Fix convs by reverting #1803 (#1882)

* Build in padding to Winograd kernels
* Add new fused Winograd kernel
* Enable weight flipping in Winograd kernels

.circleci/config.yml

@@ -13,8 +13,62 @@ parameters:
   test_release:
     type: boolean
     default: false
+  linux_release:
+    type: boolean
+    default: false
 
 jobs:
+  build_documentation:
+    parameters:
+      upload-docs:
+        type: boolean
+        default: false
+    macos:
+      xcode: "15.2.0"
+    resource_class: macos.m1.medium.gen1
+    steps:
+      - checkout
+      - run:
+          name: Install
+          command: |
+            brew install python@3.9
+            brew install doxygen
+            python3.9 -m venv env
+            source env/bin/activate
+            pip install --upgrade pip
+            pip install --upgrade cmake
+            pip install -r docs/requirements.txt
+            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install . -v
+      - when:
+          condition:
+            not: << parameters.upload-docs >>
+          steps:
+            - run:
+               name: Build documentation
+               command: |
+                 source env/bin/activate
+                 cd docs && doxygen && make html O=-W
+      - when:
+          condition: << parameters.upload-docs >>
+          steps:
+            - add_ssh_keys:
+                fingerprints:
+                  - "SHA256:OhcVVMovbT0pkgMeiVRyxMnjV9R2t+hKBsNcuxq9h+0"
+            - run:
+               name: Upload documentation
+               command: |
+                 source env/bin/activate
+                 git config user.email "mlx@group.apple.com"
+                 git config user.name "CircleCI Docs"
+                 git checkout gh-pages
+                 git rebase main
+                 cd docs
+                 git rm -rf build/html
+                 doxygen && make html O=-W
+                 git add -f build/html
+                 git commit -m "rebase"
+                 git push -f origin gh-pages
+
   linux_build_and_test:
     docker:
       - image: cimg/python:3.9
@@ -31,33 +85,35 @@ jobs:
           name: Install dependencies
           command: |
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
+            pip install nanobind==2.4.0
             pip install numpy
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" CMAKE_BUILD_PARALLEL_LEVEL="" python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" CMAKE_BUILD_PARALLEL_LEVEL="" python3 setup.py develop
+            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: |
             echo "stubs"
+            pip install typing_extensions
             python setup.py generate_stubs 
       - run:
           name: Run Python tests
           command: |
             python3 -m unittest discover python/tests -v
-      # TODO: Reenable when extension api becomes stable
-      # - run:
-      #     name: Build example extension
-      #     command: |
-      #       cd examples/extensions && python3 -m pip install . 
       - run:
           name: Build CPP only
           command: |
-            mkdir -p build && cd build && cmake .. -DMLX_BUILD_METAL=OFF && make -j
+            mkdir -p build && cd build 
+            cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
+            make -j `nproc`
       - run:
           name: Run CPP tests
           command: ./build/tests/tests
@@ -69,18 +125,19 @@ jobs:
         default: "15.2.0"
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
-            brew install python@3.8
-            python3.8 -m venv env
+            brew install python@3.9
+            brew install openmpi
+            python3.9 -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
+            pip install nanobind==2.4.0
             pip install numpy
             pip install torch
             pip install tensorflow
@@ -89,11 +146,12 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL="" pip install -e . -v
+            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
+            pip install typing_extensions
             python setup.py generate_stubs 
       - run:
           name: Run Python tests
@@ -101,23 +159,48 @@ jobs:
             source env/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
-      # TODO: Reenable when extension api becomes stable
-      # - run:
-      #     name: Build example extension
-      #     command: |
-      #       cd examples/extensions && python3.11 -m pip install . 
+            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
+      - run:
+          name: Build example extension
+          command: |
+            source env/bin/activate
+            cd examples/extensions
+            pip install -r requirements.txt
+            python setup.py build_ext -j8
       - store_test_results:
           path: test-results
       - run:
           name: Build CPP only
           command: |
             source env/bin/activate
-            mkdir -p build && cd build && cmake .. && make -j
+            mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
       - run:
           name: Run CPP tests
           command: |
             DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 ./build/tests/tests
-            DEVICE=cpu ./build/tests/tests
+      - run:
+          name: Build small binary
+          command: |
+            source env/bin/activate
+            cd build/
+            cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
+              -DBUILD_SHARED_LIBS=ON \
+              -DMLX_BUILD_CPU=OFF \
+              -DMLX_BUILD_SAFETENSORS=OFF \
+              -DMLX_BUILD_GGUF=OFF \
+              -DMLX_METAL_JIT=ON
+            make -j `sysctl -n hw.ncpu`
+      - run:
+          name: Run Python tests with JIT
+          command: |
+            source env/bin/activate
+            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
 
   build_release:
     parameters:
@@ -132,18 +215,19 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
             brew install python@<< parameters.python_version >>
+            brew install openmpi
             python<< parameters.python_version >> -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
+            pip install nanobind==2.4.0
             pip install --upgrade setuptools
             pip install numpy
             pip install twine
@@ -153,19 +237,20 @@ jobs:
           command: |
             source env/bin/activate
             DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
+            pip install typing_extensions
             python setup.py generate_stubs 
       - run:
           name: Build Python package
           command: |
             source env/bin/activate
             << parameters.build_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               python -m build -w
       - when:
           condition: << parameters.build_env >>
@@ -178,7 +263,7 @@ jobs:
       - store_artifacts:
           path: dist/
 
-  build_linux_test_release:
+  build_linux_release:
     parameters:
       python_version:
         type: string
@@ -207,21 +292,28 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
+            pip install nanobind==2.4.0
             pip install --upgrade setuptools
             pip install numpy
             pip install auditwheel
             pip install patchelf
             pip install build
+            pip install twine
             << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               pip install . -v
+            pip install typing_extensions
             python setup.py generate_stubs 
             << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               python -m build --wheel
             auditwheel show dist/*
             auditwheel repair dist/* --plat manylinux_2_31_x86_64
+      - run:
+          name: Upload package
+          command: |
+            source env/bin/activate
+            twine upload wheelhouse/*
       - store_artifacts:
           path: wheelhouse/
 
@@ -239,8 +331,9 @@ workflows:
       - mac_build_and_test:
           matrix:
             parameters:
-              xcode_version: ["15.0.0", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test
+      - build_documentation 
 
   build_pypi_release:
     when:
@@ -257,9 +350,17 @@ workflows:
               ignore: /.*/
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               xcode_version: ["15.0.0", "15.2.0"]
               build_env: ["PYPI_RELEASE=1"]
+      - build_documentation:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          upload-docs: true
+
   prb:
     when:
       matches:
@@ -274,7 +375,7 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              xcode_version: ["15.0.0", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test:
           requires: [ hold ]
   nightly_build:
@@ -286,7 +387,7 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               xcode_version: ["15.0.0", "15.2.0"]
   weekly_build:
     when:
@@ -297,17 +398,17 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["15.0.0", "15.2.0"]
+              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.test_release >>
+        - << pipeline.parameters.linux_release >>
     jobs:
-      - build_linux_test_release:
+      - build_linux_release:
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               extra_env: ["PYPI_RELEASE=1"]

.github/workflows/pull_request.yml

@@ -17,4 +17,4 @@ jobs:
           pip install pre-commit black isort clang-format
       - name: Run lint
         run: |
-          pre-commit run --all-files
+          pre-commit run --all-files

.gitignore

@@ -76,6 +76,9 @@ build/
 *.out
 *.app
 
+# Debug symbols
+*.pdb
+
 # VSCode 
 .vscode/
 .DS_Store

.pre-commit-config.yaml

@@ -1,16 +1,21 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v18.1.4
+    rev: v19.1.7
     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: 24.4.2
+    rev: 25.1.0
     hooks:
     -   id: black
+    
 -   repo: https://github.com/pycqa/isort
-    rev: 5.13.2
+    rev: 6.0.0
     hooks:
     -   id: isort
         args:
             - --profile=black
+- repo: https://github.com/cheshirekow/cmake-format-precommit
+  rev: v0.6.13
+  hooks:
+    - id: cmake-format

ACKNOWLEDGMENTS.md

@@ -7,15 +7,18 @@ 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`.
+- 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.
 - 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` and safetensor support.
+- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream`, safetensors support, `einsum`, and `einsum_path`.
 - Gabrijel Boduljak: Added `mlx.core.linalg`, implemented `norm` method and `InstanceNorm` layer. Implemented pooling layers and ``Upsample``.
 - Hinrik Snær Guðmundsson: Added `atleast_1d`, `atleast_2d`, `atleast_3d` ops.
 - Luca Arnaboldi: Added `Ceil` and `Floor` ops; implemented pickling, copy and deepcopy for mlx arrays.
 - Brian Keene & Atila Orhon, with Argmax Inc.: Added `fast.scaled_dot_product_attention`
 - AmirHossein Razlighi: Added chaining support for some of the ops in `nn.Module`. Comparison works for non array objects in `mlx.core.array`. Exception handling for invalid operations in `mlx.core.array`.
+- Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
+- Paul Paczuski: Improved stability of BCE loss calculation
+- Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />

CITATION.cff

@@ -0,0 +1,24 @@
+cff-version: 1.2.0
+title: mlx
+message: >-
+  If you use this software, please cite it using the
+  metadata from this file.
+type: software
+authors:
+  - given-names: Awni
+    family-names: Hannun
+    affiliation: Apple
+  - given-names: Jagrit
+    family-names: Digani
+    affiliation: Apple
+  - given-names: Angelos
+    family-names: Katharopoulos
+    affiliation: Apple
+  - given-names: Ronan
+    family-names: Collobert
+    affiliation: Apple
+repository-code: 'https://github.com/ml-explore'
+abstract: >-
+  MLX: efficient and flexible machine learning on Apple
+  silicon
+license: MIT

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.24)
+cmake_minimum_required(VERSION 3.25)
 
 project(mlx LANGUAGES C CXX)
 
@@ -15,37 +15,44 @@ option(MLX_BUILD_EXAMPLES "Build examples for mlx" ON)
 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_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 0.12.2)
+  set(MLX_VERSION 0.23.1)
 endif()
+add_compile_definitions("MLX_VERSION=${MLX_VERSION}")
 
 # --------------------- Processor tests -------------------------
+message(
+  STATUS
+    "Building MLX for ${CMAKE_SYSTEM_PROCESSOR} processor on ${CMAKE_SYSTEM_NAME}"
+)
 
-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(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
   if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
     if(NOT MLX_ENABLE_X64_MAC)
-      message(FATAL_ERROR
-        "Building for x86_64 on macOS is not supported."
-        " If you are on an Apple silicon system, check the build"
-        " documentation for possible fixes: "
-        "https://ml-explore.github.io/mlx/build/html/install.html#build-from-source")
+      message(
+        FATAL_ERROR
+          "Building for x86_64 on macOS is not supported."
+          " If you are on an Apple silicon system, check the build"
+          " documentation for possible fixes: "
+          "https://ml-explore.github.io/mlx/build/html/install.html#build-from-source"
+      )
     else()
+      set(MLX_BUILD_METAL OFF)
       message(WARNING "Building for x86_64 arch is not officially supported.")
     endif()
-    set(MLX_BUILD_METAL OFF)
-  elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm64")
-    set(MLX_BUILD_ARM ON)
   endif()
 
 else()
+  set(MLX_BUILD_METAL OFF)
   message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
 endif()
 
@@ -56,171 +63,232 @@ include(FetchContent)
 cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
+set_target_properties(mlx PROPERTIES COMPILE_WARNING_AS_ERROR ON)
 
-if (MLX_BUILD_METAL)
-  find_library(METAL_LIB Metal)
-  find_library(FOUNDATION_LIB Foundation)
-  find_library(QUARTZ_LIB QuartzCore)
+if(MLX_BUILD_METAL)
+  set(METAL_LIB "-framework Metal")
+  set(FOUNDATION_LIB "-framework Foundation")
+  set(QUARTZ_LIB "-framework QuartzCore")
 endif()
 
-if (MLX_BUILD_METAL AND NOT METAL_LIB)
+if(MLX_BUILD_METAL AND NOT METAL_LIB)
   message(STATUS "Metal not found. Unable to build GPU")
   set(MLX_BUILD_METAL OFF)
   set(MLX_METAL_DEBUG OFF)
-elseif (MLX_BUILD_METAL)
+elseif(MLX_BUILD_METAL)
   message(STATUS "Building METAL sources")
 
-  if (MLX_METAL_DEBUG)
+  if(MLX_METAL_DEBUG)
     add_compile_definitions(MLX_METAL_DEBUG)
   endif()
 
   # Throw an error if xcrun not found
-  execute_process(COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
-                  OUTPUT_VARIABLE MACOS_VERSION
-                  COMMAND_ERROR_IS_FATAL ANY)
+  execute_process(
+    COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
+    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
 
-  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
-
-  if (${MACOS_VERSION} GREATER_EQUAL 14.2)
-    set(METAL_CPP_PATCH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/metal.14.2.diff)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14.2_iOS17.2.zip)
-  elseif (${MACOS_VERSION} GREATER_EQUAL 14.0)
-    set(METAL_CPP_PATCH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/metal.14.0.diff)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14_iOS17-beta.zip)
-  else()
-    message(FATAL_ERROR "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON" )
+  if(${MACOS_SDK_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}")
 
-  FetchContent_Declare(
-    metal_cpp
-    URL ${METAL_CPP_URL}
-    PATCH_COMMAND /usr/bin/patch -N -i ${METAL_CPP_PATCH} || true
-  )
+  set(METAL_CPP_URL
+      https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18.zip)
+
+  if(NOT CMAKE_OSX_DEPLOYMENT_TARGET STREQUAL "")
+    set(XCRUN_FLAGS "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
+  endif()
+  execute_process(
+    COMMAND
+      zsh "-c"
+      "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal ${XCRUN_FLAGS} -E -x metal -P - | tail -1 | tr -d '\n'"
+    OUTPUT_VARIABLE MLX_METAL_VERSION COMMAND_ERROR_IS_FATAL ANY)
+  FetchContent_Declare(metal_cpp URL ${METAL_CPP_URL})
 
   FetchContent_MakeAvailable(metal_cpp)
   target_include_directories(
-    mlx PUBLIC
-    $<BUILD_INTERFACE:${metal_cpp_SOURCE_DIR}>
-    $<INSTALL_INTERFACE:include/metal_cpp>
-  )
-  target_link_libraries(
-    mlx
-    ${METAL_LIB}
-    ${FOUNDATION_LIB}
-    ${QUARTZ_LIB})
+    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()
 
-find_library(ACCELERATE_LIBRARY Accelerate)
-if (MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
-  message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
-  set(MLX_BUILD_ACCELERATE ON)
-  target_link_libraries(mlx ${ACCELERATE_LIBRARY})
-  add_compile_definitions(ACCELERATE_NEW_LAPACK)
+if(WIN32)
+  if(MSVC)
+    # GGUF does not build with MSVC.
+    set(MLX_BUILD_GGUF OFF)
+    # There is no prebuilt OpenBLAS distribution for MSVC.
+    set(MLX_BUILD_BLAS_FROM_SOURCE ON)
+  endif()
+  # Windows implementation of dlfcn.h APIs.
+  FetchContent_Declare(
+    dlfcn-win32
+    GIT_REPOSITORY https://github.com/dlfcn-win32/dlfcn-win32.git
+    GIT_TAG v1.4.1
+    EXCLUDE_FROM_ALL)
+  block()
+  set(BUILD_SHARED_LIBS OFF)
+  FetchContent_MakeAvailable(dlfcn-win32)
+  endblock()
+  target_include_directories(mlx PRIVATE "${dlfcn-win32_SOURCE_DIR}/src")
+  target_link_libraries(mlx PRIVATE dl)
+endif()
+
+if(MLX_BUILD_CPU)
+  find_library(ACCELERATE_LIBRARY Accelerate)
+  if(ACCELERATE_LIBRARY)
+    message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
+    set(MLX_BUILD_ACCELERATE ON)
+  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(${CMAKE_HOST_APPLE})
+      # The blas shipped in macOS SDK is not supported, search homebrew for
+      # openblas instead.
+      set(BLA_VENDOR OpenBLAS)
+      set(LAPACK_ROOT
+          "${LAPACK_ROOT};$ENV{LAPACK_ROOT};/usr/local/opt/openblas")
+    endif()
+    # Search and link with lapack.
+    find_package(LAPACK REQUIRED)
+    if(NOT LAPACK_FOUND)
+      message(FATAL_ERROR "Must have LAPACK installed")
+    endif()
+    find_path(LAPACK_INCLUDE_DIRS lapacke.h /usr/include /usr/local/include
+              /usr/local/opt/openblas/include)
+    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})
+    # 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)
+    if(NOT BLAS_FOUND)
+      message(FATAL_ERROR "Must have BLAS installed")
+    endif()
+    # TODO find a cleaner way to do this
+    find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include /usr/local/include
+              $ENV{BLAS_HOME}/include)
+    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 ${BLAS_LIBRARIES})
+  endif()
 else()
-  message(STATUS "Accelerate or arm neon not found, using default backend.")
   set(MLX_BUILD_ACCELERATE OFF)
-  if(${CMAKE_HOST_APPLE})
-    # The blas shipped in macOS SDK is not supported, search homebrew for
-    # openblas instead.
-    set(BLA_VENDOR OpenBLAS)
-    set(LAPACK_ROOT "${LAPACK_ROOT};$ENV{LAPACK_ROOT};/usr/local/opt/openblas")
+endif()
+
+find_package(MPI)
+if(MPI_FOUND)
+  execute_process(
+    COMMAND zsh "-c" "mpirun --version"
+    OUTPUT_VARIABLE MPI_VERSION
+    ERROR_QUIET)
+  if(${MPI_VERSION} MATCHES ".*Open MPI.*")
+    target_include_directories(mlx PRIVATE ${MPI_INCLUDE_PATH})
+  elseif(MPI_VERSION STREQUAL "")
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING "MPI found but mpirun is not available. Building without MPI.")
+  else()
+    set(MPI_FOUND FALSE)
+    message(WARNING "MPI which is not OpenMPI found. Building without MPI.")
   endif()
-  # Search and link with lapack.
-  find_package(LAPACK REQUIRED)
-  if (NOT LAPACK_FOUND)
-    message(FATAL_ERROR "Must have LAPACK installed")
-  endif()
-  find_path(LAPACK_INCLUDE_DIRS lapacke.h
-    /usr/include
-    /usr/local/include
-    /usr/local/opt/openblas/include)
-  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 ${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)
-  if (NOT BLAS_FOUND)
-    message(FATAL_ERROR "Must have BLAS installed")
-  endif()
-  # TODO find a cleaner way to do this
-  find_path(BLAS_INCLUDE_DIRS cblas.h
-    /usr/include
-    /usr/local/include
-    $ENV{BLAS_HOME}/include)
-  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 ${BLAS_LIBRARIES})
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
 
 target_include_directories(
-  mlx
-  PUBLIC
-  $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
-  $<INSTALL_INTERFACE:include>
-)
+  mlx PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
+             $<INSTALL_INTERFACE:include>)
 
-if (MLX_BUILD_PYTHON_BINDINGS)
+FetchContent_Declare(
+  fmt
+  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+  GIT_TAG 10.2.1
+  EXCLUDE_FROM_ALL)
+FetchContent_MakeAvailable(fmt)
+target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:fmt::fmt-header-only>)
+
+if(MLX_BUILD_PYTHON_BINDINGS)
   message(STATUS "Building Python bindings.")
-  find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)
+  find_package(
+    Python 3.8
+    COMPONENTS Interpreter Development.Module
+    REQUIRED)
   execute_process(
     COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
-    OUTPUT_STRIP_TRAILING_WHITESPACE OUTPUT_VARIABLE NB_DIR)
-  list(APPEND CMAKE_PREFIX_PATH "${NB_DIR}")
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+    OUTPUT_VARIABLE nanobind_ROOT)
   find_package(nanobind CONFIG REQUIRED)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/python/src)
 endif()
 
-if (MLX_BUILD_TESTS)
+if(MLX_BUILD_TESTS)
   include(CTest)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/tests)
 endif()
 
-if (MLX_BUILD_EXAMPLES)
+if(MLX_BUILD_EXAMPLES)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/examples/cpp)
 endif()
 
-if (MLX_BUILD_BENCHMARKS)
+if(MLX_BUILD_BENCHMARKS)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/benchmarks/cpp)
 endif()
 
-
-
 # ----------------------------- Installation -----------------------------
 include(GNUInstallDirs)
 
 # Install library
 install(
-    TARGETS mlx
-    EXPORT MLXTargets
-    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
-    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
-    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
-    INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
-)
-
+  TARGETS mlx
+  EXPORT MLXTargets
+  LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+  ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+  RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+  INCLUDES
+  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 
 # Install headers
 install(
-    DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/mlx
-    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
-    COMPONENT headers
-    FILES_MATCHING PATTERN "*.h"
-)
+  DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/mlx
+  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
+  COMPONENT headers
+  FILES_MATCHING
+  PATTERN "*.h"
+  PATTERN "backend/metal/kernels.h" EXCLUDE)
 
 # Install metal dependencies
-if (MLX_BUILD_METAL)
+if(MLX_BUILD_METAL)
 
   # Install metal cpp
   install(
-      DIRECTORY ${metal_cpp_SOURCE_DIR}/
-      DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/metal_cpp
-      COMPONENT metal_cpp_source
-  )
+    DIRECTORY ${metal_cpp_SOURCE_DIR}/
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/metal_cpp
+    COMPONENT metal_cpp_source)
 
 endif()
 
@@ -232,31 +300,24 @@ set(MLX_CMAKE_INSTALL_MODULE_DIR share/cmake/MLX)
 install(
   EXPORT MLXTargets
   FILE MLXTargets.cmake
-  DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR}
-)
+  DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR})
 
 include(CMakePackageConfigHelpers)
 
 write_basic_package_version_file(
   ${MLX_CMAKE_BUILD_VERSION_CONFIG}
   COMPATIBILITY SameMajorVersion
-  VERSION ${MLX_VERSION}
-)
+  VERSION ${MLX_VERSION})
 
 configure_package_config_file(
-  ${CMAKE_CURRENT_LIST_DIR}/mlx.pc.in
-  ${MLX_CMAKE_BUILD_CONFIG}
+  ${CMAKE_CURRENT_LIST_DIR}/mlx.pc.in ${MLX_CMAKE_BUILD_CONFIG}
   INSTALL_DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR}
   NO_CHECK_REQUIRED_COMPONENTS_MACRO
-  PATH_VARS CMAKE_INSTALL_LIBDIR CMAKE_INSTALL_INCLUDEDIR MLX_CMAKE_INSTALL_MODULE_DIR
-)
+  PATH_VARS CMAKE_INSTALL_LIBDIR CMAKE_INSTALL_INCLUDEDIR
+            MLX_CMAKE_INSTALL_MODULE_DIR)
 
-install(
-  FILES ${MLX_CMAKE_BUILD_CONFIG} ${MLX_CMAKE_BUILD_VERSION_CONFIG}
-  DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR}
-)
+install(FILES ${MLX_CMAKE_BUILD_CONFIG} ${MLX_CMAKE_BUILD_VERSION_CONFIG}
+        DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR})
 
-install(
-  DIRECTORY ${CMAKE_MODULE_PATH}/
-  DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR}
-)
+install(DIRECTORY ${CMAKE_MODULE_PATH}/
+        DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR})

README.md

@@ -6,7 +6,7 @@
 
 [![CircleCI](https://circleci.com/gh/ml-explore/mlx.svg?style=svg)](https://circleci.com/gh/ml-explore/mlx)
 
-MLX is an array framework for machine learning research on Apple silicon,
+MLX is an array framework for machine learning on Apple silicon,
 brought to you by Apple machine learning research.
 
 Some key features of MLX include:
@@ -88,13 +88,13 @@ for more information on building the C++ and Python APIs from source.
 
 ## Contributing 
 
-Check out the [contribution guidelines](CONTRIBUTING.md) for more information
+Check out the [contribution guidelines](https://github.com/ml-explore/mlx/tree/main/CONTRIBUTING.md) for more information
 on contributing to MLX. See the
 [docs](https://ml-explore.github.io/mlx/build/html/install.html) for more
 information on building from source, and running tests.
 
 We are grateful for all of [our
-contributors](ACKNOWLEDGMENTS.md#Individual-Contributors). If you contribute
+contributors](https://github.com/ml-explore/mlx/tree/main/ACKNOWLEDGMENTS.md#Individual-Contributors). If you contribute
 to MLX and wish to be acknowledged, please add your name to the list in your
 pull request.
 

benchmarks/cpp/autograd.cpp

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

benchmarks/cpp/irregular_strides.cpp

@@ -6,105 +6,105 @@
 #include "mlx/mlx.h"
 #include "time_utils.h"
 
-using namespace mlx::core;
+namespace mx = mlx::core;
 
 void time_irregular_binary_ops_1D() {
-  auto device = default_device();
+  auto device = mx::default_device();
   int size = 1000000;
   int step = 2;
-  auto a = random::uniform({size});
-  auto b = random::uniform({size});
-  eval(a, b);
+  auto a = mx::random::uniform({size});
+  auto b = mx::random::uniform({size});
+  mx::eval(a, b);
   a = slice(a, {0}, {size}, {step});
   b = slice(b, {0}, {size}, {step});
-  TIMEM("1D strided", add, a, b, device);
+  TIMEM("1D strided", mx::add, a, b, device);
 }
 
 void time_irregular_binary_ops_2D() {
-  auto device = default_device();
+  auto device = mx::default_device();
   int size = 2048;
-  auto a = random::uniform({size, size});
-  auto b = random::uniform({size, size});
-  eval(a, b);
-  TIMEM("2D regular", add, a, b, device);
+  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);
 
-  b = transpose(b);
-  eval(b);
-  TIMEM("2D transpose", add, a, b, device);
+  b = mx::transpose(b);
+  mx::eval(b);
+  TIMEM("2D mx::transpose", mx::add, a, b, device);
 
-  b = random::uniform({size});
-  eval(b);
-  TIMEM("2D broadcast dim 0", add, a, b, device);
+  b = mx::random::uniform({size});
+  mx::eval(b);
+  TIMEM("2D broadcast dim 0", mx::add, a, b, device);
 
-  b = reshape(b, {size, 1});
-  eval(b);
-  TIMEM("2D broadcast dim 1", add, a, b, device);
+  b = mx::reshape(b, {size, 1});
+  mx::eval(b);
+  TIMEM("2D broadcast dim 1", mx::add, a, b, device);
 }
 
 void time_irregular_binary_ops_3D() {
-  auto device = default_device();
+  auto device = mx::default_device();
   int d0 = 32;
   int d1 = 512;
   int d2 = 512;
-  auto a = random::uniform({d0, d1, d2});
-  auto b = random::uniform({d0, d1, d2});
-  TIMEM("3D regular", add, a, b, device);
+  auto a = mx::random::uniform({d0, d1, d2});
+  auto b = mx::random::uniform({d0, d1, d2});
+  TIMEM("3D regular", mx::add, a, b, device);
 
-  b = transpose(b, {0, 2, 1});
-  TIMEM("3D transpose", add, a, b, device);
+  b = mx::transpose(b, {0, 2, 1});
+  TIMEM("3D mx::transpose", mx::add, a, b, device);
 
-  b = random::uniform({d1, d2});
-  TIMEM("3D broadcast dim 0", add, a, b, device);
+  b = mx::random::uniform({d1, d2});
+  TIMEM("3D broadcast dim 0", 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, 1, d2});
+  TIMEM("3D broadcast dim 1", mx::add, a, b, device);
 
-  b = random::uniform({d0, d1, 1});
-  TIMEM("3D broadcast dim 2", add, a, b, device);
+  b = mx::random::uniform({d0, d1, 1});
+  TIMEM("3D broadcast dim 2", mx::add, a, b, device);
 
-  b = random::uniform({d2});
-  TIMEM("3D broadcast dims 0, 1", add, a, b, device);
+  b = mx::random::uniform({d2});
+  TIMEM("3D broadcast dims 0, 1", mx::add, a, b, device);
 
-  b = random::uniform({d1, 1});
-  TIMEM("3D broadcast dims 0, 2", add, a, b, device);
+  b = mx::random::uniform({d1, 1});
+  TIMEM("3D broadcast dims 0, 2", mx::add, a, b, device);
 
-  b = random::uniform({d0, 1, 1});
-  TIMEM("3D broadcast dims 1, 2", add, a, b, device);
+  b = mx::random::uniform({d0, 1, 1});
+  TIMEM("3D broadcast dims 1, 2", mx::add, a, b, device);
 }
 
 void time_irregular_binary_ops_4D() {
-  auto device = default_device();
+  auto device = mx::default_device();
   std::vector<int> shape = {8, 8, 512, 512};
-  auto a = random::uniform(shape);
-  auto b = random::uniform(shape);
+  auto a = mx::random::uniform(shape);
+  auto b = mx::random::uniform(shape);
 
-  TIMEM("4D regular", add, a, b, device);
+  TIMEM("4D regular", mx::add, a, b, device);
 
-  b = transpose(b, {0, 1, 3, 2});
-  TIMEM("4D transpose", add, a, b, device);
+  b = mx::transpose(b, {0, 1, 3, 2});
+  TIMEM("4D mx::transpose", mx::add, a, b, device);
 
   std::string om = "4D broadcast dims ";
   for (int i = 0; i < shape.size(); ++i) {
     shape[i] = 1;
-    b = random::uniform(shape);
+    b = mx::random::uniform(shape);
     std::ostringstream msg;
     msg << om << i;
-    TIMEM(msg.str(), add, a, b, device);
+    TIMEM(msg.str(), mx::add, a, b, device);
 
     for (int j = i + 1; j < shape.size(); ++j) {
       shape[j] = 1;
       std::ostringstream msg;
       msg << om << i << ", " << j;
-      b = random::uniform(shape);
-      TIMEM(msg.str(), add, a, b, device);
+      b = mx::random::uniform(shape);
+      TIMEM(msg.str(), mx::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 = random::uniform(shape);
-        TIMEM(msg.str(), add, a, b, device);
+        b = mx::random::uniform(shape);
+        TIMEM(msg.str(), mx::add, a, b, device);
         shape[k] = a.shape(k);
       }
     }
@@ -113,83 +113,83 @@ void time_irregular_binary_ops_4D() {
 }
 
 void time_irregular_reshape() {
-  auto device = default_device();
+  auto device = mx::default_device();
   std::vector<int> shape;
-  auto reshape_fn = [&shape, device](const array& a) {
-    return reshape(a, shape, device);
+  auto reshape_fn = [&shape, device](const mx::array& a) {
+    return mx::reshape(a, shape, device);
   };
 
   int size = 64;
   int d = 2 * size;
 
-  auto a = random::uniform({d, d, d});
+  auto a = mx::random::uniform({d, d, d});
 
   shape = {8 * size, size, size};
   TIMEM("3D contiguous", reshape_fn, a);
 
-  a = transpose(a);
+  a = mx::transpose(a);
   shape = {8 * size, size, size};
-  TIMEM("3D transpose", reshape_fn, a);
+  TIMEM("3D mx::transpose", reshape_fn, a);
 
-  a = transpose(a, {1, 2, 0});
+  a = mx::transpose(a, {1, 2, 0});
   shape = {8 * size, size, size};
-  TIMEM("3D transpose dims 1 2", reshape_fn, a);
+  TIMEM("3D mx::transpose dims 1 2", reshape_fn, a);
 
-  a = broadcast_to(random::uniform({d, d}), {d, d, d});
+  a = mx::broadcast_to(mx::random::uniform({d, d}), {d, d, d});
   TIMEM("3D broadcast dim 0", reshape_fn, a);
 
-  a = broadcast_to(random::uniform({d, 1, d}), {d, d, d});
+  a = mx::broadcast_to(mx::random::uniform({d, 1, d}), {d, d, d});
   TIMEM("3D broadcast dim 1", reshape_fn, a);
 
-  a = broadcast_to(random::uniform({d, d, 1}), {d, d, d});
+  a = mx::broadcast_to(mx::random::uniform({d, d, 1}), {d, d, d});
   TIMEM("3D broadcast dim 2", reshape_fn, a);
 
-  a = broadcast_to(random::uniform({d}), {d, d, d});
+  a = mx::broadcast_to(mx::random::uniform({d}), {d, d, d});
   TIMEM("3D broadcast dims 0, 1", reshape_fn, a);
 
-  a = broadcast_to(random::uniform({d, 1}), {d, d, d});
+  a = mx::broadcast_to(mx::random::uniform({d, 1}), {d, d, d});
   TIMEM("3D broadcast dims 0, 2", reshape_fn, a);
 
-  a = broadcast_to(random::uniform({d, 1, 1}), {d, d, d});
+  a = mx::broadcast_to(mx::random::uniform({d, 1, 1}), {d, d, d});
   TIMEM("3D broadcast dims 1, 2", reshape_fn, a);
 
-  a = broadcast_to(random::uniform({1, 1, 1}), {d, d, d});
+  a = mx::broadcast_to(mx::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 = default_device();
+  auto device = mx::default_device();
   int size = 1000000;
   int step = 2;
-  auto a = random::uniform({size});
+  auto a = mx::random::uniform({size});
   a = slice(a, {0}, {size}, {step});
-  TIMEM("1D strided", astype, a, int32, device);
+  TIMEM("1D strided", mx::astype, a, mx::int32, device);
 }
 
 void time_irregular_astype_2D() {
-  auto device = default_device();
+  auto device = mx::default_device();
   int size = 2048;
   std::vector<int> shape = {size, size};
 
-  auto a = random::uniform(shape);
-  TIMEM("2D regular", astype, a, int32, device);
+  auto a = mx::random::uniform(shape);
+  TIMEM("2D regular", mx::astype, a, mx::int32, device);
 
-  a = transpose(a);
-  TIMEM("2D transpose", astype, a, int32, device);
+  a = mx::transpose(a);
+  TIMEM("2D mx::transpose", 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}), shape);
+  TIMEM("2D broadcast dim 0", mx::astype, a, mx::int32, device);
 
-  a = broadcast_to(random::uniform({size, 1}), shape);
-  TIMEM("2D broadcast dim 1", 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);
 }
 
 int main(int argc, char** argv) {
   if (argc > 1) {
     bool use_gpu = !strcmp(argv[1], "gpu");
-    set_default_device(use_gpu ? Device::gpu : Device::cpu);
+    set_default_device(use_gpu ? mx::Device::gpu : mx::Device::cpu);
   }
-  std::cout << "Benchmarks for " << default_device() << std::endl;
+  std::cout << "Benchmarks for " << mx::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"
 
-using namespace mlx::core;
+namespace mx = mlx::core;
 
 void time_creation_ops() {
   int M = 2000;
   int N = 500;
   auto shape = {M, N};
-  auto full_fp32 = [&]() { return full(shape, 3.3f); };
+  auto full_fp32 = [&]() { return mx::full(shape, 3.3f); };
   TIME(full_fp32);
-  auto zeros_fp32 = [&]() { return zeros(shape, float32); };
+  auto zeros_fp32 = [&]() { return mx::zeros(shape, mx::float32); };
   TIME(zeros_fp32);
-  auto ones_fp32 = [&]() { return ones(shape, float32); };
+  auto ones_fp32 = [&]() { return mx::ones(shape, mx::float32); };
   TIME(ones_fp32);
 
-  auto arange_fp32 = [&]() { return arange(0.0, 10.0, 1e-4); };
+  auto arange_fp32 = [&]() { return mx::arange(0.0, 10.0, 1e-4); };
   TIME(arange_fp32);
 }
 
@@ -24,194 +24,196 @@ void time_type_conversions() {
   int M = 2000;
   int N = 500;
   auto shape = {M, N};
-  auto device = default_device();
+  auto device = mx::default_device();
 
-  auto a = zeros(shape, float32);
-  eval(a);
-  TIMEM("float32 to int32", astype, a, int32, device);
-  TIMEM("float32 to uint32", astype, a, uint32, 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);
 
-  a = zeros(shape, int32);
-  eval(a);
-  TIMEM("int32 to float32", astype, a, float32, 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, 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);
+  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);
 }
 
 void time_random_generation() {
   int M = 2000;
   int N = 500;
 
-  auto uniform = [&]() { return random::uniform({M, N}, float32); };
+  auto uniform = [&]() { return mx::random::uniform({M, N}, mx::float32); };
   TIME(uniform);
-  auto normal = [&]() { return random::normal({M, N}, float32); };
+  auto normal = [&]() { return mx::random::normal({M, N}, mx::float32); };
   TIME(normal);
 }
 
 void time_unary_ops() {
   int M = 2000;
   int N = 500;
-  auto device = default_device();
+  auto device = mx::default_device();
 
-  auto a = random::normal({M, N});
-  eval(a);
+  auto a = mx::random::normal({M, N});
+  mx::eval(a);
   TIME(mlx::core::abs, a, device);
-  TIME(negative, a, device);
-  TIME(sign, a, device);
-  TIME(square, a, device);
+  TIME(mx::negative, a, device);
+  TIME(mx::sign, a, device);
+  TIME(mx::square, a, device);
   TIME(mlx::core::sqrt, a, device);
-  TIME(rsqrt, a, device);
+  TIME(mx::rsqrt, a, device);
   TIME(mlx::core::exp, a, device);
 
-  a = random::uniform({M, N});
+  a = mx::random::uniform({M, N});
   TIME(mlx::core::log, a, device);
 }
 
 void time_binary_ops() {
   int M = 1000, N = 100, K = 10;
-  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);
+  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);
 
-  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);
+  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);
 
-  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::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 = 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);
+  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);
 }
 
 void time_strided_ops() {
   int M = 50, N = 50, O = 50, P = 50;
-  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);
+  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);
 }
 
 void time_comparisons() {
   int M = 1000, N = 100, K = 10;
-  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);
+  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);
 }
 
 void time_matvec() {
   int M = 2000, N = 200;
-  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); };
+  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); };
   TIME(matvec);
 
-  auto matvec_transpose = [&]() { return matmul(transpose(a), c); };
+  auto matvec_transpose = [&]() { return mx::matmul(mx::transpose(a), c); };
   TIME(matvec_transpose);
 }
 
 void time_matmul() {
   int M = 1000, N = 1000, K = 1000;
-  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 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 transpose_matmul = [&]() { return matmul(transpose(a), b); };
+  auto transpose_matmul = [&]() { return mx::matmul(mx::transpose(a), b); };
   TIME(transpose_matmul);
 }
 
 void time_reductions() {
-  auto a = random::normal({10000, 1000});
-  eval(a);
-  auto sum_all = [&a]() { return sum(a, false); };
+  auto a = mx::random::normal({10000, 1000});
+  mx::eval(a);
+  auto sum_all = [&a]() { return mx::sum(a, false); };
   TIME(sum_all);
 
-  auto sum_along_0 = [&a]() { return sum(a, 0, false); };
+  auto sum_along_0 = [&a]() { return mx::sum(a, 0, false); };
   TIME(sum_along_0);
 
-  auto sum_along_1 = [&a]() { return sum(a, 1, false); };
+  auto sum_along_1 = [&a]() { return mx::sum(a, 1, false); };
   TIME(sum_along_1);
 
-  auto prod_all = [&a]() { return prod(a, false); };
+  auto prod_all = [&a]() { return mx::prod(a, false); };
   TIME(prod_all);
 
-  auto all_true = [&a]() { return all(a, false); };
+  auto all_true = [&a]() { return mx::all(a, false); };
   TIME(all_true);
 
-  auto all_along_0 = [&a]() { return all(a, 0, false); };
+  auto all_along_0 = [&a]() { return mx::all(a, 0, false); };
   TIME(all_along_0);
 
-  auto all_along_1 = [&a]() { return all(a, 1, false); };
+  auto all_along_1 = [&a]() { return mx::all(a, 1, false); };
   TIME(all_along_1);
 
-  auto any_true = [&a]() { return any(a, false); };
+  auto any_true = [&a]() { return mx::any(a, false); };
   TIME(any_true);
 
-  auto argmin_along_0 = [&a]() { return argmin(a, 0, false); };
+  auto argmin_along_0 = [&a]() { return mx::argmin(a, 0, false); };
   TIME(argmin_along_0);
 
-  auto argmin_along_1 = [&a]() { return argmin(a, 1, false); };
+  auto argmin_along_1 = [&a]() { return mx::argmin(a, 1, false); };
   TIME(argmin_along_1);
 }
 
 void time_gather_scatter() {
-  auto a = random::normal({1000, 768});
-  eval(a);
-  auto indices = random::randint(0, 1000, {256});
-  eval(indices);
+  auto a = mx::random::normal({1000, 768});
+  mx::eval(a);
+  auto indices = mx::random::randint(0, 1000, {256});
+  mx::eval(indices);
 
-  auto embedding_lookup = [&a, &indices]() { return take(a, indices, 0); };
+  auto embedding_lookup = [&a, &indices]() { return mx::take(a, indices, 0); };
   TIME(embedding_lookup);
 
-  indices = random::randint(0, 768 * 1000, {256 * 768});
-  eval(indices);
+  indices = mx::random::randint(0, 768 * 1000, {256 * 768});
+  mx::eval(indices);
 
-  auto single_element_lookup = [&a, &indices]() { return take(a, indices); };
+  auto single_element_lookup = [&a, &indices]() {
+    return mx::take(a, indices);
+  };
   TIME(single_element_lookup);
 
-  indices = random::randint(0, 1000, {256});
-  auto updates = random::normal({256, 1, 768});
-  eval(indices, updates);
+  indices = mx::random::randint(0, 1000, {256});
+  auto updates = mx::random::normal({256, 1, 768});
+  mx::eval(indices, updates);
 
   auto embedding_update = [&a, &indices, &updates]() {
     return scatter(a, indices, updates, 0);
@@ -223,10 +225,10 @@ void time_gather_scatter() {
   };
   TIME(embedding_add);
 
-  a = reshape(a, {-1});
-  indices = random::randint(0, 768 * 1000, {768 * 256});
-  updates = random::normal({256 * 768, 1});
-  eval(a, indices, updates);
+  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);
 
   auto single_element_update = [&a, &indices, &updates]() {
     return scatter(a, indices, updates, 0);
@@ -240,21 +242,21 @@ void time_gather_scatter() {
 }
 
 void time_divmod() {
-  auto a = random::normal({1000});
-  auto b = random::normal({1000});
-  eval({a, b});
+  auto a = mx::random::normal({1000});
+  auto b = mx::random::normal({1000});
+  mx::eval({a, b});
 
-  auto divmod_fused = [&a, &b]() { return divmod(a, b); };
+  auto divmod_fused = [&a, &b]() { return mx::divmod(a, b); };
   TIME(divmod_fused);
 
   auto divmod_separate = [&a, &b]() {
-    return std::vector<array>{floor_divide(a, b), remainder(a, b)};
+    return std::vector<mx::array>{mx::floor_divide(a, b), mx::remainder(a, b)};
   };
   TIME(divmod_separate);
 }
 
 int main() {
-  std::cout << "Benchmarks for " << default_device() << std::endl;
+  std::cout << "Benchmarks for " << mx::default_device() << std::endl;
   time_creation_ops();
   time_type_conversions();
   time_unary_ops();

benchmarks/python/comparative/bench_mlx.py

@@ -144,6 +144,13 @@ def reduction(op, axis, x):
     mx.eval(ys)
 
 
+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)
+    mx.eval(z)
+
+
 def softmax(axis, x):
     ys = []
     for i in range(100):
@@ -505,5 +512,8 @@ if __name__ == "__main__":
     elif args.benchmark == "selu":
         print(bench(selu, x))
 
+    elif args.benchmark == "sum_and_add":
+        print(bench(sum_and_add, axis, *xs))
+
     else:
         raise ValueError("Unknown benchmark")

benchmarks/python/comparative/bench_torch.py

@@ -185,7 +185,7 @@ def prelu(x: torch.Tensor) -> torch.Tensor:
 def mish(x: torch.Tensor) -> torch.Tensor:
     y = x
     for _ in range(100):
-        return torch.nn.functional.mish(y)
+        y = torch.nn.functional.mish(y)
     sync_if_needed(x)
 
 
@@ -283,6 +283,14 @@ def topk(axis, x):
     sync_if_needed(x)
 
 
+@torch.no_grad()
+def step_function(x):
+    y = x
+    for i in range(100):
+        y = torch.where(y < 0, 0, 1)
+    sync_if_needed(x)
+
+
 @torch.no_grad()
 def selu(x):
     y = x
@@ -446,5 +454,11 @@ if __name__ == "__main__":
     elif args.benchmark == "topk":
         print(bench(topk, axis, x))
 
+    elif args.benchmark == "step":
+        print(bench(step_function, x))
+
+    elif args.benchmark == "selu":
+        print(bench(selu, x))
+
     else:
-        raise ValueError("Unknown benchmark")
+        raise ValueError(f"Unknown benchmark `{args.benchmark}`.")

benchmarks/python/comparative/compare.py

@@ -16,7 +16,9 @@ def run_or_raise(*args, **kwargs):
         result = run(*args, capture_output=True, **kwargs)
         return float(result.stdout)
     except ValueError:
-        raise ValueError(f"stdout: {result.stdout}\nstderr: {result.stderr}")
+        raise ValueError(
+            f"stdout: {result.stdout.decode()}\nstderr: {result.stderr.decode()}"
+        )
 
 
 def compare(args):

benchmarks/python/compile_bench.py

@@ -9,7 +9,6 @@ from time_utils import time_fn
 
 
 def bench_gelu():
-
     def gelu(x):
         return x * (1 + mx.erf(x / math.sqrt(2))) / 2
 
@@ -51,7 +50,6 @@ def bench_gelu():
 
 
 def bench_layernorm():
-
     weight = mx.random.uniform(shape=(4096,)).astype(mx.float16)
     bias = mx.random.uniform(shape=(4096,)).astype(mx.float16)
     mx.eval(weight, bias)

benchmarks/python/conv2d_bench_cpu.py

@@ -0,0 +1,127 @@
+import argparse
+import math
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_warmup = 1
+N_iter_bench = 10
+N_iter_func = 5
+mx.set_default_device(mx.cpu)
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+
+    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)
+        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("cpu")
+    b_pt = torch.from_numpy(b_np.transpose((0, 3, 1, 2))).to("cpu")
+
+    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__":
+    parser = argparse.ArgumentParser(description="Run conv benchmarks")
+
+    dtypes = ("float32",)
+    shapes = (
+        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2), 1),
+        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        # (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 2),
+        # (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 16),
+        # (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 64),
+        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2), 1),
+    )
+
+    for dtype in dtypes:
+        print(
+            "(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  groups, diff%"
+        )
+        for N, H, W, C, kH, kW, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, H, W, C, kH, kW, O, strides, padding, groups, np_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}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
+            )
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

benchmarks/python/conv2d_train_bench_cpu.py

@@ -0,0 +1,143 @@
+import time
+
+import mlx.core as mx
+import mlx.nn
+import mlx.optimizers as opt
+import torch
+
+
+def bench_mlx(steps: int = 20) -> float:
+    mx.set_default_device(mx.cpu)
+
+    class BenchNetMLX(mlx.nn.Module):
+        # simple encoder-decoder net
+
+        def __init__(self, in_channels, hidden_channels=32):
+            super().__init__()
+
+            self.net = mlx.nn.Sequential(
+                mlx.nn.Conv2d(in_channels, hidden_channels, kernel_size=3, padding=1),
+                mlx.nn.ReLU(),
+                mlx.nn.Conv2d(
+                    hidden_channels, 2 * hidden_channels, kernel_size=3, padding=1
+                ),
+                mlx.nn.ReLU(),
+                mlx.nn.ConvTranspose2d(
+                    2 * hidden_channels, hidden_channels, kernel_size=3, padding=1
+                ),
+                mlx.nn.ReLU(),
+                mlx.nn.ConvTranspose2d(
+                    hidden_channels, in_channels, kernel_size=3, padding=1
+                ),
+            )
+
+        def __call__(self, input):
+            return self.net(input)
+
+    benchNet = BenchNetMLX(3)
+    mx.eval(benchNet.parameters())
+    optim = opt.Adam(learning_rate=1e-3)
+
+    inputs = mx.random.normal([10, 256, 256, 3])
+
+    params = benchNet.parameters()
+    optim.init(params)
+
+    state = [benchNet.state, optim.state]
+
+    def loss_fn(params, image):
+        benchNet.update(params)
+        pred_image = benchNet(image)
+        return (pred_image - image).abs().mean()
+
+    def step(params, image):
+        loss, grads = mx.value_and_grad(loss_fn)(params, image)
+        optim.update(benchNet, grads)
+        return loss
+
+    total_time = 0.0
+    print("MLX:")
+    for i in range(steps):
+        start_time = time.perf_counter()
+
+        step(benchNet.parameters(), inputs)
+        mx.eval(state)
+        end_time = time.perf_counter()
+
+        print(f"{i:3d}, time={(end_time-start_time) * 1000:7.2f} ms")
+        total_time += (end_time - start_time) * 1000
+
+    return total_time
+
+
+def bench_torch(steps: int = 20) -> float:
+    device = torch.device("cpu")
+
+    class BenchNetTorch(torch.nn.Module):
+        # simple encoder-decoder net
+
+        def __init__(self, in_channels, hidden_channels=32):
+            super().__init__()
+
+            self.net = torch.nn.Sequential(
+                torch.nn.Conv2d(in_channels, hidden_channels, kernel_size=3, padding=1),
+                torch.nn.ReLU(),
+                torch.nn.Conv2d(
+                    hidden_channels, 2 * hidden_channels, kernel_size=3, padding=1
+                ),
+                torch.nn.ReLU(),
+                torch.nn.ConvTranspose2d(
+                    2 * hidden_channels, hidden_channels, kernel_size=3, padding=1
+                ),
+                torch.nn.ReLU(),
+                torch.nn.ConvTranspose2d(
+                    hidden_channels, in_channels, kernel_size=3, padding=1
+                ),
+            )
+
+        def forward(self, input):
+            return self.net(input)
+
+    benchNet = BenchNetTorch(3).to(device)
+    optim = torch.optim.Adam(benchNet.parameters(), lr=1e-3)
+
+    inputs = torch.randn(10, 3, 256, 256, device=device)
+
+    def loss_fn(pred_image, image):
+        return (pred_image - image).abs().mean()
+
+    total_time = 0.0
+    print("PyTorch:")
+    for i in range(steps):
+        start_time = time.perf_counter()
+
+        optim.zero_grad()
+        pred_image = benchNet(inputs)
+        loss = loss_fn(pred_image, inputs)
+        loss.backward()
+        optim.step()
+
+        end_time = time.perf_counter()
+
+        print(f"{i:3d}, time={(end_time-start_time) * 1000:7.2f} ms")
+        total_time += (end_time - start_time) * 1000
+
+    return total_time
+
+
+def main():
+    steps = 20
+    time_mlx = bench_mlx(steps)
+    time_torch = bench_torch(steps)
+
+    print(f"average time of MLX:     {time_mlx/steps:9.2f} ms")
+    print(f"total time of MLX:       {time_mlx:9.2f} ms")
+    print(f"average time of PyTorch: {time_torch/steps:9.2f} ms")
+    print(f"total time of PyTorch:   {time_torch:9.2f} ms")
+
+    diff = time_torch / time_mlx - 1.0
+    print(f"torch/mlx diff: {100. * diff:+5.2f}%")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/python/conv2d_transpose_bench_cpu.py

@@ -0,0 +1,129 @@
+import argparse
+import math
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_warmup = 1
+N_iter_bench = 10
+N_iter_func = 5
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+
+    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_transpose_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    def mx_conv_transpose_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv_transpose2d(
+                a, b, stride=strides, padding=padding, groups=groups, stream=mx.cpu
+            )
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_transpose_2D
+
+
+def make_pt_conv_transpose_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_transpose_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv_transpose2d(
+                a, b, stride=strides, padding=padding, groups=groups
+            )
+            ys.append(y)
+        return ys
+
+    return pt_conv_transpose_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, (int(O / groups), kH, kW, C)).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("cpu")
+    b_pt = torch.from_numpy(b_np.transpose((3, 0, 1, 2))).to("cpu")
+
+    f_mx = make_mx_conv_transpose_2D(strides, padding, groups)
+    f_pt = make_pt_conv_transpose_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.conv_transpose2d(
+        a_mx, b_mx, stride=strides, padding=padding, groups=groups, stream=mx.cpu
+    )
+    out_pt = torch.conv_transpose2d(
+        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__":
+    parser = argparse.ArgumentParser(description="Run conv benchmarks")
+
+    dtypes = ("float32",)
+    shapes = (
+        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2), 1),
+        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2), 1),
+    )
+
+    for dtype in dtypes:
+        print(
+            "(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  groups, diff%"
+        )
+        for N, H, W, C, kH, kW, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, H, W, C, kH, kW, O, strides, padding, groups, np_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}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
+            )
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

benchmarks/python/conv3d_bench_cpu.py

@@ -0,0 +1,110 @@
+import argparse
+import math
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_warmup = 1
+N_iter_bench = 10
+N_iter_func = 5
+mx.set_default_device(mx.cpu)
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+
+    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_3D(strides=(1, 1), padding=(0, 0), groups=1):
+    def mx_conv_3D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv3d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_3D
+
+
+def make_pt_conv_3D(strides=(1, 1), padding=(0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_3D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv3d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        return ys
+
+    return pt_conv_3D
+
+
+def bench_shape(N, D, H, W, C, kD, kH, kW, O, strides, padding, groups, np_dtype):
+    scale = 1.0 / math.sqrt(kD * kH * kW * C)
+    a_np = np.random.uniform(0, 0.5, (N, D, H, W, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kD, 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, 4, 1, 2, 3))).to("cpu")
+    b_pt = torch.from_numpy(b_np.transpose((0, 4, 1, 2, 3))).to("cpu")
+
+    f_mx = make_mx_conv_3D(strides, padding, groups)
+    f_pt = make_pt_conv_3D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv3d(a_mx, b_mx, stride=strides, padding=padding, groups=groups)
+    out_pt = torch.conv3d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 3, 4, 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, D, H, W, C)}, {(O, kD, 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__":
+    parser = argparse.ArgumentParser(description="Run conv benchmarks")
+
+    dtypes = ("float32",)
+    shapes = (
+        (4, 16, 16, 16, 16, 5, 5, 5, 16, (1, 1, 1), (2, 2, 2), 1),
+        (4, 16, 16, 16, 32, 5, 5, 5, 32, (1, 1, 1), (2, 2, 2), 1),
+    )
+
+    for dtype in dtypes:
+        print(
+            "(N,   D,   H,   W,   C), (  O, kD, kH, kW,   C),   dtype,    stride,      pads,  groups, diff%"
+        )
+        for N, D, H, W, C, kD, kH, kW, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, D, H, W, C, kD, kH, kW, O, strides, padding, groups, np_dtype
+            )
+            diff = time_torch / time_mlx - 1.0
+
+            print(
+                f"({N}, {D:3d}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kD:2d}, {kH:2d}, {kW:2d}, {C:3d}), {dtype}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
+            )
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

benchmarks/python/conv3d_train_bench_cpu.py

@@ -0,0 +1,143 @@
+import time
+
+import mlx.core as mx
+import mlx.nn
+import mlx.optimizers as opt
+import torch
+
+
+def bench_mlx(steps: int = 20, shape=(10, 32, 32, 32, 3)) -> float:
+    mx.set_default_device(mx.cpu)
+
+    class BenchNetMLX(mlx.nn.Module):
+        # simple encoder-decoder net
+
+        def __init__(self, in_channels, hidden_channels=16):
+            super().__init__()
+
+            self.net = mlx.nn.Sequential(
+                mlx.nn.Conv3d(in_channels, hidden_channels, kernel_size=3, padding=1),
+                mlx.nn.ReLU(),
+                mlx.nn.Conv3d(
+                    hidden_channels, 2 * hidden_channels, kernel_size=3, padding=1
+                ),
+                mlx.nn.ReLU(),
+                mlx.nn.ConvTranspose3d(
+                    2 * hidden_channels, hidden_channels, kernel_size=3, padding=1
+                ),
+                mlx.nn.ReLU(),
+                mlx.nn.ConvTranspose3d(
+                    hidden_channels, in_channels, kernel_size=3, padding=1
+                ),
+            )
+
+        def __call__(self, input):
+            return self.net(input)
+
+    benchNet = BenchNetMLX(3)
+    mx.eval(benchNet.parameters())
+    optim = opt.Adam(learning_rate=1e-3)
+
+    inputs = mx.random.normal(shape)
+
+    params = benchNet.parameters()
+    optim.init(params)
+
+    state = [benchNet.state, optim.state]
+
+    def loss_fn(params, image):
+        benchNet.update(params)
+        pred_image = benchNet(image)
+        return (pred_image - image).abs().mean()
+
+    def step(params, image):
+        loss, grads = mx.value_and_grad(loss_fn)(params, image)
+        optim.update(benchNet, grads)
+        return loss
+
+    total_time = 0.0
+    print("MLX:")
+    for i in range(steps):
+        start_time = time.perf_counter()
+
+        step(benchNet.parameters(), inputs)
+        mx.eval(state)
+        end_time = time.perf_counter()
+
+        print(f"{i:3d}, time={(end_time-start_time) * 1000:7.2f} ms")
+        total_time += (end_time - start_time) * 1000
+
+    return total_time
+
+
+def bench_torch(steps: int = 20, shape=(10, 3, 32, 32, 32)) -> float:
+    device = torch.device("cpu")
+
+    class BenchNetTorch(torch.nn.Module):
+        # simple encoder-decoder net
+
+        def __init__(self, in_channels, hidden_channels=16):
+            super().__init__()
+
+            self.net = torch.nn.Sequential(
+                torch.nn.Conv3d(in_channels, hidden_channels, kernel_size=3, padding=1),
+                torch.nn.ReLU(),
+                torch.nn.Conv3d(
+                    hidden_channels, 2 * hidden_channels, kernel_size=3, padding=1
+                ),
+                torch.nn.ReLU(),
+                torch.nn.ConvTranspose3d(
+                    2 * hidden_channels, hidden_channels, kernel_size=3, padding=1
+                ),
+                torch.nn.ReLU(),
+                torch.nn.ConvTranspose3d(
+                    hidden_channels, in_channels, kernel_size=3, padding=1
+                ),
+            )
+
+        def forward(self, input):
+            return self.net(input)
+
+    benchNet = BenchNetTorch(3).to(device)
+    optim = torch.optim.Adam(benchNet.parameters(), lr=1e-3)
+
+    inputs = torch.randn(*shape, device=device)
+
+    def loss_fn(pred_image, image):
+        return (pred_image - image).abs().mean()
+
+    total_time = 0.0
+    print("PyTorch:")
+    for i in range(steps):
+        start_time = time.perf_counter()
+
+        optim.zero_grad()
+        pred_image = benchNet(inputs)
+        loss = loss_fn(pred_image, inputs)
+        loss.backward()
+        optim.step()
+
+        end_time = time.perf_counter()
+
+        print(f"{i:3d}, time={(end_time-start_time) * 1000:7.2f} ms")
+        total_time += (end_time - start_time) * 1000
+
+    return total_time
+
+
+def main():
+    steps = 10
+    time_mlx = bench_mlx(steps)
+    time_torch = bench_torch(steps)
+
+    print(f"average time of MLX:     {time_mlx/steps:9.2f} ms")
+    print(f"total time of MLX:       {time_mlx:9.2f} ms")
+    print(f"average time of PyTorch: {time_torch/steps:9.2f} ms")
+    print(f"total time of PyTorch:   {time_torch:9.2f} ms")
+
+    diff = time_torch / time_mlx - 1.0
+    print(f"torch/mlx diff: {100. * diff:+5.2f}%")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/python/conv3d_transpose_bench_cpu.py

@@ -0,0 +1,116 @@
+import argparse
+import math
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_warmup = 1
+N_iter_bench = 10
+N_iter_func = 5
+mx.set_default_device(mx.cpu)
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+
+    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_3D(strides=(1, 1, 1), padding=(0, 0, 0), groups=1):
+    def mx_conv_3D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv_transpose3d(
+                a, b, stride=strides, padding=padding, groups=groups
+            )
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_3D
+
+
+def make_pt_conv_3D(strides=(1, 1, 1), padding=(0, 0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_3D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv_transpose3d(
+                a, b, stride=strides, padding=padding, groups=groups
+            )
+            ys.append(y)
+        return ys
+
+    return pt_conv_3D
+
+
+def bench_shape(N, D, H, W, C, kD, kH, kW, O, strides, padding, groups, np_dtype):
+    scale = 1.0 / math.sqrt(kD * kH * kW * C)
+    a_np = np.random.uniform(0, 0.5, (N, D, H, W, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kD, 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, 4, 1, 2, 3))).to("cpu")
+    b_pt = torch.from_numpy(b_np.transpose((4, 0, 1, 2, 3))).to("cpu")
+
+    f_mx = make_mx_conv_3D(strides, padding, groups)
+    f_pt = make_pt_conv_3D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv_transpose3d(
+        a_mx, b_mx, stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.conv_transpose3d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 3, 4, 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, D, H, W, C)}, {(O, kD, 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__":
+    parser = argparse.ArgumentParser(description="Run conv benchmarks")
+
+    dtypes = ("float32",)
+    shapes = (
+        (4, 16, 16, 16, 16, 5, 5, 5, 16, (1, 1, 1), (2, 2, 2), 1),
+        (4, 16, 16, 16, 32, 5, 5, 5, 32, (1, 1, 1), (2, 2, 2), 1),
+    )
+
+    for dtype in dtypes:
+        print(
+            "(N,   D,   H,   W,   C), (  O, kD, kH, kW,   C),   dtype,    stride,      pads,  groups, diff%"
+        )
+        for N, D, H, W, C, kD, kH, kW, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, D, H, W, C, kD, kH, kW, O, strides, padding, groups, np_dtype
+            )
+            diff = time_torch / time_mlx - 1.0
+
+            print(
+                f"({N}, {D:3d}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kD:2d}, {kH:2d}, {kW:2d}, {C:3d}), {dtype}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
+            )
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

benchmarks/python/conv_bench.py

@@ -28,11 +28,11 @@ def bench(f, a, b):
     return (e - s) * 1e-9
 
 
-def make_mx_conv_2D(strides=(1, 1), padding=(0, 0)):
+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)
+            y = mx.conv2d(a, b, stride=strides, padding=padding, groups=groups)
             ys.append(y)
         mx.eval(ys)
         return ys
@@ -40,12 +40,12 @@ def make_mx_conv_2D(strides=(1, 1), padding=(0, 0)):
     return mx_conv_2D
 
 
-def make_pt_conv_2D(strides=(1, 1), padding=(0, 0)):
+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)
+            y = torch.conv2d(a, b, stride=strides, padding=padding, groups=groups)
             ys.append(y)
         torch.mps.synchronize()
         return ys
@@ -53,11 +53,12 @@ def make_pt_conv_2D(strides=(1, 1), padding=(0, 0)):
     return pt_conv_2D
 
 
-def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
-
+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, 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)
@@ -67,15 +68,15 @@ def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
 
     torch.mps.synchronize()
 
-    f_mx = make_mx_conv_2D(strides, padding)
-    f_pt = make_pt_conv_2D(strides, padding)
+    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)
+    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
+        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)
@@ -84,7 +85,7 @@ def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
 
     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}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
+            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
@@ -95,35 +96,40 @@ if __name__ == "__main__":
 
     dtypes = ("float32",)
     shapes = (
-        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2)),
-        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2)),
-        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2)),
-        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2)),
-        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2)),
-        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2)),
-        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2)),
-        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2)),
-        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2)),
-        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2)),
-        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2)),
-        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2)),
+        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2), 1),
+        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 2),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 16),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 64),
+        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2), 1),
     )
 
     for dtype in dtypes:
-        print("(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  diff%")
-        for N, H, W, C, kH, kW, O, strides, padding in shapes:
+        print(
+            "(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  groups, diff%"
+        )
+        for N, H, W, C, kH, kW, O, strides, padding, groups in shapes:
             np_dtype = getattr(np, dtype)
             time_mlx, time_torch = bench_shape(
-                N, H, W, C, kH, kW, O, strides, padding, np_dtype
+                N, H, W, C, kH, kW, O, strides, padding, groups, np_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}, {strides}, {padding}, {100. * diff:+5.2f}%"
+                f"({N}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kH:2d}, {kW:2d}, {C:3d}), {dtype}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
             )
             if time_mlx >= 2.0 * time_torch:
                 print("ATTENTION ^^^^^^^")

benchmarks/python/conv_transpose_bench.py

@@ -0,0 +1,135 @@
+import argparse
+import math
+import os
+import subprocess
+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_transpose_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    def mx_conv_transpose_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv_transpose2d(
+                a, b, stride=strides, padding=padding, groups=groups
+            )
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_transpose_2D
+
+
+def make_pt_conv_transpose_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_transpose_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv_transpose2d(
+                a, b, stride=strides, padding=padding, groups=groups
+            )
+            ys.append(y)
+        torch.mps.synchronize()
+        return ys
+
+    return pt_conv_transpose_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((3, 0, 1, 2))).to("mps")
+
+    torch.mps.synchronize()
+
+    f_mx = make_mx_conv_transpose_2D(strides, padding, groups)
+    f_pt = make_pt_conv_transpose_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.conv_transpose2d(
+        a_mx, b_mx, stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.conv_transpose2d(
+        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__":
+    parser = argparse.ArgumentParser(description="Run conv benchmarks")
+
+    dtypes = ("float32",)
+    shapes = (
+        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2), 1),
+        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2), 1),
+    )
+
+    for dtype in dtypes:
+        print(
+            "(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  groups, diff%"
+        )
+        for N, H, W, C, kH, kW, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, H, W, C, kH, kW, O, strides, padding, groups, np_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}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
+            )
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

benchmarks/python/distributed_bench.py

@@ -0,0 +1,66 @@
+# Copyright © 2024 Apple Inc.
+
+"""
+Run with:
+    mpirun -n 2 python /path/to/distributed_bench.py
+"""
+
+import time
+
+import mlx.core as mx
+
+
+def time_fn(fn, *args, **kwargs):
+    msg = kwargs.pop("msg", None)
+    world = mx.distributed.init()
+    if world.rank() == 0:
+        if msg:
+            print(f"Timing {msg} ...", end=" ")
+        else:
+            print(f"Timing {fn.__name__} ...", end=" ")
+
+    # warmup
+    for _ in range(5):
+        mx.eval(fn(*args, **kwargs))
+
+    num_iters = 100
+    tic = time.perf_counter()
+    for _ in range(num_iters):
+        x = mx.eval(fn(*args, **kwargs))
+    toc = time.perf_counter()
+
+    msec = 1e3 * (toc - tic) / num_iters
+    if world.rank() == 0:
+        print(f"{msec:.5f} msec")
+
+
+def time_all_sum():
+    shape = (4096,)
+    x = mx.random.uniform(shape=shape)
+    mx.eval(x)
+
+    def sine(x):
+        for _ in range(20):
+            x = mx.sin(x)
+        return x
+
+    time_fn(sine, x)
+
+    def all_sum_plain(x):
+        for _ in range(20):
+            x = mx.distributed.all_sum(x)
+        return x
+
+    time_fn(all_sum_plain, x)
+
+    def all_sum_with_sine(x):
+        for _ in range(20):
+            x = mx.sin(x)
+            x = mx.distributed.all_sum(x)
+        return x
+
+    time_fn(all_sum_with_sine, x)
+
+
+if __name__ == "__main__":
+    time_all_sum()

benchmarks/python/einsum_bench.py

@@ -0,0 +1,84 @@
+# Copyright © 2024 Apple Inc.
+
+import time
+
+import mlx.core as mx
+import numpy as np
+
+
+def timeit(fn, its=100, args=[]):
+    for _ in range(5):
+        fn(*args)
+    tic = time.perf_counter()
+    for _ in range(its):
+        fn(*args)
+    toc = time.perf_counter()
+    return 1e3 * (toc - tic) / its
+
+
+def time_little_einsum_path():
+    subscripts = "ik,kj->ij"
+    x = mx.ones((32, 32))
+    y = mx.ones((32, 32))
+    mx_time = timeit(mx.einsum_path, args=(subscripts, x, y))
+
+    x = np.array(x)
+    y = np.array(y)
+    np_time = timeit(np.einsum_path, args=(subscripts, x, y))
+    print("Timing little einsum path...")
+    print(f"MLX ... {mx_time:.3f} ms")
+    print(f"NumPy... {np_time:.3f} ms")
+
+
+def time_big_einsum_path():
+    chars = list("abcdefgh")
+    char_to_dim = {c: v for v, c in enumerate(chars)}
+
+    num_inputs = 10
+    inputs = []
+    subscripts = []
+    for _ in range(num_inputs):
+        subscript = np.random.choice(chars, size=5, replace=False).tolist()
+        subscripts.append("".join(subscript))
+        inputs.append(np.ones(list(char_to_dim[c] for c in subscript)))
+    subscripts = ",".join(subscripts)
+
+    np_time = timeit(np.einsum_path, args=(subscripts, *inputs))
+
+    inputs = [mx.array(x) for x in inputs]
+    mx_time = timeit(mx.einsum_path, args=(subscripts, *inputs))
+    print("Timing big einsum path...")
+    print(f"MLX ... {mx_time:.3f} ms")
+    print(f"NumPy... {np_time:.3f} ms")
+
+
+def time_attention():
+    def regular_attention(x):
+        # shape [batch, sequence, num_heads, head_dim]
+        queries, keys, values = x, x, x
+        scores = queries.transpose(0, 2, 1, 3) @ keys.transpose(0, 2, 3, 1)
+        scores = mx.softmax(scores, axis=-1)
+        output = (scores @ values.transpose(0, 2, 1, 3)).swapaxes(1, 2)
+        mx.eval(output)
+
+    def einsum_attention(x):
+        # shape [batch, sequence, num_heads, head_dim]
+        queries, keys, values = x, x, x
+        scores = mx.einsum("itjk,iujk->ijtu", queries, keys)
+        scores = mx.softmax(scores, axis=-1)
+        output = mx.einsum("ijtu,iujk->itjk", scores, values)
+        mx.eval(output)
+
+    x = mx.random.uniform(shape=(8, 512, 32, 128))
+
+    regular_time = timeit(regular_attention, args=(x,))
+    ein_time = timeit(einsum_attention, args=(x,))
+    print("Timing einsum attention...")
+    print(f"Regular ... {regular_time:.3f} ms")
+    print(f"Einsum ... {ein_time:.3f} ms")
+
+
+if __name__ == "__main__":
+    time_little_einsum_path()
+    time_big_einsum_path()
+    time_attention()

benchmarks/python/fft_bench.py

@@ -3,6 +3,8 @@
 import matplotlib
 import mlx.core as mx
 import numpy as np
+import sympy
+import torch
 from time_utils import measure_runtime
 
 matplotlib.use("Agg")
@@ -16,41 +18,100 @@ def bandwidth_gb(runtime_ms, system_size):
     return system_size * bytes_per_fft / runtime_ms * ms_per_s / bytes_per_gb
 
 
-def run_bench(system_size):
-    def fft(x):
-        out = mx.fft.fft(x)
+def run_bench(system_size, fft_sizes, backend="mlx", dim=1):
+    def fft_mlx(x):
+        if dim == 1:
+            out = mx.fft.fft(x)
+        elif dim == 2:
+            out = mx.fft.fft2(x)
         mx.eval(out)
         return out
 
-    bandwidths = []
-    for k in range(4, 12):
-        n = 2**k
-        x = mx.random.uniform(shape=(system_size // n, n)).astype(mx.float32)
-        x = x.astype(mx.complex64)
-        mx.eval(x)
-        runtime_ms = measure_runtime(fft, x=x)
-        bandwidths.append(bandwidth_gb(runtime_ms, system_size))
+    def fft_mps(x):
+        if dim == 1:
+            out = torch.fft.fft(x)
+        elif dim == 2:
+            out = torch.fft.fft2(x)
+        torch.mps.synchronize()
+        return out
 
-    return bandwidths
+    bandwidths = []
+    for n in fft_sizes:
+        batch_size = system_size // n**dim
+        shape = [batch_size] + [n for _ in range(dim)]
+        if backend == "mlx":
+            x_np = np.random.uniform(size=(system_size // n, n)).astype(np.complex64)
+            x = mx.array(x_np)
+            mx.eval(x)
+            fft = fft_mlx
+        elif backend == "mps":
+            x_np = np.random.uniform(size=(system_size // n, n)).astype(np.complex64)
+            x = torch.tensor(x_np, device="mps")
+            torch.mps.synchronize()
+            fft = fft_mps
+        else:
+            raise NotImplementedError()
+        runtime_ms = measure_runtime(fft, x=x)
+        bandwidth = bandwidth_gb(runtime_ms, np.prod(shape))
+        print(n, bandwidth)
+        bandwidths.append(bandwidth)
+
+    return np.array(bandwidths)
 
 
 def time_fft():
+    x = np.array(range(2, 512))
+    system_size = int(2**26)
 
-    with mx.stream(mx.cpu):
-        cpu_bandwidths = run_bench(system_size=int(2**22))
-
+    print("MLX GPU")
     with mx.stream(mx.gpu):
-        gpu_bandwidths = run_bench(system_size=int(2**29))
+        gpu_bandwidths = run_bench(system_size=system_size, fft_sizes=x)
 
-    # plot bandwidths
-    x = [2**k for k in range(4, 12)]
-    plt.scatter(x, gpu_bandwidths, color="green", label="GPU")
-    plt.scatter(x, cpu_bandwidths, color="red", label="CPU")
-    plt.title("MLX FFT Benchmark")
-    plt.xlabel("N")
-    plt.ylabel("Bandwidth (GB/s)")
-    plt.legend()
-    plt.savefig("fft_plot.png")
+    print("MPS GPU")
+    mps_bandwidths = run_bench(system_size=system_size, fft_sizes=x, backend="mps")
+
+    print("CPU")
+    system_size = int(2**20)
+    with mx.stream(mx.cpu):
+        cpu_bandwidths = run_bench(system_size=system_size, fft_sizes=x)
+
+    x = np.array(x)
+
+    all_indices = x - x[0]
+    radix_2to13 = (
+        np.array([i for i in x if all(p <= 13 for p in sympy.primefactors(i))]) - x[0]
+    )
+    bluesteins = (
+        np.array([i for i in x if any(p > 13 for p in sympy.primefactors(i))]) - x[0]
+    )
+
+    for indices, name in [
+        (all_indices, "All"),
+        (radix_2to13, "Radix 2-13"),
+        (bluesteins, "Bluestein's"),
+    ]:
+        # plot bandwidths
+        print(name)
+        plt.scatter(x[indices], gpu_bandwidths[indices], color="green", label="GPU")
+        plt.scatter(x[indices], mps_bandwidths[indices], color="blue", label="MPS")
+        plt.scatter(x[indices], cpu_bandwidths[indices], color="red", label="CPU")
+        plt.title(f"MLX FFT Benchmark -- {name}")
+        plt.xlabel("N")
+        plt.ylabel("Bandwidth (GB/s)")
+        plt.legend()
+        plt.savefig(f"{name}.png")
+        plt.clf()
+
+    av_gpu_bandwidth = np.mean(gpu_bandwidths)
+    av_mps_bandwidth = np.mean(mps_bandwidths)
+    av_cpu_bandwidth = np.mean(cpu_bandwidths)
+    print("Average bandwidths:")
+    print("GPU:", av_gpu_bandwidth)
+    print("MPS:", av_mps_bandwidth)
+    print("CPU:", av_cpu_bandwidth)
+
+    portion_faster = len(np.where(gpu_bandwidths > mps_bandwidths)[0]) / len(x)
+    print("Percent MLX faster than MPS: ", portion_faster * 100)
 
 
 if __name__ == "__main__":

benchmarks/python/hadamard_bench.py

@@ -0,0 +1,70 @@
+import argparse
+
+import matplotlib
+import mlx.core as mx
+import numpy as np
+from time_utils import measure_runtime
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+
+def had(x):
+    y = mx.hadamard_transform(x)
+    mx.eval(y)
+
+
+def copy(x):
+    y = x + 1.0
+    mx.eval(y)
+
+
+def run(dtype):
+    system_size = 2**26
+    outputs = {}
+    for test_fn in (had, copy):
+        for m in [1, 12, 20, 28]:
+            if test_fn == copy:
+                key = "copy"
+            elif m == 1:
+                key = "had_2^k"
+            else:
+                key = "had_m*2^k"
+            outputs.setdefault(key, {})
+            for k in range(7, 14):
+                n = m * 2**k
+                if n > 2**15:
+                    continue
+                x_np = np.random.normal(size=(system_size // n, n)).astype(dtype)
+                x = mx.array(x_np)
+                runtime_ms = measure_runtime(test_fn, x=x)
+                bytes_per_gb = 1e9
+                ms_per_s = 1e3
+                bytes_per_had = np.dtype(x_np.dtype).itemsize * 2
+                bandwidth_gb = (
+                    system_size * bytes_per_had / runtime_ms * ms_per_s / bytes_per_gb
+                )
+                print(n, bandwidth_gb)
+                outputs[key][n] = bandwidth_gb
+
+    colors = {
+        "copy": "black",
+        "had_2^k": "steelblue",
+        "had_m*2^k": "skyblue",
+    }
+    for key, output in outputs.items():
+        plt.scatter(output.keys(), output.values(), color=colors[key], label=key)
+    plt.title(f"MLX Hadamard Benchmark -- {dtype.__name__}")
+    plt.xlabel("N")
+    plt.ylabel("Bandwidth (GB/s)")
+    plt.legend()
+    plt.savefig(f"bench_{dtype.__name__}.png")
+    plt.clf()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--fp16", action="store_true")
+    args = parser.parse_args()
+    dtype = np.float16 if args.fp16 else np.float32
+    run(dtype)

benchmarks/python/scatter_bench.py

@@ -9,7 +9,7 @@ from time_utils import measure_runtime
 
 def benchmark_scatter_mlx(dst_shape, x_shape, idx_shapes):
     def scatter(dst, x, idx):
-        dst[*idx] = x
+        dst[tuple(idx)] = x
         mx.eval(dst)
 
     idx = []
@@ -23,8 +23,8 @@ def benchmark_scatter_mlx(dst_shape, x_shape, idx_shapes):
 
 
 def benchmark_scatter_torch(dst_shape, x_shape, idx_shapes, device):
-    def gather(dst, x, idx, device):
-        dst[*idx] = x
+    def scatter(dst, x, idx, device):
+        dst[tuple(idx)] = x
         if device == torch.device("mps"):
             torch.mps.synchronize()
 
@@ -34,7 +34,7 @@ def benchmark_scatter_torch(dst_shape, x_shape, idx_shapes, device):
     x = torch.randn(x_shape, dtype=torch.float32).to(device)
     dst = torch.randn(dst_shape, dtype=torch.float32).to(device)
 
-    runtime = measure_runtime(gather, dst=dst, x=x, idx=idx, device=device)
+    runtime = measure_runtime(scatter, dst=dst, x=x, idx=idx, device=device)
     print(f"PyTorch: {runtime:.3f}ms")
 
 
@@ -54,7 +54,7 @@ if __name__ == "__main__":
         (100_000, 64),
         (1_000_000, 64),
         (100_000,),
-        (2_000_00,),
+        (200_000,),
         (20_000_000,),
         (10000, 64),
         (100, 64),
@@ -91,6 +91,6 @@ if __name__ == "__main__":
 
     for dst_shape, x_shape, idx_shape in zip(dst_shapes, x_shapes, idx_shapes):
         print("=" * 20)
-        print(f"X {x_shape}, Indices {idx_shape}")
+        print(f"Dst: {dst_shape}, X {x_shape}, Indices {idx_shape}")
         benchmark_scatter_mlx(dst_shape, x_shape, idx_shape)
         benchmark_scatter_torch(dst_shape, x_shape, idx_shape, device=device)

benchmarks/python/sdpa_bench.py

@@ -0,0 +1,189 @@
+# Copyright © 2024 Apple Inc.
+
+import argparse
+import math
+import os
+import subprocess
+import time
+
+import mlx.core as mx
+import numpy as np
+
+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
+
+
+def bench(f, *args):
+    for i in range(N_warmup):
+        f(*args)
+
+    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 mlx_sdpa_fused_inner(q, k, v, scale):
+    return mx.fast.scaled_dot_product_attention(q, k, v, scale=scale, mask=None)
+
+
+def mlx_sdpa_unfused_inner(q, k, v, scale, f32softmax=False):
+    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]
+
+    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 f32softmax:
+        scores = mx.softmax(scores.astype(mx.float32), axis=-1).astype(q_dtype)
+    else:
+        scores = mx.softmax(scores, axis=-1)
+
+    out = scores @ v
+    if n_repeats > 1:
+        out = mx.reshape(out, [B, n_q_heads, L, -1])
+
+    return out
+
+
+def mlx_spda_unfused(q, k, v, scale, transpose):
+    q_out = q
+    if transpose:
+        k = mx.transpose(k, (0, 2, 1, 3))
+        v = mx.transpose(v, (0, 2, 1, 3))
+
+    for i in range(N_iter_func):
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+        q_out = mlx_sdpa_unfused_inner(q_out, k, v, scale)
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+
+    mx.eval(q_out)
+    return q_out
+
+
+def mlx_spda_fused(q, k, v, scale, transpose):
+    q_out = q
+    if transpose:
+        k = mx.transpose(k, (0, 2, 1, 3))
+        v = mx.transpose(v, (0, 2, 1, 3))
+
+    for i in range(N_iter_func):
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+        q_out = mlx_sdpa_fused_inner(q_out, k, v, scale)
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+
+    mx.eval(q_out)
+    return q_out
+
+
+def bench_shape(B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, np_dtype, transpose=True):
+    shape_q = (
+        (B, qsl, n_q_heads, head_dim) if transpose else (B, n_q_heads, qsl, head_dim)
+    )
+    shape_kv = (
+        (B, ksl, n_kv_heads, head_dim) if transpose else (B, n_kv_heads, ksl, head_dim)
+    )
+
+    q_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_q).astype(np_dtype)
+    k_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_kv).astype(np_dtype)
+    v_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_kv).astype(np_dtype)
+
+    scale = math.sqrt(1.0 / head_dim)
+
+    q_mx = mx.array(q_np)
+    k_mx = mx.array(k_np)
+    v_mx = mx.array(v_np)
+
+    time_mlx_unfused = bench(mlx_spda_unfused, q_mx, k_mx, v_mx, scale, transpose)
+    time_mlx_fused = bench(mlx_spda_fused, q_mx, k_mx, v_mx, scale, transpose)
+
+    if transpose:
+        q_mx = mx.transpose(q_mx, (0, 2, 1, 3))
+        k_mx = mx.transpose(k_mx, (0, 2, 1, 3))
+        v_mx = mx.transpose(v_mx, (0, 2, 1, 3))
+
+    o_mlx_fused = mlx_sdpa_fused_inner(q_mx, k_mx, v_mx, scale)
+    o_mlx_unfused = mlx_sdpa_unfused_inner(q_mx, k_mx, v_mx, scale, f32softmax=True)
+
+    atol = 1e-5 if np_dtype == np.float32 else 1e-4
+
+    if not mx.allclose(o_mlx_fused, o_mlx_unfused, atol=atol):
+        print(
+            f"Failed at (B: {B}, qsl: {qsl}, ksl: {ksl}, head_dim: {head_dim}, n_qh: {n_q_heads}, n_kvh: {n_kv_heads}) [tpose = {transpose}] with max(|a - b|) = {mx.max(mx.abs(o_mlx_unfused - o_mlx_fused)):3.2e}"
+        )
+
+    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)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run gemm benchmarks")
+
+    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,    32),
+          (  1,  2048,  2048,       64,   32,    32),
+          (  1,  4096,  4096,       64,   32,    32),
+    )
+
+    shapes_80 = (
+        # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
+          (  1,  1024,  1024,       80,   32,    32),
+          (  1,  2048,  2048,       80,   32,    32),
+          (  1,  4096,  4096,       80,   32,    32),
+    )
+
+    shapes_128 = (
+        # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
+          (  1,  1024,  1024,      128,   32,    32),
+          (  1,  2048,  2048,      128,   32,    32),
+          (  1,  4096,  4096,      128,   32,    32),
+    )
+    # fmt: on
+
+    shapes = shapes_64 + shapes_80 + shapes_128
+
+    print("  B,   qsl,   ksl, hdim, n_qh, n_kvh, tpose,   dtype, t_unfs, t_fuse, diff%")
+
+    for dtype in dtypes:
+        for transpose in transposes:
+            for B, qsl, ksl, head_dim, n_q_heads, n_kv_heads in shapes:
+                np_dtype = getattr(np, dtype)
+                time_mlx_fused, time_mlx_unfused = bench_shape(
+                    B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, np_dtype, transpose
+                )
+                diff = time_mlx_unfused / time_mlx_fused - 1.0
+                t_str = 1 if transpose else 0
+                print(
+                    f"{B:3d}, {qsl:5d}, {ksl:5d}, {head_dim:4d}, {n_q_heads:4d}, {n_kv_heads:5d}, {t_str:5d}, {dtype}, {time_mlx_unfused: 2.3f}, {time_mlx_fused: 2.3f}, {100. * diff:+5.2f}%"
+                )

benchmarks/python/sdpa_vector_bench.py

@@ -0,0 +1,95 @@
+import argparse
+import math
+
+import mlx.core as mx
+from time_utils import time_fn
+
+L = 16384
+H = 32
+H_k = H // 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, 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 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)
+
+
+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)
+
+
+if __name__ == "__main__":
+    time_self_attention_sdpa()
+    time_self_attention_primitives()
+    time_self_attention_sdpa_with_mask()

benchmarks/python/synchronize_bench.py

@@ -0,0 +1,55 @@
+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

@@ -1,56 +1,41 @@
 include(CMakeParseArguments)
 
-###############################################################################
+# ##############################################################################
 # Build metal library
 #
 # Adds a custom target ${TARGET} to build ${OUTPUT_DIRECTORY}/{TITLE}.metallib
 # from list ${SOURCES}, including list ${INCLUDE_DIRS}, depends on list ${DEPS}
 #
-# 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)
+# Args: TARGET: Custom target to be added for the metal library TITLE: Name of
+# the .metallib OUTPUT_DIRECTORY: Where to place ${TITLE}.metallib SOURCES: List
+# of source files INCLUDE_DIRS: List of include dirs DEPS: List of dependency
+# files (like headers)
 #
 macro(mlx_build_metallib)
   # Parse args
   set(oneValueArgs TARGET TITLE OUTPUT_DIRECTORY)
   set(multiValueArgs SOURCES INCLUDE_DIRS DEPS)
-  cmake_parse_arguments(
-      MTLLIB 
-      ""
-      "${oneValueArgs}"
-      "${multiValueArgs}" 
-      ${ARGN}
-  )
+  cmake_parse_arguments(MTLLIB "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
   # Set output
   set(MTLLIB_BUILD_TARGET "${MTLLIB_OUTPUT_DIRECTORY}/${MTLLIB_TITLE}.metallib")
 
-  # Collect compile options 
+  # Collect compile options
   set(MTLLIB_COMPILE_OPTIONS -Wall -Wextra -fno-fast-math)
 
   # Prepare metallib build command
   add_custom_command(
     OUTPUT ${MTLLIB_BUILD_TARGET}
-    COMMAND xcrun -sdk macosx metal 
-                  "$<LIST:TRANSFORM,${MTLLIB_INCLUDE_DIRS},PREPEND,-I>"
-                  ${MTLLIB_COMPILE_OPTIONS}
-                  ${MTLLIB_SOURCES}
-                  -o ${MTLLIB_BUILD_TARGET}
+    COMMAND
+      xcrun -sdk macosx metal
+      "$<LIST:TRANSFORM,${MTLLIB_INCLUDE_DIRS},PREPEND,-I>"
+      ${MTLLIB_COMPILE_OPTIONS} ${MTLLIB_SOURCES} -o ${MTLLIB_BUILD_TARGET}
     DEPENDS ${MTLLIB_DEPS} ${MTLLIB_SOURCES}
     COMMAND_EXPAND_LISTS
     COMMENT "Building ${MTLLIB_TITLE}.metallib"
-    VERBATIM
-  )
+    VERBATIM)
 
   # Add metallib custom target
-  add_custom_target(
-    ${MTLLIB_TARGET}
-    DEPENDS
-    ${MTLLIB_BUILD_TARGET}
-  )
+  add_custom_target(${MTLLIB_TARGET} DEPENDS ${MTLLIB_BUILD_TARGET})
 
-endmacro(mlx_build_metallib)
+endmacro(mlx_build_metallib)

cmake/metal.14.0.diff

@@ -1,36 +0,0 @@
-diff -ur Metal/MTLEvent.hpp MetalNew/MTLEvent.hpp
---- Metal/MTLEvent.hpp	2023-06-01 12:18:26
-+++ MetalNew/MTLEvent.hpp	2024-04-15 07:36:59
-@@ -62,6 +62,7 @@
- 
-     uint64_t                 signaledValue() const;
-     void                     setSignaledValue(uint64_t signaledValue);
-+    bool                     waitUntilSignaledValue(uint64_t signaledValue, uint64_t timeoutMS);
- };
- 
- class SharedEventHandle : public NS::SecureCoding<SharedEventHandle>
-@@ -138,6 +139,11 @@
- _MTL_INLINE void MTL::SharedEvent::setSignaledValue(uint64_t signaledValue)
- {
-     Object::sendMessage<void>(this, _MTL_PRIVATE_SEL(setSignaledValue_), signaledValue);
-+}
-+
-+// method: waitUntilSignaledValue
-+_MTL_INLINE bool MTL::SharedEvent::waitUntilSignaledValue(uint64_t signaledValue, uint64_t timeoutMS) {
-+    return Object::sendMessage<bool>(this, _MTL_PRIVATE_SEL(waitUntilSignaledValue_timeoutMS_), signaledValue, timeoutMS);
- }
- 
- // static method: alloc
-diff -ur Metal/MTLHeaderBridge.hpp MetalNew/MTLHeaderBridge.hpp
---- Metal/MTLHeaderBridge.hpp	2023-06-01 12:18:26
-+++ MetalNew/MTLHeaderBridge.hpp	2024-04-15 07:37:29
-@@ -1906,6 +1906,9 @@
-     "setShouldMaximizeConcurrentCompilation:");
- _MTL_PRIVATE_DEF_SEL(setSignaledValue_,
-     "setSignaledValue:");
-+_MTL_PRIVATE_DEF_SEL(
-+    waitUntilSignaledValue_timeoutMS_,
-+    "waitUntilSignaledValue:timeoutMS:");
- _MTL_PRIVATE_DEF_SEL(setSize_,
-     "setSize:");
- _MTL_PRIVATE_DEF_SEL(setSlice_,

cmake/metal.14.2.diff

@@ -1,36 +0,0 @@
-diff -ur Metal/MTLEvent.hpp MetalNew/MTLEvent.hpp
---- Metal/MTLEvent.hpp	2024-04-15 07:12:10
-+++ MetalNew/MTLEvent.hpp	2024-04-15 07:15:50
-@@ -62,6 +62,7 @@
- 
-     uint64_t                 signaledValue() const;
-     void                     setSignaledValue(uint64_t signaledValue);
-+    bool                     waitUntilSignaledValue(uint64_t signaledValue, uint64_t timeoutMS);
- };
- 
- class SharedEventHandle : public NS::SecureCoding<SharedEventHandle>
-@@ -138,6 +139,11 @@
- _MTL_INLINE void MTL::SharedEvent::setSignaledValue(uint64_t signaledValue)
- {
-     Object::sendMessage<void>(this, _MTL_PRIVATE_SEL(setSignaledValue_), signaledValue);
-+}
-+
-+// method: waitUntilSignaledValue
-+_MTL_INLINE bool MTL::SharedEvent::waitUntilSignaledValue(uint64_t signaledValue, uint64_t timeoutMS) {
-+    return Object::sendMessage<bool>(this, _MTL_PRIVATE_SEL(waitUntilSignaledValue_timeoutMS_), signaledValue, timeoutMS);
- }
- 
- // static method: alloc
-diff -ur Metal/MTLHeaderBridge.hpp MetalNew/MTLHeaderBridge.hpp
---- Metal/MTLHeaderBridge.hpp	2024-04-15 07:12:10
-+++ MetalNew/MTLHeaderBridge.hpp	2024-04-15 07:16:15
-@@ -1918,6 +1918,9 @@
-     "setShouldMaximizeConcurrentCompilation:");
- _MTL_PRIVATE_DEF_SEL(setSignaledValue_,
-     "setSignaledValue:");
-+_MTL_PRIVATE_DEF_SEL(
-+    waitUntilSignaledValue_timeoutMS_,
-+    "waitUntilSignaledValue:timeoutMS:");
- _MTL_PRIVATE_DEF_SEL(setSize_,
-     "setSize:");
- _MTL_PRIVATE_DEF_SEL(setSlice_,

docs/requirements.txt

@@ -1,3 +1,4 @@
 sphinx
 breathe
 sphinx-book-theme
+mlx

docs/src/conf.py

@@ -60,6 +60,7 @@ html_theme_options = {
     },
 }
 
+html_favicon = html_theme_options["logo"]["image_light"]
 
 # -- Options for HTMLHelp output ---------------------------------------------
 
@@ -83,3 +84,15 @@ def setup(app):
 # -- Options for LaTeX output ------------------------------------------------
 
 latex_documents = [(main_doc, "MLX.tex", "MLX Documentation", author, "manual")]
+latex_elements = {
+    "preamble": r"""
+    \usepackage{enumitem}
+    \setlistdepth{5}
+    \setlist[itemize,1]{label=$\bullet$}
+    \setlist[itemize,2]{label=$\bullet$}
+    \setlist[itemize,3]{label=$\bullet$}
+    \setlist[itemize,4]{label=$\bullet$}
+    \setlist[itemize,5]{label=$\bullet$}
+    \renewlist{itemize}{itemize}{5}
+""",
+}

docs/src/dev/custom_metal_kernels.rst

@@ -0,0 +1,427 @@
+.. _custom_metal_kernels:
+
+Custom Metal Kernels
+====================
+
+MLX supports writing custom Metal kernels through the Python and C++ APIs.
+
+Simple Example
+--------------
+
+Let's write a custom kernel that computes ``exp`` elementwise:
+
+.. code-block:: python
+
+  def exp_elementwise(a: mx.array):
+      source = """
+          uint elem = thread_position_in_grid.x;
+          T tmp = inp[elem];
+          out[elem] = metal::exp(tmp);
+      """
+
+      kernel = mx.fast.metal_kernel(
+          name="myexp",
+          input_names=["inp"],
+          output_names=["out"],
+          source=source,
+      )
+      outputs = kernel(
+          inputs=[a],
+          template=[("T", mx.float32)],
+          grid=(a.size, 1, 1),
+          threadgroup=(256, 1, 1),
+          output_shapes=[a.shape],
+          output_dtypes=[a.dtype],
+      )
+      return outputs[0]
+
+  a = mx.random.normal(shape=(4, 16)).astype(mx.float16)
+  b = exp_elementwise(a)
+  assert mx.allclose(b, mx.exp(a))
+
+.. note::
+    We are only required to pass the body of the Metal kernel in ``source``.
+
+The full function signature will be generated using:
+
+* The shapes/dtypes of ``inputs``
+    In the above, ``a`` is an ``mx.array`` of type ``mx.float16`` and we pass it with the key ``inp``
+    so we will add ``const device float16_t* inp`` to the signature.
+    ``inp_shape``, ``inp_strides`` and ``inp_ndim`` are also added for convenience if they are present
+    in ``source``.
+* The list of ``output_dtypes``
+    In the above, ``out`` is an ``mx.array`` of type ``mx.float16``
+    so we add ``device float16_t* out``.
+* Template parameters passed using ``template``
+    In the above, ``template=[("T", mx.float32)]`` adds a template of ``template <typename T>`` to the function
+    and instantiates the template with ``custom_kernel_myexp_float<float>``.
+    Template parameters can be ``mx.core.Dtype``, ``int`` or ``bool``.
+* Metal attributes used in ``source`` such as ``[[thread_position_in_grid]]``
+    These will be added as function arguments.
+    All the attributes defined in Table 5.8 of the `Metal Shading Language Specification <https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf>`_ are supported.
+
+Putting this all together, the generated function signature for ``myexp`` is as follows:
+
+.. code-block:: cpp
+
+  template <typename T>
+  [[kernel]] void custom_kernel_myexp_float(
+    const device float16_t* inp [[buffer(0)]],
+    device float16_t* out [[buffer(1)]],
+    uint3 thread_position_in_grid [[thread_position_in_grid]]) {
+
+          uint elem = thread_position_in_grid.x;
+          T tmp = inp[elem];
+          out[elem] = metal::exp(tmp);
+
+  }
+
+  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.
+
+Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
+
+Using Shape/Strides
+-------------------
+
+``mx.fast.metal_kernel`` supports an argument ``ensure_row_contiguous`` which is ``True`` by default.
+This will copy the ``mx.array`` inputs if needed before the kernel is launched to ensure that the memory layout is row contiguous.
+Generally this makes writing the kernel easier, since we don't have to worry about gaps or the ordering of the dims
+when indexing.
+
+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``:
+
+.. code-block:: python
+
+  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)],
+          grid=(a.size, 1, 1),
+          threadgroup=(256, 1, 1),
+          output_shapes=[a.shape],
+          output_dtypes=[a.dtype],
+          ensure_row_contiguous=False,
+      )
+      return outputs[0]
+
+  a = mx.random.normal(shape=(4, 16)).astype(mx.float16)
+  # make non-contiguous
+  a = a[::2]
+  b = exp_elementwise(a)
+  assert mx.allclose(b, mx.exp(a))
+
+Complex Example
+-----------------------------
+
+Let's implement a more complex example: ``grid_sample`` in ``"bilinear"`` mode.
+
+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
+
+        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_sw = ix_nw
+        iy_sw = 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)
+
+        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)
+
+        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
+
+        return output
+
+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
+
+    @mx.custom_function
+    def grid_sample(x, grid):
+
+        assert x.ndim == 4, "`x` must be 4D."
+        assert grid.ndim == 4, "`grid` must be 4D."
+
+        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."
+
+        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 w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
+
+            uint grid_idx = elem / C * 2;
+            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+
+            int ix_nw = floor(ix);
+            int iy_nw = floor(iy);
+
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
+
+            int ix_sw = ix_nw;
+            int iy_sw = iy_nw + 1;
+
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
+
+            T nw = (ix_se - ix)    * (iy_se - iy);
+            T ne = (ix    - ix_sw) * (iy_sw - iy);
+            T sw = (ix_ne - ix)    * (iy    - iy_ne);
+            T se = (ix    - ix_nw) * (iy    - iy_nw);
+
+            int batch_idx = elem / C / gH / gW * b_stride;
+            int channel_idx = elem % C;
+            int base_idx = batch_idx + channel_idx;
+
+            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];
+
+            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;
+
+            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)
+
+On an M1 Max, we see a big performance improvement:
+
+``55.7ms -> 6.7ms => 8x speed up``
+
+Grid Sample VJP
+---------------
+
+Since we decorated ``grid_sample`` with ``mx.custom_function``, we can now define
+its custom vjp transform so MLX can differentiate it.
+
+The backwards pass requires atomically updating ``x_grad``/``grid_grad`` and so
+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.
+
+* ``atomic_outputs=True``
+    Designate all of the kernel outputs as ``atomic`` in the function signature. 
+    This means we can use Metal's ``atomic`` features to simultaneously update the ``x_grad`` and ``grid_grad`` arrays from multiple threadgroups. 
+    See section 6.15 of the `Metal Shading Language Specification <https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf>`_ for more details.
+
+We can then implement the backwards pass as follows:
+
+.. code-block:: python
+
+    @grid_sample.vjp
+    def grid_sample_vjp(primals, cotangent, _):
+        x, grid = primals
+        B, _, _, C = x.shape
+        _, gN, gM, D = grid.shape
+
+        assert D == 2, "Last dim of `grid` must be size 2."
+
+        source = """
+            uint elem = thread_position_in_grid.x;
+            int H = x_shape[1];
+            int W = x_shape[2];
+            int C = x_shape[3];
+            // Pad C to the nearest larger simdgroup size multiple
+            int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
+
+            int gH = grid_shape[1];
+            int gW = grid_shape[2];
+
+            int w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
+
+            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_nw = floor(ix);
+            int iy_nw = floor(iy);
+
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
+
+            int ix_sw = ix_nw;
+            int iy_sw = iy_nw + 1;
+
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
+
+            T nw = (ix_se - ix)    * (iy_se - iy);
+            T ne = (ix    - ix_sw) * (iy_sw - iy);
+            T sw = (ix_ne - ix)    * (iy    - iy_ne);
+            T se = (ix    - ix_nw) * (iy    - iy_nw);
+
+            int batch_idx = elem / C_padded / gH / gW * b_stride;
+            int channel_idx = elem % C_padded;
+            int base_idx = batch_idx + channel_idx;
+
+            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_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_ne = x[offset];
+                    gix += I_ne * (iy_sw - iy) * cot;
+                    giy -= I_ne * (ix - ix_sw) * cot;
+                }
+                if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
+                    int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
+
+                    T I_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_se = x[offset];
+                    gix += I_se * (iy - iy_nw) * cot;
+                    giy += I_se * (ix - ix_nw) * cot;
+                }
+            }
+
+            T gix_mult = W / 2;
+            T giy_mult = H / 2;
+
+            // Reduce across each simdgroup first.
+            // This is much faster than relying purely on atomics.
+            gix = simd_sum(gix);
+            giy = simd_sum(giy);
+
+            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:
+
+``676.4ms -> 16.7ms => 40x speed up``

docs/src/dev/extensions.rst

@@ -1,5 +1,5 @@
-Developer Documentation
-=======================
+Custom Extensions in MLX
+========================
 
 You can extend MLX with custom operations on the CPU or GPU. This guide
 explains how to do that with a simple example.
@@ -420,8 +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 size_t* x_strides [[buffer(6)]],
-            constant const size_t* y_strides [[buffer(7)]],
+            constant const int64_t* x_strides [[buffer(6)]],
+            constant const int64_t* y_strides [[buffer(7)]],
             constant const int& ndim [[buffer(8)]],
             uint index [[thread_position_in_grid]]) {
         // Convert linear indices to offsets in array
@@ -438,24 +438,10 @@ each instantiation a unique host name so we can identify it.
 
 .. code-block:: C++
 
-    #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);
+    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)
 
 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
@@ -486,16 +472,15 @@ below.
         std::ostringstream kname;
         kname << "axpby_" << "general_" << type_to_name(out);
 
-        // Make sure the metal library is available and look for it
-        // in the same folder as this executable if needed
-        d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
+        // 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(), "mlx_ext");
 
         // Prepare to encode kernel
-        auto compute_encoder = d.get_command_encoder(s.index);
-        compute_encoder->setComputePipelineState(kernel);
+        auto& compute_encoder = d.get_command_encoder(s.index);
+        compute_encoder.set_compute_pipeline_state(kernel);
 
         // Kernel parameters are registered with buffer indices corresponding to
         // those in the kernel declaration at axpby.metal
@@ -503,21 +488,21 @@ below.
         size_t nelem = out.size();
 
         // Encode input arrays to kernel
-        set_array_buffer(compute_encoder, x, 0);
-        set_array_buffer(compute_encoder, y, 1);
+        compute_encoder.set_input_array(x, 0);
+        compute_encoder.set_input_array(y, 1);
 
         // Encode output arrays to kernel
-        set_array_buffer(compute_encoder, out, 2);
+        compute_encoder.set_output_array(out, 2);
 
         // Encode alpha and beta
-        compute_encoder->setBytes(&alpha_, sizeof(float), 3);
-        compute_encoder->setBytes(&beta_, sizeof(float), 4);
+        compute_encoder.set_bytes(alpha_, 3);
+        compute_encoder.set_bytes(beta_, 4);
 
         // Encode shape, strides and ndim
-        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);
+        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);
 
         // 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
@@ -531,7 +516,7 @@ below.
 
         // Launch the grid with the given number of threads divided among
         // the given threadgroups
-        compute_encoder->dispatchThreads(grid_dims, group_dims);
+        compute_encoder.dispatch_threads(grid_dims, group_dims);
     }
 
 We can now call the :meth:`axpby` operation on both the CPU and the GPU!
@@ -825,7 +810,7 @@ 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 correctness: {mx.all(c == 6.0).item()}")
+    print(f"c correct: {mx.all(c == 6.0).item()}")
 
 Output:
 

docs/src/dev/mlx_in_cpp.rst

@@ -0,0 +1,121 @@
+.. _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/examples/llama-inference.rst

@@ -15,7 +15,7 @@ module to concisely define the model architecture.
 Attention layer
 ^^^^^^^^^^^^^^^^
 
-We will start with the llama attention layer which notably uses the RoPE
+We will start with the Llama attention layer which notably uses the RoPE
 positional encoding. [1]_ In addition, our attention layer will optionally use a
 key/value cache that will be concatenated with the provided keys and values to
 support efficient inference.

docs/src/examples/mlp.rst

@@ -64,7 +64,7 @@ set:
 Next, setup the problem parameters and load the data. To load the data, you need our
 `mnist data loader
 <https://github.com/ml-explore/mlx-examples/blob/main/mnist/mnist.py>`_, which
-we will import as `mnist`.
+we will import as ``mnist``.
 
 .. code-block:: python
 

docs/src/index.rst

@@ -43,7 +43,9 @@ are the CPU and GPU.
    usage/function_transforms
    usage/compile
    usage/numpy
+   usage/distributed
    usage/using_streams
+   usage/export
 
 .. toctree::
    :caption: Examples
@@ -60,6 +62,7 @@ are the CPU and GPU.
    python/array
    python/data_types
    python/devices_and_streams
+   python/export
    python/ops
    python/random
    python/transforms
@@ -69,6 +72,7 @@ are the CPU and GPU.
    python/metal
    python/nn
    python/optimizers
+   python/distributed
    python/tree_utils
 
 .. toctree::
@@ -83,3 +87,5 @@ are the CPU and GPU.
 
    dev/extensions
    dev/metal_debugger
+   dev/custom_metal_kernels
+   dev/mlx_in_cpp

docs/src/install.rst

@@ -1,3 +1,5 @@
+.. _build_and_install:
+
 Build and Install
 =================
 
@@ -14,7 +16,7 @@ silicon computer is
 To install from PyPI you must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
-- Using a native Python >= 3.8
+- Using a native Python >= 3.9
 - macOS >= 13.5
 
 .. note::
@@ -53,7 +55,7 @@ Build Requirements
 ^^^^^^^^^^^^^^^^^^
 
 - A C++ compiler with C++17 support (e.g. Clang >= 5.0)
-- `cmake <https://cmake.org/>`_ -- version 3.24 or later, and ``make``
+- `cmake <https://cmake.org/>`_ -- version 3.25 or later, and ``make``
 - Xcode >= 15.0 and macOS SDK >= 14.0
 
 .. note::
@@ -70,36 +72,36 @@ To build and install the MLX python library from source, first, clone MLX from
 
    git clone git@github.com:ml-explore/mlx.git mlx && cd mlx
 
-Install `nanobind <https://nanobind.readthedocs.io/en/latest/>`_ with:
-
-.. code-block:: shell
-
-    pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
-
 Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-   env CMAKE_BUILD_PARALLEL_LEVEL="" pip install .
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
 
-For developing use an editable install:
+For developing, install the package with development dependencies, and use an
+editable install:
 
 .. code-block:: shell
 
-  env CMAKE_BUILD_PARALLEL_LEVEL="" pip install -e .
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
 
-To make sure the install is working run the tests with:
+Once the development dependencies are installed, you can build faster with:
+
+.. code-block:: shell
+
+ CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
+
+Run the tests with:
 
 .. code-block:: shell
 
-  pip install ".[testing]"
   python -m unittest discover python/tests
 
-Optional: Install stubs to enable auto completions and type checking from your IDE:
+Optional: Install stubs to enable auto completions and type checking from your
+IDE:
 
 .. code-block:: shell
 
-  pip install ".[dev]"
   python setup.py generate_stubs
 
 C++ API
@@ -153,11 +155,18 @@ should point to the path to the built metal library.
      - OFF
    * - MLX_BUILD_METAL
      - ON
+   * - MLX_BUILD_CPU
+     - ON
    * - MLX_BUILD_PYTHON_BINDINGS
      - OFF
    * - MLX_METAL_DEBUG
      - OFF
-
+   * - MLX_BUILD_SAFETENSORS
+     - ON
+   * - MLX_BUILD_GGUF
+     - ON
+   * - MLX_METAL_JIT
+     - OFF
 
 .. note::
 
@@ -176,10 +185,37 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
+Binary Size Minimization
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+To produce a smaller binary use the CMake flags ``CMAKE_BUILD_TYPE=MinSizeRel``
+and ``BUILD_SHARED_LIBS=ON``.
+
+The MLX CMake build has several additional options to make smaller binaries.
+For example, if you don't need the CPU backend or support for safetensors and
+GGUF, you can do:
+
+.. code-block:: shell
+
+  cmake .. \
+    -DCMAKE_BUILD_TYPE=MinSizeRel \
+    -DBUILD_SHARED_LIBS=ON \
+    -DMLX_BUILD_CPU=OFF \
+    -DMLX_BUILD_SAFETENSORS=OFF \
+    -DMLX_BUILD_GGUF=OFF \
+    -DMLX_METAL_JIT=ON
+
+THE ``MLX_METAL_JIT`` flag minimizes the size of the MLX Metal library which
+contains pre-built GPU kernels. This substantially reduces the size of the
+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.
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 
-
 Metal not found
 ~~~~~~~~~~~~~~~
 
@@ -206,7 +242,7 @@ x86 Shell
 
 .. _build shell:
 
-If the ouptut of ``uname -p``  is ``x86`` then your shell is running as x86 via
+If the output of ``uname -p``  is ``x86`` then your shell is running as x86 via
 Rosetta instead of natively.
 
 To fix this, find the application in Finder (``/Applications`` for iTerm,
@@ -230,4 +266,4 @@ Also check that cmake is using the correct architecture:
 
 If you see ``"x86_64"``, try re-installing ``cmake``. If you see ``"arm64"``
 but the build errors out with "Building for x86_64 on macOS is not supported."
-wipe your build cahce with ``rm -rf build/`` and try again.
+wipe your build cache with ``rm -rf build/`` and try again.

docs/src/python/array.rst

@@ -24,6 +24,7 @@ Array
     array.any
     array.argmax
     array.argmin
+    array.conj
     array.cos
     array.cummax
     array.cummin
@@ -52,8 +53,10 @@ Array
     array.sqrt
     array.square
     array.squeeze
-    array.swapaxes
+    array.std
     array.sum
+    array.swapaxes
     array.transpose
     array.T
     array.var
+    array.view

docs/src/python/data_types.rst

@@ -51,11 +51,20 @@ 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.
@@ -66,3 +75,4 @@ documentation for more information. Use :func:`issubdtype` to determine if one
    Dtype
    DtypeCategory
    issubdtype
+   finfo

docs/src/python/distributed.rst

@@ -0,0 +1,22 @@
+.. _distributed:
+
+.. currentmodule:: mlx.core.distributed
+
+Distributed Communication
+==========================
+
+MLX provides a distributed communication package using MPI. The MPI library is
+loaded at runtime; if MPI is available then distributed communication is also
+made available.
+
+.. autosummary::
+   :toctree: _autosummary
+
+    Group
+    is_available
+    init
+    all_sum
+    all_gather
+    send
+    recv
+    recv_like

docs/src/python/export.rst

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

docs/src/python/fast.rst

@@ -12,3 +12,4 @@ Fast
   layer_norm
   rope
   scaled_dot_product_attention
+  metal_kernel

docs/src/python/linalg.rst

@@ -5,8 +5,20 @@ Linear Algebra
 
 .. currentmodule:: mlx.core.linalg
 
-.. autosummary:: 
-   :toctree: _autosummary 
+.. autosummary::
+   :toctree: _autosummary
 
+    inv
+    tri_inv
     norm
+    cholesky
+    cholesky_inv
+    cross
     qr
+    svd
+    eigvalsh
+    eigh
+    lu
+    lu_factor
+    solve
+    solve_triangular

docs/src/python/metal.rst

@@ -14,6 +14,7 @@ Metal
   get_cache_memory
   set_memory_limit
   set_cache_limit
+  set_wired_limit
   clear_cache
   start_capture
   stop_capture

docs/src/python/nn/functions.rst

@@ -13,10 +13,13 @@ simple functions.
    :template: nn-module-template.rst
 
    elu
+   celu
    gelu
    gelu_approx
    gelu_fast_approx
    glu
+   hard_shrink
+   hard_tanh
    hardswish
    leaky_relu
    log_sigmoid
@@ -29,6 +32,7 @@ simple functions.
    sigmoid
    silu
    softmax
+   softmin
    softplus
    softshrink
    step

docs/src/python/nn/layers.rst

@@ -12,22 +12,37 @@ Layers
    ALiBi
    AvgPool1d
    AvgPool2d
+   AvgPool3d
    BatchNorm
+   CELU
    Conv1d
    Conv2d
+   Conv3d
+   ConvTranspose1d
+   ConvTranspose2d
+   ConvTranspose3d
    Dropout
    Dropout2d
    Dropout3d
    Embedding
+   ELU
    GELU
+   GLU
    GroupNorm
    GRU
+   HardShrink
+   HardTanh
+   Hardswish
    InstanceNorm
    LayerNorm
+   LeakyReLU
    Linear
+   LogSigmoid
+   LogSoftmax
    LSTM
    MaxPool1d
    MaxPool2d
+   MaxPool3d
    Mish
    MultiHeadAttention
    PReLU
@@ -35,13 +50,20 @@ Layers
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU6
    RNN
    RoPE
    SELU
    Sequential
+   Sigmoid
    SiLU
    SinusoidalPositionalEncoding
+   Softmin
    Softshrink
+   Softsign
+   Softmax
+   Softplus
    Step
+   Tanh
    Transformer
    Upsample

docs/src/python/ops.rst

@@ -10,6 +10,7 @@ Operations
 
    abs
    add
+   addmm
    all
    allclose
    any
@@ -19,27 +20,35 @@ Operations
    arcsin
    arcsinh
    arctan
+   arctan2
    arctanh
    argmax
    argmin
    argpartition
    argsort
    array_equal
+   as_strided
    atleast_1d
    atleast_2d
    atleast_3d
    bitwise_and
+   bitwise_invert
    bitwise_or
    bitwise_xor
    block_masked_mm
-   block_sparse_mm
    broadcast_to
    ceil
    clip
    concatenate
+   conj
+   conjugate
    convolve
    conv1d
    conv2d
+   conv3d
+   conv_transpose1d
+   conv_transpose2d
+   conv_transpose3d
    conv_general
    cos
    cosh
@@ -53,6 +62,8 @@ Operations
    diagonal
    divide
    divmod
+   einsum
+   einsum_path
    equal
    erf
    erfinv
@@ -64,15 +75,22 @@ Operations
    floor
    floor_divide
    full
+   gather_mm
+   gather_qmm
    greater
    greater_equal
+   hadamard_transform
    identity
+   imag
    inner
+   isfinite
    isclose
    isinf
    isnan
    isneginf
    isposinf
+   issubdtype
+   kron
    left_shift
    less
    less_equal
@@ -96,6 +114,7 @@ Operations
    minimum
    moveaxis
    multiply
+   nan_to_num
    negative
    not_equal
    ones
@@ -103,14 +122,19 @@ Operations
    outer
    partition
    pad
+   power
    prod
+   put_along_axis
    quantize
    quantized_matmul
    radians
+   real
    reciprocal
+   remainder
    repeat
    reshape
    right_shift
+   roll
    round
    rsqrt
    save
@@ -122,6 +146,8 @@ Operations
    sign
    sin
    sinh
+   slice
+   slice_update
    softmax
    sort
    split
@@ -141,11 +167,14 @@ Operations
    tensordot
    tile
    topk
+   trace
    transpose
    tri
    tril
    triu
+   unflatten
    var
+   view
    where
    zeros
    zeros_like

docs/src/python/optimizers.rst

@@ -31,6 +31,41 @@ model's parameters and the **optimizer state**.
             # Compute the new parameters but also the optimizer state.
             mx.eval(model.parameters(), optimizer.state)
 
+Saving and Loading
+------------------
+
+To serialize an optimizer, save its state. To load an optimizer, load and set
+the saved state. Here's a simple example:
+
+.. code-block:: python
+
+   import mlx.core as mx
+   from mlx.utils import tree_flatten, tree_unflatten
+   import mlx.optimizers as optim
+
+   optimizer = optim.Adam(learning_rate=1e-2)
+
+   # Perform some updates with the optimizer
+   model = {"w" : mx.zeros((5, 5))}
+   grads = {"w" : mx.ones((5, 5))}
+   optimizer.update(model, grads)
+
+   # Save the state
+   state = tree_flatten(optimizer.state)
+   mx.save_safetensors("optimizer.safetensors", dict(state))
+
+   # Later on, for example when loading from a checkpoint,
+   # recreate the optimizer and load the state
+   optimizer = optim.Adam(learning_rate=1e-2)
+
+   state = tree_unflatten(list(mx.load("optimizer.safetensors").items()))
+   optimizer.state = state
+
+Note, not every optimizer configuation parameter is saved in the state. For
+example, for Adam the learning rate is saved but the ``betas`` and ``eps``
+parameters are not. A good rule of thumb is if the parameter can be scheduled
+then it will be included in the optimizer state.
+
 .. toctree::
 
    optimizers/optimizer

docs/src/python/random.rst

@@ -44,3 +44,5 @@ we use a splittable version of Threefry, which is a counter-based PRNG.
    split
    truncated_normal
    uniform
+   laplace
+   permutation

docs/src/python/transforms.rst

@@ -10,6 +10,7 @@ Transforms
 
    eval
    compile
+   custom_function
    disable_compile
    enable_compile
    grad

docs/src/usage/compile.rst

@@ -33,12 +33,12 @@ Let's start with a simple example:
   # Compile the function
   compiled_fun = mx.compile(fun)
 
-  # Prints: array(2.36788, dtype=float32) 
+  # Prints: array(2.36788, dtype=float32)
   print(compiled_fun(x, y))
 
 The output of both the regular function and the compiled function is the same
 up to numerical precision.
-   
+
 The first time you call a compiled function, MLX will build the compute
 graph, optimize it, and generate and compile code. This can be relatively
 slow. However, MLX will cache compiled functions, so calling a compiled
@@ -96,7 +96,7 @@ element-wise operations:
 
 .. code-block:: python
 
-  def gelu(x):  
+  def gelu(x):
       return x * (1 + mx.erf(x / math.sqrt(2))) / 2
 
 If you use this function with small arrays, it will be overhead bound. If you
@@ -136,13 +136,6 @@ Now make an array, and benchmark both functions:
 On an M1 Max the times are 15.5 and 3.1 milliseconds. The compiled ``gelu`` is
 five times faster.
 
-.. note::
-
-  As of the latest MLX, CPU functions are not fully compiled. Compiling CPU
-  functions can still be helpful, but won't typically result in as large a
-  speedup as compiling operations that run on the GPU.
-
-
 Debugging
 ---------
 
@@ -287,7 +280,7 @@ to the function. In some cases this can be pretty inconvenient. Hence,
   print(fun(mx.array(1.0)))
 
 
-Compiling Training Graphs 
+Compiling Training Graphs
 -------------------------
 
 This section will step through how to use :func:`compile` with a simple example
@@ -297,7 +290,7 @@ full forward, backward, and update with :func:`compile`.
 
 To start, here is the simple example without any compilation:
 
-.. code-block:: python 
+.. code-block:: python
 
   import mlx.core as mx
   import mlx.nn as nn
@@ -330,7 +323,7 @@ To start, here is the simple example without any compilation:
 To compile the update we can put it all in a function and compile it with the
 appropriate input and output captures. Here's the same example but compiled:
 
-.. code-block:: python 
+.. code-block:: python
 
   import mlx.core as mx
   import mlx.nn as nn
@@ -355,7 +348,7 @@ appropriate input and output captures. Here's the same example but compiled:
 
   # The state that will be captured as input and output
   state = [model.state, optimizer.state]
-      
+
   @partial(mx.compile, inputs=state, outputs=state)
   def step(x, y):
       loss_and_grad_fn = nn.value_and_grad(model, loss_fn)
@@ -410,7 +403,7 @@ Compiling transformed functions works just as expected:
 
    In order to compile as much as possible, a transformation of a compiled
    function will not by default be compiled. To compile the transformed
-   function simply pass it through :func:`compile`. 
+   function simply pass it through :func:`compile`.
 
 You can also compile functions which themselves call compiled functions. A
 good practice is to compile the outer most function to give :func:`compile`
@@ -428,3 +421,77 @@ 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

@@ -0,0 +1,167 @@
+.. _usage_distributed:
+
+Distributed Communication
+=========================
+
+.. currentmodule:: mlx.core.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::
+   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
+---------------
+
+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
+
+    import mlx.core as mx
+
+    world = mx.distributed.init()
+    x = mx.distributed.all_sum(mx.ones(10))
+    print(world.rank(), x)
+
+The program above sums the array ``mx.ones(10)`` across all
+distributed processes. If simply run with ``python``, however, only one
+process is launched and no distributed communication takes place.
+
+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
+
+    $ 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)
+
+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
+
+    $ 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``.
+
+.. code:: shell
+
+    $ 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::
+  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``.
+
+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::
+
+    host1 slots=1
+    host2 slots=1
+
+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
+----------------
+
+In this section we will adapt an MLX training loop to support data parallel
+distributed training. Namely, we will average the gradients across a set of
+hosts before applying them to the model.
+
+Our training loop looks like the following code snippet if we omit the model,
+dataset and optimizer initialization.
+
+.. code:: python
+
+    model = ...
+    optimizer = ...
+    dataset = ...
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        optimizer.update(model, grads)
+        return loss
+
+    for x, y in dataset:
+        loss = step(model, x, y)
+        mx.eval(loss, model.parameters())
+
+All we have to do to average the gradients across machines is perform an
+:func:`all_sum` and divide by the size of the :class:`Group`. Namely we
+have to :func:`mlx.utils.tree_map` the gradients with following function.
+
+.. code:: python
+
+    def all_avg(x):
+        return mx.distributed.all_sum(x) / mx.distributed.init().size()
+
+Putting everything together our training loop step looks as follows with
+everything else remaining the same.
+
+.. code:: python
+
+    from mlx.utils import tree_map
+
+    def all_reduce_grads(grads):
+        N = mx.distributed.init().size()
+        if N == 1:
+            return grads
+        return tree_map(
+            lambda x: mx.distributed.all_sum(x) / N,
+            grads
+        )
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        grads = all_reduce_grads(grads)  # <--- This line was added
+        optimizer.update(model, grads)
+        return loss
+
+Tuning All Reduce
+-----------------
+
+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:
+
+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

@@ -0,0 +1,288 @@
+.. _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

@@ -25,7 +25,7 @@ Here is a simple example:
 
 The output of :func:`grad` on :func:`sin` is simply another function. In this
 case it is the gradient of the sine function which is exactly the cosine
-function. To get the second derivative you can do: 
+function. To get the second derivative you can do:
 
 .. code-block:: shell
 
@@ -50,7 +50,7 @@ Automatic Differentiation
 .. _auto diff:
 
 Automatic differentiation in MLX works on functions rather than on implicit
-graphs. 
+graphs.
 
 .. note::
 
@@ -114,7 +114,7 @@ way to do that is the following:
 
    def loss_fn(params, x, y):
       w, b = params["weight"], params["bias"]
-      h = w * x + b 
+      h = w * x + b
       return mx.mean(mx.square(h - y))
 
    params = {"weight": mx.array(1.0), "bias": mx.array(0.0)}
@@ -132,7 +132,7 @@ way to do that is the following:
 
 Notice the tree structure of the parameters is preserved in the gradients.
 
-In some cases you may want to stop gradients from propagating through a 
+In some cases you may want to stop gradients from propagating through a
 part of the function. You can use the :func:`stop_gradient` for that.
 
 
@@ -161,19 +161,19 @@ A naive way to add the elements from two sets of vectors is with a loop:
   ys = mx.random.uniform(shape=(100, 4096))
 
   def naive_add(xs, ys):
-      return [xs[i] + ys[:, i] for i in range(xs.shape[1])]
+      return [xs[i] + ys[:, i] for i in range(xs.shape[0])]
 
 Instead you can use :func:`vmap` to automatically vectorize the addition:
 
 .. code-block:: python
-   
+
    # Vectorize over the second dimension of x and the
    # first dimension of y
-   vmap_add = mx.vmap(lambda x, y: x + y, in_axes=(1, 0))
+   vmap_add = mx.vmap(lambda x, y: x + y, in_axes=(0, 1))
 
 The ``in_axes`` parameter can be used to specify which dimensions of the
 corresponding input to vectorize over. Similarly, use ``out_axes`` to specify
-where the vectorized axes should be in the outputs. 
+where the vectorized axes should be in the outputs.
 
 Let's time these two different versions:
 
@@ -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 ``0.390`` seconds whereas the
-vectorized version takes only ``0.025`` seconds, more than ten times faster.
+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.
 
 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

@@ -51,7 +51,7 @@ You can also use an :obj:`array` to index another :obj:`array`:
 .. code-block:: shell
 
   >>> arr = mx.arange(10)
-  >>> idx = mx.array([5, 7]) 
+  >>> idx = mx.array([5, 7])
   >>> arr[idx]
   array([5, 7], dtype=int32)
 
@@ -77,12 +77,12 @@ from the GPU. Performing bounds checking for array indices before launching the
 kernel would be extremely inefficient.
 
 Indexing with boolean masks is something that MLX may support in the future. In
-general, MLX has limited support for operations for which outputs
+general, MLX has limited support for operations for which output
 *shapes* are dependent on input *data*. Other examples of these types of
 operations which MLX does not yet support include :func:`numpy.nonzero` and the
 single input version of :func:`numpy.where`.
 
-In Place Updates 
+In Place Updates
 ----------------
 
 In place updates to indexed arrays are possible in MLX. For example:

docs/src/usage/lazy_evaluation.rst

@@ -13,7 +13,7 @@ compute graph is recorded. The actual computation only happens if an
 :func:`eval` is performed.
 
 MLX uses lazy evaluation because it has some nice features, some of which we
-describe below. 
+describe below.
 
 Transforming Compute Graphs
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -109,14 +109,14 @@ Here is a concrete example:
 
 An important behavior to be aware of is when the graph will be implicitly
 evaluated. Anytime you ``print`` an array, convert it to an
-:obj:`numpy.ndarray`, or otherwise access it's memory via :obj:`memoryview`,
+:obj:`numpy.ndarray`, or otherwise access its memory via :obj:`memoryview`,
 the graph will be evaluated. Saving arrays via :func:`save` (or any other MLX
 saving functions) will also evaluate the array.
 
 
 Calling :func:`array.item` on a scalar array will also evaluate it. In the
 example above, printing the loss (``print(loss)``) or adding the loss scalar to
-a list (``losses.append(loss.item())``) would cause a graph evaluation. If 
+a list (``losses.append(loss.item())``) would cause a graph evaluation. If
 these lines are before ``mx.eval(loss, model.parameters())`` then this
 will be a partial evaluation, computing only the forward pass.
 

docs/src/usage/numpy.rst

@@ -3,7 +3,11 @@
 Conversion to NumPy and Other Frameworks
 ========================================
 
-MLX array implements the `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_.
+MLX array supports conversion between other frameworks with either:
+
+* The `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_.
+* `DLPack <https://dmlc.github.io/dlpack/latest/>`_.
+
 Let's convert an array to NumPy and back.
 
 .. code-block:: python
@@ -17,11 +21,13 @@ 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
 
@@ -31,10 +37,16 @@ By default, NumPy copies data to a new array. This can be prevented by creating
   a_view[0] = 1
   print(a[0].item()) # 1
 
-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.
+.. note::
 
-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.
+    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.
 
 Let's demonstrate this in an example:
 
@@ -52,22 +64,24 @@ 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
 -------
 
-.. warning:: 
+.. warning::
 
    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
 
@@ -78,7 +92,8 @@ PyTorch supports the buffer protocol, but it requires an explicit :obj:`memoryvi
   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
 ---
@@ -96,7 +111,8 @@ 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
 

docs/src/usage/quick_start.rst

@@ -64,4 +64,4 @@ Other gradient transformations include :func:`vjp` for vector-Jacobian products
 and :func:`jvp` for Jacobian-vector products.
 
 Use :func:`value_and_grad` to efficiently compute both a function's output and
-gradient with respect to the function's input. 
+gradient with respect to the function's input.

docs/src/usage/saving_and_loading.rst

@@ -8,33 +8,33 @@ Saving and Loading Arrays
 MLX supports multiple array serialization formats.
 
 .. list-table:: Serialization Formats
-   :widths: 20 8 25 25 
+   :widths: 20 8 25 25
    :header-rows: 1
 
-   * - Format 
-     - Extension 
+   * - Format
+     - Extension
      - Function
-     - Notes 
-   * - NumPy 
-     - ``.npy`` 
+     - Notes
+   * - NumPy
+     - ``.npy``
      - :func:`save`
      - Single arrays only
-   * - NumPy archive 
-     - ``.npz`` 
+   * - NumPy archive
+     - ``.npz``
      - :func:`savez` and :func:`savez_compressed`
-     - Multiple arrays 
+     - Multiple arrays
    * - Safetensors
-     - ``.safetensors`` 
+     - ``.safetensors``
      - :func:`save_safetensors`
-     - Multiple arrays 
-   * - GGUF 
-     - ``.gguf`` 
+     - Multiple arrays
+   * - GGUF
+     - ``.gguf``
      - :func:`save_gguf`
      - Multiple arrays
 
 The :func:`load` function will load any of the supported serialization
 formats. It determines the format from the extensions. The output of
-:func:`load` depends on the format. 
+:func:`load` depends on the format.
 
 Here's an example of saving a single array to a file:
 

docs/src/usage/unified_memory.rst

@@ -20,7 +20,7 @@ Both ``a`` and ``b`` live in unified memory.
 
 In MLX, rather than moving arrays to devices, you specify the device when you
 run the operation. Any device can perform any operation on ``a`` and ``b``
-without needing to move them from one memory location to another. For example: 
+without needing to move them from one memory location to another. For example:
 
 .. code-block:: python
 

examples/cmake_project/CMakeLists.txt

@@ -0,0 +1,22 @@
+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

@@ -0,0 +1,26 @@
+## 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

@@ -0,0 +1,14 @@
+// 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/CMakeLists.txt

@@ -9,3 +9,4 @@ build_example(tutorial.cpp)
 build_example(linear_regression.cpp)
 build_example(logistic_regression.cpp)
 build_example(metal_capture.cpp)
+build_example(distributed.cpp)

examples/cpp/distributed.cpp

@@ -0,0 +1,22 @@
+// Copyright © 2024 Apple Inc.
+
+#include <iostream>
+
+#include "mlx/mlx.h"
+
+namespace mx = mlx::core;
+
+int main() {
+  if (!mx::distributed::is_available()) {
+    std::cout << "No communication backend found" << std::endl;
+    return 1;
+  }
+
+  auto global_group = mx::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);
+
+  std::cout << out << std::endl;
+}

examples/cpp/linear_regression.cpp

@@ -10,7 +10,7 @@
 /**
  * An example of linear regression with MLX.
  */
-using namespace mlx::core;
+namespace mx = mlx::core;
 
 int main() {
   int num_features = 100;
@@ -19,35 +19,35 @@ int main() {
   float learning_rate = 0.01;
 
   // True parameters
-  auto w_star = random::normal({num_features});
+  auto w_star = mx::random::normal({num_features});
 
   // The input examples (design matrix)
-  auto X = random::normal({num_examples, num_features});
+  auto X = mx::random::normal({num_examples, num_features});
 
   // Noisy labels
-  auto eps = 1e-2 * random::normal({num_examples});
-  auto y = matmul(X, w_star) + eps;
+  auto eps = 1e-2 * mx::random::normal({num_examples});
+  auto y = mx::matmul(X, w_star) + eps;
 
   // Initialize random parameters
-  array w = 1e-2 * random::normal({num_features});
+  mx::array w = 1e-2 * mx::random::normal({num_features});
 
-  auto loss_fn = [&](array w) {
-    auto yhat = matmul(X, w);
-    return (0.5f / num_examples) * sum(square(yhat - y));
+  auto loss_fn = [&](mx::array w) {
+    auto yhat = mx::matmul(X, w);
+    return (0.5f / num_examples) * mx::sum(mx::square(yhat - y));
   };
 
-  auto grad_fn = grad(loss_fn);
+  auto grad_fn = mx::grad(loss_fn);
 
   auto tic = timer::time();
   for (int it = 0; it < num_iters; ++it) {
-    auto grad = grad_fn(w);
-    w = w - learning_rate * grad;
-    eval(w);
+    auto grads = grad_fn(w);
+    w = w - learning_rate * grads;
+    mx::eval(w);
   }
   auto toc = timer::time();
 
   auto loss = loss_fn(w);
-  auto error_norm = std::sqrt(sum(square(w - w_star)).item<float>());
+  auto error_norm = std::sqrt(mx::sum(mx::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.
  */
-using namespace mlx::core;
+namespace mx = mlx::core;
 
 int main() {
   int num_features = 100;
@@ -19,35 +19,35 @@ int main() {
   float learning_rate = 0.1;
 
   // True parameters
-  auto w_star = random::normal({num_features});
+  auto w_star = mx::random::normal({num_features});
 
   // The input examples
-  auto X = random::normal({num_examples, num_features});
+  auto X = mx::random::normal({num_examples, num_features});
 
   // Labels
-  auto y = matmul(X, w_star) > 0;
+  auto y = mx::matmul(X, w_star) > 0;
 
   // Initialize random parameters
-  array w = 1e-2 * random::normal({num_features});
+  mx::array w = 1e-2 * mx::random::normal({num_features});
 
-  auto loss_fn = [&](array w) {
-    auto logits = matmul(X, w);
+  auto loss_fn = [&](mx::array w) {
+    auto logits = mx::matmul(X, w);
     auto scale = (1.0f / num_examples);
-    return scale * sum(logaddexp(array(0.0f), logits) - y * logits);
+    return scale * mx::sum(mx::logaddexp(mx::array(0.0f), logits) - y * logits);
   };
 
-  auto grad_fn = grad(loss_fn);
+  auto grad_fn = mx::grad(loss_fn);
 
   auto tic = timer::time();
   for (int it = 0; it < num_iters; ++it) {
-    auto grad = grad_fn(w);
-    w = w - learning_rate * grad;
-    eval(w);
+    auto grads = grad_fn(w);
+    w = w - learning_rate * grads;
+    mx::eval(w);
   }
   auto toc = timer::time();
 
   auto loss = loss_fn(w);
-  auto acc = sum((matmul(X, w) > 0) == y) / num_examples;
+  auto acc = mx::sum((mx::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"
 
-using namespace mlx::core;
+namespace mx = 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.
-  metal::start_capture("mlx_trace.gputrace");
+  mx::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(Device::gpu);
-  auto s3 = new_stream(Device::gpu);
+  auto s2 = new_stream(mx::Device::gpu);
+  auto s3 = new_stream(mx::Device::gpu);
 
-  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);
+  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);
 
   // The multiply will happen on the default stream.
-  std::cout << multiply(x, y) << std::endl;
+  std::cout << mx::multiply(x, y) << std::endl;
 
-  metal::stop_capture();
+  mx::metal::stop_capture();
 }

examples/cpp/tutorial.cpp

@@ -5,11 +5,11 @@
 
 #include "mlx/mlx.h"
 
-using namespace mlx::core;
+namespace mx = mlx::core;
 
 void array_basics() {
   // Make a scalar array:
-  array x(1.0);
+  mx::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 == float32);
+  assert(dtype == mx::float32);
 
   // Specify the dtype when constructing the array:
-  x = array(1, int32);
-  assert(x.dtype() == int32);
+  x = mx::array(1, mx::int32);
+  assert(x.dtype() == mx::int32);
   x.item<int>(); // OK
   // x.item<float>();  // Undefined!
 
   // Make a multidimensional array:
-  x = array({1.0f, 2.0f, 3.0f, 4.0f}, {2, 2});
+  x = mx::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 = ones({2, 2});
+  auto y = mx::ones({2, 2});
 
   // Pointwise add x and y:
-  auto z = add(x, y);
+  auto z = mx::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() == float32);
+  assert(z.dtype() == mx::float32);
   assert(z.shape(0) == 2);
   assert(z.shape(1) == 2);
 
@@ -63,34 +63,34 @@ 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.
-  eval(z);
+  mx::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:
-  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:
+  mx::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 = ones({1});
+  z = mx::ones({1});
   z.item<float>(); // implicit evaluation
 
-  z = ones({2, 2});
+  z = mx::ones({2, 2});
   std::cout << z << std::endl; // implicit evaluation
 }
 
 void automatic_differentiation() {
-  auto fn = [](array x) { return square(x); };
+  auto fn = [](mx::array x) { return mx::square(x); };
 
   // Computing the derivative function of a function
-  auto grad_fn = grad(fn);
+  auto grad_fn = mx::grad(fn);
   // Call grad_fn on the input to get the derivative
-  auto x = array(1.5);
+  auto x = mx::array(1.5);
   auto dfdx = grad_fn(x);
   // dfdx is 2 * x
 
   // Get the second derivative by composing grad with grad
-  auto df2dx2 = grad(grad(fn))(x);
-  // df2dx2 is 2
+  auto d2fdx2 = mx::grad(mx::grad(fn))(x);
+  // d2fdx2 is 2
 }
 
 int main() {

examples/export/CMakeLists.txt

@@ -0,0 +1,22 @@
+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

@@ -0,0 +1,49 @@
+## 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

@@ -0,0 +1,25 @@
+// 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

@@ -0,0 +1,52 @@
+# 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

@@ -0,0 +1,35 @@
+// 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

@@ -0,0 +1,76 @@
+# 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

@@ -11,11 +11,14 @@ 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 REQUIRED)
+find_package(
+  Python 3.8
+  COMPONENTS Interpreter Development.Module
+  REQUIRED)
 execute_process(
   COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
-  OUTPUT_STRIP_TRAILING_WHITESPACE OUTPUT_VARIABLE NB_DIR)
-list(APPEND CMAKE_PREFIX_PATH "${NB_DIR}")
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+  OUTPUT_VARIABLE nanobind_ROOT)
 find_package(nanobind CONFIG REQUIRED)
 
 # ----------------------------- Extensions -----------------------------
@@ -24,16 +27,10 @@ find_package(nanobind CONFIG REQUIRED)
 add_library(mlx_ext)
 
 # Add sources
-target_sources(
-  mlx_ext
-  PUBLIC
-  ${CMAKE_CURRENT_LIST_DIR}/axpby/axpby.cpp
-)
+target_sources(mlx_ext PUBLIC ${CMAKE_CURRENT_LIST_DIR}/axpby/axpby.cpp)
 
 # Add include headers
-target_include_directories(
-  mlx_ext PUBLIC ${CMAKE_CURRENT_LIST_DIR}
-)
+target_include_directories(mlx_ext PUBLIC ${CMAKE_CURRENT_LIST_DIR})
 
 # Link to mlx
 target_link_libraries(mlx_ext PUBLIC mlx)
@@ -43,27 +40,32 @@ target_link_libraries(mlx_ext PUBLIC mlx)
 # Build metallib
 if(MLX_BUILD_METAL)
   mlx_build_metallib(
-    TARGET mlx_ext_metallib
-    TITLE mlx_ext
-    SOURCES ${CMAKE_CURRENT_LIST_DIR}/axpby/axpby.metal
-    INCLUDE_DIRS ${PROJECT_SOURCE_DIR} ${MLX_INCLUDE_DIRS}
-    OUTPUT_DIRECTORY ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}
-  )
-
-  add_dependencies(
-    mlx_ext
+    TARGET
     mlx_ext_metallib
-  )
+    TITLE
+    mlx_ext
+    SOURCES
+    ${CMAKE_CURRENT_LIST_DIR}/axpby/axpby.metal
+    INCLUDE_DIRS
+    ${PROJECT_SOURCE_DIR}
+    ${MLX_INCLUDE_DIRS}
+    OUTPUT_DIRECTORY
+    ${CMAKE_LIBRARY_OUTPUT_DIRECTORY})
+
+  add_dependencies(mlx_ext mlx_ext_metallib)
 
 endif()
 
 # ----------------------------- Python Bindings -----------------------------
 nanobind_add_module(
   _ext
-  NB_STATIC STABLE_ABI LTO NOMINSIZE
-  NB_DOMAIN mlx 
-  ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
-)
+  NB_STATIC
+  STABLE_ABI
+  LTO
+  NOMINSIZE
+  NB_DOMAIN
+  mlx
+  ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp)
 target_link_libraries(_ext PRIVATE mlx_ext)
 
 if(BUILD_SHARED_LIBS)

examples/extensions/README.md

@@ -1,5 +1,5 @@
 
-## Build the extensions
+## Build
 
 ```
 pip install -e .
@@ -16,3 +16,9 @@ And then run:
 ```
 python setup.py build_ext -j8 --inplace
 ```
+
+## Test
+
+```
+python test.py
+```

examples/extensions/axpby/axpby.cpp

@@ -6,6 +6,7 @@
 
 #include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/copy.h"
 #include "mlx/utils.h"
 
 #include "axpby/axpby.h"
@@ -19,7 +20,7 @@
 #include "mlx/backend/metal/utils.h"
 #endif
 
-namespace mlx::core {
+namespace my_ext {
 
 ///////////////////////////////////////////////////////////////////////////////
 // Operation Implementation
@@ -32,24 +33,24 @@ namespace mlx::core {
  *  Follow numpy style broadcasting between x and y
  *  Inputs are upcasted to floats if needed
  **/
-array axpby(
-    const array& x, // Input array x
-    const array& y, // Input array y
+mx::array axpby(
+    const mx::array& x, // Input mx::array x
+    const mx::array& y, // Input mx::array y
     const float alpha, // Scaling factor for x
     const float beta, // Scaling factor for y
-    StreamOrDevice s /* = {} */ // Stream on which to schedule the operation
+    mx::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 = issubdtype(promoted_dtype, float32)
+  auto out_dtype = mx::issubdtype(promoted_dtype, mx::float32)
       ? promoted_dtype
-      : promote_types(promoted_dtype, float32);
+      : promote_types(promoted_dtype, mx::float32);
 
   // Cast x and y up to the determined dtype (on the same stream s)
-  auto x_casted = astype(x, out_dtype, s);
-  auto y_casted = astype(y, out_dtype, s);
+  auto x_casted = mx::astype(x, out_dtype, s);
+  auto y_casted = mx::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);
@@ -57,12 +58,12 @@ array axpby(
 
   // Construct the array as the output of the Axpby primitive
   // with the broadcasted and upcasted arrays as inputs
-  return array(
-      /* const std::vector<int>& shape = */ out_shape,
-      /* Dtype dtype = */ out_dtype,
-      /* std::unique_ptr<Primitive> primitive = */
+  return mx::array(
+      /* const mx::Shape& shape = */ out_shape,
+      /* mx::Dtype dtype = */ out_dtype,
+      /* std::shared_ptr<mx::Primitive> primitive = */
       std::make_shared<Axpby>(to_stream(s), alpha, beta),
-      /* const std::vector<array>& inputs = */ broadcasted_inputs);
+      /* const std::vector<mx::array>& inputs = */ broadcasted_inputs);
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -71,16 +72,16 @@ array axpby(
 
 template <typename T>
 void axpby_impl(
-    const array& x,
-    const array& y,
-    array& out,
+    const mx::array& x,
+    const mx::array& y,
+    mx::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()));
+  out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
 
   // Collect input and output data pointers
   const T* x_ptr = x.data<T>();
@@ -94,8 +95,8 @@ void axpby_impl(
   // 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());
+    auto x_offset = mx::elem_to_loc(out_idx, x.shape(), x.strides());
+    auto y_offset = mx::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
@@ -105,8 +106,8 @@ void axpby_impl(
 
 /** Fall back implementation for evaluation on CPU */
 void Axpby::eval(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
+    const std::vector<mx::array>& inputs,
+    std::vector<mx::array>& outputs) {
   // Check the inputs (registered in the op while constructing the out array)
   assert(inputs.size() == 2);
   auto& x = inputs[0];
@@ -114,14 +115,14 @@ void Axpby::eval(
   auto& out = outputs[0];
 
   // Dispatch to the correct dtype
-  if (out.dtype() == float32) {
+  if (out.dtype() == mx::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 if (out.dtype() == mx::float16) {
+    return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_);
+  } else if (out.dtype() == mx::bfloat16) {
+    return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_);
+  } else if (out.dtype() == mx::complex64) {
+    return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_);
   } else {
     throw std::runtime_error(
         "Axpby is only supported for floating point types.");
@@ -136,9 +137,9 @@ void Axpby::eval(
 
 template <typename T>
 void axpby_impl_accelerate(
-    const array& x,
-    const array& y,
-    array& out,
+    const mx::array& x,
+    const mx::array& y,
+    mx::array& out,
     float alpha_,
     float beta_) {
   // Accelerate library provides catlas_saxpby which does
@@ -150,10 +151,10 @@ void axpby_impl_accelerate(
   // 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()));
+  out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
 
   // We then copy over the elements using the contiguous vector specialization
-  copy_inplace(y, out, CopyType::Vector);
+  copy_inplace(y, out, mx::CopyType::Vector);
 
   // Get x and y pointers for catlas_saxpby
   const T* x_ptr = x.data<T>();
@@ -175,15 +176,15 @@ void axpby_impl_accelerate(
 
 /** Evaluate primitive on CPU using accelerate specializations */
 void Axpby::eval_cpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
+    const std::vector<mx::array>& inputs,
+    std::vector<mx::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 &&
+  if (out.dtype() == mx::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_);
@@ -198,8 +199,8 @@ void Axpby::eval_cpu(
 
 /** Evaluate primitive on CPU falling back to common backend */
 void Axpby::eval_cpu(
-    const std::vector<array>& inputs,
-    const std::vector<array>& outputs) {
+    const std::vector<mx::array>& inputs,
+    std::vector<mx::array>& outputs) {
   eval(inputs, outputs);
 }
 
@@ -213,8 +214,8 @@ void Axpby::eval_cpu(
 
 /** Evaluate primitive on GPU */
 void Axpby::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
+    const std::vector<mx::array>& inputs,
+    std::vector<mx::array>& outputs) {
   // Prepare inputs
   assert(inputs.size() == 2);
   auto& x = inputs[0];
@@ -225,7 +226,7 @@ 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 = metal::device(s.device);
+  auto& d = mx::metal::device(s.device);
 
   // Prepare to specialize based on contiguity
   bool contiguous_kernel =
@@ -235,12 +236,12 @@ void Axpby::eval_gpu(
   // Allocate output memory with strides based on specialization
   if (contiguous_kernel) {
     out.set_data(
-        allocator::malloc_or_wait(x.data_size() * out.itemsize()),
+        mx::allocator::malloc_or_wait(x.data_size() * out.itemsize()),
         x.data_size(),
         x.strides(),
         x.flags());
   } else {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
   }
 
   // Resolve name of kernel (corresponds to axpby.metal)
@@ -249,16 +250,15 @@ void Axpby::eval_gpu(
   kname << (contiguous_kernel ? "contiguous_" : "general_");
   kname << type_to_name(out);
 
-  // Make sure the metal library is available and look for it
-  // in the same folder as this executable if needed
-  d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
+  // 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(), "mlx_ext");
 
   // Prepare to encode kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
-  compute_encoder->setComputePipelineState(kernel);
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder.set_compute_pipeline_state(kernel);
 
   // Kernel parameters are registered with buffer indices corresponding to
   // those in the kernel declaration at axpby.metal
@@ -266,22 +266,22 @@ void Axpby::eval_gpu(
   size_t nelem = out.size();
 
   // Encode input arrays to kernel
-  set_array_buffer(compute_encoder, x, 0);
-  set_array_buffer(compute_encoder, y, 1);
+  compute_encoder.set_input_array(x, 0);
+  compute_encoder.set_input_array(y, 1);
 
   // Encode output arrays to kernel
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode alpha and beta
-  compute_encoder->setBytes(&alpha_, sizeof(float), 3);
-  compute_encoder->setBytes(&beta_, sizeof(float), 4);
+  compute_encoder.set_bytes(alpha_, 3);
+  compute_encoder.set_bytes(beta_, 4);
 
   // Encode shape, strides and ndim if needed
   if (!contiguous_kernel) {
-    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);
+    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);
   }
 
   // We launch 1 thread for each input and make sure that the number of
@@ -296,15 +296,15 @@ void Axpby::eval_gpu(
 
   // Launch the grid with the given number of threads divided among
   // the given threadgroups
-  compute_encoder->dispatchThreads(grid_dims, group_dims);
+  compute_encoder.dispatch_threads(grid_dims, group_dims);
 }
 
 #else // Metal is not available
 
 /** Fail evaluation on GPU */
 void Axpby::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& out) {
+    const std::vector<mx::array>& inputs,
+    std::vector<mx::array>& out) {
   throw std::runtime_error("Axpby has no GPU implementation.");
 }
 
@@ -315,9 +315,9 @@ void Axpby::eval_gpu(
 ///////////////////////////////////////////////////////////////////////////////
 
 /** The Jacobian-vector product. */
-std::vector<array> Axpby::jvp(
-    const std::vector<array>& primals,
-    const std::vector<array>& tangents,
+std::vector<mx::array> Axpby::jvp(
+    const std::vector<mx::array>& primals,
+    const std::vector<mx::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
@@ -329,8 +329,8 @@ std::vector<array> Axpby::jvp(
   // scaled by beta
   if (argnums.size() > 1) {
     auto scale = argnums[0] == 0 ? alpha_ : beta_;
-    auto scale_arr = array(scale, tangents[0].dtype());
-    return {multiply(scale_arr, tangents[0], stream())};
+    auto scale_arr = mx::array(scale, tangents[0].dtype());
+    return {mx::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
@@ -340,24 +340,24 @@ std::vector<array> Axpby::jvp(
 }
 
 /** The vector-Jacobian product. */
-std::vector<array> Axpby::vjp(
-    const std::vector<array>& primals,
-    const std::vector<array>& cotangents,
+std::vector<mx::array> Axpby::vjp(
+    const std::vector<mx::array>& primals,
+    const std::vector<mx::array>& cotangents,
     const std::vector<int>& argnums,
-    const std::vector<array>&) {
+    const std::vector<mx::array>&) {
   // Reverse mode diff
-  std::vector<array> vjps;
+  std::vector<mx::array> vjps;
   for (auto arg : argnums) {
     auto scale = arg == 0 ? alpha_ : beta_;
-    auto scale_arr = array(scale, cotangents[0].dtype());
-    vjps.push_back(multiply(scale_arr, cotangents[0], stream()));
+    auto scale_arr = mx::array(scale, cotangents[0].dtype());
+    vjps.push_back(mx::multiply(scale_arr, cotangents[0], stream()));
   }
   return vjps;
 }
 
 /** Vectorize primitive along given axis */
-std::pair<std::vector<array>, std::vector<int>> Axpby::vmap(
-    const std::vector<array>& inputs,
+std::pair<std::vector<mx::array>, std::vector<int>> Axpby::vmap(
+    const std::vector<mx::array>& inputs,
     const std::vector<int>& axes) {
   throw std::runtime_error("Axpby has no vmap implementation.");
 }
@@ -368,4 +368,4 @@ bool Axpby::is_equivalent(const Primitive& other) const {
   return alpha_ == r_other.alpha_ && beta_ == r_other.beta_;
 }
 
-} // namespace mlx::core
+} // namespace my_ext

examples/extensions/axpby/axpby.h

@@ -5,7 +5,9 @@
 #include "mlx/ops.h"
 #include "mlx/primitives.h"
 
-namespace mlx::core {
+namespace mx = mlx::core;
+
+namespace my_ext {
 
 ///////////////////////////////////////////////////////////////////////////////
 // Operation
@@ -18,22 +20,22 @@ namespace mlx::core {
  *  Follow numpy style broadcasting between x and y
  *  Inputs are upcasted to floats if needed
  **/
-array axpby(
-    const array& x, // Input array x
-    const array& y, // Input array y
+mx::array axpby(
+    const mx::array& x, // Input array x
+    const mx::array& y, // Input array y
     const float alpha, // Scaling factor for x
     const float beta, // Scaling factor for y
-    StreamOrDevice s = {} // Stream on which to schedule the operation
+    mx::StreamOrDevice s = {} // Stream on which to schedule the operation
 );
 
 ///////////////////////////////////////////////////////////////////////////////
 // Primitive
 ///////////////////////////////////////////////////////////////////////////////
 
-class Axpby : public Primitive {
+class Axpby : public mx::Primitive {
  public:
-  explicit Axpby(Stream stream, float alpha, float beta)
-      : Primitive(stream), alpha_(alpha), beta_(beta) {};
+  explicit Axpby(mx::Stream stream, float alpha, float beta)
+      : mx::Primitive(stream), alpha_(alpha), beta_(beta) {};
 
   /**
    * A primitive must know how to evaluate itself on the CPU/GPU
@@ -42,23 +44,25 @@ class Axpby : public Primitive {
    * To avoid unnecessary allocations, the evaluation function
    * is responsible for allocating space for the array.
    */
-  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;
+  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;
 
   /** The Jacobian-vector product. */
-  std::vector<array> jvp(
-      const std::vector<array>& primals,
-      const std::vector<array>& tangents,
+  std::vector<mx::array> jvp(
+      const std::vector<mx::array>& primals,
+      const std::vector<mx::array>& tangents,
       const std::vector<int>& argnums) override;
 
   /** The vector-Jacobian product. */
-  std::vector<array> vjp(
-      const std::vector<array>& primals,
-      const std::vector<array>& cotangents,
+  std::vector<mx::array> vjp(
+      const std::vector<mx::array>& primals,
+      const std::vector<mx::array>& cotangents,
       const std::vector<int>& argnums,
-      const std::vector<array>& outputs) override;
+      const std::vector<mx::array>& outputs) override;
 
   /**
    * The primitive must know how to vectorize itself across
@@ -66,8 +70,8 @@ class Axpby : public Primitive {
    * representing the vectorized computation and the axis which
    * corresponds to the output vectorized dimension.
    */
-  std::pair<std::vector<array>, std::vector<int>> vmap(
-      const std::vector<array>& inputs,
+  std::pair<std::vector<mx::array>, std::vector<int>> vmap(
+      const std::vector<mx::array>& inputs,
       const std::vector<int>& axes) override;
 
   /** Print the primitive. */
@@ -76,14 +80,16 @@ class Axpby : public Primitive {
   }
 
   /** Equivalence check **/
-  bool is_equivalent(const Primitive& other) const override;
+  bool is_equivalent(const mx::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);
+  void eval(
+      const std::vector<mx::array>& inputs,
+      std::vector<mx::array>& outputs);
 };
 
-} // namespace mlx::core
+} // namespace my_ext

examples/extensions/axpby/axpby.metal

@@ -2,7 +2,6 @@
 
 #include <metal_stdlib>
 
-#include "mlx/backend/metal/kernels/bf16.h"
 #include "mlx/backend/metal/kernels/utils.h"
 
 template <typename T>
@@ -13,8 +12,8 @@ template <typename T>
     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 int64_t* x_strides [[buffer(6)]],
+    constant const int64_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);
@@ -35,29 +34,14 @@ template <typename T>
       static_cast<T>(alpha) * x[index] + static_cast<T>(beta) * y[index];
 }
 
-#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]]);
+// 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)
 
 instantiate_axpby(float32, float);
 instantiate_axpby(float16, half);
 instantiate_axpby(bfloat16, bfloat16_t);
-instantiate_axpby(complex64, complex64_t);
+instantiate_axpby(complex64, complex64_t);
+// clang-format on

examples/extensions/bindings.cpp

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

examples/extensions/mlx_sample_extensions/__init__.py

@@ -2,4 +2,4 @@
 
 import mlx.core as mx
 
-from .mlx_sample_extensions import *
+from ._ext import axpby

examples/extensions/pyproject.toml

@@ -1,8 +1,8 @@
 [build-system]
 requires = [
   "setuptools>=42",
-  "cmake>=3.24",
-  "mlx>=0.9.0",
-  "nanobind@git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4",
+  "cmake>=3.25",
+  "mlx>=0.18.0",
+  "nanobind==2.4.0",
 ]
 build-backend = "setuptools.build_meta"

examples/extensions/requirements.txt

@@ -1,4 +1,4 @@
 setuptools>=42
-cmake>=3.24
-mlx>=0.9.0
-nanobind@git+https://github.com/wjakob/nanobind.git#egg=4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+cmake>=3.25
+mlx>=0.21.0
+nanobind==2.2.0

examples/extensions/setup.py

@@ -13,7 +13,6 @@ if __name__ == "__main__":
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-        extras_require={"dev": []},
         zip_safe=False,
         python_requires=">=3.8",
     )

examples/extensions/test.py

@@ -0,0 +1,10 @@
+import mlx.core as mx
+from mlx_sample_extensions import axpby
+
+a = mx.ones((3, 4))
+b = mx.ones((3, 4))
+c = axpby(a, b, 4.0, 2.0, stream=mx.cpu)
+
+print(f"c shape: {c.shape}")
+print(f"c dtype: {c.dtype}")
+print(f"c correct: {mx.all(c == 6.0).item()}")

mlx.pc.in

@@ -28,10 +28,19 @@ endif()
 if (@MLX_BUILD_METAL@)
     set(MLX_BUILD_METAL @MLX_BUILD_METAL@)
     set(MLX_CXX_FLAGS ${MLX_CXX_FLAGS} -D_METAL_)
-    set_and_check(MLX_INCLUDE_DIRS 
-        ${MLX_INCLUDE_DIRS} 
+    set(MLX_INCLUDE_DIRS 
+        "${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
@@ -40,4 +49,4 @@ set_target_properties(mlx PROPERTIES
 )
 
 include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(MLX DEFAULT_MSG MLX_LIBRARY MLX_INCLUDE_DIRS)
+find_package_handle_standard_args(MLX DEFAULT_MSG MLX_LIBRARY MLX_INCLUDE_DIRS)

mlx/CMakeLists.txt

@@ -1,37 +1,46 @@
 target_sources(
   mlx
-  PRIVATE
-  ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/array.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/compile.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/dtype.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/fast.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/ops.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/graph_utils.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/random.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/scheduler.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/transforms.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/linalg.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/metal.h
-)
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/array.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/compile.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/dtype.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
+          ${CMAKE_CURRENT_SOURCE_DIR}/ops.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/graph_utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/random.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/scheduler.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/transforms.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/linalg.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/metal.h)
 
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
-if (MLX_BUILD_ACCELERATE)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/accelerate)
-else()
-  target_sources(
-    mlx
-    PRIVATE
-    ${CMAKE_CURRENT_SOURCE_DIR}/backend/common/default_primitives.cpp
-  )
+if(MSVC)
+  # Disable some MSVC warnings to speed up compilation.
+  target_compile_options(mlx PUBLIC /wd4068 /wd4244 /wd4267 /wd4804)
 endif()
 
-if (MLX_BUILD_METAL)
+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)
+else()
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_cpu)
+endif()
+
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/distributed)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
+
+if(MLX_BUILD_METAL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/metal)
 else()
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_metal)

mlx/allocator.cpp

@@ -19,15 +19,26 @@ Buffer malloc(size_t size) {
 }
 
 void free(Buffer buffer) {
-  return allocator().free(buffer);
+  allocator().free(buffer);
 }
 
 Buffer CommonAllocator::malloc(size_t size, bool) {
-  return Buffer{std::malloc(size)};
+  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.raw_ptr());
+  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) {

mlx/allocator.h

@@ -41,6 +41,7 @@ class Allocator {
  public:
   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;
 
   Allocator() = default;
   Allocator(const Allocator& other) = delete;
@@ -57,6 +58,7 @@ class CommonAllocator : public 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;

mlx/array.cpp

@@ -1,5 +1,6 @@
 // Copyright © 2023-2024 Apple Inc.
 #include <functional>
+#include <unordered_map>
 
 #include "mlx/array.h"
 #include "mlx/ops.h"
@@ -9,24 +10,14 @@
 
 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();
-}
-
-} // namespace
-
 array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
-    : array_desc_(std::make_shared<ArrayDesc>(std::vector<int>{}, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(Shape{}, dtype)) {
   auto cval = static_cast<complex64_t>(val);
   init(&cval);
 }
 
 array::array(
-    std::vector<int> shape,
+    Shape shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
     std::vector<array> inputs)
@@ -34,10 +25,21 @@ array::array(
           std::move(shape),
           dtype,
           std::move(primitive),
-          std::move(inputs))) {}
+          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::vector<array> array::make_arrays(
-    std::vector<std::vector<int>> shapes,
+    std::vector<Shape> shapes,
     const std::vector<Dtype>& dtypes,
     const std::shared_ptr<Primitive>& primitive,
     const std::vector<array>& inputs) {
@@ -56,28 +58,36 @@ std::vector<array> array::make_arrays(
 
 array::array(std::initializer_list<float> data)
     : array_desc_(std::make_shared<ArrayDesc>(
-          std::vector<int>{static_cast<int>(data.size())},
+          Shape{static_cast<ShapeElem>(data.size())},
           float32)) {
   init(data.begin());
 }
 
 array::array(std::initializer_list<int> data, Dtype dtype)
     : array_desc_(std::make_shared<ArrayDesc>(
-          std::vector<int>{static_cast<int>(data.size())},
+          Shape{static_cast<ShapeElem>(data.size())},
           dtype)) {
   init(data.begin());
 }
 
 /* Build an array from a shared buffer */
-array::array(
-    allocator::Buffer data,
-    std::vector<int> shape,
-    Dtype dtype,
-    deleter_t deleter)
+array::array(allocator::Buffer data, Shape shape, Dtype dtype, Deleter deleter)
     : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   set_data(data, deleter);
 }
 
+array::array(
+    allocator::Buffer data,
+    Shape shape,
+    Dtype dtype,
+    Strides strides,
+    size_t data_size,
+    Flags flags,
+    Deleter deleter)
+    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
+  set_data(data, data_size, std::move(strides), flags, deleter);
+}
+
 void array::detach() {
   for (auto& s : array_desc_->siblings) {
     s.array_desc_->inputs.clear();
@@ -91,21 +101,38 @@ void array::detach() {
   array_desc_->primitive = nullptr;
 }
 
-void array::eval() {
-  // Ensure the array is ready to be read
-  if (status() == Status::scheduled) {
+bool array::is_available() const {
+  if (status() == Status::available) {
+    return true;
+  } else if (status() == Status::evaluated && event().is_signaled()) {
+    set_status(Status::available);
+    return true;
+  }
+  return false;
+}
+
+void array::wait() {
+  if (!is_available()) {
     event().wait();
     set_status(Status::available);
-  } else if (status() == Status::unscheduled) {
+  }
+}
+
+void array::eval() {
+  // Ensure the array is ready to be read
+  if (status() == Status::unscheduled) {
     mlx::core::eval({*this});
+  } else {
+    wait();
   }
 }
 
 bool array::is_tracer() const {
-  return array_desc_->is_tracer && in_tracing();
+  return (array_desc_->is_tracer && detail::in_tracing()) ||
+      detail::retain_graph();
 }
 
-void array::set_data(allocator::Buffer buffer, deleter_t d) {
+void array::set_data(allocator::Buffer buffer, Deleter d) {
   array_desc_->data = std::make_shared<Data>(buffer, d);
   array_desc_->data_ptr = buffer.raw_ptr();
   array_desc_->data_size = size();
@@ -118,9 +145,9 @@ void array::set_data(allocator::Buffer buffer, deleter_t d) {
 void array::set_data(
     allocator::Buffer buffer,
     size_t data_size,
-    std::vector<size_t> strides,
+    Strides strides,
     Flags flags,
-    deleter_t d) {
+    Deleter d) {
   array_desc_->data = std::make_shared<Data>(buffer, d);
   array_desc_->data_ptr = buffer.raw_ptr();
   array_desc_->data_size = data_size;
@@ -130,7 +157,7 @@ void array::set_data(
 
 void array::copy_shared_buffer(
     const array& other,
-    const std::vector<size_t>& strides,
+    const Strides& strides,
     Flags flags,
     size_t data_size,
     size_t offset /* = 0 */) {
@@ -149,7 +176,7 @@ void array::copy_shared_buffer(const array& other) {
 
 void array::move_shared_buffer(
     array other,
-    const std::vector<size_t>& strides,
+    const Strides& strides,
     Flags flags,
     size_t data_size,
     size_t offset /* = 0 */) {
@@ -158,8 +185,10 @@ void array::move_shared_buffer(
   array_desc_->flags = flags;
   array_desc_->data_size = data_size;
   auto char_offset = sizeof(char) * itemsize() * offset;
-  array_desc_->data_ptr = static_cast<void*>(
-      static_cast<char*>(other.array_desc_->data_ptr) + char_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) {
@@ -171,10 +200,11 @@ array::~array() {
     return;
   }
 
-  // Ignore arrays that will be detached
-  if (status() != array::Status::unscheduled) {
+  // Ignore arrays that might be detached during eval
+  if (status() == array::Status::scheduled) {
     return;
   }
+
   // Break circular reference for non-detached arrays with siblings
   if (auto n = siblings().size(); n > 0) {
     bool do_detach = true;
@@ -191,6 +221,8 @@ 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();
@@ -206,18 +238,18 @@ void array::ArrayDesc::init() {
     strides[i] = size;
     size *= shape[i];
   }
-  for (auto& in : inputs) {
+  for (const auto& in : inputs) {
     is_tracer |= in.is_tracer();
   }
 }
 
-array::ArrayDesc::ArrayDesc(std::vector<int> shape, Dtype dtype)
+array::ArrayDesc::ArrayDesc(Shape shape, Dtype dtype)
     : shape(std::move(shape)), dtype(dtype), status(Status::available) {
   init();
 }
 
 array::ArrayDesc::ArrayDesc(
-    std::vector<int> shape,
+    Shape shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
     std::vector<array> inputs)
@@ -231,31 +263,64 @@ array::ArrayDesc::ArrayDesc(
 
 array::ArrayDesc::~ArrayDesc() {
   // When an array description is destroyed it will delete a bunch of arrays
-  // that may also destory their corresponding descriptions and so on and so
+  // that may also destroy their corresponding descriptions and so on and so
   // forth.
   //
   // This calls recursively the destructor and can result in stack overflow, we
   // instead put them in a vector and destroy them one at a time resulting in a
   // max stack depth of 2.
+  if (inputs.empty()) {
+    return;
+  }
+
   std::vector<std::shared_ptr<ArrayDesc>> for_deletion;
 
-  for (array& a : inputs) {
-    if (a.array_desc_.use_count() == 1) {
-      for_deletion.push_back(std::move(a.array_desc_));
+  auto append_deletable_inputs = [&for_deletion](ArrayDesc& ad) {
+    std::unordered_map<std::uintptr_t, array> input_map;
+    for (array& a : ad.inputs) {
+      if (a.array_desc_) {
+        input_map.insert({a.id(), a});
+        for (auto& s : a.siblings()) {
+          input_map.insert({s.id(), s});
+        }
+      }
     }
-  }
+    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) {
+        for_deletion.push_back(std::move(a.array_desc_));
+      }
+    }
+  };
+
+  append_deletable_inputs(*this);
 
   while (!for_deletion.empty()) {
     // top is going to be deleted at the end of the block *after* the arrays
     // with inputs have been moved into the vector
     auto top = std::move(for_deletion.back());
     for_deletion.pop_back();
+    append_deletable_inputs(*top);
 
-    for (array& a : top->inputs) {
-      if (a.array_desc_.use_count() == 1) {
-        for_deletion.push_back(std::move(a.array_desc_));
-      }
+    // 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();
   }
 }
 
@@ -267,7 +332,7 @@ array::ArrayIterator::ArrayIterator(const array& arr, int idx)
 }
 
 array::ArrayIterator::reference array::ArrayIterator::operator*() const {
-  auto start = std::vector<int>(arr.ndim(), 0);
+  auto start = Shape(arr.ndim(), 0);
   auto end = arr.shape();
   auto shape = arr.shape();
   shape.erase(shape.begin());

mlx/array.h

@@ -15,7 +15,11 @@ namespace mlx::core {
 
 // Forward declaration
 class Primitive;
-using deleter_t = std::function<void(allocator::Buffer)>;
+
+using Deleter = std::function<void(allocator::Buffer)>;
+using ShapeElem = int32_t;
+using Shape = std::vector<ShapeElem>;
+using Strides = std::vector<int64_t>;
 
 class array {
   /* An array is really a node in a graph. It contains a shared ArrayDesc
@@ -31,33 +35,33 @@ class array {
   explicit array(const std::complex<float>& val, Dtype dtype = complex64);
 
   template <typename It>
-  array(
+  explicit array(
       It data,
-      std::vector<int> shape,
+      Shape shape,
       Dtype dtype =
           TypeToDtype<typename std::iterator_traits<It>::value_type>());
 
   template <typename T>
-  array(std::initializer_list<T> data, Dtype dtype = TypeToDtype<T>());
+  explicit array(std::initializer_list<T> data, Dtype dtype = TypeToDtype<T>());
 
   /* Special case so empty lists default to float32. */
-  array(std::initializer_list<float> data);
+  explicit array(std::initializer_list<float> data);
 
   /* Special case so array({}, type) is an empty array. */
-  array(std::initializer_list<int> data, Dtype dtype);
+  explicit array(std::initializer_list<int> data, Dtype dtype);
 
   template <typename T>
-  array(
+  explicit array(
       std::initializer_list<T> data,
-      std::vector<int> shape,
+      Shape shape,
       Dtype dtype = TypeToDtype<T>());
 
   /* Build an array from a buffer */
-  array(
+  explicit array(
       allocator::Buffer data,
-      std::vector<int> shape,
+      Shape shape,
       Dtype dtype,
-      deleter_t deleter = allocator::free);
+      Deleter deleter = allocator::free);
 
   /** Assignment to rvalue does not compile. */
   array& operator=(const array& other) && = delete;
@@ -73,60 +77,60 @@ class array {
       this->array_desc_ = other.array_desc_;
     }
     return *this;
-  };
+  }
 
   /** The size of the array's datatype in bytes. */
   size_t itemsize() const {
     return size_of(dtype());
-  };
+  }
 
   /** The number of elements in the array. */
   size_t size() const {
     return array_desc_->size;
-  };
+  }
 
   /** The number of bytes in the array. */
   size_t nbytes() const {
     return size() * itemsize();
-  };
+  }
 
   /** The number of dimensions of the array. */
   size_t ndim() const {
     return array_desc_->shape.size();
-  };
+  }
 
   /** The shape of the array as a vector of integers. */
-  const std::vector<int>& shape() const {
+  const Shape& shape() const {
     return array_desc_->shape;
-  };
+  }
 
   /**
    *  Get the size of the corresponding dimension.
    *
    *  This function supports negative indexing and provides
    *  bounds checking. */
-  int shape(int dim) const {
+  auto shape(int dim) const {
     return shape().at(dim < 0 ? dim + ndim() : dim);
-  };
+  }
 
   /** The strides of the array. */
-  const std::vector<size_t>& strides() const {
+  const Strides& strides() const {
     return array_desc_->strides;
-  };
+  }
 
   /**
    *  Get the stride of the corresponding dimension.
    *
    *  This function supports negative indexing and provides
    *  bounds checking. */
-  size_t strides(int dim) const {
+  auto strides(int dim) const {
     return strides().at(dim < 0 ? dim + ndim() : dim);
-  };
+  }
 
   /** Get the arrays data type. */
   Dtype dtype() const {
     return array_desc_->dtype;
-  };
+  }
 
   /** Evaluate the array. */
   void eval();
@@ -160,10 +164,10 @@ class array {
 
     friend bool operator==(const ArrayIterator& a, const ArrayIterator& b) {
       return a.arr.id() == b.arr.id() && a.idx == b.idx;
-    };
+    }
     friend bool operator!=(const ArrayIterator& a, const ArrayIterator& b) {
       return !(a == b);
-    };
+    }
 
    private:
     const array& arr;
@@ -184,13 +188,13 @@ class array {
    */
 
   array(
-      std::vector<int> shape,
+      Shape shape,
       Dtype dtype,
       std::shared_ptr<Primitive> primitive,
       std::vector<array> inputs);
 
   static std::vector<array> make_arrays(
-      std::vector<std::vector<int>> shapes,
+      std::vector<Shape> shapes,
       const std::vector<Dtype>& dtypes,
       const std::shared_ptr<Primitive>& primitive,
       const std::vector<array>& inputs);
@@ -207,9 +211,9 @@ class array {
 
   struct Data {
     allocator::Buffer buffer;
-    deleter_t d;
-    Data(allocator::Buffer buffer, deleter_t d = allocator::free)
-        : buffer(buffer), d(d) {};
+    Deleter d;
+    Data(allocator::Buffer buffer, Deleter d = allocator::free)
+        : buffer(buffer), d(d) {}
     // Not copyable
     Data(const Data& d) = delete;
     Data& operator=(const Data& d) = delete;
@@ -219,33 +223,57 @@ class array {
   };
 
   struct Flags {
-    // True if there are no gaps in the underlying data. Each item
+    // True iff there are no gaps in the underlying data. Each item
     // in the underlying data buffer belongs to at least one index.
+    //
+    // True iff:
+    // prod(shape[i] for i in range(ndim) if strides[i] > 0) == data_size()
     bool contiguous : 1;
 
+    // True iff:
+    // strides[-1] == 1 and
+    // all(strides[i] == (shape[i+1]*strides[i+1]) or shape[i] == 1 for i in
+    // range(ndim - 1))
     bool row_contiguous : 1;
+
+    // True iff:
+    // strides[0] == 1 and
+    // all(strides[i] == (shape[i-1]*strides[i-1]) or shape[i] == 1 for i in
+    // range(1, ndim))
     bool col_contiguous : 1;
   };
 
+  /** Build an array from all the info held by the array description. Including
+   * the buffer, strides, flags.
+   */
+  explicit array(
+      allocator::Buffer data,
+      Shape shape,
+      Dtype dtype,
+      Strides strides,
+      size_t data_size,
+      Flags flags,
+      Deleter deleter = allocator::free);
+
   /** The array's primitive. */
   Primitive& primitive() const {
     return *(array_desc_->primitive);
-  };
+  }
 
   /** A shared pointer to the array's primitive. */
   std::shared_ptr<Primitive>& primitive_ptr() const {
     return array_desc_->primitive;
-  };
+  }
 
   /** Check if the array has an attached primitive or is a leaf node. */
   bool has_primitive() const {
     return array_desc_->primitive != nullptr;
-  };
+  }
 
   /** The array's inputs. */
   const std::vector<array>& inputs() const {
     return array_desc_->inputs;
-  };
+  }
 
   std::vector<array>& inputs() {
     return array_desc_->inputs;
@@ -259,12 +287,12 @@ class array {
   /** The array's siblings. */
   const std::vector<array>& siblings() const {
     return array_desc_->siblings;
-  };
+  }
 
   /** The array's siblings. */
   std::vector<array>& siblings() {
     return array_desc_->siblings;
-  };
+  }
 
   void set_siblings(std::vector<array> siblings, uint16_t position) {
     array_desc_->siblings = std::move(siblings);
@@ -281,7 +309,7 @@ class array {
     outputs.push_back(*this);
     outputs.insert(outputs.end(), siblings().begin() + idx, siblings().end());
     return outputs;
-  };
+  }
 
   /** Detach the array from the graph. */
   void detach();
@@ -289,19 +317,32 @@ class array {
   /** Get the Flags bit-field. */
   const Flags& flags() const {
     return array_desc_->flags;
-  };
+  }
 
-  /** The size (in elements) of the underlying buffer the array points to. */
+  /** The size (in elements) of the underlying buffer the array points to.
+   *
+   * This can be different than the actual size of the array if the array has
+   * been broadcast or irregularly strided.  If ``first`` is the offset into
+   * the data buffer of the first element of the array (i.e. the offset
+   * corresponding to ``arr[0, 0, ...]``) and last is the offset into the
+   * data buffer of the last element of the array (i.e. the offset
+   * corresponding to ``arr[-1, -1, ...]``) then ``data_size = last - first``.
+   * Note, ``data_size`` is in units of ``item_size`` (not bytes).
+   **/
   size_t data_size() const {
     return array_desc_->data_size;
-  };
+  }
 
   allocator::Buffer& buffer() {
     return array_desc_->data->buffer;
-  };
+  }
   const allocator::Buffer& buffer() const {
     return array_desc_->data->buffer;
-  };
+  }
+
+  size_t buffer_size() const {
+    return allocator::allocator().size(buffer());
+  }
 
   // Return a copy of the shared pointer
   // to the array::Data struct
@@ -312,19 +353,42 @@ class array {
   template <typename T>
   T* data() {
     return static_cast<T*>(array_desc_->data_ptr);
-  };
+  }
 
   template <typename T>
   const T* data() const {
     return static_cast<T*>(array_desc_->data_ptr);
+  }
+
+  enum Status {
+    // 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.
+    evaluated,
+
+    // If the array is the output of a computation then the computation
+    // is complete. Constant arrays are always available (e.g. `array({1, 2,
+    // 3})`)
+    available
   };
 
-  enum Status { unscheduled, scheduled, available };
+  // Check if the array is safe to read.
+  bool is_available() const;
 
-  bool is_available() const {
-    return status() == Status::available;
-  }
-  const Status status() const {
+  // Wait on the array to be available. After this `is_available` returns
+  // `true`.
+  void wait();
+
+  Status status() const {
     return array_desc_->status;
   }
 
@@ -349,18 +413,18 @@ class array {
   // Check if the array is a tracer array
   bool is_tracer() const;
 
-  void set_data(allocator::Buffer buffer, deleter_t d = allocator::free);
+  void set_data(allocator::Buffer buffer, Deleter d = allocator::free);
 
   void set_data(
       allocator::Buffer buffer,
       size_t data_size,
-      std::vector<size_t> strides,
+      Strides strides,
       Flags flags,
-      deleter_t d = allocator::free);
+      Deleter d = allocator::free);
 
   void copy_shared_buffer(
       const array& other,
-      const std::vector<size_t>& strides,
+      const Strides& strides,
       Flags flags,
       size_t data_size,
       size_t offset = 0);
@@ -369,7 +433,7 @@ class array {
 
   void move_shared_buffer(
       array other,
-      const std::vector<size_t>& strides,
+      const Strides& strides,
       Flags flags,
       size_t data_size,
       size_t offset = 0);
@@ -388,8 +452,8 @@ class array {
   void init(const It src);
 
   struct ArrayDesc {
-    std::vector<int> shape;
-    std::vector<size_t> strides;
+    Shape shape;
+    Strides strides;
     size_t size;
     Dtype dtype;
     std::shared_ptr<Primitive> primitive;
@@ -411,8 +475,6 @@ class array {
     void* data_ptr{nullptr};
 
     // The size in elements of the data buffer the array accesses
-    // This can be different than the actual size of the array if it
-    // has been broadcast or irregularly strided.
     size_t data_size;
 
     // Contains useful meta data about the array
@@ -425,10 +487,10 @@ class array {
     // The arrays position in the output list
     uint32_t position{0};
 
-    explicit ArrayDesc(std::vector<int> shape, Dtype dtype);
+    explicit ArrayDesc(Shape shape, Dtype dtype);
 
     explicit ArrayDesc(
-        std::vector<int> shape,
+        Shape shape,
         Dtype dtype,
         std::shared_ptr<Primitive> primitive,
         std::vector<array> inputs);
@@ -449,14 +511,14 @@ class array {
 
 template <typename T>
 array::array(T val, Dtype dtype /* = TypeToDtype<T>() */)
-    : array_desc_(std::make_shared<ArrayDesc>(std::vector<int>{}, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(Shape{}, dtype)) {
   init(&val);
 }
 
 template <typename It>
 array::array(
   It data,
-  std::vector<int> shape,
+  Shape shape,
   Dtype dtype /* = TypeToDtype<typename std::iterator_traits<It>::value_type>() */) :
     array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   init(data);
@@ -467,7 +529,7 @@ array::array(
     std::initializer_list<T> data,
     Dtype dtype /* = TypeToDtype<T>() */)
     : array_desc_(std::make_shared<ArrayDesc>(
-          std::vector<int>{static_cast<int>(data.size())},
+          Shape{static_cast<ShapeElem>(data.size())},
           dtype)) {
   init(data.begin());
 }
@@ -475,7 +537,7 @@ array::array(
 template <typename T>
 array::array(
     std::initializer_list<T> data,
-    std::vector<int> shape,
+    Shape shape,
     Dtype dtype /* = TypeToDtype<T>() */)
     : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   if (data.size() != size()) {
@@ -544,6 +606,9 @@ 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

@@ -1,10 +0,0 @@
-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
-)