mlx.utils
-
-
-class MLP(nn.Module):
-    """A simple MLP."""
-
-    def __init__(
-        self, num_layers: int, input_dim: int, hidden_dim: int, output_dim: int
-    ):
-        super().__init__()
-        layer_sizes = [input_dim] + [hidden_dim] * num_layers + [output_dim]
-        self.layers = [
-            nn.Linear(idim, odim)
-            for idim, odim in zip(layer_sizes[:-1], layer_sizes[1:])
-        ]
-
-    def __call__(self, x):
-        for l in self.layers[:-1]:
-            x = nn.relu(l(x))
-        return self.layers[-1](x)
-
-
-if __name__ == "__main__":
-
-    batch_size = 8
-    input_dim = 32
-    output_dim = 10
-
-    def init():
-        # Seed for the parameter initialization
-        mx.random.seed(0)
-        model = MLP(
-            num_layers=3, input_dim=input_dim, hidden_dim=64, output_dim=output_dim
-        )
-        optimizer = optim.SGD(learning_rate=1e-1)
-        optimizer.init(model.parameters())
-        state = [model.parameters(), optimizer.state]
-        tree_structure, state = zip(*mlx.utils.tree_flatten(state))
-        return model, optimizer, tree_structure, state
-
-    # Export the model parameter initialization
-    model, optimizer, tree_structure, state = init()
-    mx.eval(state)
-    mx.export_function("init_mlp.mlxfn", lambda: init()[-1])
-
-    def loss_fn(params, X, y):
-        model.update(params)
-        return nn.losses.cross_entropy(model(X), y, reduction="mean")
-
-    def step(*inputs):
-        *state, X, y = inputs
-        params, opt_state = mlx.utils.tree_unflatten(list(zip(tree_structure, state)))
-        optimizer.state = opt_state
-        loss, grads = mx.value_and_grad(loss_fn)(params, X, y)
-        params = optimizer.apply_gradients(grads, params)
-        _, state = zip(*mlx.utils.tree_flatten([params, optimizer.state]))
-        return *state, loss
-
-    # Make some random data
-    mx.random.seed(42)
-    example_X = mx.random.normal(shape=(batch_size, input_dim))
-    example_y = mx.random.randint(low=0, high=output_dim, shape=(batch_size,))
-    mx.export_function("train_mlp.mlxfn", step, *state, example_X, example_y)
-
-    # Export one step of SGD
-    imported_step = mx.import_function("train_mlp.mlxfn")
-
-    for it in range(100):
-        *state, loss = imported_step(*state, example_X, example_y)
-        if it % 10 == 0:
-            print(f"Loss {loss.item():.6}")

examples/extensions/CMakeLists.txt

@@ -10,6 +10,7 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 option(BUILD_SHARED_LIBS "Build extensions as a shared library" ON)
 
 # ----------------------------- Dependencies -----------------------------
+find_package(MLX CONFIG REQUIRED)
 find_package(
   Python 3.8
   COMPONENTS Interpreter Development.Module
@@ -17,15 +18,10 @@ find_package(
 execute_process(
   COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
   OUTPUT_STRIP_TRAILING_WHITESPACE
-  OUTPUT_VARIABLE nanobind_ROOT)
+  OUTPUT_VARIABLE NB_DIR)
+list(APPEND CMAKE_PREFIX_PATH "${NB_DIR}")
 find_package(nanobind CONFIG REQUIRED)
 
-execute_process(
-  COMMAND "${Python_EXECUTABLE}" -m mlx --cmake-dir
-  OUTPUT_STRIP_TRAILING_WHITESPACE
-  OUTPUT_VARIABLE MLX_ROOT)
-find_package(MLX CONFIG REQUIRED)
-
 # ----------------------------- Extensions -----------------------------
 
 # Add library

examples/extensions/axpby/axpby.cpp

@@ -1,20 +1,25 @@
-// Copyright © 2023-2025 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
+#include <cassert>
 #include <iostream>
 #include <sstream>
 
+#include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/cpu/encoder.h"
 #include "mlx/utils.h"
 
 #include "axpby/axpby.h"
 
+#ifdef ACCELERATE_NEW_LAPACK
+#include <vecLib/cblas_new.h>
+#endif
+
 #ifdef _METAL_
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/utils.h"
 #endif
 
-namespace my_ext {
+namespace mlx::core {
 
 ///////////////////////////////////////////////////////////////////////////////
 // Operation Implementation
@@ -27,24 +32,24 @@ namespace my_ext {
  *  Follow numpy style broadcasting between x and y
  *  Inputs are upcasted to floats if needed
  **/
-mx::array axpby(
-    const mx::array& x, // Input mx::array x
-    const mx::array& y, // Input mx::array y
+array axpby(
+    const array& x, // Input array x
+    const array& y, // Input array y
     const float alpha, // Scaling factor for x
     const float beta, // Scaling factor for y
-    mx::StreamOrDevice s /* = {} */ // Stream on which to schedule the operation
+    StreamOrDevice s /* = {} */ // Stream on which to schedule the operation
 ) {
   // Promote dtypes between x and y as needed
   auto promoted_dtype = promote_types(x.dtype(), y.dtype());
 
   // Upcast to float32 for non-floating point inputs x and y
-  auto out_dtype = mx::issubdtype(promoted_dtype, mx::float32)
+  auto out_dtype = issubdtype(promoted_dtype, float32)
       ? promoted_dtype
-      : promote_types(promoted_dtype, mx::float32);
+      : promote_types(promoted_dtype, float32);
 
   // Cast x and y up to the determined dtype (on the same stream s)
-  auto x_casted = mx::astype(x, out_dtype, s);
-  auto y_casted = mx::astype(y, out_dtype, s);
+  auto x_casted = astype(x, out_dtype, s);
+  auto y_casted = astype(y, out_dtype, s);
 
   // Broadcast the shapes of x and y (on the same stream s)
   auto broadcasted_inputs = broadcast_arrays({x_casted, y_casted}, s);
@@ -52,12 +57,12 @@ mx::array axpby(
 
   // Construct the array as the output of the Axpby primitive
   // with the broadcasted and upcasted arrays as inputs
-  return mx::array(
-      /* const mx::Shape& shape = */ out_shape,
-      /* mx::Dtype dtype = */ out_dtype,
-      /* std::shared_ptr<mx::Primitive> primitive = */
+  return array(
+      /* const std::vector<int>& shape = */ out_shape,
+      /* Dtype dtype = */ out_dtype,
+      /* std::unique_ptr<Primitive> primitive = */
       std::make_shared<Axpby>(to_stream(s), alpha, beta),
-      /* const std::vector<mx::array>& inputs = */ broadcasted_inputs);
+      /* const std::vector<array>& inputs = */ broadcasted_inputs);
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -66,69 +71,140 @@ mx::array axpby(
 
 template <typename T>
 void axpby_impl(
-    const mx::array& x,
-    const mx::array& y,
-    mx::array& out,
+    const array& x,
+    const array& y,
+    array& out,
     float alpha_,
-    float beta_,
-    mx::Stream stream) {
-  out.set_data(mx::allocator::malloc(out.nbytes()));
+    float beta_) {
+  // We only allocate memory when we are ready to fill the output
+  // malloc_or_wait synchronously allocates available memory
+  // There may be a wait executed here if the allocation is requested
+  // under memory-pressured conditions
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
-  // Get the CPU command encoder and register input and output arrays
-  auto& encoder = mx::cpu::get_command_encoder(stream);
-  encoder.set_input_array(x);
-  encoder.set_input_array(y);
-  encoder.set_output_array(out);
+  // Collect input and output data pointers
+  const T* x_ptr = x.data<T>();
+  const T* y_ptr = y.data<T>();
+  T* out_ptr = out.data<T>();
 
-  // Launch the CPU kernel
-  encoder.dispatch([x_ptr = x.data<T>(),
-                    y_ptr = y.data<T>(),
-                    out_ptr = out.data<T>(),
-                    size = out.size(),
-                    shape = out.shape(),
-                    x_strides = x.strides(),
-                    y_strides = y.strides(),
-                    alpha_,
-                    beta_]() {
-    // Cast alpha and beta to the relevant types
-    T alpha = static_cast<T>(alpha_);
-    T beta = static_cast<T>(beta_);
+  // Cast alpha and beta to the relevant types
+  T alpha = static_cast<T>(alpha_);
+  T beta = static_cast<T>(beta_);
 
-    // Do the element-wise operation for each output
-    for (size_t out_idx = 0; out_idx < size; out_idx++) {
-      // Map linear indices to offsets in x and y
-      auto x_offset = mx::elem_to_loc(out_idx, shape, x_strides);
-      auto y_offset = mx::elem_to_loc(out_idx, shape, y_strides);
+  // Do the element-wise operation for each output
+  for (size_t out_idx = 0; out_idx < out.size(); out_idx++) {
+    // Map linear indices to offsets in x and y
+    auto x_offset = elem_to_loc(out_idx, x.shape(), x.strides());
+    auto y_offset = elem_to_loc(out_idx, y.shape(), y.strides());
 
-      // We allocate the output to be contiguous and regularly strided
-      // (defaults to row major) and hence it doesn't need additional mapping
-      out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
-    }
-  });
+    // We allocate the output to be contiguous and regularly strided
+    // (defaults to row major) and hence it doesn't need additional mapping
+    out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
+  }
 }
 
-void Axpby::eval_cpu(
-    const std::vector<mx::array>& inputs,
-    std::vector<mx::array>& outputs) {
+/** Fall back implementation for evaluation on CPU */
+void Axpby::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  // Check the inputs (registered in the op while constructing the out array)
+  assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
   auto& out = outputs[0];
 
   // Dispatch to the correct dtype
-  if (out.dtype() == mx::float32) {
-    return axpby_impl<float>(x, y, out, alpha_, beta_, stream());
-  } else if (out.dtype() == mx::float16) {
-    return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_, stream());
-  } else if (out.dtype() == mx::bfloat16) {
-    return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_, stream());
-  } else if (out.dtype() == mx::complex64) {
-    return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_, stream());
+  if (out.dtype() == float32) {
+    return axpby_impl<float>(x, y, out, alpha_, beta_);
+  } else if (out.dtype() == float16) {
+    return axpby_impl<float16_t>(x, y, out, alpha_, beta_);
+  } else if (out.dtype() == bfloat16) {
+    return axpby_impl<bfloat16_t>(x, y, out, alpha_, beta_);
+  } else if (out.dtype() == complex64) {
+    return axpby_impl<complex64_t>(x, y, out, alpha_, beta_);
   } else {
     throw std::runtime_error(
         "Axpby is only supported for floating point types.");
   }
 }
 
+///////////////////////////////////////////////////////////////////////////////
+// Primitive Accelerate Backend Implementation
+///////////////////////////////////////////////////////////////////////////////
+
+#ifdef ACCELERATE_NEW_LAPACK
+
+template <typename T>
+void axpby_impl_accelerate(
+    const array& x,
+    const array& y,
+    array& out,
+    float alpha_,
+    float beta_) {
+  // Accelerate library provides catlas_saxpby which does
+  // Y = (alpha * X) + (beta * Y) in place
+  // To use it, we first copy the data in y over to the output array
+
+  // This specialization requires both x and y be contiguous in the same mode
+  // i.e: corresponding linear indices in both point to corresponding elements
+  // The data in the output array is allocated to match the strides in y
+  // such that x, y, and out are contiguous in the same mode and
+  // no transposition is needed
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  // We then copy over the elements using the contiguous vector specialization
+  copy_inplace(y, out, CopyType::Vector);
+
+  // Get x and y pointers for catlas_saxpby
+  const T* x_ptr = x.data<T>();
+  T* y_ptr = out.data<T>();
+
+  T alpha = static_cast<T>(alpha_);
+  T beta = static_cast<T>(beta_);
+
+  // Call the inplace accelerate operator
+  catlas_saxpby(
+      /* N = */ out.size(),
+      /* ALPHA = */ alpha,
+      /* X = */ x_ptr,
+      /* INCX = */ 1,
+      /* BETA = */ beta,
+      /* Y = */ y_ptr,
+      /* INCY = */ 1);
+}
+
+/** Evaluate primitive on CPU using accelerate specializations */
+void Axpby::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 2);
+  auto& x = inputs[0];
+  auto& y = inputs[1];
+  auto& out = outputs[0];
+
+  // Accelerate specialization for contiguous single precision float arrays
+  if (out.dtype() == float32 &&
+      ((x.flags().row_contiguous && y.flags().row_contiguous) ||
+       (x.flags().col_contiguous && y.flags().col_contiguous))) {
+    axpby_impl_accelerate<float>(x, y, out, alpha_, beta_);
+    return;
+  }
+
+  // Fall back to common backend if specializations are not available
+  eval(inputs, outputs);
+}
+
+#else // Accelerate not available
+
+/** Evaluate primitive on CPU falling back to common backend */
+void Axpby::eval_cpu(
+    const std::vector<array>& inputs,
+    const std::vector<array>& outputs) {
+  eval(inputs, outputs);
+}
+
+#endif
+
 ///////////////////////////////////////////////////////////////////////////////
 // Primitive Metal Backend Implementation
 ///////////////////////////////////////////////////////////////////////////////
@@ -137,9 +213,10 @@ void Axpby::eval_cpu(
 
 /** Evaluate primitive on GPU */
 void Axpby::eval_gpu(
-    const std::vector<mx::array>& inputs,
-    std::vector<mx::array>& outputs) {
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
   // Prepare inputs
+  assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
   auto& out = outputs[0];
@@ -148,7 +225,7 @@ void Axpby::eval_gpu(
   // and each stream carries its device identifiers
   auto& s = stream();
   // We get the needed metal device using the stream
-  auto& d = mx::metal::device(s.device);
+  auto& d = metal::device(s.device);
 
   // Prepare to specialize based on contiguity
   bool contiguous_kernel =
@@ -158,12 +235,12 @@ void Axpby::eval_gpu(
   // Allocate output memory with strides based on specialization
   if (contiguous_kernel) {
     out.set_data(
-        mx::allocator::malloc(x.data_size() * out.itemsize()),
+        allocator::malloc_or_wait(x.data_size() * out.itemsize()),
         x.data_size(),
         x.strides(),
         x.flags());
   } else {
-    out.set_data(mx::allocator::malloc(out.nbytes()));
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
   }
 
   // Resolve name of kernel (corresponds to axpby.metal)
@@ -172,15 +249,15 @@ void Axpby::eval_gpu(
   kname << (contiguous_kernel ? "contiguous_" : "general_");
   kname << type_to_name(out);
 
-  // Load the metal library
-  auto lib = d.get_library("mlx_ext");
+  // Make sure the metal library is available
+  d.register_library("mlx_ext");
 
   // Make a kernel from this metal library
-  auto kernel = d.get_kernel(kname.str(), lib);
+  auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
   // Prepare to encode kernel
   auto& compute_encoder = d.get_command_encoder(s.index);
-  compute_encoder.set_compute_pipeline_state(kernel);
+  compute_encoder->setComputePipelineState(kernel);
 
   // Kernel parameters are registered with buffer indices corresponding to
   // those in the kernel declaration at axpby.metal
@@ -195,15 +272,15 @@ void Axpby::eval_gpu(
   compute_encoder.set_output_array(out, 2);
 
   // Encode alpha and beta
-  compute_encoder.set_bytes(alpha_, 3);
-  compute_encoder.set_bytes(beta_, 4);
+  compute_encoder->setBytes(&alpha_, sizeof(float), 3);
+  compute_encoder->setBytes(&beta_, sizeof(float), 4);
 
   // Encode shape, strides and ndim if needed
   if (!contiguous_kernel) {
-    compute_encoder.set_vector_bytes(x.shape(), 5);
-    compute_encoder.set_vector_bytes(x.strides(), 6);
-    compute_encoder.set_vector_bytes(y.strides(), 7);
-    compute_encoder.set_bytes(ndim, 8);
+    compute_encoder->setBytes(x.shape().data(), ndim * sizeof(int), 5);
+    compute_encoder->setBytes(x.strides().data(), ndim * sizeof(size_t), 6);
+    compute_encoder->setBytes(y.strides().data(), ndim * sizeof(size_t), 7);
+    compute_encoder->setBytes(&ndim, sizeof(int), 8);
   }
 
   // We launch 1 thread for each input and make sure that the number of
@@ -218,15 +295,15 @@ void Axpby::eval_gpu(
 
   // Launch the grid with the given number of threads divided among
   // the given threadgroups
-  compute_encoder.dispatch_threads(grid_dims, group_dims);
+  compute_encoder.dispatchThreads(grid_dims, group_dims);
 }
 
 #else // Metal is not available
 
 /** Fail evaluation on GPU */
 void Axpby::eval_gpu(
-    const std::vector<mx::array>& inputs,
-    std::vector<mx::array>& out) {
+    const std::vector<array>& inputs,
+    std::vector<array>& out) {
   throw std::runtime_error("Axpby has no GPU implementation.");
 }
 
@@ -237,9 +314,9 @@ void Axpby::eval_gpu(
 ///////////////////////////////////////////////////////////////////////////////
 
 /** The Jacobian-vector product. */
-std::vector<mx::array> Axpby::jvp(
-    const std::vector<mx::array>& primals,
-    const std::vector<mx::array>& tangents,
+std::vector<array> Axpby::jvp(
+    const std::vector<array>& primals,
+    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
@@ -251,8 +328,8 @@ std::vector<mx::array> Axpby::jvp(
   // scaled by beta
   if (argnums.size() > 1) {
     auto scale = argnums[0] == 0 ? alpha_ : beta_;
-    auto scale_arr = mx::array(scale, tangents[0].dtype());
-    return {mx::multiply(scale_arr, tangents[0], stream())};
+    auto scale_arr = array(scale, tangents[0].dtype());
+    return {multiply(scale_arr, tangents[0], stream())};
   }
   // If, argnums = {0, 1}, we take contributions from both
   // which gives us jvp = tangent_x * alpha + tangent_y * beta
@@ -262,24 +339,24 @@ std::vector<mx::array> Axpby::jvp(
 }
 
 /** The vector-Jacobian product. */
-std::vector<mx::array> Axpby::vjp(
-    const std::vector<mx::array>& primals,
-    const std::vector<mx::array>& cotangents,
+std::vector<array> Axpby::vjp(
+    const std::vector<array>& primals,
+    const std::vector<array>& cotangents,
     const std::vector<int>& argnums,
-    const std::vector<mx::array>&) {
+    const std::vector<array>&) {
   // Reverse mode diff
-  std::vector<mx::array> vjps;
+  std::vector<array> vjps;
   for (auto arg : argnums) {
     auto scale = arg == 0 ? alpha_ : beta_;
-    auto scale_arr = mx::array(scale, cotangents[0].dtype());
-    vjps.push_back(mx::multiply(scale_arr, cotangents[0], stream()));
+    auto scale_arr = array(scale, cotangents[0].dtype());
+    vjps.push_back(multiply(scale_arr, cotangents[0], stream()));
   }
   return vjps;
 }
 
 /** Vectorize primitive along given axis */
-std::pair<std::vector<mx::array>, std::vector<int>> Axpby::vmap(
-    const std::vector<mx::array>& inputs,
+std::pair<std::vector<array>, std::vector<int>> Axpby::vmap(
+    const std::vector<array>& inputs,
     const std::vector<int>& axes) {
   throw std::runtime_error("Axpby has no vmap implementation.");
 }
@@ -290,4 +367,4 @@ bool Axpby::is_equivalent(const Primitive& other) const {
   return alpha_ == r_other.alpha_ && beta_ == r_other.beta_;
 }
 
-} // namespace my_ext
+} // namespace mlx::core

examples/extensions/axpby/axpby.h

@@ -1,13 +1,11 @@
-// Copyright © 2023-2025 Apple Inc.
+// Copyright © 2023 Apple Inc.
 
 #pragma once
 
 #include "mlx/ops.h"
 #include "mlx/primitives.h"
 
-namespace mx = mlx::core;
-
-namespace my_ext {
+namespace mlx::core {
 
 ///////////////////////////////////////////////////////////////////////////////
 // Operation
@@ -20,22 +18,22 @@ namespace my_ext {
  *  Follow numpy style broadcasting between x and y
  *  Inputs are upcasted to floats if needed
  **/
-mx::array axpby(
-    const mx::array& x, // Input array x
-    const mx::array& y, // Input array y
+array axpby(
+    const array& x, // Input array x
+    const array& y, // Input array y
     const float alpha, // Scaling factor for x
     const float beta, // Scaling factor for y
-    mx::StreamOrDevice s = {} // Stream on which to schedule the operation
+    StreamOrDevice s = {} // Stream on which to schedule the operation
 );
 
 ///////////////////////////////////////////////////////////////////////////////
 // Primitive
 ///////////////////////////////////////////////////////////////////////////////
 
-class Axpby : public mx::Primitive {
+class Axpby : public Primitive {
  public:
-  explicit Axpby(mx::Stream stream, float alpha, float beta)
-      : mx::Primitive(stream), alpha_(alpha), beta_(beta) {};
+  explicit Axpby(Stream stream, float alpha, float beta)
+      : Primitive(stream), alpha_(alpha), beta_(beta) {};
 
   /**
    * A primitive must know how to evaluate itself on the CPU/GPU
@@ -44,25 +42,23 @@ class Axpby : public mx::Primitive {
    * To avoid unnecessary allocations, the evaluation function
    * is responsible for allocating space for the array.
    */
-  void eval_cpu(
-      const std::vector<mx::array>& inputs,
-      std::vector<mx::array>& outputs) override;
-  void eval_gpu(
-      const std::vector<mx::array>& inputs,
-      std::vector<mx::array>& outputs) override;
+  void eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs)
+      override;
+  void eval_gpu(const std::vector<array>& inputs, std::vector<array>& outputs)
+      override;
 
   /** The Jacobian-vector product. */
-  std::vector<mx::array> jvp(
-      const std::vector<mx::array>& primals,
-      const std::vector<mx::array>& tangents,
+  std::vector<array> jvp(
+      const std::vector<array>& primals,
+      const std::vector<array>& tangents,
       const std::vector<int>& argnums) override;
 
   /** The vector-Jacobian product. */
-  std::vector<mx::array> vjp(
-      const std::vector<mx::array>& primals,
-      const std::vector<mx::array>& cotangents,
+  std::vector<array> vjp(
+      const std::vector<array>& primals,
+      const std::vector<array>& cotangents,
       const std::vector<int>& argnums,
-      const std::vector<mx::array>& outputs) override;
+      const std::vector<array>& outputs) override;
 
   /**
    * The primitive must know how to vectorize itself across
@@ -70,21 +66,24 @@ class Axpby : public mx::Primitive {
    * representing the vectorized computation and the axis which
    * corresponds to the output vectorized dimension.
    */
-  std::pair<std::vector<mx::array>, std::vector<int>> vmap(
-      const std::vector<mx::array>& inputs,
+  std::pair<std::vector<array>, std::vector<int>> vmap(
+      const std::vector<array>& inputs,
       const std::vector<int>& axes) override;
 
-  /** The name of primitive. */
-  const char* name() const override {
-    return "Axpby";
+  /** Print the primitive. */
+  void print(std::ostream& os) override {
+    os << "Axpby";
   }
 
   /** Equivalence check **/
-  bool is_equivalent(const mx::Primitive& other) const override;
+  bool is_equivalent(const Primitive& other) const override;
 
  private:
   float alpha_;
   float beta_;
+
+  /** Fall back implementation for evaluation on CPU */
+  void eval(const std::vector<array>& inputs, std::vector<array>& outputs);
 };
 
-} // namespace my_ext
+} // namespace mlx::core

examples/extensions/axpby/axpby.metal

@@ -1,7 +1,8 @@
-// Copyright © 2023-2025 Apple Inc.
+// Copyright © 2023 Apple Inc.
 
 #include <metal_stdlib>
 
+#include "mlx/backend/metal/kernels/bf16.h"
 #include "mlx/backend/metal/kernels/utils.h"
 
 template <typename T>
@@ -12,8 +13,8 @@ template <typename T>
     constant const float& alpha [[buffer(3)]],
     constant const float& beta [[buffer(4)]],
     constant const int* shape [[buffer(5)]],
-    constant const int64_t* x_strides [[buffer(6)]],
-    constant const int64_t* y_strides [[buffer(7)]],
+    constant const size_t* x_strides [[buffer(6)]],
+    constant const size_t* y_strides [[buffer(7)]],
     constant const int& ndim [[buffer(8)]],
     uint index [[thread_position_in_grid]]) {
   auto x_offset = elem_to_loc(index, shape, x_strides, ndim);
@@ -34,14 +35,29 @@ template <typename T>
       static_cast<T>(alpha) * x[index] + static_cast<T>(beta) * y[index];
 }
 
-// clang-format off
-#define instantiate_axpby(type_name, type)                             \
-  instantiate_kernel("axpby_general_" #type_name, axpby_general, type) \
-  instantiate_kernel(                                                  \
-          "axpby_contiguous_" #type_name, axpby_contiguous, type)
+#define instantiate_axpby(type_name, type)                               \
+  template [[host_name("axpby_general_" #type_name)]] [[kernel]] void    \
+  axpby_general<type>(                                                   \
+      device const type* x [[buffer(0)]],                                \
+      device const type* y [[buffer(1)]],                                \
+      device type* out [[buffer(2)]],                                    \
+      constant const float& alpha [[buffer(3)]],                         \
+      constant const float& beta [[buffer(4)]],                          \
+      constant const int* shape [[buffer(5)]],                           \
+      constant const size_t* x_strides [[buffer(6)]],                    \
+      constant const size_t* y_strides [[buffer(7)]],                    \
+      constant const int& ndim [[buffer(8)]],                            \
+      uint index [[thread_position_in_grid]]);                           \
+  template [[host_name("axpby_contiguous_" #type_name)]] [[kernel]] void \
+  axpby_contiguous<type>(                                                \
+      device const type* x [[buffer(0)]],                                \
+      device const type* y [[buffer(1)]],                                \
+      device type* out [[buffer(2)]],                                    \
+      constant const float& alpha [[buffer(3)]],                         \
+      constant const float& beta [[buffer(4)]],                          \
+      uint index [[thread_position_in_grid]]);
 
 instantiate_axpby(float32, float);
 instantiate_axpby(float16, half);
 instantiate_axpby(bfloat16, bfloat16_t);
 instantiate_axpby(complex64, complex64_t);
-// clang-format on

examples/extensions/bindings.cpp

@@ -8,12 +8,14 @@
 namespace nb = nanobind;
 using namespace nb::literals;
 
+using namespace mlx::core;
+
 NB_MODULE(_ext, m) {
   m.doc() = "Sample extension for MLX";
 
   m.def(
       "axpby",
-      &my_ext::axpby,
+      &axpby,
       "x"_a,
       "y"_a,
       "alpha"_a,

examples/extensions/pyproject.toml

@@ -1,8 +1,8 @@
 [build-system]
 requires = [
   "setuptools>=42",
-  "cmake>=3.25",
+  "cmake>=3.24",
   "mlx>=0.18.0",
-  "nanobind==2.4.0",
+  "nanobind==2.2.0",
 ]
 build-backend = "setuptools.build_meta"

examples/extensions/requirements.txt

@@ -1,4 +1,4 @@
 setuptools>=42
-cmake>=3.25
-mlx>=0.21.0
+cmake>=3.24
+mlx>=0.18.1
 nanobind==2.2.0

mlx.pc.in

@@ -28,19 +28,10 @@ endif()
 if (@MLX_BUILD_METAL@)
     set(MLX_BUILD_METAL @MLX_BUILD_METAL@)
     set(MLX_CXX_FLAGS ${MLX_CXX_FLAGS} -D_METAL_)
-    set(MLX_INCLUDE_DIRS 
-        "${MLX_INCLUDE_DIRS};"
+    set_and_check(MLX_INCLUDE_DIRS 
+        ${MLX_INCLUDE_DIRS} 
         @PACKAGE_CMAKE_INSTALL_INCLUDEDIR@/metal_cpp
     )
-    if(@MLX_METAL_VERSION@ GREATER_EQUAL 310)
-      set(MLX_INCLUDE_DIRS
-        "${MLX_INCLUDE_DIRS};"
-        @PACKAGE_CMAKE_INSTALL_INCLUDEDIR@/mlx/backend/metal/kernels/metal_3_1)
-    else()
-      set(MLX_INCLUDE_DIRS
-        "${MLX_INCLUDE_DIRS};"
-        @PACKAGE_CMAKE_INSTALL_INCLUDEDIR@/mlx/backend/metal/kernels/metal_3_0)
-    endif()
 endif()
 
 set_target_properties(mlx PROPERTIES

mlx/CMakeLists.txt

@@ -5,8 +5,6 @@ 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
           ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
@@ -20,48 +18,24 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/linalg.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/metal.h)
 
-# Define MLX_VERSION only in the version.cpp file.
-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 $<BUILD_INTERFACE:mlx_version>)
-
-if(MSVC)
-  # Disable some MSVC warnings to speed up compilation.
-  target_compile_options(mlx PUBLIC /wd4068 /wd4244 /wd4267 /wd4804)
-endif()
-
-if(WIN32)
-  # Export symbols by default to behave like macOS/linux.
-  set_target_properties(mlx PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS TRUE)
-endif()
-
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
-
 if(MLX_BUILD_CPU)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/cpu)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
 else()
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_cpu)
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/distributed)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
+if(MLX_BUILD_ACCELERATE)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/accelerate)
+elseif(MLX_BUILD_CPU)
+  target_sources(
+    mlx
+    PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/common/default_primitives.cpp)
+endif()
 
 if(MLX_BUILD_METAL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/metal)
 else()
-  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(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_gpu)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_metal)
 endif()

mlx/allocator.cpp

@@ -4,11 +4,12 @@
 #include <sstream>
 
 #include "mlx/allocator.h"
+#include "mlx/scheduler.h"
 
 namespace mlx::core::allocator {
 
 Buffer malloc(size_t size) {
-  auto buffer = allocator().malloc(size);
+  auto buffer = allocator().malloc(size, /* allow_swap */ true);
   if (size && !buffer.ptr()) {
     std::ostringstream msg;
     msg << "[malloc] Unable to allocate " << size << " bytes.";
@@ -18,7 +19,48 @@ Buffer malloc(size_t size) {
 }
 
 void free(Buffer buffer) {
-  allocator().free(buffer);
+  return 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

@@ -32,10 +32,14 @@ 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) = 0;
+  virtual Buffer malloc(size_t size, bool allow_swap = false) = 0;
   virtual void free(Buffer buffer) = 0;
   virtual size_t size(Buffer buffer) const = 0;
 
@@ -49,4 +53,16 @@ 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.cpp

@@ -1,6 +1,5 @@
 // Copyright © 2023-2024 Apple Inc.
 #include <functional>
-#include <unordered_map>
 
 #include "mlx/array.h"
 #include "mlx/ops.h"
@@ -10,14 +9,28 @@
 
 namespace mlx::core {
 
+namespace {
+
+/** Return true if we are currently performing a function transformation in
+ * order to keep the graph when evaluating tracer arrays. */
+bool in_tracing() {
+  return detail::InTracing::in_tracing();
+}
+
+bool retain_graph() {
+  return detail::RetainGraph::retain_graph();
+}
+
+} // namespace
+
 array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
-    : array_desc_(std::make_shared<ArrayDesc>(Shape{}, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(std::vector<int>{}, dtype)) {
   auto cval = static_cast<complex64_t>(val);
   init(&cval);
 }
 
 array::array(
-    Shape shape,
+    std::vector<int> shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
     std::vector<array> inputs)
@@ -25,21 +38,10 @@ array::array(
           std::move(shape),
           dtype,
           std::move(primitive),
-          std::move(inputs))) {
-  if (has_primitive() && this->primitive().stream().device == Device::gpu) {
-    for (auto& in : this->inputs()) {
-      if (in.dtype() == float64) {
-        throw std::invalid_argument("float64 is not supported on the GPU");
-      }
-    }
-    if (this->dtype() == float64) {
-      throw std::invalid_argument("float64 is not supported on the GPU");
-    }
-  }
-}
+          std::move(inputs))) {}
 
 std::vector<array> array::make_arrays(
-    std::vector<Shape> shapes,
+    std::vector<std::vector<int>> shapes,
     const std::vector<Dtype>& dtypes,
     const std::shared_ptr<Primitive>& primitive,
     const std::vector<array>& inputs) {
@@ -56,59 +58,47 @@ std::vector<array> array::make_arrays(
   return outputs;
 }
 
-array array::unsafe_weak_copy(const array& other) {
-  auto cpy = array(other.shape(), other.dtype(), nullptr, {});
-  cpy.set_data(
-      other.buffer(),
-      other.data_size(),
-      other.strides(),
-      other.flags(),
-      [](auto) {});
-  cpy.array_desc_->data_ptr = other.array_desc_->data_ptr;
-  return cpy;
-}
-
 array::array(std::initializer_list<float> data)
     : array_desc_(std::make_shared<ArrayDesc>(
-          Shape{static_cast<ShapeElem>(data.size())},
+          std::vector<int>{static_cast<int>(data.size())},
           float32)) {
   init(data.begin());
 }
 
 array::array(std::initializer_list<int> data, Dtype dtype)
     : array_desc_(std::make_shared<ArrayDesc>(
-          Shape{static_cast<ShapeElem>(data.size())},
+          std::vector<int>{static_cast<int>(data.size())},
           dtype)) {
   init(data.begin());
 }
 
 /* Build an array from a shared buffer */
-array::array(allocator::Buffer data, Shape shape, Dtype dtype, Deleter deleter)
+array::array(
+    allocator::Buffer data,
+    std::vector<int> shape,
+    Dtype dtype,
+    deleter_t deleter)
     : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   set_data(data, deleter);
 }
 
 void array::detach() {
-  array_desc_->primitive = nullptr;
-  for (auto& s : array_desc_->siblings) {
-    s.array_desc_->primitive = nullptr;
-  }
   for (auto& s : array_desc_->siblings) {
     s.array_desc_->inputs.clear();
     s.array_desc_->siblings.clear();
     s.array_desc_->position = 0;
+    s.array_desc_->primitive = nullptr;
   }
   array_desc_->inputs.clear();
   array_desc_->siblings.clear();
   array_desc_->position = 0;
+  array_desc_->primitive = nullptr;
 }
 
 bool array::is_available() const {
   if (status() == Status::available) {
     return true;
-  } else if (
-      status() == Status::evaluated &&
-      (!event().valid() || event().is_signaled())) {
+  } else if (status() == Status::evaluated && event().is_signaled()) {
     set_status(Status::available);
     return true;
   }
@@ -117,10 +107,7 @@ bool array::is_available() const {
 
 void array::wait() {
   if (!is_available()) {
-    if (event().valid()) {
-      event().wait();
-      detach_event();
-    }
+    event().wait();
     set_status(Status::available);
   }
 }
@@ -135,11 +122,10 @@ void array::eval() {
 }
 
 bool array::is_tracer() const {
-  return (array_desc_->is_tracer && detail::in_tracing()) ||
-      detail::retain_graph();
+  return array_desc_->is_tracer && in_tracing() || retain_graph();
 }
 
-void array::set_data(allocator::Buffer buffer, Deleter d) {
+void array::set_data(allocator::Buffer buffer, deleter_t d) {
   array_desc_->data = std::make_shared<Data>(buffer, d);
   array_desc_->data_ptr = buffer.raw_ptr();
   array_desc_->data_size = size();
@@ -152,9 +138,9 @@ void array::set_data(allocator::Buffer buffer, Deleter d) {
 void array::set_data(
     allocator::Buffer buffer,
     size_t data_size,
-    Strides strides,
+    std::vector<size_t> strides,
     Flags flags,
-    Deleter d) {
+    deleter_t d) {
   array_desc_->data = std::make_shared<Data>(buffer, d);
   array_desc_->data_ptr = buffer.raw_ptr();
   array_desc_->data_size = data_size;
@@ -164,7 +150,7 @@ void array::set_data(
 
 void array::copy_shared_buffer(
     const array& other,
-    const Strides& strides,
+    const std::vector<size_t>& strides,
     Flags flags,
     size_t data_size,
     size_t offset /* = 0 */) {
@@ -181,13 +167,34 @@ void array::copy_shared_buffer(const array& other) {
   copy_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
+void array::move_shared_buffer(
+    array other,
+    const std::vector<size_t>& strides,
+    Flags flags,
+    size_t data_size,
+    size_t offset /* = 0 */) {
+  array_desc_->data = std::move(other.array_desc_->data);
+  array_desc_->strides = strides;
+  array_desc_->flags = flags;
+  array_desc_->data_size = data_size;
+  auto char_offset = sizeof(char) * itemsize() * offset;
+  auto data_ptr = other.array_desc_->data_ptr;
+  other.array_desc_->data_ptr = nullptr;
+  array_desc_->data_ptr =
+      static_cast<void*>(static_cast<char*>(data_ptr) + char_offset);
+}
+
+void array::move_shared_buffer(array other) {
+  move_shared_buffer(other, other.strides(), other.flags(), other.data_size());
+}
+
 array::~array() {
   if (array_desc_ == nullptr) {
     return;
   }
 
-  // Detached/detaching
-  if (array_desc_->primitive == nullptr) {
+  // Ignore arrays that might be detached during eval
+  if (status() == array::Status::scheduled) {
     return;
   }
 
@@ -207,8 +214,6 @@ array::~array() {
     if (do_detach) {
       for (auto& s : siblings()) {
         for (auto& ss : s.siblings()) {
-          // Set to null here to avoid descending into array destructor
-          // for siblings
           ss.array_desc_ = nullptr;
         }
         s.array_desc_->siblings.clear();
@@ -229,13 +234,13 @@ void array::ArrayDesc::init() {
   }
 }
 
-array::ArrayDesc::ArrayDesc(Shape shape, Dtype dtype)
+array::ArrayDesc::ArrayDesc(std::vector<int> shape, Dtype dtype)
     : shape(std::move(shape)), dtype(dtype), status(Status::available) {
   init();
 }
 
 array::ArrayDesc::ArrayDesc(
-    Shape shape,
+    std::vector<int> shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
     std::vector<array> inputs)
@@ -273,19 +278,7 @@ array::ArrayDesc::~ArrayDesc() {
     }
     ad.inputs.clear();
     for (auto& [_, a] : input_map) {
-      bool is_deletable =
-          (a.array_desc_.use_count() <= a.siblings().size() + 1);
-      // An array with siblings is deletable only if all of its siblings
-      // are deletable
-      for (auto& s : a.siblings()) {
-        if (!is_deletable) {
-          break;
-        }
-        int is_input = (input_map.find(s.id()) != input_map.end());
-        is_deletable &=
-            s.array_desc_.use_count() <= a.siblings().size() + is_input;
-      }
-      if (is_deletable) {
+      if (a.array_desc_.use_count() <= a.siblings().size() + 1) {
         for_deletion.push_back(std::move(a.array_desc_));
       }
     }
@@ -299,14 +292,6 @@ array::ArrayDesc::~ArrayDesc() {
     auto top = std::move(for_deletion.back());
     for_deletion.pop_back();
     append_deletable_inputs(*top);
-
-    // Clear out possible siblings to break circular references
-    for (auto& s : top->siblings) {
-      // Set to null here to avoid descending into top-level
-      // array destructor for siblings
-      s.array_desc_ = nullptr;
-    }
-    top->siblings.clear();
   }
 }
 
@@ -318,7 +303,7 @@ array::ArrayIterator::ArrayIterator(const array& arr, int idx)
 }
 
 array::ArrayIterator::reference array::ArrayIterator::operator*() const {
-  auto start = Shape(arr.ndim(), 0);
+  auto start = std::vector<int>(arr.ndim(), 0);
   auto end = arr.shape();
   auto shape = arr.shape();
   shape.erase(shape.begin());

mlx/array.h

@@ -15,11 +15,7 @@ namespace mlx::core {
 
 // Forward declaration
 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 deleter_t = std::function<void(allocator::Buffer)>;
 
 class array {
   /* An array is really a node in a graph. It contains a shared ArrayDesc
@@ -35,33 +31,33 @@ class array {
   explicit array(const std::complex<float>& val, Dtype dtype = complex64);
 
   template <typename It>
-  explicit array(
+  array(
       It data,
-      Shape shape,
+      std::vector<int> shape,
       Dtype dtype =
           TypeToDtype<typename std::iterator_traits<It>::value_type>());
 
   template <typename T>
-  explicit array(std::initializer_list<T> data, Dtype dtype = TypeToDtype<T>());
+  array(std::initializer_list<T> data, Dtype dtype = TypeToDtype<T>());
 
   /* Special case so empty lists default to float32. */
-  explicit array(std::initializer_list<float> data);
+  array(std::initializer_list<float> data);
 
   /* Special case so array({}, type) is an empty array. */
-  explicit array(std::initializer_list<int> data, Dtype dtype);
+  array(std::initializer_list<int> data, Dtype dtype);
 
   template <typename T>
-  explicit array(
+  array(
       std::initializer_list<T> data,
-      Shape shape,
+      std::vector<int> shape,
       Dtype dtype = TypeToDtype<T>());
 
   /* Build an array from a buffer */
-  explicit array(
+  array(
       allocator::Buffer data,
-      Shape shape,
+      std::vector<int> shape,
       Dtype dtype,
-      Deleter deleter = allocator::free);
+      deleter_t deleter = allocator::free);
 
   /** Assignment to rvalue does not compile. */
   array& operator=(const array& other) && = delete;
@@ -100,7 +96,7 @@ class array {
   }
 
   /** The shape of the array as a vector of integers. */
-  const Shape& shape() const {
+  const std::vector<int>& shape() const {
     return array_desc_->shape;
   }
 
@@ -109,12 +105,12 @@ class array {
    *
    *  This function supports negative indexing and provides
    *  bounds checking. */
-  auto shape(int dim) const {
+  int shape(int dim) const {
     return shape().at(dim < 0 ? dim + ndim() : dim);
   }
 
   /** The strides of the array. */
-  const Strides& strides() const {
+  const std::vector<size_t>& strides() const {
     return array_desc_->strides;
   }
 
@@ -123,7 +119,7 @@ class array {
    *
    *  This function supports negative indexing and provides
    *  bounds checking. */
-  auto strides(int dim) const {
+  size_t strides(int dim) const {
     return strides().at(dim < 0 ? dim + ndim() : dim);
   }
 
@@ -188,24 +184,17 @@ class array {
    */
 
   array(
-      Shape shape,
+      std::vector<int> shape,
       Dtype dtype,
       std::shared_ptr<Primitive> primitive,
       std::vector<array> inputs);
 
   static std::vector<array> make_arrays(
-      std::vector<Shape> shapes,
+      std::vector<std::vector<int>> shapes,
       const std::vector<Dtype>& dtypes,
       const std::shared_ptr<Primitive>& primitive,
       const std::vector<array>& inputs);
 
-  /**
-   * Get a new array that refers to the same data as the input but with a
-   * non-owning pointer to it. Note the array is detached from the graph and has
-   * no inputs, siblings or primitive.
-   */
-  static array unsafe_weak_copy(const array& other);
-
   /** A unique identifier for an array. */
   std::uintptr_t id() const {
     return reinterpret_cast<std::uintptr_t>(array_desc_.get());
@@ -218,16 +207,12 @@ class array {
 
   struct Data {
     allocator::Buffer buffer;
-    Deleter d;
-    Data(allocator::Buffer buffer, Deleter d = allocator::free)
+    deleter_t d;
+    Data(allocator::Buffer buffer, deleter_t d = allocator::free)
         : buffer(buffer), d(d) {}
     // 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);
     }
@@ -343,11 +328,11 @@ class array {
     return allocator::allocator().size(buffer());
   }
 
-  // Return the shared pointer to the array::Data struct
-  const std::shared_ptr<Data>& data_shared_ptr() const {
+  // Return a copy of the shared pointer
+  // to the array::Data struct
+  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() {
@@ -360,10 +345,15 @@ class array {
   }
 
   enum Status {
-    // The output of a computation which has not been scheduled.
+    // The ouptut of a computation which has not been scheduled.
     // For example, the status of `x` in `auto x = a + b`.
     unscheduled,
 
+    // The ouptut of a computation which has been scheduled but `eval_*` has
+    // not yet been called on the array's primitive. A possible
+    // status of `x` in `auto x = a + b; eval(x);`
+    scheduled,
+
     // The array's `eval_*` function has been run, but the computation is not
     // necessarily complete. The array will have memory allocated and if it is
     // not a tracer then it will be detached from the graph.
@@ -400,10 +390,6 @@ class array {
     array_desc_->event = std::move(e);
   }
 
-  void detach_event() const {
-    array_desc_->event = Event{};
-  }
-
   // Mark the array as a tracer array (true) or not.
   void set_tracer(bool is_tracer) {
     array_desc_->is_tracer = is_tracer;
@@ -411,24 +397,33 @@ class array {
   // Check if the array is a tracer array
   bool is_tracer() const;
 
-  void set_data(allocator::Buffer buffer, Deleter d = allocator::free);
+  void set_data(allocator::Buffer buffer, deleter_t d = allocator::free);
 
   void set_data(
       allocator::Buffer buffer,
       size_t data_size,
-      Strides strides,
+      std::vector<size_t> strides,
       Flags flags,
-      Deleter d = allocator::free);
+      deleter_t d = allocator::free);
 
   void copy_shared_buffer(
       const array& other,
-      const Strides& strides,
+      const std::vector<size_t>& strides,
       Flags flags,
       size_t data_size,
       size_t offset = 0);
 
   void copy_shared_buffer(const array& other);
 
+  void move_shared_buffer(
+      array other,
+      const std::vector<size_t>& strides,
+      Flags flags,
+      size_t data_size,
+      size_t offset = 0);
+
+  void move_shared_buffer(array other);
+
   void overwrite_descriptor(const array& other) {
     array_desc_ = other.array_desc_;
   }
@@ -441,8 +436,8 @@ class array {
   void init(const It src);
 
   struct ArrayDesc {
-    Shape shape;
-    Strides strides;
+    std::vector<int> shape;
+    std::vector<size_t> strides;
     size_t size;
     Dtype dtype;
     std::shared_ptr<Primitive> primitive;
@@ -476,10 +471,10 @@ class array {
     // The arrays position in the output list
     uint32_t position{0};
 
-    explicit ArrayDesc(Shape shape, Dtype dtype);
+    explicit ArrayDesc(std::vector<int> shape, Dtype dtype);
 
     explicit ArrayDesc(
-        Shape shape,
+        std::vector<int> shape,
         Dtype dtype,
         std::shared_ptr<Primitive> primitive,
         std::vector<array> inputs);
@@ -500,14 +495,14 @@ class array {
 
 template <typename T>
 array::array(T val, Dtype dtype /* = TypeToDtype<T>() */)
-    : array_desc_(std::make_shared<ArrayDesc>(Shape{}, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(std::vector<int>{}, dtype)) {
   init(&val);
 }
 
 template <typename It>
 array::array(
   It data,
-  Shape shape,
+  std::vector<int> shape,
   Dtype dtype /* = TypeToDtype<typename std::iterator_traits<It>::value_type>() */) :
     array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   init(data);
@@ -518,7 +513,7 @@ array::array(
     std::initializer_list<T> data,
     Dtype dtype /* = TypeToDtype<T>() */)
     : array_desc_(std::make_shared<ArrayDesc>(
-          Shape{static_cast<ShapeElem>(data.size())},
+          std::vector<int>{static_cast<int>(data.size())},
           dtype)) {
   init(data.begin());
 }
@@ -526,7 +521,7 @@ array::array(
 template <typename T>
 array::array(
     std::initializer_list<T> data,
-    Shape shape,
+    std::vector<int> shape,
     Dtype dtype /* = TypeToDtype<T>() */)
     : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   if (data.size() != size()) {
@@ -595,9 +590,6 @@ void array::init(It src) {
     case float32:
       std::copy(src, src + size(), data<float>());
       break;
-    case float64:
-      std::copy(src, src + size(), data<double>());
-      break;
     case bfloat16:
       std::copy(src, src + size(), data<bfloat16_t>());
       break;

mlx/backend/accelerate/CMakeLists.txt

@@ -0,0 +1,8 @@
+target_sources(
+  mlx
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp)

mlx/backend/accelerate/conv.cpp

@@ -0,0 +1,20 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <cassert>
+
+#include <Accelerate/Accelerate.h>
+#include <simd/vector.h>
+
+#include "mlx/backend/common/copy.h"
+#include "mlx/primitives.h"
+#include "mlx/utils.h"
+
+namespace mlx::core {
+
+void Convolution::eval_cpu(const std::vector<array>& inputs, array& out) {
+  eval(inputs, out);
+
+  // TODO: Add accelerate based optimizations for CPU conv
+}
+
+} // namespace mlx::core

mlx/backend/accelerate/matmul.cpp

@@ -0,0 +1,253 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <cassert>
+
+#include <Accelerate/Accelerate.h>
+
+#include "mlx/backend/accelerate/utils.h"
+#include "mlx/backend/common/copy.h"
+#include "mlx/primitives.h"
+#include "mlx/utils.h"
+
+namespace mlx::core {
+
+namespace {
+
+std::tuple<bool, size_t, array> check_transpose(const array& arr) {
+  auto stx = arr.strides()[arr.ndim() - 2];
+  auto sty = arr.strides()[arr.ndim() - 1];
+  if (stx == arr.shape(-1) && sty == 1) {
+    return std::make_tuple(false, stx, arr);
+  } else if (stx == 1 && sty == arr.shape(-2)) {
+    return std::make_tuple(true, sty, arr);
+  } else {
+    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+    copy(arr, arr_copy, CopyType::General);
+    size_t stx = arr.shape(-1);
+    return std::make_tuple(false, stx, arr_copy);
+  }
+}
+
+inline void matmul_cblas_general(
+    const array& a_pre,
+    const array& b_pre,
+    array& out,
+    float alpha = 1.0f,
+    float beta = 0.0f) {
+  if (out.dtype() != float32) {
+    throw std::runtime_error(
+        "[matmul_cblas] on CPU currently only supports float32");
+  }
+
+  auto [a_transposed, lda, a] = check_transpose(a_pre);
+  auto [b_transposed, ldb, b] = check_transpose(b_pre);
+  size_t M = a.shape(-2);
+  size_t N = b.shape(-1);
+  size_t K = a.shape(-1);
+
+  if (M == 0 || N == 0) {
+    return;
+  }
+  if (K == 0) {
+    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
+    return;
+  }
+
+  for (int i = 0; i < (a.size() / (M * K)); ++i) {
+    cblas_sgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        alpha, // alpha
+        a.data<float>() + elem_to_loc(M * K * i, a.shape(), a.strides()),
+        lda,
+        b.data<float>() + elem_to_loc(K * N * i, b.shape(), b.strides()),
+        ldb,
+        beta, // beta
+        out.data<float>() + M * N * i,
+        out.shape(-1) // ldc
+    );
+  }
+}
+
+inline void matmul_cblas(const array& a_pre, const array& b_pre, array& out) {
+  if (out.dtype() != float32) {
+    throw std::runtime_error(
+        "[matmul_cblas] on CPU currently only supports float32");
+  }
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  return matmul_cblas_general(a_pre, b_pre, out);
+}
+
+inline void matmul_bnns_general(
+    const array& a_pre,
+    const array& b_pre,
+    array& out,
+    float alpha = 1.0f,
+    float beta = 0.0f) {
+  // TODO: Update to utilize BNNS broadcasting
+
+  auto [a_transposed, lda, a] = check_transpose(a_pre);
+  auto [b_transposed, ldb, b] = check_transpose(b_pre);
+  size_t M = a.shape(-2);
+  size_t N = b.shape(-1);
+  size_t K = a.shape(-1);
+
+  if (M == 0 || N == 0) {
+    return;
+  }
+  if (K == 0) {
+    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
+    return;
+  }
+
+  BNNSDataType bnns_dtype = to_bnns_dtype(out.dtype());
+
+  const BNNSLayerParametersBroadcastMatMul gemm_params{
+      /* float alpha = */ alpha,
+      /* float beta = */ beta,
+      /* bool transA = */ a_transposed,
+      /* bool transB = */ b_transposed,
+      /* bool quadratic = */ false,
+      /* bool a_is_weights = */ false,
+      /* bool b_is_weights = */ false,
+      /* BNNSNDArrayDescriptor iA_desc = */
+      BNNSNDArrayDescriptor{
+          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
+          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
+
+          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
+          {lda, (M * K) / lda, 0, 0, 0, 0, 0, 0},
+          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
+          {1, lda, 0, 0, 0, 0, 0, 0},
+
+          /* void * _Nullable data = */ nullptr,
+          /* BNNSDataType data_type = */ bnns_dtype,
+
+          /* void * _Nullable table_data = */ nullptr,
+          /* BNNSDataType table_data_type = */ bnns_dtype,
+
+          /* float data_scale = */ 1.0,
+          /* float data_bias = */ 0.0,
+      },
+      /* BNNSNDArrayDescriptor iB_desc = */
+      BNNSNDArrayDescriptor{
+          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
+          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
+
+          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
+          {ldb, (K * N) / ldb, 0, 0, 0, 0, 0, 0},
+          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
+          {1, ldb, 0, 0, 0, 0, 0, 0},
+
+          /* void * _Nullable data = */ nullptr,
+          /* BNNSDataType data_type = */ bnns_dtype,
+
+          /* void * _Nullable table_data = */ nullptr,
+          /* BNNSDataType table_data_type = */ bnns_dtype,
+
+          /* float data_scale = */ 1.0,
+          /* float data_bias = */ 0.0,
+      },
+      /* BNNSNDArrayDescriptor o_desc = */
+      BNNSNDArrayDescriptor{
+          /* BNNSNDArrayFlags flags = */ BNNSNDArrayFlagBackpropSet,
+          /* BNNSDataLayout layout = */ BNNSDataLayoutRowMajorMatrix,
+
+          /* size_t size[BNNS_MAX_TENSOR_DIMENSION] = */
+          {N, M, 0, 0, 0, 0, 0, 0},
+          /* size_t stride[BNNS_MAX_TENSOR_DIMENSION] = */
+          {1, N, 0, 0, 0, 0, 0, 0},
+
+          /* void * _Nullable data = */ nullptr,
+          /* BNNSDataType data_type = */ bnns_dtype,
+
+          /* void * _Nullable table_data = */ nullptr,
+          /* BNNSDataType table_data_type = */ bnns_dtype,
+
+          /* float data_scale = */ 1.0,
+          /* float data_bias = */ 0.0,
+      },
+  };
+
+  auto bnns_filter =
+      BNNSFilterCreateLayerBroadcastMatMul(&gemm_params, nullptr);
+
+  for (int i = 0; i < (a.size() / (M * K)); ++i) {
+    BNNSFilterApplyTwoInput(
+        bnns_filter,
+        a.data<uint8_t>() +
+            elem_to_loc(M * K * i, a.shape(), a.strides()) * a.itemsize(),
+        b.data<uint8_t>() +
+            elem_to_loc(K * N * i, b.shape(), b.strides()) * b.itemsize(),
+        out.data<uint8_t>() + M * N * i * out.itemsize());
+  }
+
+  BNNSFilterDestroy(bnns_filter);
+}
+
+inline void matmul_bnns(const array& a_pre, const array& b_pre, array& out) {
+  // TODO: Update to utilize BNNS broadcasting
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  return matmul_bnns_general(a_pre, b_pre, out);
+}
+
+template <typename T>
+inline void mask_matrix(
+    T* data,
+    const bool* mask,
+    int tile_size,
+    const int X,
+    const int Y,
+    const size_t X_data_str,
+    const size_t Y_data_str,
+    const size_t X_mask_str,
+    const size_t Y_mask_str) {
+  int tX = (X + tile_size - 1) / tile_size;
+  int tY = (Y + tile_size - 1) / tile_size;
+
+  for (int i = 0; i < tX; i++) {
+    for (int j = 0; j < tY; j++) {
+      bool do_mask = mask[i * X_mask_str + j * Y_mask_str];
+      if (!do_mask) {
+        int loc_x = i * tile_size;
+        int loc_y = j * tile_size;
+        T* data_block = data + loc_x * X_data_str + loc_y * Y_data_str;
+
+        int size_x = std::min(tile_size, X - loc_x);
+        int size_y = std::min(tile_size, Y - loc_y);
+        for (int ii = 0; ii < size_x; ii++) {
+          for (int jj = 0; jj < size_y; jj++) {
+            data_block[ii * X_data_str + jj * Y_data_str] = T(0.);
+          }
+        }
+      }
+    }
+  }
+}
+
+} // namespace
+
+void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
+  if (out.dtype() == float32) {
+    return matmul_cblas(inputs[0], inputs[1], out);
+  }
+  return matmul_bnns(inputs[0], inputs[1], out);
+}
+
+void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  // Fill output with C
+  auto& c = inputs[2];
+  CopyType ctype = c.data_size() == 1 ? CopyType::Scalar : CopyType::General;
+  copy(c, out, ctype);
+
+  if (out.dtype() == float32) {
+    return matmul_cblas_general(inputs[0], inputs[1], out, alpha_, beta_);
+  }
+  return matmul_bnns_general(inputs[0], inputs[1], out, alpha_, beta_);
+}
+
+} // namespace mlx::core

mlx/backend/accelerate/primitives.cpp

@@ -0,0 +1,601 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <cassert>
+#include <cmath>
+
+#include <Accelerate/Accelerate.h>
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/common/copy.h"
+#include "mlx/backend/common/unary.h"
+#include "mlx/primitives.h"
+
+#define DEFAULT(primitive)                                                 \
+  void primitive::eval_cpu(const std::vector<array>& inputs, array& out) { \
+    primitive::eval(inputs, out);                                          \
+  }
+
+#define DEFAULT_MULTI(primitive)                                       \
+  void primitive::eval_cpu(                                            \
+      const std::vector<array>& inputs, std::vector<array>& outputs) { \
+    primitive::eval(inputs, outputs);                                  \
+  }
+
+namespace mlx::core {
+
+// Use the default implementation for the following primitives
+DEFAULT(Arange)
+DEFAULT(ArgPartition)
+DEFAULT(ArgReduce)
+DEFAULT(ArgSort)
+DEFAULT(AsStrided)
+DEFAULT(BlockMaskedMM)
+DEFAULT(Broadcast)
+DEFAULT(Ceil)
+DEFAULT(Concatenate)
+DEFAULT(Conjugate)
+DEFAULT(Copy)
+DEFAULT_MULTI(CustomTransforms)
+DEFAULT_MULTI(Depends)
+DEFAULT_MULTI(DivMod)
+DEFAULT(NumberOfElements)
+DEFAULT(Equal)
+DEFAULT(Erf)
+DEFAULT(ErfInv)
+DEFAULT(FFT)
+DEFAULT(Floor)
+DEFAULT(Gather)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
+DEFAULT(Greater)
+DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
+DEFAULT(Less)
+DEFAULT(LessEqual)
+DEFAULT(Load)
+DEFAULT(LogicalNot)
+DEFAULT(LogicalAnd)
+DEFAULT(LogicalOr)
+DEFAULT(LogAddExp)
+DEFAULT(Maximum)
+DEFAULT(Minimum)
+DEFAULT(NotEqual)
+DEFAULT(Pad)
+DEFAULT(Partition)
+DEFAULT_MULTI(QRF)
+DEFAULT(RandomBits)
+DEFAULT(Reshape)
+DEFAULT(Remainder)
+DEFAULT(Round)
+DEFAULT(Scatter)
+DEFAULT(Select)
+DEFAULT(Sigmoid)
+DEFAULT(Sign)
+DEFAULT(Slice)
+DEFAULT(SliceUpdate)
+DEFAULT_MULTI(Split)
+DEFAULT(Sort)
+DEFAULT(StopGradient)
+DEFAULT_MULTI(SVD)
+DEFAULT(Transpose)
+DEFAULT(Inverse)
+DEFAULT(Cholesky)
+DEFAULT_MULTI(Eigh)
+
+void Abs::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  if (in.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    vDSP_vabs(in.data<float>(), 1, out.data<float>(), 1, in.data_size());
+  } else if (in.dtype() == int32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    vDSP_vabsi(in.data<int>(), 1, out.data<int>(), 1, in.data_size());
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+
+  if (a.dtype() == float32) {
+    binary_op<float>(
+        a,
+        b,
+        out,
+        [](auto x, auto y) { return x + y; },
+        [](const auto* s, const auto* vec, auto* o, auto n) {
+          vDSP_vsadd((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
+        },
+        [](const auto* vec, const auto* s, auto* o, auto n) {
+          vDSP_vsadd((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
+        },
+        [](const auto* a, const auto* b, auto* o, auto n) {
+          vDSP_vadd((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
+        });
+  } else if (a.dtype() == int32) {
+    binary_op<int>(
+        a,
+        b,
+        out,
+        [](auto x, auto y) { return x + y; },
+        [](const auto* s, const auto* vec, auto* o, auto n) {
+          vDSP_vsaddi((const int*)vec, 1, (const int*)s, (int*)o, 1, n);
+        },
+        [](const auto* vec, const auto* s, auto* o, auto n) {
+          vDSP_vsaddi((const int*)vec, 1, (const int*)s, (int*)o, 1, n);
+        },
+        [](const auto* a, const auto* b, auto* o, auto n) {
+          vDSP_vaddi((const int*)a, 1, (const int*)b, 1, (int*)o, 1, n);
+        });
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void ArcCos::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvacosf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void ArcCosh::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvacoshf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void ArcSin::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvasinf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void ArcSinh::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvasinhf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void ArcTan::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvatanf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  if (out.dtype() == float32 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    if (a.is_donatable()) {
+      out.copy_shared_buffer(a);
+    } else if (b.is_donatable()) {
+      out.copy_shared_buffer(b);
+    } else {
+      out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    }
+    int size = a.data_size();
+    vvatan2f(out.data<float>(), a.data<float>(), b.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void ArcTanh::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvatanhf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void AsType::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+
+  if (in.flags().contiguous) {
+    // Use accelerate functions if possible
+    if (in.dtype() == float32 && out.dtype() == uint32) {
+      set_unary_output_data(in, out);
+      vDSP_vfixu32(
+          in.data<float>(), 1, out.data<uint32_t>(), 1, in.data_size());
+      return;
+    } else if (in.dtype() == float32 && out.dtype() == int32) {
+      set_unary_output_data(in, out);
+      vDSP_vfix32(in.data<float>(), 1, out.data<int32_t>(), 1, in.data_size());
+      return;
+    } else if (in.dtype() == uint32 && out.dtype() == float32) {
+      set_unary_output_data(in, out);
+      vDSP_vfltu32(
+          in.data<uint32_t>(), 1, out.data<float>(), 1, in.data_size());
+      return;
+    } else if (in.dtype() == int32 && out.dtype() == float32) {
+      set_unary_output_data(in, out);
+      vDSP_vflt32(in.data<int32_t>(), 1, out.data<float>(), 1, in.data_size());
+      return;
+    }
+  }
+  eval(inputs, out);
+}
+
+void Cos::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvcosf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Cosh::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvcoshf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+
+  if (a.dtype() == int32) {
+    binary_op<int>(
+        a,
+        b,
+        out,
+        [](auto x, auto y) { return x / y; },
+        UseDefaultBinaryOp(),
+        [](const auto* vec, const auto* s, auto* o, auto n) {
+          vDSP_vsdivi((const int*)vec, 1, (const int*)s, (int*)o, 1, n);
+        },
+        [](const auto* a, const auto* b, auto* o, auto n) {
+          vDSP_vdivi((const int*)b, 1, (const int*)a, 1, (int*)o, 1, n);
+        });
+  } else if (a.dtype() == float32) {
+    binary_op<float>(
+        a,
+        b,
+        out,
+        [](auto x, auto y) { return x / y; },
+        [](const auto* s, const auto* vec, auto* o, auto n) {
+          vDSP_svdiv((const float*)s, (const float*)vec, 1, (float*)o, 1, n);
+        },
+        [](const auto* vec, const auto* s, auto* o, auto n) {
+          vDSP_vsdiv((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
+        },
+        [](const auto* a, const auto* b, auto* o, auto n) {
+          vDSP_vdiv((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
+        });
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vvexpf(out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Expm1::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vvexpm1f(
+        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  assert(in.dtype() == out.dtype());
+  if (in.data_size() == 1 && out.dtype() == float32) {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    vDSP_vfill(in.data<float>(), out.data<float>(), 1, out.size());
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Log::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    switch (base_) {
+      case Base::e:
+        vvlogf(
+            out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+        break;
+      case Base::two:
+        vvlog2f(
+            out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+        break;
+      case Base::ten:
+        vvlog10f(
+            out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+        break;
+    }
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Log1p::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vvlog1pf(
+        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+
+  if (a.dtype() == float32) {
+    binary_op<float>(
+        a,
+        b,
+        out,
+        [](auto x, auto y) { return x * y; },
+        [](const auto* s, const auto* vec, auto* o, auto n) {
+          vDSP_vsmul((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
+        },
+        [](const auto* vec, const auto* s, auto* o, auto n) {
+          vDSP_vsmul((const float*)vec, 1, (const float*)s, (float*)o, 1, n);
+        },
+        [](const auto* a, const auto* b, auto* o, auto n) {
+          vDSP_vmul((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
+        });
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Negative::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  if (in.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    vDSP_vneg(in.data<float>(), 1, out.data<float>(), 1, in.data_size());
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Power::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  if (out.dtype() == float32 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    int size = a.size();
+    if (a.is_donatable() && a.itemsize() == out.itemsize()) {
+      out.copy_shared_buffer(a);
+    } else if (b.is_donatable() && b.itemsize() == out.itemsize()) {
+      out.copy_shared_buffer(b);
+    } else {
+      out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    }
+    vvpowf(out.data<float>(), b.data<float>(), a.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (reduce_type_ == Scan::Sum && out.dtype() == float32 &&
+      in.flags().row_contiguous && in.strides()[axis_] == 1 && !inclusive_) {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    int stride = in.shape(axis_);
+    int count = in.size() / stride;
+    const float* input = in.data<float>();
+    float* output = out.data<float>();
+    float s = 1.0;
+    if (!reverse_) {
+      for (int i = 0; i < count; i++) {
+        vDSP_vrsum(input - 1, 1, &s, output, 1, stride);
+        input += stride;
+        output += stride;
+      }
+    } else {
+      for (int i = 0; i < count; i++) {
+        input += stride - 1;
+        output += stride - 1;
+        vDSP_vrsum(input + 1, -1, &s, output, -1, stride);
+        input++;
+        output++;
+      }
+    }
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Sin::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvsinf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Sinh::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvsinhf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Square::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  if (in.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vDSP_vsq(in.data<float>(), 1, out.data<float>(), 1, size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Sqrt::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  if (in.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    if (recip_) {
+      vvrsqrtf(out.data<float>(), in.data<float>(), &size);
+    } else {
+      vvsqrtf(out.data<float>(), in.data<float>(), &size);
+    }
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+
+  if (a.dtype() == float32) {
+    binary_op<float>(
+        a,
+        b,
+        out,
+        [](auto x, auto y) { return x - y; },
+        [](const auto* s, const auto* vec, auto* o, auto n) {
+          float minus_1 = -1;
+          vDSP_vsmsa(
+              (const float*)vec, 1, &minus_1, (const float*)s, (float*)o, 1, n);
+        },
+        [](const auto* vec, const auto* s, auto* o, auto n) {
+          float val = -(*s);
+          vDSP_vsadd((const float*)vec, 1, &val, (float*)o, 1, n);
+        },
+        [](const auto* a, const auto* b, auto* o, auto n) {
+          vDSP_vsub((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
+        });
+  } else if (a.dtype() == int32) {
+    binary_op<int>(
+        a,
+        b,
+        out,
+        [](auto x, auto y) { return x - y; },
+        UseDefaultBinaryOp(),
+        [](const auto* vec, const auto* s, auto* o, auto n) {
+          int val = -(*s);
+          vDSP_vsaddi((const int*)vec, 1, &val, (int*)o, 1, n);
+        },
+        UseDefaultBinaryOp());
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Tan::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvtanf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+void Tanh::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    int size = in.data_size();
+    vvtanhf(out.data<float>(), in.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/accelerate/quantized.cpp

@@ -0,0 +1,117 @@
+// Copyright © 2023 Apple Inc.
+
+#include <cassert>
+
+#include <simd/vector.h>
+
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+namespace {
+
+void _qmm_t_4_64(
+    float* result,
+    const float* x,
+    const uint32_t* w,
+    const float* scales,
+    const float* biases,
+    int M,
+    int N,
+    int K,
+    int B,
+    bool batched_w) {
+  constexpr int bits = 4;
+  constexpr int group_size = 64;
+  constexpr int bitmask = (1 << bits) - 1;
+  constexpr int pack_factor = 32 / bits;
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  int w_els = N * K / pack_factor;
+  int g_els = w_els * pack_factor / group_size;
+
+  for (int i = 0; i < B; i++) {
+    for (int m = 0; m < M; m++) {
+      const uint32_t* w_local = w;
+      const float* scales_local = scales;
+      const float* biases_local = biases;
+
+      for (int n = 0; n < N; n++) {
+        const simd_float16* x_local = (simd_float16*)x;
+        simd_float16 sum = 0;
+        for (int k = 0; k < K; k += group_size) {
+          float scale = *scales_local++;
+          float bias = *biases_local++;
+
+          for (int kw = 0; kw < packs_in_group; kw += 2) {
+            // TODO: vectorize this properly
+            simd_uint16 wi;
+            for (int e = 0; e < 2; e++) {
+              uint32_t wii = *w_local++;
+              for (int p = 0; p < 8; p++) {
+                wi[e * 8 + p] = wii & bitmask;
+                wii >>= bits;
+              }
+            }
+            simd_float16 wf = simd_float(wi);
+            wf *= scale;
+            wf += bias;
+
+            sum += (*x_local) * wf;
+            x_local++;
+          }
+        }
+
+        *result = simd_reduce_add(sum);
+        result++;
+      }
+
+      x += K;
+    }
+    if (batched_w) {
+      w += w_els;
+      scales += g_els;
+      biases += g_els;
+    }
+  }
+}
+
+} // namespace
+
+void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 4);
+
+  auto& x = inputs[0];
+  auto& w = inputs[1];
+  auto& scales = inputs[2];
+  auto& biases = inputs[3];
+
+  bool condition =
+      (transpose_ && x.flags().row_contiguous && w.flags().row_contiguous &&
+       scales.flags().row_contiguous && biases.flags().row_contiguous &&
+       x.dtype() == float32 && bits_ == 4 && group_size_ == 64);
+
+  if (condition) {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    int K = x.shape(-1);
+    int M = x.shape(-2);
+    int N = out.shape(-1);
+    int B = x.size() / K / M;
+    bool batched_w = w.ndim() > 2;
+    _qmm_t_4_64(
+        out.data<float>(),
+        x.data<float>(),
+        w.data<uint32_t>(),
+        scales.data<float>(),
+        biases.data<float>(),
+        M,
+        N,
+        K,
+        B,
+        batched_w);
+  } else {
+    eval(inputs, out);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/accelerate/reduce.cpp

@@ -0,0 +1,139 @@
+// Copyright © 2023 Apple Inc.
+
+#include <cassert>
+
+#include <Accelerate/Accelerate.h>
+#include <simd/vector.h>
+
+#include "mlx/backend/common/reduce.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename T, typename VT>
+struct MinReduction {
+  T operator()(const T& a, const T& b) {
+    return std::min(a, b);
+  }
+
+  VT operator()(VT a, VT b) {
+    return simd_min(a, b);
+  }
+};
+
+template <typename T, typename VT>
+struct MaxReduction {
+  T operator()(const T& a, const T& b) {
+    return std::max(a, b);
+  }
+
+  VT operator()(VT a, VT b) {
+    return simd_max(a, b);
+  }
+};
+
+template <typename T, typename VT>
+struct SumReduction {
+  T operator()(const T& a, const T& b) {
+    return a + b;
+  }
+
+  VT operator()(VT a, VT b) {
+    return a + b;
+  }
+};
+
+template <typename T, typename VT, int N, typename Reduction>
+struct StridedReduce {
+  void operator()(const T* x, T* accum, int size, size_t stride) {
+    Reduction op;
+
+    for (int i = 0; i < size; i++) {
+      size_t s = stride;
+      T* a = accum;
+      while (s >= N) {
+        *(VT*)a = op((*(VT*)x), (*(VT*)a));
+        x += N;
+        a += N;
+        s -= N;
+      }
+      while (s-- > 0) {
+        *a = op(*a, *x);
+        a++;
+        x++;
+      }
+    }
+  }
+};
+
+} // namespace
+
+void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+
+  if (in.dtype() == float32) {
+    if (reduce_type_ == Reduce::Sum) {
+      reduction_op<float, float>(
+          in,
+          out,
+          axes_,
+          0,
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              SumReduction<float, simd_float16>>(),
+          [](const auto* x, auto* accum, int size) {
+            float acc;
+            vDSP_sve((const float*)x, 1, &acc, size);
+            (*accum) += acc;
+          },
+          [](auto* accum, auto x) { *accum += x; });
+      return;
+    } else if (reduce_type_ == Reduce::Max) {
+      reduction_op<float, float>(
+          in,
+          out,
+          axes_,
+          -std::numeric_limits<float>::infinity(),
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              MaxReduction<float, simd_float16>>(),
+          [](const auto* x, auto* accum, int size) {
+            float max;
+            vDSP_maxv((const float*)x, 1, &max, size);
+            (*accum) = (*accum < max) ? max : *accum;
+          },
+          [](auto* accum, auto x) { (*accum) = (*accum < x) ? x : *accum; });
+      return;
+    } else if (reduce_type_ == Reduce::Min) {
+      reduction_op<float, float>(
+          in,
+          out,
+          axes_,
+          std::numeric_limits<float>::infinity(),
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              MinReduction<float, simd_float16>>(),
+          [](const auto* x, auto* accum, int size) {
+            float min;
+            vDSP_minv((const float*)x, 1, &min, size);
+            (*accum) = (*accum > min) ? min : *accum;
+          },
+          [](auto* accum, auto x) { (*accum) = (*accum > x) ? x : *accum; });
+      return;
+    }
+  }
+  // TODO: Add integer addition and min/max using the templates above and
+  //       simd_int16 and friends.
+  eval(inputs, out);
+}
+
+} // namespace mlx::core

mlx/backend/accelerate/softmax.cpp

@@ -0,0 +1,393 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <cassert>
+#include <limits>
+
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#include <arm_neon.h>
+#endif
+
+#include <simd/math.h>
+#include <simd/vector.h>
+
+#include "mlx/backend/common/copy.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+namespace {
+
+/**
+ * Compute exp(x) in an optimizer friendly way as follows:
+ *
+ * First change the problem to computing 2**y where y = x / ln(2).
+ *
+ * Now we will compute 2**y as 2**y1 * 2**y2 where y1 is the integer part
+ * `ipart` and y2 is fractional part. For the integer part we perform bit
+ * shifting and for the fractional part we use a polynomial approximation.
+ *
+ * The algorithm and constants of the polynomial taken from
+ * https://github.com/akohlmey/fastermath/blob/master/src/exp.c which took them
+ * from Cephes math library.
+ *
+ * Note: The implementation below is a general fast exp. There could be faster
+ *       implementations for numbers strictly < 0.
+ */
+inline simd_float16 simd_fast_exp(simd_float16 x_init) {
+  auto x = x_init * 1.442695; // multiply with log_2(e)
+  simd_float16 ipart, fpart;
+  simd_int16 epart;
+  x = simd_clamp(x, -80, 80);
+  ipart = simd::floor(x + 0.5);
+  fpart = x - ipart;
+
+  x = 1.535336188319500e-4f;
+  x = x * fpart + 1.339887440266574e-3f;
+  x = x * fpart + 9.618437357674640e-3f;
+  x = x * fpart + 5.550332471162809e-2f;
+  x = x * fpart + 2.402264791363012e-1f;
+  x = x * fpart + 6.931472028550421e-1f;
+  x = x * fpart + 1.000000000000000f;
+
+  // generate 2**ipart in the floating point representation using integer
+  // bitshifting
+  epart = (simd_int(ipart) + 127) << 23;
+
+  // Avoid supressing NaNs
+  simd_int16 eq = (x_init == x_init);
+  return simd_bitselect(x_init, (*(simd_float16*)&epart) * x, eq);
+}
+
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+/**
+ * The ARM neon equivalent of the fast exp above.
+ */
+inline float16x8_t neon_fast_exp(float16x8_t x) {
+  x = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
+  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f))); // clamp under with -14
+  x = vminq_f16(x, vdupq_n_f16(float16_t(14.f))); // clamp over with 14
+
+  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(float16_t(0.5f))));
+  float16x8_t fpart = vsubq_f16(x, ipart);
+
+  x = vdupq_n_f16(float16_t(1.535336188319500e-4f));
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.000000000000000f)), x, fpart);
+
+  // generate 2**ipart in the floating point representation using integer
+  // bitshifting
+  int16x8_t epart = vcvtq_s16_f16(ipart);
+  epart = vaddq_s16(epart, vdupq_n_s16(15));
+  epart = vshlq_n_s16(epart, 10);
+
+  return vmulq_f16(vreinterpretq_f16_s16(epart), x);
+}
+
+/**
+ * Implementation of folding maximum for ARM neon. This should possibly be
+ * refactored out of softmax.cpp at some point.
+ */
+inline float16_t neon_reduce_max(float16x8_t x) {
+  float16x4_t y;
+  y = vpmax_f16(vget_low_f16(x), vget_high_f16(x));
+  y = vpmax_f16(y, y);
+  y = vpmax_f16(y, y);
+  return vget_lane_f16(y, 0);
+}
+
+/**
+ * Implementation of folding sum for ARM neon. This should possibly be
+ * refactored out of softmax.cpp at some point.
+ */
+inline float16_t neon_reduce_add(float16x8_t x) {
+  float16x4_t y;
+  float16x4_t zero = vdup_n_f16(0);
+  y = vpadd_f16(vget_low_f16(x), vget_high_f16(x));
+  y = vpadd_f16(y, zero);
+  y = vpadd_f16(y, zero);
+  return vget_lane_f16(y, 0);
+}
+
+template <typename T, typename VT>
+struct NeonFp16SimdOps {
+  VT init(T a) {
+    return vdupq_n_f16(a);
+  }
+
+  VT load(const T* a) {
+    return vld1q_f16(a);
+  }
+
+  void store(T* dst, VT x) {
+    vst1q_f16(dst, x);
+  }
+
+  VT max(VT a, VT b) {
+    return vmaxq_f16(a, b);
+  }
+
+  VT exp(VT x) {
+    return neon_fast_exp(x);
+  }
+
+  VT add(VT a, VT b) {
+    return vaddq_f16(a, b);
+  }
+
+  VT sub(VT a, T b) {
+    return vsubq_f16(a, vdupq_n_f16(b));
+  }
+
+  VT mul(VT a, VT b) {
+    return vmulq_f16(a, b);
+  }
+
+  VT mul(VT a, T b) {
+    return vmulq_f16(a, vdupq_n_f16(b));
+  }
+
+  T reduce_max(VT x) {
+    return neon_reduce_max(x);
+  }
+
+  T reduce_add(VT x) {
+    return neon_reduce_add(x);
+  }
+};
+
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+template <typename T, typename VT>
+struct AccelerateSimdOps {
+  VT init(T a) {
+    return a;
+  }
+
+  VT load(const T* a) {
+    return *(VT*)a;
+  }
+
+  void store(T* dst, VT x) {
+    *(VT*)dst = x;
+  }
+
+  VT max(VT a, VT b) {
+    return simd_max(a, b);
+  }
+
+  VT exp(VT x) {
+    return simd_fast_exp(x);
+  }
+
+  VT add(VT a, VT b) {
+    return a + b;
+  }
+
+  VT sub(VT a, T b) {
+    return a - b;
+  }
+
+  VT mul(VT a, VT b) {
+    return a * b;
+  }
+
+  VT mul(VT a, T b) {
+    return a * b;
+  }
+
+  T reduce_max(VT x) {
+    return simd_reduce_max(x);
+  }
+
+  T reduce_add(VT x) {
+    return simd_reduce_add(x);
+  }
+};
+
+template <typename T, typename AccT, typename VT, typename Ops, int N>
+void softmax(const array& in, array& out) {
+  Ops ops;
+
+  const T* in_ptr = in.data<T>();
+  T* out_ptr = out.data<T>();
+  int M = in.shape().back();
+  int L = in.data_size() / M;
+  const T* current_in_ptr;
+  T* current_out_ptr;
+
+  for (int i = 0; i < L; i++, in_ptr += M, out_ptr += M) {
+    // Find the maximum
+    current_in_ptr = in_ptr;
+    VT vmaximum = ops.init(-std::numeric_limits<float>::infinity());
+    size_t s = M;
+    while (s >= N) {
+      VT vals;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vals = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vals[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vmaximum = ops.max(vals, vmaximum);
+      current_in_ptr += N;
+      s -= N;
+    }
+    AccT maximum = ops.reduce_max(vmaximum);
+    while (s-- > 0) {
+      maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
+      current_in_ptr++;
+    }
+
+    // Compute the normalizer and the exponentials
+    VT vnormalizer = ops.init(0.0);
+    current_out_ptr = out_ptr;
+    current_in_ptr = in_ptr;
+    s = M;
+    while (s >= N) {
+      VT vexp;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vexp = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vexp = ops.exp(ops.sub(vexp, maximum));
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, vexp);
+      }
+      vnormalizer = ops.add(vnormalizer, vexp);
+      current_in_ptr += N;
+      current_out_ptr += N;
+      s -= N;
+    }
+    AccT normalizer = ops.reduce_add(vnormalizer);
+    while (s-- > 0) {
+      AccT _exp = std::exp(*current_in_ptr - maximum);
+      if (std::is_same<T, AccT>::value) {
+        *current_out_ptr = _exp;
+      }
+      normalizer += _exp;
+      current_in_ptr++;
+      current_out_ptr++;
+    }
+    normalizer = 1 / normalizer;
+
+    // Normalize
+    current_out_ptr = out_ptr;
+    current_in_ptr = in_ptr;
+    s = M;
+    while (s >= N) {
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
+      } else {
+        VT vexp;
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+        vexp = ops.mul(ops.exp(ops.sub(vexp, maximum)), normalizer);
+        for (int i = 0; i < N; ++i) {
+          current_out_ptr[i] = vexp[i];
+        }
+        current_in_ptr += N;
+      }
+      current_out_ptr += N;
+      s -= N;
+    }
+    while (s-- > 0) {
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr *= normalizer;
+      } else {
+        AccT _exp = std::exp(*current_in_ptr - maximum);
+        *current_out_ptr = static_cast<T>(_exp * normalizer);
+        current_in_ptr++;
+      }
+      current_out_ptr++;
+    }
+  }
+}
+
+} // namespace
+
+void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+
+  // Make sure that the last dimension is contiguous
+  auto check_input = [](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) {
+      return x;
+    } else {
+      array x_copy(x.shape(), x.dtype(), nullptr, {});
+      copy(x, x_copy, CopyType::General);
+      return x_copy;
+    }
+  };
+  array in = check_input(std::move(inputs[0]));
+  out.set_data(
+      allocator::malloc_or_wait(in.data_size() * in.itemsize()),
+      in.data_size(),
+      in.strides(),
+      in.flags());
+
+  switch (in.dtype()) {
+    case bool_:
+    case uint8:
+    case uint16:
+    case uint32:
+    case uint64:
+    case int8:
+    case int16:
+    case int32:
+    case int64:
+      throw std::invalid_argument(
+          "Softmax is defined only for floating point types");
+      break;
+    case float32:
+      softmax<
+          float,
+          float,
+          simd_float16,
+          AccelerateSimdOps<float, simd_float16>,
+          16>(in, out);
+      break;
+    case float16:
+      if (precise_) {
+        softmax<
+            float16_t,
+            float,
+            simd_float16,
+            AccelerateSimdOps<float, simd_float16>,
+            16>(in, out);
+      } else {
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        softmax<
+            float16_t,
+            float16_t,
+            float16x8_t,
+            NeonFp16SimdOps<float16_t, float16x8_t>,
+            8>(in, out);
+#else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        eval(inputs, out); // Redirect to common backend for consistency
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+      }
+      break;
+    case bfloat16:
+      eval(inputs, out);
+      break;
+    case complex64:
+      eval(inputs, out);
+      break;
+  }
+}
+
+} // namespace mlx::core

mlx/backend/accelerate/utils.h

@@ -0,0 +1,28 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#pragma once
+
+#include <Accelerate/Accelerate.h>
+#include "mlx/dtype.h"
+
+namespace mlx::core {
+
+BNNSDataType to_bnns_dtype(Dtype mlx_dtype) {
+  uint32_t size_bits = size_of(mlx_dtype) * 8;
+  switch (kindof(mlx_dtype)) {
+    case Dtype::Kind::b:
+      return BNNSDataTypeBoolean;
+    case Dtype::Kind::u:
+      return BNNSDataType(BNNSDataTypeUIntBit | size_bits);
+    case Dtype::Kind::i:
+      return BNNSDataType(BNNSDataTypeIntBit | size_bits);
+    case Dtype::Kind::f:
+      return BNNSDataType(BNNSDataTypeFloatBit | size_bits);
+    case Dtype::Kind::V:
+      return BNNSDataTypeBFloat16;
+    case Dtype::Kind::c:
+      throw std::invalid_argument("BNNS does not support complex types");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/CMakeLists.txt

@@ -1,9 +1,62 @@
+if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+  set(COMPILER ${CMAKE_C_COMPILER})
+  set(CLANG TRUE)
+else()
+  set(COMPILER ${CMAKE_CXX_COMPILER})
+endif()
+
+add_custom_command(
+  OUTPUT compiled_preamble.cpp
+  COMMAND
+    /bin/bash ${CMAKE_CURRENT_SOURCE_DIR}/make_compiled_preamble.sh
+    ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp ${COMPILER}
+    ${PROJECT_SOURCE_DIR} ${CLANG}
+  DEPENDS make_compiled_preamble.sh
+          compiled_preamble.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/half_types.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/fp16.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/bf16.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/complex.h
+          ops.h)
+
+add_custom_target(cpu_compiled_preamble DEPENDS compiled_preamble.cpp)
+
+add_dependencies(mlx cpu_compiled_preamble)
+
 target_sources(
   mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/broadcasting.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/common.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eigh.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/erf.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/masked_mm.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce_utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp)
+          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/svd.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/cholesky.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
+          ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp)
+
+if(IOS)
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled_nocpu.cpp)
+else()
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled_cpu.cpp)
+endif()

mlx/backend/common/arange.h

@@ -0,0 +1,74 @@
+// Copyright © 2023 Apple Inc.
+
+#pragma once
+
+#include "mlx/allocator.h"
+#include "mlx/array.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename T>
+void arange(T start, T next, array& out, size_t size) {
+  auto ptr = out.data<T>();
+  auto step_size = next - start;
+  for (int i = 0; i < size; ++i) {
+    ptr[i] = start;
+    start += step_size;
+  }
+}
+
+} // namespace
+
+void arange(
+    const std::vector<array>& inputs,
+    array& out,
+    double start,
+    double step) {
+  assert(inputs.size() == 0);
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  switch (out.dtype()) {
+    case bool_:
+      throw std::runtime_error("Bool type unsupported for arange.");
+      break;
+    case uint8:
+      arange<uint8_t>(start, start + step, out, out.size());
+      break;
+    case uint16:
+      arange<uint16_t>(start, start + step, out, out.size());
+      break;
+    case uint32:
+      arange<uint32_t>(start, start + step, out, out.size());
+      break;
+    case uint64:
+      arange<uint64_t>(start, start + step, out, out.size());
+      break;
+    case int8:
+      arange<int8_t>(start, start + step, out, out.size());
+      break;
+    case int16:
+      arange<int16_t>(start, start + step, out, out.size());
+      break;
+    case int32:
+      arange<int32_t>(start, start + step, out, out.size());
+      break;
+    case int64:
+      arange<int64_t>(start, start + step, out, out.size());
+      break;
+    case float16:
+      arange<float16_t>(start, start + step, out, out.size());
+      break;
+    case float32:
+      arange<float>(start, start + step, out, out.size());
+      break;
+    case bfloat16:
+      arange<bfloat16_t>(start, start + step, out, out.size());
+      break;
+    case complex64:
+      arange<complex64_t>(start, start + step, out, out.size());
+      break;
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/arg_reduce.cpp

@@ -0,0 +1,112 @@
+// Copyright © 2023 Apple Inc.
+
+#include <cassert>
+
+#include "mlx/primitives.h"
+#include "utils.h"
+
+namespace mlx::core {
+
+namespace {
+
+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];
+  std::vector<size_t> strides = in.strides();
+  std::vector<int> shape = in.shape();
+  strides.erase(strides.begin() + axis);
+  shape.erase(shape.begin() + axis);
+  for (uint32_t i = 0; i < out.size(); ++i) {
+    auto loc = elem_to_loc(i, shape, strides);
+    auto in_ptr = in.data<InT>() + loc;
+    uint32_t ind_v = 0;
+    InT v = (*in_ptr);
+    for (uint32_t j = 0; j < axis_size; ++j, in_ptr += axis_stride) {
+      op(j, (*in_ptr), &ind_v, &v);
+    }
+    out.data<uint32_t>()[i] = ind_v;
+  }
+}
+
+template <typename InT>
+void arg_reduce_dispatch(
+    const array& in,
+    array& out,
+    ArgReduce::ReduceType rtype,
+    int axis) {
+  switch (rtype) {
+    case ArgReduce::ArgMin: {
+      auto op = [](auto ind_x, auto x, auto ind_y, auto y) {
+        if (x < (*y)) {
+          (*y) = x;
+          (*ind_y) = ind_x;
+        }
+      };
+      arg_reduce<InT>(in, out, op, axis);
+      break;
+    }
+    case ArgReduce::ArgMax: {
+      auto op = [](auto ind_x, auto x, auto ind_y, auto y) {
+        if (x > (*y)) {
+          (*y) = x;
+          (*ind_y) = ind_x;
+        }
+      };
+      arg_reduce<InT>(in, out, op, axis);
+      break;
+    }
+  }
+}
+
+} // namespace
+
+void ArgReduce::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  switch (in.dtype()) {
+    case bool_:
+      arg_reduce_dispatch<bool>(in, out, reduce_type_, axis_);
+      break;
+    case uint8:
+      arg_reduce_dispatch<uint8_t>(in, out, reduce_type_, axis_);
+      break;
+    case uint16:
+      arg_reduce_dispatch<uint16_t>(in, out, reduce_type_, axis_);
+      break;
+    case uint32:
+      arg_reduce_dispatch<uint32_t>(in, out, reduce_type_, axis_);
+      break;
+    case uint64:
+      arg_reduce_dispatch<uint64_t>(in, out, reduce_type_, axis_);
+      break;
+    case int8:
+      arg_reduce_dispatch<int8_t>(in, out, reduce_type_, axis_);
+      break;
+    case int16:
+      arg_reduce_dispatch<int16_t>(in, out, reduce_type_, axis_);
+      break;
+    case int32:
+      arg_reduce_dispatch<int32_t>(in, out, reduce_type_, axis_);
+      break;
+    case int64:
+      arg_reduce_dispatch<int64_t>(in, out, reduce_type_, axis_);
+      break;
+    case float16:
+      arg_reduce_dispatch<float16_t>(in, out, reduce_type_, axis_);
+      break;
+    case float32:
+      arg_reduce_dispatch<float>(in, out, reduce_type_, axis_);
+      break;
+    case bfloat16:
+      arg_reduce_dispatch<bfloat16_t>(in, out, reduce_type_, axis_);
+      break;
+    case complex64:
+      arg_reduce_dispatch<complex64_t>(in, out, reduce_type_, axis_);
+      break;
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/binary.cpp

@@ -0,0 +1,331 @@
+// Copyright © 2023 Apple Inc.
+
+#include <cassert>
+#include <cmath>
+#include <sstream>
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/common/binary_two.h"
+#include "mlx/backend/common/ops.h"
+#include "mlx/primitives.h"
+#include "mlx/utils.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename T, typename U, typename Op>
+void comparison_op(const array& a, const array& b, array& out, Op op) {
+  DefaultScalarVector<T, U, Op> opsv(op);
+  DefaultVectorScalar<T, U, Op> opvs(op);
+  DefaultVectorVector<T, U, Op> opvv(op);
+  binary_op<T, U>(a, b, out, op, opsv, opvs, opvv);
+}
+
+template <typename Op>
+void comparison_op(const array& a, const array& b, array& out, Op op) {
+  switch (a.dtype()) {
+    case bool_:
+      comparison_op<bool, bool>(a, b, out, op);
+      break;
+    case uint8:
+      comparison_op<uint8_t, bool>(a, b, out, op);
+      break;
+    case uint16:
+      comparison_op<uint16_t, bool>(a, b, out, op);
+      break;
+    case uint32:
+      comparison_op<uint32_t, bool>(a, b, out, op);
+      break;
+    case uint64:
+      comparison_op<uint64_t, bool>(a, b, out, op);
+      break;
+    case int8:
+      comparison_op<int8_t, bool>(a, b, out, op);
+      break;
+    case int16:
+      comparison_op<int16_t, bool>(a, b, out, op);
+      break;
+    case int32:
+      comparison_op<int32_t, bool>(a, b, out, op);
+      break;
+    case int64:
+      comparison_op<int64_t, bool>(a, b, out, op);
+      break;
+    case float16:
+      comparison_op<float16_t, bool>(a, b, out, op);
+      break;
+    case float32:
+      comparison_op<float, bool>(a, b, out, op);
+      break;
+    case bfloat16:
+      comparison_op<bfloat16_t, bool>(a, b, out, op);
+      break;
+    case complex64:
+      comparison_op<complex64_t, bool>(a, b, out, op);
+      break;
+  }
+}
+
+} // namespace
+
+void Add::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Add());
+}
+
+void DivMod::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto integral_op = [](auto x, auto y) {
+    return std::make_pair(x / y, x % y);
+  };
+  auto float_op = [](auto x, auto y) {
+    return std::make_pair(std::trunc(x / y), std::fmod(x, y));
+  };
+  switch (outputs[0].dtype()) {
+    case bool_:
+      binary_op<bool>(a, b, outputs, integral_op);
+    case uint8:
+      binary_op<uint8_t>(a, b, outputs, integral_op);
+      break;
+    case uint16:
+      binary_op<uint16_t>(a, b, outputs, integral_op);
+      break;
+    case uint32:
+      binary_op<uint32_t>(a, b, outputs, integral_op);
+      break;
+    case uint64:
+      binary_op<uint64_t>(a, b, outputs, integral_op);
+      break;
+    case int8:
+      binary_op<int8_t>(a, b, outputs, integral_op);
+      break;
+    case int16:
+      binary_op<int16_t>(a, b, outputs, integral_op);
+      break;
+    case int32:
+      binary_op<int32_t>(a, b, outputs, integral_op);
+      break;
+    case int64:
+      binary_op<int64_t>(a, b, outputs, integral_op);
+      break;
+    case float16:
+      binary_op<float16_t>(a, b, outputs, float_op);
+      break;
+    case float32:
+      binary_op<float>(a, b, outputs, float_op);
+      break;
+    case bfloat16:
+      binary_op<bfloat16_t>(a, b, outputs, float_op);
+      break;
+    case complex64:
+      // Should never get here
+      throw std::runtime_error("[DivMod] Complex type not supported");
+      break;
+  }
+}
+
+void Divide::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Divide());
+}
+
+void Remainder::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Remainder());
+}
+
+void Equal::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  if (equal_nan_) {
+    comparison_op(inputs[0], inputs[1], out, detail::NaNEqual());
+  } else {
+    comparison_op(inputs[0], inputs[1], out, detail::Equal());
+  }
+}
+
+void Greater::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::Greater());
+}
+
+void GreaterEqual::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::GreaterEqual());
+}
+
+void Less::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::Less());
+}
+
+void LessEqual::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::LessEqual());
+}
+
+void LogAddExp::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  if (out.dtype() == float32) {
+    binary_op<float>(a, b, out, detail::LogAddExp());
+  } else if (out.dtype() == float16) {
+    binary_op<float16_t>(a, b, out, detail::LogAddExp());
+  } else if (out.dtype() == bfloat16) {
+    binary_op<bfloat16_t>(a, b, out, detail::LogAddExp());
+  } else if (issubdtype(out.dtype(), inexact)) {
+    std::ostringstream err;
+    err << "[logaddexp] Does not support " << out.dtype();
+    throw std::invalid_argument(err.str());
+  } else {
+    throw std::invalid_argument(
+        "[logaddexp] Cannot compute logaddexp for arrays with"
+        " non floating point type.");
+  }
+}
+
+void LogicalAnd::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalAnd());
+}
+
+void LogicalOr::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalOr requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalOr());
+}
+
+void Maximum::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Maximum());
+}
+
+void Minimum::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Minimum());
+}
+
+void Multiply::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Multiply());
+}
+
+void NotEqual::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  comparison_op(inputs[0], inputs[1], out, detail::NotEqual());
+}
+
+void Power::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Power());
+}
+
+void Subtract::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  binary(a, b, out, detail::Subtract());
+}
+
+void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto dispatch_type = [&a, &b, &out](auto op) {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool>(a, b, out, op);
+      case uint8:
+        binary_op<uint8_t>(a, b, out, op);
+        break;
+      case uint16:
+        binary_op<uint16_t>(a, b, out, op);
+        break;
+      case uint32:
+        binary_op<uint32_t>(a, b, out, op);
+        break;
+      case uint64:
+        binary_op<uint64_t>(a, b, out, op);
+        break;
+      case int8:
+        binary_op<int8_t>(a, b, out, op);
+        break;
+      case int16:
+        binary_op<int16_t>(a, b, out, op);
+        break;
+      case int32:
+        binary_op<int32_t>(a, b, out, op);
+        break;
+      case int64:
+        binary_op<int64_t>(a, b, out, op);
+        break;
+      default:
+        throw std::runtime_error(
+            "[BitwiseBinary::eval_cpu] Type not supported");
+        break;
+    }
+  };
+  switch (op_) {
+    case BitwiseBinary::And:
+      dispatch_type(detail::BitwiseAnd());
+      break;
+    case BitwiseBinary::Or:
+      dispatch_type(detail::BitwiseOr());
+      break;
+    case BitwiseBinary::Xor:
+      dispatch_type(detail::BitwiseXor());
+      break;
+    case BitwiseBinary::LeftShift:
+      dispatch_type(detail::LeftShift());
+      break;
+    case BitwiseBinary::RightShift:
+      dispatch_type(detail::RightShift());
+      break;
+  }
+}
+
+void ArcTan2::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  if (out.dtype() == float32) {
+    binary_op<float>(a, b, out, detail::ArcTan2());
+  } else if (out.dtype() == float16) {
+    binary_op<float16_t>(a, b, out, detail::ArcTan2());
+  } else if (out.dtype() == bfloat16) {
+    binary_op<bfloat16_t>(a, b, out, detail::ArcTan2());
+  } else if (issubdtype(out.dtype(), inexact)) {
+    std::ostringstream err;
+    err << "[arctan2] Does not support " << out.dtype();
+    throw std::invalid_argument(err.str());
+  } else {
+    throw std::invalid_argument(
+        "[arctan2] Cannot compute inverse tangent for arrays"
+        " with non floating point type.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/binary.h

@@ -1,6 +1,7 @@
 // Copyright © 2023 Apple Inc.
 
 #pragma once
+#include <cassert>
 
 #include "mlx/allocator.h"
 #include "mlx/array.h"
@@ -8,6 +9,8 @@
 
 namespace mlx::core {
 
+namespace {
+
 enum class BinaryOpType {
   ScalarScalar,
   ScalarVector,
@@ -16,7 +19,7 @@ enum class BinaryOpType {
   General,
 };
 
-inline BinaryOpType get_binary_op_type(const array& a, const array& b) {
+BinaryOpType get_binary_op_type(const array& a, const array& b) {
   BinaryOpType bopt;
   if (a.data_size() == 1 && b.data_size() == 1) {
     bopt = BinaryOpType::ScalarScalar;
@@ -25,8 +28,8 @@ inline BinaryOpType get_binary_op_type(const array& a, const array& b) {
   } else if (b.data_size() == 1 && a.flags().contiguous) {
     bopt = BinaryOpType::VectorScalar;
   } else if (
-      (a.flags().row_contiguous && b.flags().row_contiguous) ||
-      (a.flags().col_contiguous && b.flags().col_contiguous)) {
+      a.flags().row_contiguous && b.flags().row_contiguous ||
+      a.flags().col_contiguous && b.flags().col_contiguous) {
     bopt = BinaryOpType::VectorVector;
   } else {
     bopt = BinaryOpType::General;
@@ -34,24 +37,29 @@ inline BinaryOpType get_binary_op_type(const array& a, const array& b) {
   return bopt;
 }
 
-inline void set_binary_op_output_data(
+void set_binary_op_output_data(
     const array& a,
     const array& b,
     array& out,
-    BinaryOpType bopt) {
+    BinaryOpType bopt,
+    bool donate_with_move = false) {
   bool b_donatable = is_donatable(b, out);
   bool a_donatable = is_donatable(a, out);
   switch (bopt) {
     case BinaryOpType::ScalarScalar:
       out.set_data(
-          allocator::malloc(out.itemsize()), 1, a.strides(), a.flags());
+          allocator::malloc_or_wait(out.itemsize()), 1, a.strides(), a.flags());
       break;
     case BinaryOpType::ScalarVector:
       if (b_donatable) {
-        out.copy_shared_buffer(b);
+        if (donate_with_move) {
+          out.move_shared_buffer(b);
+        } else {
+          out.copy_shared_buffer(b);
+        }
       } else {
         out.set_data(
-            allocator::malloc(b.data_size() * out.itemsize()),
+            allocator::malloc_or_wait(b.data_size() * out.itemsize()),
             b.data_size(),
             b.strides(),
             b.flags());
@@ -59,10 +67,14 @@ inline void set_binary_op_output_data(
       break;
     case BinaryOpType::VectorScalar:
       if (a_donatable) {
-        out.copy_shared_buffer(a);
+        if (donate_with_move) {
+          out.move_shared_buffer(a);
+        } else {
+          out.copy_shared_buffer(a);
+        }
       } else {
         out.set_data(
-            allocator::malloc(a.data_size() * out.itemsize()),
+            allocator::malloc_or_wait(a.data_size() * out.itemsize()),
             a.data_size(),
             a.strides(),
             a.flags());
@@ -70,12 +82,20 @@ inline void set_binary_op_output_data(
       break;
     case BinaryOpType::VectorVector:
       if (a_donatable) {
-        out.copy_shared_buffer(a);
+        if (donate_with_move) {
+          out.move_shared_buffer(a);
+        } else {
+          out.copy_shared_buffer(a);
+        }
       } else if (b_donatable) {
-        out.copy_shared_buffer(b);
+        if (donate_with_move) {
+          out.move_shared_buffer(b);
+        } else {
+          out.copy_shared_buffer(b);
+        }
       } else {
         out.set_data(
-            allocator::malloc(a.data_size() * out.itemsize()),
+            allocator::malloc_or_wait(a.data_size() * out.itemsize()),
             a.data_size(),
             a.strides(),
             a.flags());
@@ -83,15 +103,428 @@ inline void set_binary_op_output_data(
       break;
     case BinaryOpType::General:
       if (a_donatable && a.flags().row_contiguous && a.size() == out.size()) {
-        out.copy_shared_buffer(a);
+        if (donate_with_move) {
+          out.move_shared_buffer(a);
+        } else {
+          out.copy_shared_buffer(a);
+        }
       } else if (
           b_donatable && b.flags().row_contiguous && b.size() == out.size()) {
-        out.copy_shared_buffer(b);
+        if (donate_with_move) {
+          out.move_shared_buffer(b);
+        } else {
+          out.copy_shared_buffer(b);
+        }
       } else {
-        out.set_data(allocator::malloc(out.nbytes()));
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
       }
       break;
   }
 }
 
+struct UseDefaultBinaryOp {};
+
+template <typename T, typename U, typename Op>
+struct DefaultVectorScalar {
+  Op op;
+
+  DefaultVectorScalar(Op op_) : op(op_) {}
+
+  void operator()(const T* a, const T* b, U* dst, int size) {
+    T scalar = *b;
+    while (size-- > 0) {
+      *dst = op(*a, scalar);
+      dst++;
+      a++;
+    }
+  }
+};
+
+template <typename T, typename U, typename Op>
+struct DefaultScalarVector {
+  Op op;
+
+  DefaultScalarVector(Op op_) : op(op_) {}
+
+  void operator()(const T* a, const T* b, U* dst, int size) {
+    T scalar = *a;
+    while (size-- > 0) {
+      *dst = op(scalar, *b);
+      dst++;
+      b++;
+    }
+  }
+};
+
+template <typename T, typename U, typename Op>
+struct DefaultVectorVector {
+  Op op;
+
+  DefaultVectorVector(Op op_) : op(op_) {}
+
+  void operator()(const T* a, const T* b, U* dst, int size) {
+    while (size-- > 0) {
+      *dst = op(*a, *b);
+      dst++;
+      a++;
+      b++;
+    }
+  }
+};
+
+template <typename T, typename U, typename Op, int D, bool Strided>
+void binary_op_dims(
+    const T* a,
+    const T* b,
+    U* out,
+    Op op,
+    const std::vector<int>& shape,
+    const std::vector<size_t>& a_strides,
+    const std::vector<size_t>& b_strides,
+    const std::vector<size_t>& out_strides,
+    int axis) {
+  auto stride_a = a_strides[axis];
+  auto stride_b = b_strides[axis];
+  auto stride_out = out_strides[axis];
+  auto N = shape[axis];
+
+  for (int i = 0; i < N; i++) {
+    if constexpr (D > 1) {
+      binary_op_dims<T, U, Op, D - 1, Strided>(
+          a, b, out, op, shape, a_strides, b_strides, out_strides, axis + 1);
+    } else {
+      if constexpr (Strided) {
+        op(a, b, out, stride_out);
+      } else {
+        *out = op(*a, *b);
+      }
+    }
+    out += stride_out;
+    a += stride_a;
+    b += stride_b;
+  }
+}
+
+template <typename T, typename U, bool Strided, typename Op>
+void binary_op_dispatch_dims(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    int dim,
+    const std::vector<int>& shape,
+    const std::vector<size_t>& a_strides,
+    const std::vector<size_t>& b_strides,
+    const std::vector<size_t>& out_strides) {
+  const T* a_ptr = a.data<T>();
+  const T* b_ptr = b.data<T>();
+  U* out_ptr = out.data<U>();
+  switch (dim) {
+    case 1:
+      binary_op_dims<T, U, Op, 1, Strided>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          op,
+          shape,
+          a_strides,
+          b_strides,
+          out_strides,
+          0);
+      return;
+    case 2:
+      binary_op_dims<T, U, Op, 2, Strided>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          op,
+          shape,
+          a_strides,
+          b_strides,
+          out_strides,
+          0);
+      return;
+    case 3:
+      binary_op_dims<T, U, Op, 3, Strided>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          op,
+          shape,
+          a_strides,
+          b_strides,
+          out_strides,
+          0);
+      return;
+  }
+
+  ContiguousIterator<size_t> a_it(shape, a_strides, dim - 3);
+  ContiguousIterator<size_t> b_it(shape, b_strides, dim - 3);
+  size_t stride = out_strides[dim - 4];
+  for (size_t elem = 0; elem < a.size(); elem += stride) {
+    binary_op_dims<T, U, Op, 3, Strided>(
+        a_ptr + a_it.loc,
+        b_ptr + b_it.loc,
+        out_ptr + elem,
+        op,
+        shape,
+        a_strides,
+        b_strides,
+        out_strides,
+        dim - 3);
+    a_it.step();
+    b_it.step();
+  }
+}
+
+template <
+    typename T,
+    typename U,
+    typename Op,
+    typename OpSV,
+    typename OpVS,
+    typename OpVV>
+void binary_op(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    OpSV opsv,
+    OpVS opvs,
+    OpVV opvv) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  // The full computation is scalar scalar so call the base op once
+  if (bopt == BinaryOpType::ScalarScalar) {
+    *(out.data<U>()) = op(*a.data<T>(), *b.data<T>());
+    return;
+  }
+
+  // The full computation is scalar vector so delegate to the op
+  if (bopt == BinaryOpType::ScalarVector) {
+    opsv(a.data<T>(), b.data<T>(), out.data<U>(), b.data_size());
+    return;
+  }
+
+  // The full computation is vector scalar so delegate to the op
+  if (bopt == BinaryOpType::VectorScalar) {
+    opvs(a.data<T>(), b.data<T>(), out.data<U>(), a.data_size());
+    return;
+  }
+
+  // The full computation is vector vector so delegate to the op
+  if (bopt == BinaryOpType::VectorVector) {
+    opvv(a.data<T>(), b.data<T>(), out.data<U>(), out.size());
+    return;
+  }
+
+  // General computation so let's try to optimize
+  auto [new_shape, new_strides] = collapse_contiguous_dims(
+      a.shape(), {a.strides(), b.strides(), out.strides()});
+  const auto& a_strides = new_strides[0];
+  const auto& b_strides = new_strides[1];
+  const auto& strides = new_strides[2];
+
+  // Get the left-most dim such that the array is row contiguous after
+  auto leftmost_rc_dim = [&strides](const std::vector<size_t>& arr_strides) {
+    int d = arr_strides.size() - 1;
+    for (; d >= 0 && arr_strides[d] == strides[d]; d--) {
+    }
+    return d + 1;
+  };
+  auto a_rc_dim = leftmost_rc_dim(a_strides);
+  auto b_rc_dim = leftmost_rc_dim(b_strides);
+
+  // Get the left-most dim such that the array is a broadcasted "scalar" after
+  auto leftmost_s_dim = [](const std::vector<size_t>& arr_strides) {
+    int d = arr_strides.size() - 1;
+    for (; d >= 0 && arr_strides[d] == 0; d--) {
+    }
+    return d + 1;
+  };
+  auto a_s_dim = leftmost_s_dim(a_strides);
+  auto b_s_dim = leftmost_s_dim(b_strides);
+
+  auto ndim = new_shape.size();
+
+  // Case 1: LxM and FxM where L and F are broadcastable and M is row contiguous
+  int dim = ndim;
+  if (int d = std::max(a_rc_dim, b_rc_dim); d < ndim) {
+    bopt = BinaryOpType::VectorVector;
+    dim = d;
+    // Case 2: LxM and Fx1 where L and F are broadcastable and M is row
+    // contiguous
+  } else if (int d = std::max(a_rc_dim, b_s_dim); d < ndim) {
+    bopt = BinaryOpType::VectorScalar;
+    dim = d;
+    // Case 3: Lx1 and FxM where L and F are broadcastable and M is row
+    // contiguous
+  } else if (int d = std::max(a_s_dim, b_rc_dim); d < ndim) {
+    bopt = BinaryOpType::ScalarVector;
+    dim = d;
+  }
+
+  // Can be sure dim > 0 since otherwise we would have used one of the fully
+  // contiguous methods above. Except for the case that the flags do not
+  // correspond to the underlying contiguity.
+  if (dim == 0 || strides[dim - 1] < 16) {
+    bopt = BinaryOpType::General;
+    dim = ndim;
+  }
+
+  switch (bopt) {
+    case BinaryOpType::VectorVector:
+      binary_op_dispatch_dims<T, U, true>(
+          a, b, out, opvv, dim, new_shape, a_strides, b_strides, strides);
+      break;
+    case BinaryOpType::VectorScalar:
+      binary_op_dispatch_dims<T, U, true>(
+          a, b, out, opvs, dim, new_shape, a_strides, b_strides, strides);
+      break;
+    case BinaryOpType::ScalarVector:
+      binary_op_dispatch_dims<T, U, true>(
+          a, b, out, opsv, dim, new_shape, a_strides, b_strides, strides);
+      break;
+    default:
+      binary_op_dispatch_dims<T, U, false>(
+          a, b, out, op, dim, new_shape, a_strides, b_strides, strides);
+      break;
+  }
+}
+
+template <typename T, typename Op, typename OpSV, typename OpVS, typename OpVV>
+void binary_op(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    OpSV opsv,
+    OpVS opvs,
+    OpVV opvv) {
+  // TODO: The following mess of constexpr evaluations can probably be achieved
+  //       with template specializations and overloading. Would it be simpler?
+
+  if constexpr (std::is_same<decltype(opsv), UseDefaultBinaryOp>::value) {
+    if constexpr (std::is_same<decltype(opvs), UseDefaultBinaryOp>::value) {
+      if constexpr (std::is_same<decltype(opvv), UseDefaultBinaryOp>::value) {
+        // All ops are UseDefaultBinaryOp (why oh why would someone call that?)
+        binary_op<T, T>(
+            a,
+            b,
+            out,
+            op,
+            DefaultScalarVector<T, T, Op>(op),
+            DefaultVectorScalar<T, T, Op>(op),
+            DefaultVectorVector<T, T, Op>(op));
+      } else {
+        // opsv and opvs were UseDefaultBinaryOp
+        binary_op<T, T>(
+            a,
+            b,
+            out,
+            op,
+            DefaultScalarVector<T, T, Op>(op),
+            DefaultVectorScalar<T, T, Op>(op),
+            opvv);
+      }
+    } else if constexpr (std::is_same<decltype(opvv), UseDefaultBinaryOp>::
+                             value) {
+      // opsv and opvv were UseDefaultBinaryOp
+      binary_op<T, T>(
+          a,
+          b,
+          out,
+          op,
+          DefaultScalarVector<T, T, Op>(op),
+          opvs,
+          DefaultVectorVector<T, T, Op>(op));
+    } else {
+      // opsv was UseDefaultBinaryOp
+      binary_op<T, T>(
+          a, b, out, op, DefaultScalarVector<T, T, Op>(op), opvs, opvv);
+    }
+  } else if constexpr (std::is_same<decltype(opvs), UseDefaultBinaryOp>::
+                           value) {
+    if (std::is_same<decltype(opvv), UseDefaultBinaryOp>::value) {
+      // opvs and opvv were UseDefaultBinaryOp
+      binary_op<T, T>(
+          a,
+          b,
+          out,
+          op,
+          opsv,
+          DefaultVectorScalar<T, T, Op>(op),
+          DefaultVectorVector<T, T, Op>(op));
+    } else {
+      // opvs was UseDefaultBinaryOp
+      binary_op<T, T>(
+          a, b, out, op, opsv, DefaultVectorScalar<T, T, Op>(op), opvv);
+    }
+  } else if constexpr (std::is_same<decltype(opvv), UseDefaultBinaryOp>::
+                           value) {
+    // opvv was UseDefaultBinaryOp
+    binary_op<T, T>(
+        a, b, out, op, opsv, opvs, DefaultVectorVector<T, T, Op>(op));
+  } else {
+    // All ops provided
+    binary_op<T, T>(a, b, out, op, opsv, opvs, opvv);
+  }
+}
+
+template <typename T, typename Op>
+void binary_op(const array& a, const array& b, array& out, Op op) {
+  DefaultScalarVector<T, T, Op> opsv(op);
+  DefaultVectorScalar<T, T, Op> opvs(op);
+  DefaultVectorVector<T, T, Op> opvv(op);
+  binary_op<T, T>(a, b, out, op, opsv, opvs, opvv);
+}
+
+template <typename... Ops>
+void binary(const array& a, const array& b, array& out, Ops... ops) {
+  switch (out.dtype()) {
+    case bool_:
+      binary_op<bool>(a, b, out, ops...);
+      break;
+    case uint8:
+      binary_op<uint8_t>(a, b, out, ops...);
+      break;
+    case uint16:
+      binary_op<uint16_t>(a, b, out, ops...);
+      break;
+    case uint32:
+      binary_op<uint32_t>(a, b, out, ops...);
+      break;
+    case uint64:
+      binary_op<uint64_t>(a, b, out, ops...);
+      break;
+    case int8:
+      binary_op<int8_t>(a, b, out, ops...);
+      break;
+    case int16:
+      binary_op<int16_t>(a, b, out, ops...);
+      break;
+    case int32:
+      binary_op<int32_t>(a, b, out, ops...);
+      break;
+    case int64:
+      binary_op<int64_t>(a, b, out, ops...);
+      break;
+    case float16:
+      binary_op<float16_t>(a, b, out, ops...);
+      break;
+    case float32:
+      binary_op<float>(a, b, out, ops...);
+      break;
+    case bfloat16:
+      binary_op<bfloat16_t>(a, b, out, ops...);
+      break;
+    case complex64:
+      binary_op<complex64_t>(a, b, out, ops...);
+      break;
+  }
+}
+
+} // namespace
+
 } // namespace mlx::core

mlx/backend/common/binary_two.h

@@ -2,8 +2,8 @@
 
 #pragma once
 
+#include "mlx/backend/common/binary.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/cpu/binary.h"
 
 namespace mlx::core {
 
@@ -16,10 +16,10 @@ void binary_op_dims(
     U* out_a,
     U* out_b,
     Op op,
-    const Shape& shape,
-    const Strides& a_strides,
-    const Strides& b_strides,
-    const Strides& out_strides,
+    const std::vector<int>& shape,
+    const std::vector<size_t>& a_strides,
+    const std::vector<size_t>& b_strides,
+    const std::vector<size_t>& out_strides,
     int axis) {
   auto stride_a = a_strides[axis];
   auto stride_b = b_strides[axis];
@@ -58,14 +58,14 @@ void binary_op_dispatch_dims(
     Op op) {
   auto [shape, strides] = collapse_contiguous_dims(
       a.shape(), {a.strides(), b.strides(), out_a.strides()});
+  const auto& a_strides = strides[0];
+  const auto& b_strides = strides[1];
+  const auto& out_strides = strides[2];
   const T* a_ptr = a.data<T>();
   const T* b_ptr = b.data<T>();
   U* out_a_ptr = out_a.data<U>();
   U* out_b_ptr = out_b.data<U>();
 
-  const auto& a_strides = strides[0];
-  const auto& b_strides = strides[1];
-  const auto& out_strides = strides[2];
   int ndim = shape.size();
   switch (ndim) {
     case 1:
@@ -96,9 +96,9 @@ void binary_op_dispatch_dims(
       return;
   }
 
-  ContiguousIterator a_it(shape, a_strides, ndim - 2);
-  ContiguousIterator b_it(shape, b_strides, ndim - 2);
-  auto stride = out_strides[ndim - 3];
+  ContiguousIterator<size_t> a_it(shape, a_strides, ndim - 2);
+  ContiguousIterator<size_t> b_it(shape, b_strides, ndim - 2);
+  size_t stride = out_strides[ndim - 3];
   for (size_t elem = 0; elem < a.size(); elem += stride) {
     binary_op_dims<T, U, Op, 2>(
         a_ptr + a_it.loc,
@@ -120,10 +120,14 @@ template <typename T, typename U = T, typename Op>
 void binary_op(
     const array& a,
     const array& b,
-    array& out_a,
-    array& out_b,
-    Op op,
-    BinaryOpType bopt) {
+    std::vector<array>& outputs,
+    Op op) {
+  auto bopt = get_binary_op_type(a, b);
+  auto& out_a = outputs[0];
+  auto& out_b = outputs[1];
+  set_binary_op_output_data(a, b, out_a, bopt);
+  set_binary_op_output_data(a, b, out_b, bopt);
+
   // The full computation is scalar scalar so call the base op once
   if (bopt == BinaryOpType::General) {
     binary_op_dispatch_dims<T, U, Op>(a, b, out_a, out_b, op);
@@ -137,14 +141,14 @@ void binary_op(
   if (bopt == BinaryOpType::ScalarScalar) {
     std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
   } else if (bopt == BinaryOpType::ScalarVector) {
-    for (size_t i = 0; i < b.data_size(); ++i) {
+    for (size_t i = 0; i < b.size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
       b_ptr++;
     }
   } else if (bopt == BinaryOpType::VectorScalar) {
-    for (size_t i = 0; i < a.data_size(); ++i) {
+    for (size_t i = 0; i < a.size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
@@ -161,6 +165,55 @@ void binary_op(
   }
 }
 
+template <typename Op>
+void binary(
+    const array& a,
+    const array& b,
+    std::vector<array>& outputs,
+    Op op) {
+  switch (outputs[0].dtype()) {
+    case bool_:
+      binary_op<bool>(a, b, outputs, op);
+      break;
+    case uint8:
+      binary_op<uint8_t>(a, b, outputs, op);
+      break;
+    case uint16:
+      binary_op<uint16_t>(a, b, outputs, op);
+      break;
+    case uint32:
+      binary_op<uint32_t>(a, b, outputs, op);
+      break;
+    case uint64:
+      binary_op<uint64_t>(a, b, outputs, op);
+      break;
+    case int8:
+      binary_op<int8_t>(a, b, outputs, op);
+      break;
+    case int16:
+      binary_op<int16_t>(a, b, outputs, op);
+      break;
+    case int32:
+      binary_op<int32_t>(a, b, outputs, op);
+      break;
+    case int64:
+      binary_op<int64_t>(a, b, outputs, op);
+      break;
+    case float16:
+      binary_op<float16_t>(a, b, outputs, op);
+      break;
+    case float32:
+      binary_op<float>(a, b, outputs, op);
+      break;
+    case bfloat16:
+      binary_op<bfloat16_t>(a, b, outputs, op);
+      break;
+    case complex64:
+      binary_op<complex64_t>(a, b, outputs, op);
+      break;
+  }
+}
+
 } // namespace
 
 } // namespace mlx::core

mlx/backend/common/broadcasting.cpp

@@ -1,24 +0,0 @@
-// 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

@@ -1,11 +0,0 @@
-// 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

@@ -1,157 +0,0 @@
-// 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/cholesky.cpp

@@ -0,0 +1,74 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/copy.h"
+#include "mlx/backend/common/lapack.h"
+#include "mlx/linalg.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+void cholesky_impl(const array& a, array& factor, bool upper) {
+  // Lapack uses the column-major convention. We take advantage of the fact that
+  // the matrix should be symmetric:
+  //   (A)ᵀ = A
+  // and that a column-major lower triangular matrix is a row-major upper
+  // triangular matrix, so uplo is the opposite of what we would expect from
+  // upper
+
+  char uplo = (upper) ? 'L' : 'U';
+
+  // The decomposition is computed in place, so just copy the input to the
+  // output.
+  copy(
+      a,
+      factor,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+
+  const int N = a.shape(-1);
+  const size_t num_matrices = a.size() / (N * N);
+
+  float* matrix = factor.data<float>();
+
+  for (int i = 0; i < num_matrices; i++) {
+    // Compute Cholesky factorization.
+    int info;
+    MLX_LAPACK_FUNC(spotrf)
+    (
+        /* uplo = */ &uplo,
+        /* n = */ &N,
+        /* a = */ matrix,
+        /* lda = */ &N,
+        /* info = */ &info);
+
+    // TODO: We do nothing when the matrix is not positive semi-definite
+    // because throwing an error would result in a crash. If we figure out how
+    // to catch errors from the implementation we should throw.
+    if (info < 0) {
+      std::stringstream msg;
+      msg << "[cholesky] Cholesky decomposition failed with error code "
+          << info;
+      throw std::runtime_error(msg.str());
+    }
+
+    // Zero out the upper/lower triangle while advancing the pointer to the
+    // next matrix at the same time.
+    for (int row = 0; row < N; row++) {
+      if (upper) {
+        std::fill(matrix, matrix + row, 0);
+      } else {
+        std::fill(matrix + row + 1, matrix + N, 0);
+      }
+      matrix += N;
+    }
+  }
+}
+
+void Cholesky::eval(const std::vector<array>& inputs, array& output) {
+  if (inputs[0].dtype() != float32) {
+    throw std::runtime_error("[Cholesky::eval] only supports float32.");
+  }
+  cholesky_impl(inputs[0], output, upper_);
+}
+
+} // namespace mlx::core

mlx/backend/common/common.cpp

@@ -1,7 +1,6 @@
 // Copyright © 2024 Apple Inc.
 #include <cassert>
 
-#include "mlx/backend/common/broadcasting.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/primitives.h"
 
@@ -44,11 +43,22 @@ void AsStrided::eval(const std::vector<array>& inputs, array& out) {
 }
 
 void Broadcast::eval(const std::vector<array>& inputs, array& out) {
-  broadcast(inputs[0], out);
-}
-
-void BroadcastAxes::eval(const std::vector<array>& inputs, array& out) {
-  broadcast(inputs[0], out);
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+  std::vector<size_t> 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 Copy::eval(const std::vector<array>& inputs, array& out) {
@@ -75,19 +85,9 @@ void Depends::eval(
   }
 }
 
-void ExpandDims::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  auto strides = in.strides();
-  for (auto ax : axes_) {
-    strides.insert(strides.begin() + ax, 1);
-  }
-  out.copy_shared_buffer(in, strides, in.flags(), in.data_size());
-}
-
 void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  out.set_data(allocator::malloc(out.nbytes()));
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
   double numel = 1;
   for (auto ax : axes_) {
@@ -135,16 +135,15 @@ void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
     case bfloat16:
       *out.data<bfloat16_t>() = static_cast<bfloat16_t>(numel);
       break;
-    case float64:
-      *out.data<double>() = static_cast<double>(numel);
-      break;
     case complex64:
       *out.data<complex64_t>() = static_cast<complex64_t>(numel);
       break;
   }
 }
 
-std::pair<bool, Strides> prepare_reshape(const array& in, const array& out) {
+std::pair<bool, std::vector<size_t>> Reshape::prepare_reshape(
+    const array& in,
+    const array& out) {
   // Special case for empty arrays or row contiguous arrays
   if (in.size() == 0 || in.flags().row_contiguous) {
     return {false, out.strides()};
@@ -152,7 +151,8 @@ std::pair<bool, Strides> prepare_reshape(const array& in, const array& out) {
 
   // Special case for scalars
   if (in.ndim() == 0) {
-    return {false, Strides(out.ndim(), 0)};
+    std::vector<size_t> out_strides(out.ndim(), 0);
+    return {false, out_strides};
   }
 
   // Firstly let's collapse all the contiguous dimensions of the input
@@ -160,7 +160,7 @@ std::pair<bool, Strides> prepare_reshape(const array& in, const array& out) {
 
   // If shapes fit exactly in the contiguous dims then no copy is necessary so
   // let's check.
-  Strides out_strides;
+  std::vector<size_t> out_strides;
   bool copy_necessary = false;
   int j = 0;
   for (int i = 0; i < out.ndim(); i++) {
@@ -181,9 +181,9 @@ std::pair<bool, Strides> prepare_reshape(const array& in, const array& out) {
   return {copy_necessary, out_strides};
 }
 
-void shared_buffer_reshape(
+void Reshape::shared_buffer_reshape(
     const array& in,
-    const Strides& out_strides,
+    const std::vector<size_t>& out_strides,
     array& out) {
   auto flags = in.flags();
   if (flags.row_contiguous) {
@@ -249,18 +249,16 @@ void Split::eval(
   }
 }
 
-void Squeeze::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  Strides strides;
-  for (int i = 0, j = 0; i < in.ndim(); ++i) {
-    if (j < axes_.size() && i == axes_[j]) {
-      j++;
-    } else {
-      strides.push_back(in.strides(i));
-    }
+std::tuple<int64_t, std::vector<int64_t>> SliceUpdate::prepare_slice(
+    const array& in) {
+  int64_t data_offset = 0;
+  std::vector<int64_t> inp_strides(in.ndim(), 0);
+  for (int i = 0; i < in.ndim(); ++i) {
+    data_offset += start_indices_[i] * in.strides()[i];
+    inp_strides[i] = in.strides()[i] * strides_[i];
   }
-  out.copy_shared_buffer(in, strides, in.flags(), in.data_size());
+
+  return std::make_tuple(data_offset, inp_strides);
 }
 
 void StopGradient::eval(const std::vector<array>& inputs, array& out) {
@@ -270,7 +268,7 @@ void StopGradient::eval(const std::vector<array>& inputs, array& out) {
 
 void Transpose::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  Strides out_strides(out.ndim());
+  std::vector<size_t> out_strides(out.ndim());
   auto& in = inputs[0];
   for (int ax = 0; ax < axes_.size(); ++ax) {
     out_strides[ax] = in.strides()[axes_[ax]];
@@ -287,8 +285,8 @@ void Transpose::eval(const std::vector<array>& inputs, array& out) {
   //   true, they stay true)
   auto flags = in.flags();
   if (flags.contiguous && in.data_size() == in.size()) {
-    int64_t f_stride = 1;
-    int64_t b_stride = 1;
+    size_t f_stride = 1;
+    size_t b_stride = 1;
     flags.col_contiguous = true;
     flags.row_contiguous = true;
     for (int i = 0, ri = out.ndim() - 1; i < out.ndim(); ++i, --ri) {

mlx/backend/common/compiled.cpp

@@ -1,7 +1,8 @@
 // Copyright © 2023-2024 Apple Inc.
 
 #include "mlx/backend/common/compiled.h"
-#include "mlx/backend/common/utils.h"
+#include "mlx/graph_utils.h"
+#include "mlx/primitives.h"
 #include "mlx/utils.h"
 
 namespace mlx::core {
@@ -14,8 +15,6 @@ 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:
@@ -52,8 +51,6 @@ 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_:
@@ -82,9 +79,58 @@ 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) {
+    const std::vector<int>& shape) {
   bool contiguous = true;
   bool all_contig = true;
   bool all_row_contig = true;
@@ -113,11 +159,13 @@ bool compiled_check_contiguity(
 void compiled_allocate_outputs(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::function<bool(size_t)>& is_constant,
-    bool contiguous) {
+    const std::vector<array>& inputs_,
+    const std::unordered_set<uintptr_t>& constant_ids_,
+    bool contiguous,
+    bool move_buffers /* = false */) {
   if (contiguous) {
     int o = 0;
-    Strides strides;
+    std::vector<size_t> strides;
     size_t data_size;
     array::Flags flags;
     for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
@@ -128,8 +176,13 @@ void compiled_allocate_outputs(
       // - Donatable
       // - Not a constant
       if (in.itemsize() == outputs[o].itemsize() && !is_scalar(in) &&
-          in.is_donatable() && is_constant(i)) {
-        outputs[o++].copy_shared_buffer(in);
+          in.is_donatable() &&
+          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+        if (move_buffers) {
+          outputs[o++].move_shared_buffer(in);
+        } else {
+          outputs[o++].copy_shared_buffer(in);
+        }
       }
       // Get representative input flags to properly set non-donated outputs
       if (strides.empty() && in.size() == outputs[0].size()) {
@@ -140,7 +193,7 @@ void compiled_allocate_outputs(
     }
     for (; o < outputs.size(); ++o) {
       outputs[o].set_data(
-          allocator::malloc(data_size * outputs[o].itemsize()),
+          allocator::malloc_or_wait(data_size * outputs[o].itemsize()),
           data_size,
           strides,
           flags);
@@ -156,86 +209,21 @@ 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() &&
-          is_constant(i)) {
-        outputs[o].copy_shared_buffer(
-            in, outputs[o].strides(), in.flags(), in.data_size());
+          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+        if (move_buffers) {
+          outputs[o].move_shared_buffer(
+              in, outputs[o].strides(), in.flags(), in.data_size());
+        } else {
+          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(outputs[o].nbytes()));
+      outputs[o].set_data(allocator::malloc_or_wait(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

@@ -1,8 +1,9 @@
 // Copyright © 2023-2024 Apple Inc.
 #pragma once
 
-#include <functional>
 #include <iomanip>
+#include <sstream>
+#include <unordered_set>
 
 #include "mlx/array.h"
 #include "mlx/primitives.h"
@@ -10,20 +11,24 @@
 namespace mlx::core {
 
 inline bool is_static_cast(const Primitive& p) {
-  return (typeid(p) == typeid(Broadcast) || typeid(p) == typeid(AsType));
+  return (
+      typeid(p) == typeid(Broadcast) || typeid(p) == typeid(Copy) ||
+      typeid(p) == typeid(StopGradient) || 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();
-  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);
+  os << std::setprecision(std::numeric_limits<float>::digits10 + 1)
+     << x.item<T>() << std::setprecision(old_precision);
 }
 
 template <typename T>
@@ -51,25 +56,15 @@ inline bool is_scalar(const array& x) {
 // Check if we can use a contiguous operation given inputs and the output shape
 bool compiled_check_contiguity(
     const std::vector<array>& inputs,
-    const Shape& shape);
+    const std::vector<int>& shape);
 
 // Allocate space for the outputs possibly with input donation
 void compiled_allocate_outputs(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    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);
+    const std::vector<array>& inputs_,
+    const std::unordered_set<uintptr_t>& constant_ids_,
+    bool contiguous,
+    bool move_buffers = false);
 
 } // namespace mlx::core

mlx/backend/common/compiled_cpu.cpp

@@ -7,12 +7,8 @@
 #include <mutex>
 #include <shared_mutex>
 
-#include <fmt/format.h>
-
 #include "mlx/backend/common/compiled.h"
-#include "mlx/backend/cpu/compiled_preamble.h"
-#include "mlx/backend/cpu/encoder.h"
-#include "mlx/backend/cpu/jit_compiler.h"
+#include "mlx/backend/common/compiled_preamble.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
 
@@ -40,10 +36,7 @@ struct CompilerCache {
   std::shared_mutex mtx;
 };
 
-static CompilerCache& cache() {
-  static CompilerCache cache_;
-  return cache_;
-};
+static CompilerCache cache{};
 
 // GPU compile is always available if the GPU is available and since we are in
 // this file CPU compile is also available.
@@ -51,53 +44,48 @@ namespace detail {
 bool compile_available_for_device(const Device& device) {
   return true;
 }
-
 } // namespace detail
 
+std::string get_temp_file(const std::string& name) {
+  return std::filesystem::temp_directory_path().append(name);
+}
+
 // Return a pointer to a compiled function
 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();
   std::string kernel_file_name;
 
-  // Deal with long kernel names. Maximum length for filename on macOS is 255
-  // characters, and on Windows the maximum length for whole path is 260. Clip
-  // file name with a little extra room and append a 16 character hash.
-#ifdef _WIN32
-  constexpr int max_file_name_length = 140;
-#else
+  // Deal with long kernel names. Maximum length for files on macOS is 255
+  // characters. Clip file name with a little extra room and append a 16
+  // character hash.
   constexpr int max_file_name_length = 245;
-#endif
   if (kernel_name.size() > max_file_name_length) {
     std::ostringstream file_name;
     file_name
         << std::string_view(kernel_name).substr(0, max_file_name_length - 16);
-    auto file_id =
-        std::hash<std::string>{}(kernel_name.substr(max_file_name_length - 16));
+    auto file_id = std::hash<std::string>{}(kernel_name);
     file_name << "_" << std::hex << std::setw(16) << file_id << std::dec;
     kernel_file_name = file_name.str();
   } else {
     kernel_file_name = kernel_name;
   }
 
-  auto output_dir = std::filesystem::temp_directory_path();
-
-  std::string shared_lib_name = "lib" + kernel_file_name + ".so";
-  auto shared_lib_path = (output_dir / shared_lib_name).string();
+  std::ostringstream shared_lib_name;
+  shared_lib_name << "lib" << kernel_file_name << ".so";
+  auto shared_lib_path = get_temp_file(shared_lib_name.str());
   bool lib_exists = false;
   {
     std::ifstream f(shared_lib_path.c_str());
@@ -106,29 +94,32 @@ void* compile(
 
   if (!lib_exists) {
     // Open source file and write source code to it
-    std::string source_file_name = kernel_file_name + ".cpp";
-    auto source_file_path = (output_dir / source_file_name).string();
+    std::ostringstream source_file_name;
+    source_file_name << kernel_file_name << ".cpp";
+    auto source_file_path = get_temp_file(source_file_name.str());
 
     std::ofstream source_file(source_file_path);
     source_file << source_code;
     source_file.close();
 
-    try {
-      JitCompiler::exec(JitCompiler::build_command(
-          output_dir, source_file_name, shared_lib_name));
-    } catch (const std::exception& error) {
-      throw std::runtime_error(fmt::format(
-          "[Compile::eval_cpu] Failed to compile function {0}: {1}",
-          kernel_name,
-          error.what()));
+    std::ostringstream build_command;
+    build_command << "g++ -std=c++17 -O3 -Wall -fPIC -shared '"
+                  << source_file_path << "' -o '" << shared_lib_path << "'";
+    std::string build_command_str = build_command.str();
+    auto return_code = system(build_command_str.c_str());
+    if (return_code) {
+      std::ostringstream msg;
+      msg << "[Compile::eval_cpu] Failed to compile function " << kernel_name
+          << " with error code " << return_code << "." << std::endl;
+      throw std::runtime_error(msg.str());
     }
   }
 
   // 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 "
@@ -136,7 +127,7 @@ void* compile(
         << dlerror();
     throw std::runtime_error(msg.str());
   }
-  cache().kernels.insert({kernel_name, fun});
+  cache.kernels.insert({kernel_name, fun});
   return fun;
 }
 
@@ -146,30 +137,33 @@ inline void build_kernel(
     const std::vector<array>& inputs,
     const std::vector<array>& outputs,
     const std::vector<array>& tape,
-    const std::function<bool(size_t)>& is_constant,
+    const std::unordered_set<uintptr_t>& constant_ids,
     bool contiguous,
     int ndim) {
-  NodeNamer namer;
+  // 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();
 
-#ifdef _MSC_VER
-  // Export the symbol
-  os << "__declspec(dllexport) ";
-#endif
+  // 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;
 
   // Start the kernel
   os << "void " << kernel_name << "(void** args) {" << std::endl;
 
   // Add the input arguments
   int cnt = 0;
-  for (size_t i = 0; i < inputs.size(); ++i) {
+  for (auto& x : inputs) {
+    auto& xname = namer.get_name(x);
+
     // Skip constants from the input list
-    if (is_constant(i)) {
+    if (is_constant(x)) {
       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;
@@ -203,11 +197,10 @@ inline void build_kernel(
   }
 
   // Read the inputs in tmps
-  for (size_t i = 0; i < inputs.size(); ++i) {
-    const auto& x = inputs[i];
+  for (auto& x : inputs) {
     auto& xname = namer.get_name(x);
 
-    if (is_constant(i)) {
+    if (is_constant(x)) {
       os << "  " << get_type_string(x.dtype()) << " tmp_" << xname << " = ";
       print_constant(os, x);
       os << ";" << std::endl;
@@ -231,7 +224,7 @@ inline void build_kernel(
       os << "static_cast<" << get_type_string(x.dtype()) << ">(tmp_"
          << namer.get_name(x.inputs()[0]) << ");" << std::endl;
     } else {
-      os << x.primitive().name();
+      x.primitive().print(os);
       os << "()(";
       for (int i = 0; i < x.inputs().size() - 1; i++) {
         os << "tmp_" << namer.get_name(x.inputs()[i]) << ", ";
@@ -257,9 +250,8 @@ inline void build_kernel(
   } else {
     for (int d = ndim - 1; d >= 0; --d) {
       // Update pointers
-      for (size_t i = 0; i < inputs.size(); ++i) {
-        const auto& x = inputs[i];
-        if (is_constant(i) || is_scalar(x)) {
+      for (auto& x : inputs) {
+        if (is_constant(x) || is_scalar(x)) {
           continue;
         }
         auto& xname = namer.get_name(x);
@@ -281,37 +273,63 @@ inline void build_kernel(
 void Compiled::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
-  auto& encoder = cpu::get_command_encoder(stream());
+  if (kernel_lib_.empty()) {
+    kernel_lib_ = build_lib_name(inputs_, outputs_, tape_, constant_ids_);
+  }
 
-  // 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_);
+  // Figure out which kernel we are using
+  auto& shape = outputs[0].shape();
+  bool contiguous = compiled_check_contiguity(inputs, shape);
 
-  // Collect function input arguments.
+  // Handle all broadcasting and collect function input arguments
   std::vector<void*> args;
-  int strides_index = 1;
-  for (size_t i = 0; i < inputs.size(); ++i) {
-    if (is_constant_(i)) {
+  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()) {
       continue;
     }
-    const auto& x = inputs[i];
-    encoder.set_input_array(x);
+    auto& x = inputs[i];
     args.push_back((void*)x.data<void>());
-    if (!contiguous && !is_scalar(x)) {
-      args.push_back(strides[strides_index++].data());
+
+    if (contiguous || is_scalar(x)) {
+      continue;
     }
+
+    // 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(ndim);
+    kernel_name += std::to_string(shape.size());
   }
 
   // Get the function
-  auto fn_ptr = compile(kernel_name, [&, contiguous = contiguous]() {
+  auto fn_ptr = compile(kernel_name, [&]() {
     std::ostringstream kernel;
     kernel << get_kernel_preamble() << std::endl;
     kernel << "extern \"C\"  {" << std::endl;
@@ -321,7 +339,7 @@ void Compiled::eval_cpu(
         inputs_,
         outputs_,
         tape_,
-        is_constant_,
+        constant_ids_,
         contiguous,
         ndim);
     // Close extern "C"
@@ -329,22 +347,19 @@ void Compiled::eval_cpu(
     return kernel.str();
   });
 
-  compiled_allocate_outputs(inputs, outputs, is_constant_, contiguous);
+  compiled_allocate_outputs(
+      inputs, outputs, inputs_, constant_ids_, contiguous, false);
 
   for (auto& x : outputs) {
     args.push_back(x.data<void>());
-    encoder.set_output_array(x);
   }
   if (!contiguous) {
-    args.push_back((void*)shape.data());
+    args.push_back((void*)outputs[0].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),
-                    shape = std::move(shape)]() mutable { fun(args.data()); });
+  fun(args.data());
 }
 
 } // namespace mlx::core

mlx/backend/common/compiled_nocpu.cpp

@@ -1,7 +1,6 @@
 // Copyright © 2023-2024 Apple Inc.
 
-#include "mlx/compile_impl.h"
-#include "mlx/primitives.h"
+#include "mlx/backend/common/compiled.h"
 
 namespace mlx::core {
 
@@ -18,7 +17,7 @@ void Compiled::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   throw std::runtime_error(
-      "[Compiled::eval_cpu] CPU compilation not supported on the platform.");
+      "[Compiled::eval_cpu] CPU compialtion not supported on the platform.");
 }
 
 } // namespace mlx::core