patch bump (#1528)

* bfs order

* try fix event issue

.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
@@ -31,53 +85,59 @@ jobs:
           name: Install dependencies
           command: |
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
-            pip install pybind11-stubgen
+            pip install nanobind==2.2.0
             pip install numpy
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
       - run:
           name: Install Python package
           command: |
-            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: |
-            python3 setup.py generate_stubs
+            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
 
   mac_build_and_test:
+    parameters:
+      xcode_version:
+        type: string
+        default: "15.2.0"
     macos:
-      xcode: "15.2.0"
-    resource_class: macos.m1.large.gen1
+      xcode: << parameters.xcode_version >>
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
             brew install python@3.9
+            brew install openmpi
             python3.9 -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
-            pip install pybind11-stubgen
+            pip install nanobind==2.2.0
             pip install numpy
             pip install torch
             pip install tensorflow
@@ -86,35 +146,60 @@ 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
-            python setup.py generate_stubs
+            pip install typing_extensions
+            python setup.py generate_stubs 
       - run:
           name: Run Python tests
           command: |
             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 python3.9 -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 . 
+            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
+            mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
+      - 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:
@@ -129,20 +214,20 @@ 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 --upgrade pybind11[global]
+            pip install nanobind==2.2.0
             pip install --upgrade setuptools
-            pip install pybind11-stubgen
             pip install numpy
             pip install twine
             pip install build
@@ -151,19 +236,20 @@ 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
-            python setup.py generate_stubs
+            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 >>
@@ -176,7 +262,7 @@ jobs:
       - store_artifacts:
           path: dist/
 
-  build_linux_test_release:
+  build_linux_release:
     parameters:
       python_version:
         type: string
@@ -205,22 +291,28 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
+            pip install nanobind==2.2.0
             pip install --upgrade setuptools
-            pip install pybind11-stubgen
             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
-            python setup.py generate_stubs
+            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/
 
@@ -235,8 +327,12 @@ workflows:
         - not: << pipeline.parameters.weekly_build >>
         - not: << pipeline.parameters.test_release >>
     jobs:
-      - mac_build_and_test
+      - mac_build_and_test:
+          matrix:
+            parameters:
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test
+      - build_documentation 
 
   build_pypi_release:
     when:
@@ -253,9 +349,17 @@ workflows:
               ignore: /.*/
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["14.3.1", "15.2.0"]
+              python_version: ["3.9", "3.10", "3.11", "3.12"]
+              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:
@@ -268,6 +372,9 @@ workflows:
           context: pr-approval
       - mac_build_and_test:
           requires: [ hold ]
+          matrix:
+            parameters:
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test:
           requires: [ hold ]
   nightly_build:
@@ -279,8 +386,8 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["14.3.1", "15.2.0"]
+              python_version: ["3.9", "3.10", "3.11", "3.12"]
+              xcode_version: ["15.0.0", "15.2.0"]
   weekly_build:
     when:
       and:
@@ -290,17 +397,17 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["14.3.1", "15.2.0"]
+              python_version: ["3.9", "3.10", "3.11", "3.12"]
+              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"]
               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

.pre-commit-config.yaml

@@ -1,11 +1,11 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v17.0.6
+    rev: v18.1.8
     hooks:
     -   id: clang-format
 # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 24.2.0
+    rev: 24.8.0
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort
@@ -14,3 +14,7 @@ repos:
     -   id: isort
         args:
             - --profile=black
+- repo: https://github.com/cheshirekow/cmake-format-precommit
+  rev: v0.6.13
+  hooks:
+    - id: cmake-format

ACKNOWLEDGMENTS.md

@@ -7,14 +7,19 @@ 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.
+- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops. Added `clip_grad_norm` along with `tree_reduce`. Added `cross`.
 - Juarez Bochi: Fixed bug in cross attention.
 - Justin Deschenaux: Sine, Cosine, arange, randint, truncated normal, bernoulli, lion optimizer, Dropout2d, linear and logistic regression python example.
-- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream` 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" />
 </a>

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

@@ -15,38 +15,52 @@ 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_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 0.5.1)
+  set(MLX_VERSION 0.19.1)
 endif()
 
 # --------------------- Processor tests -------------------------
 
-message(STATUS "Building MLX for ${CMAKE_HOST_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_HOST_SYSTEM_PROCESSOR} MATCHES "x86_64" AND ${CMAKE_HOST_APPLE})
-    message(FATAL_ERROR
-      "Building for x86_64 on macOS is not supported."
-      " If you are on an Apple silicon system, check the build"
-      " documentation for possible fixes: "
-      "https://ml-explore.github.io/mlx/build/html/install.html#build-from-source")
-  elseif (${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "x86_64")
-    message(WARNING
-      "Building for x86_64 on macOS is not supported."
-      " If you are on an Apple silicon system, "
-      " make sure you are building for arm64.")
-  elseif(${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "arm64")
-    set(MLX_BUILD_ARM ON)
+if(${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"
+      )
+    else()
+      set(MLX_BUILD_METAL OFF)
+      message(WARNING "Building for x86_64 arch is not officially supported.")
+    endif()
   endif()
 
 else()
+  set(MLX_BUILD_METAL OFF)
   message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
 endif()
 
+if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm64")
+  set(MLX_BUILD_ARM ON)
+endif()
+
 # ----------------------------- Lib -----------------------------
 
 include(FetchContent)
@@ -55,150 +69,192 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-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)
-elseif (MLX_BUILD_METAL)
+  set(MLX_METAL_DEBUG OFF)
+elseif(MLX_BUILD_METAL)
   message(STATUS "Building METAL sources")
-  # 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)
 
-  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
-
-  if (${MACOS_VERSION} GREATER_EQUAL 14.2)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14.2_iOS17.2.zip)
-  elseif (${MACOS_VERSION} GREATER_EQUAL 14.0)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14_iOS17-beta.zip)
-  elseif (${MACOS_VERSION} GREATER_EQUAL 13.3)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS13.3_iOS16.4.zip)
-  else()
-    message(FATAL_ERROR "MLX requires macOS >= 13.4 to be built with MLX_BUILD_METAL=ON" )
+  if(MLX_METAL_DEBUG)
+    add_compile_definitions(MLX_METAL_DEBUG)
   endif()
 
-  FetchContent_Declare(
-    metal_cpp
-    URL ${METAL_CPP_URL}
+  # Throw an error if xcrun not found
+  execute_process(
+    COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
+    OUTPUT_VARIABLE MACOS_VERSION COMMAND_ERROR_IS_FATAL ANY)
+
+  if(${MACOS_VERSION} LESS 14.0)
+    message(
+      FATAL_ERROR
+        "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON")
+  endif()
+  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
+
+  set(METAL_CPP_URL
+      https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18-beta.zip
   )
+  # Get the metal version
+  execute_process(
+    COMMAND
+      zsh "-c"
+      "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal -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} ${QUARTZ_LIB})
+
+  add_compile_definitions("MLX_METAL_VERSION=${MLX_METAL_VERSION}")
 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)
-else()
-  message(STATUS "Accelerate or arm neon not found, using default backend.")
-  set(MLX_BUILD_ACCELERATE OFF)
-  #set(BLA_VENDOR Generic)
-  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})
-  find_package(LAPACK REQUIRED)
-  if (NOT LAPACK_FOUND)
+if(MLX_BUILD_CPU)
+  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 PUBLIC ${ACCELERATE_LIBRARY})
+    add_compile_definitions(ACCELERATE_NEW_LAPACK)
+  else()
+    message(STATUS "Accelerate or arm neon not found, using default backend.")
+    set(MLX_BUILD_ACCELERATE OFF)
+    if(${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 PUBLIC ${LAPACK_LIBRARIES})
+    # List blas after lapack otherwise we may accidentally incldue an old
+    # version of lapack.h from the include dirs of blas.
+    find_package(BLAS REQUIRED)
+    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 PUBLIC ${BLAS_LIBRARIES})
+  endif()
+else()
+  set(MLX_BUILD_ACCELERATE OFF)
+endif()
+
+find_package(MPI)
+if(MPI_FOUND)
+  execute_process(
+    COMMAND zsh "-c" "mpirun --version"
+    OUTPUT_VARIABLE MPI_VERSION
+    ERROR_QUIET)
+  if(${MPI_VERSION} MATCHES ".*Open MPI.*")
+    target_include_directories(mlx PRIVATE ${MPI_INCLUDE_PATH})
+  elseif(MPI_VERSION STREQUAL "")
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING "MPI found but mpirun is not available. Building without MPI.")
+  else()
+    set(MPI_FOUND FALSE)
+    message(WARNING "MPI which is not OpenMPI found. Building without MPI.")
   endif()
-  find_path(LAPACK_INCLUDE_DIRS lapacke.h
-    /usr/include
-    /usr/local/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})
 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 $<BUILD_INTERFACE:fmt::fmt-header-only>)
+
+if(MLX_BUILD_PYTHON_BINDINGS)
   message(STATUS "Building Python bindings.")
-  find_package(Python COMPONENTS Interpreter Development)
-  find_package(pybind11 CONFIG REQUIRED)
+  find_package(
+    Python 3.8
+    COMPONENTS Interpreter Development.Module
+    REQUIRED)
+  execute_process(
+    COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+    OUTPUT_VARIABLE NB_DIR)
+  list(APPEND CMAKE_PREFIX_PATH "${NB_DIR}")
+  find_package(nanobind CONFIG REQUIRED)
   add_subdirectory(${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()
 
@@ -210,31 +266,24 @@ 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

@@ -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/time_utils.h

@@ -17,14 +17,13 @@
             << std::setprecision(5) << time_fn(FUNC, ##__VA_ARGS__) << " msec" \
             << std::endl;
 
-#define TIMEM(MSG, FUNC, ...)                                                  \
-  std::cout << "Timing "                                                       \
-            << "(" << MSG << ") " << #FUNC << " ... " << std::flush            \
-            << std::setprecision(5) << time_fn(FUNC, ##__VA_ARGS__) << " msec" \
-            << std::endl;
+#define TIMEM(MSG, FUNC, ...)                                      \
+  std::cout << "Timing " << "(" << MSG << ") " << #FUNC << " ... " \
+            << std::flush << std::setprecision(5)                  \
+            << time_fn(FUNC, ##__VA_ARGS__) << " msec" << std::endl;
 
 template <typename F, typename... Args>
-double time_fn(F fn, Args... args) {
+double time_fn(F fn, Args&&... args) {
   // warmup
   for (int i = 0; i < 5; ++i) {
     eval(fn(std::forward<Args>(args)...));

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/conv1d_bench.py

@@ -0,0 +1,123 @@
+import argparse
+import math
+import os
+import subprocess
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+device_name = subprocess.check_output(["sysctl", "-n", "machdep.cpu.brand_string"])
+device_name = device_name.decode("utf-8").strip("\n")
+
+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_1D(strides=1, padding=0, groups=1):
+    def mx_conv_1D(a, b):
+        ys = []
+        for _ in range(N_iter_func):
+            y = mx.conv1d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_1D
+
+
+def make_pt_conv_1D(strides=1, padding=0, groups=1):
+    @torch.no_grad()
+    def pt_conv_1D(a, b):
+        ys = []
+        for _ in range(N_iter_func):
+            y = torch.conv1d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        torch.mps.synchronize()
+        return ys
+
+    return pt_conv_1D
+
+
+def bench_shape(N, iH, C, wH, O, strides, padding, np_dtype, groups):
+    scale = 1.0 / math.sqrt(wH * C)
+    a_np = np.random.uniform(0, 0.5, (N, iH, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, wH, 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, 2, 1))).to("mps")
+    b_pt = torch.from_numpy(b_np.transpose((0, 2, 1))).to("mps")
+
+    torch.mps.synchronize()
+
+    f_mx = make_mx_conv_1D(strides, padding, groups)
+    f_pt = make_pt_conv_1D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv1d(a_mx, b_mx, stride=strides, padding=padding, groups=groups)
+    out_pt = torch.conv1d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 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, iH, C)}, {(O, wH, 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, 5, 32, 1, 2, 1),
+        (4, 32, 32, 5, 32, 1, 2, 2),
+        (4, 32, 32, 5, 32, 1, 2, 4),
+        (4, 32, 32, 5, 32, 1, 2, 8),
+        (4, 32, 32, 5, 32, 1, 2, 8),
+        (4, 32, 32, 5, 32, 1, 2, 16),
+        (4, 32, 32, 5, 32, 1, 2, 32),
+        (4, 32, 256, 5, 512, 1, 2, 2),
+        (4, 32, 256, 5, 512, 1, 2, 128),
+        (4, 32, 256, 5, 512, 1, 2, 256),
+    )
+
+    for dtype in dtypes:
+        print("(N,  iH,  C),  (O,  wH,  C),   dtype,  stride, pads, groups, diff%")
+        for N, iH, C, wH, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, iH, C, wH, O, strides, padding, np_dtype, groups
+            )
+            diff = time_torch / time_mlx - 1.0
+
+            print(
+                f"({N}, {iH:3d}, {C:3d}), ({O:3d}, {wH:2d}, {C:3d}), {dtype}, {strides:5d}, {padding:4d}, {groups:6d}, {100. * diff:+5.2f}%"
+            )
+
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

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

@@ -0,0 +1,118 @@
+# Copyright © 2024 Apple Inc.
+
+import matplotlib
+import mlx.core as mx
+import numpy as np
+import sympy
+import torch
+from time_utils import measure_runtime
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+
+def bandwidth_gb(runtime_ms, system_size):
+    bytes_per_fft = np.dtype(np.complex64).itemsize * 2
+    bytes_per_gb = 1e9
+    ms_per_s = 1e3
+    return system_size * bytes_per_fft / runtime_ms * ms_per_s / bytes_per_gb
+
+
+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
+
+    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
+
+    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)
+
+    print("MLX GPU")
+    with mx.stream(mx.gpu):
+        gpu_bandwidths = run_bench(system_size=system_size, fft_sizes=x)
+
+    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__":
+    time_fft()

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/layer_norm_bench.py

@@ -0,0 +1,41 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import mlx.core as mx
+import mlx.nn as nn
+from time_utils import time_fn
+
+
+def layer_norm(x, w, b, eps):
+    ot = x.dtype
+    x = x.astype(mx.float32)
+    mu = mx.mean(x, -1, keepdims=True)
+    v = mx.var(x, -1, keepdims=True)
+    return (x - mu) * mx.rsqrt(v + eps) * w + b
+
+
+def time_layer_norm():
+    f1 = lambda x, w, b, y: (layer_norm(x, w, b, 1e-5) * y).sum()
+    f2 = lambda x, w, b, y: (mx.fast.layer_norm(x, w, b, 1e-5) * y).sum()
+    g1 = mx.grad(f1, argnums=(0, 1, 2))
+    g2 = mx.grad(f2, argnums=(0, 1, 2))
+
+    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    b = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    mx.eval(x, w, b, y)
+
+    def layer_norm_loop(g, x, w, b):
+        gx, gw, gb = x, w, b
+        for _ in range(32):
+            gx, gw, gb = g(gx, gw, gb, y)
+        return gx, gw, gb
+
+    time_fn(layer_norm_loop, g1, x, w, b)
+    time_fn(layer_norm_loop, g2, x, w, b)
+    time_fn(layer_norm_loop, mx.compile(g1), x, w, b)
+    time_fn(layer_norm_loop, mx.compile(g2), x, w, b)
+
+
+if __name__ == "__main__":
+    time_layer_norm()

benchmarks/python/rms_norm_bench.py

@@ -0,0 +1,39 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import mlx.core as mx
+import mlx.nn as nn
+from time_utils import time_fn
+
+
+def rms_norm(x, w, eps):
+    ot = x.dtype
+    x = x.astype(mx.float32)
+    n = mx.rsqrt(x.square().mean(-1, keepdims=True) + eps)
+    return (x * n).astype(ot) * w
+
+
+def time_rms_norm():
+    f1 = lambda x, w, y: (rms_norm(x, w, 1e-5) * y).sum()
+    f2 = lambda x, w, y: (mx.fast.rms_norm(x, w, 1e-5) * y).sum()
+    g1 = mx.grad(f1, argnums=(0, 1))
+    g2 = mx.grad(f2, argnums=(0, 1))
+
+    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    mx.eval(x, w, y)
+
+    def rms_norm_loop(g, x, w):
+        gx, gw = x, w
+        for _ in range(32):
+            gx, gw = g(gx, gw, y)
+        return gx, gw
+
+    time_fn(rms_norm_loop, g1, x, w)
+    time_fn(rms_norm_loop, g2, x, w)
+    time_fn(rms_norm_loop, mx.compile(g1), x, w)
+    time_fn(rms_norm_loop, mx.compile(g2), x, w)
+
+
+if __name__ == "__main__":
+    time_rms_norm()

benchmarks/python/rope_bench.py

@@ -6,21 +6,21 @@ from time_utils import time_fn
 
 
 def time_rope():
-    rope = nn.RoPE(4096)
+    rope = nn.RoPE(64)
 
     # vec
-    x = mx.random.uniform(shape=(1, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(1, 32, 1, 128)).astype(mx.float16)
     mx.eval(x)
 
     def rope_vec(x):
         for _ in range(32):
-            x = rope(x)
+            x = rope(x, offset=100)
         return x
 
     time_fn(rope_vec, x)
 
     # matrix
-    x = mx.random.uniform(shape=(1024, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(1, 32, 1024, 128)).astype(mx.float16)
     mx.eval(x)
 
     def rope_mat(x):

benchmarks/python/sdpa_bench.py

@@ -0,0 +1,62 @@
+import argparse
+import math
+
+import mlx.core as mx
+from time_utils import time_fn
+
+MAX_SEQ = 300
+START_SEQ = 100
+SEQ_INCREMENT = 50
+
+
+def time_self_attention_primitives():
+    mx.random.seed(3)
+    B = 2
+    H = 38
+    D = 64
+    for R in range(START_SEQ, MAX_SEQ, SEQ_INCREMENT):
+        q = mx.random.uniform(shape=(B, H, R, D))
+        k = mx.random.uniform(shape=(B, H, R, D))
+        v = mx.random.uniform(shape=(B, H, R, D))
+        scale = 1.0 / math.sqrt(float(D))
+        mx.eval(q, k, v)
+
+        def sdpa_primitives(qs, ks, vs, alpha):
+            s = (alpha * qs) @ ks.transpose(0, 1, 3, 2)
+            p = mx.softmax(s.astype(mx.float32), axis=-1).astype(s.dtype)
+            o = p @ vs
+            return o
+
+        time_fn(sdpa_primitives, q, k, v, scale)
+
+
+def time_self_attention_sdpa():
+    mx.random.seed(3)
+    B = 2
+    H = 38
+    D = 64
+    for R in range(START_SEQ, MAX_SEQ, SEQ_INCREMENT):
+        q = mx.random.uniform(shape=(B, H, R, D))
+        k = mx.random.uniform(shape=(B, H, R, D))
+        v = mx.random.uniform(shape=(B, H, R, D))
+        scale = 1.0 / math.sqrt(float(D))
+        mx.eval(q, k, v)
+
+        def sdpa_fused(qs, ks, vs, alpha):
+            o = mx.fast.scaled_dot_product_attention(qs, ks, vs, scale=alpha)
+            return o
+
+        time_fn(sdpa_fused, q, k, v, scale)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("MLX benchmarks.")
+    parser.add_argument("--gpu", action="store_true", help="Use the Metal back-end.")
+    args = parser.parse_args()
+    if args.gpu:
+        mx.set_default_device(mx.gpu)
+    else:
+        mx.set_default_device(mx.cpu)
+
+    time_self_attention_sdpa()
+    time_self_attention_primitives()

benchmarks/python/sdpa_vector_bench.py

@@ -0,0 +1,49 @@
+import argparse
+import math
+
+import mlx.core as mx
+from time_utils import time_fn
+
+L = 1024
+H = 32
+H_k = 32 // 4
+D = 128
+
+
+def attention(q, k, v):
+    B, Hq, L, D = q.shape
+    _, Hk, S, _ = k.shape
+    q = q.reshape(B, Hk, Hq // Hk, L, D)
+    k = k[:, :, None, :, :]
+    v = v[:, :, None, :, :]
+    s = q @ k.transpose(0, 1, 2, 4, 3)
+    p = mx.softmax(s.astype(mx.float32), axis=-1).astype(s.dtype)
+    o = p @ v
+    return o.reshape(B, Hq, L, D)
+
+
+def sdpa(q, k, v):
+    return mx.fast.scaled_dot_product_attention(q, k, v, scale=1.0)
+
+
+def time_self_attention_primitives():
+    mx.random.seed(3)
+    q = mx.random.uniform(shape=(1, H, 1, D))
+    k = mx.random.uniform(shape=(1, H_k, L, D))
+    v = mx.random.uniform(shape=(1, H_k, L, D))
+    mx.eval(q, k, v)
+    time_fn(attention, q, k, v)
+
+
+def time_self_attention_sdpa():
+    mx.random.seed(3)
+    q = mx.random.uniform(shape=(1, H, 1, D))
+    k = mx.random.uniform(shape=(1, H_k, L, D))
+    v = mx.random.uniform(shape=(1, H_k, L, D))
+    mx.eval(q, k, v)
+    time_fn(sdpa, q, k, v)
+
+
+if __name__ == "__main__":
+    time_self_attention_sdpa()
+    time_self_attention_primitives()

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)

docs/Doxyfile

@@ -0,0 +1,50 @@
+################################################################################
+# Primary project setup.                                                       #
+################################################################################
+
+PROJECT_NAME           = "MLX"
+OUTPUT_DIRECTORY       = build
+XML_OUTPUT             = xml
+HTML_OUTPUT            = html
+STRIP_FROM_PATH        = ../
+INPUT                  = ../mlx
+FILE_PATTERNS          = *.h
+EXCLUDE_PATTERNS       = */private/*
+CREATE_SUBDIRS         = NO
+FULL_PATH_NAMES        = YES
+RECURSIVE              = YES
+GENERATE_HTML          = YES
+GENERATE_LATEX         = NO
+GENERATE_XML           = YES
+XML_PROGRAMLISTING     = YES
+
+################################################################################
+# Doxygen preprocessor / parser control.                                       #
+################################################################################
+
+ENABLE_PREPROCESSING   = YES
+MACRO_EXPANSION        = YES
+EXPAND_ONLY_PREDEF     = NO
+SKIP_FUNCTION_MACROS   = NO
+
+################################################################################
+# Compound extraction control.                                                 #
+################################################################################
+
+EXTRACT_ALL            = YES
+EXTRACT_PACKAGE        = YES
+EXTRACT_STATIC         = YES
+CASE_SENSE_NAMES       = NO
+
+################################################################################
+# Docstring control / customization.                                           #
+################################################################################
+
+JAVADOC_AUTOBRIEF      = YES
+
+################################################################################
+# Warning suppression.                                                         #
+################################################################################
+
+QUIET                  = YES
+WARN_IF_UNDOCUMENTED   = NO

docs/README.md

@@ -2,12 +2,16 @@
 
 ### Setup (do once)
 
-Install [sphinx](https://www.sphinx-doc.org/en/master/usage/installation.html)
-for example with `conda`:
+Install Doxygen:
 
 ```
-conda install sphinx
-pip install sphinx-book-theme
+brew install doxygen
+```
+
+Install Python packages:
+
+```
+pip install -r requirements.txt
 ```
 
 ### Build
@@ -15,7 +19,7 @@ pip install sphinx-book-theme
 Build the docs from `mlx/docs/`
 
 ```
-make html
+doxygen && make html
 ```
 
 View the docs by running a server in `mlx/docs/build/html/`:

docs/requirements.txt

@@ -0,0 +1,4 @@
+sphinx
+breathe
+sphinx-book-theme
+mlx

docs/src/_templates/nn-module-template.rst

@@ -0,0 +1,20 @@
+{{ fullname | escape | underline}}
+
+.. currentmodule:: {{ module }}
+
+.. autoclass:: {{ objname }}
+
+   {% block methods %}
+
+   {% if methods %}
+   .. rubric:: {{ _('Methods') }}
+
+   .. autosummary::
+   {% for item in methods %}
+      {%- if item not in inherited_members and item != "__init__" %}
+         ~{{ name }}.{{ item }}
+      {%- endif %}
+   {%- endfor %}
+   {% endif %}
+   {% endblock %}
+

docs/src/conf.py

@@ -22,6 +22,7 @@ extensions = [
     "sphinx.ext.autosummary",
     "sphinx.ext.intersphinx",
     "sphinx.ext.napoleon",
+    "breathe",
 ]
 
 python_use_unqualified_type_names = True
@@ -29,16 +30,20 @@ autosummary_generate = True
 autosummary_filename_map = {"mlx.core.Stream": "stream_class"}
 
 intersphinx_mapping = {
-    "https://docs.python.org/3": None,
-    "https://numpy.org/doc/stable/": None,
+    "python": ("https://docs.python.org/3", None),
+    "numpy": ("https://numpy.org/doc/stable/", None),
 }
 
+breathe_projects = {"mlx": "../build/xml"}
+breathe_default_project = "mlx"
+
 templates_path = ["_templates"]
 html_static_path = ["_static"]
 source_suffix = ".rst"
-master_doc = "index"
+main_doc = "index"
 highlight_language = "python"
 pygments_style = "sphinx"
+add_module_names = False
 
 # -- Options for HTML output -------------------------------------------------
 
@@ -59,3 +64,34 @@ html_theme_options = {
 # -- Options for HTMLHelp output ---------------------------------------------
 
 htmlhelp_basename = "mlx_doc"
+
+
+def setup(app):
+    from sphinx.util import inspect
+
+    wrapped_isfunc = inspect.isfunction
+
+    def isfunc(obj):
+        type_name = str(type(obj))
+        if "nanobind.nb_method" in type_name or "nanobind.nb_func" in type_name:
+            return True
+        return wrapped_isfunc(obj)
+
+    inspect.isfunction = isfunc
+
+
+# -- 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/cpp/ops.rst

@@ -3,4 +3,5 @@
 Operations
 ==========
 
-
+.. doxygengroup:: ops
+   :content-only:

docs/src/dev/custom_metal_kernels.rst

@@ -0,0 +1,421 @@
+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>;
+
+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

@@ -1,24 +1,16 @@
-Developer Documentation
-=======================
+Custom Extensions in MLX
+========================
 
-MLX provides a open and flexible backend to which users may add operations 
-and specialized implementations without much hassle. While the library supplies
-efficient operations that can be used and composed for any number of 
-applications, there may arise cases where new functionalities or highly 
-optimized implementations are needed. For such cases, you may design and 
-implement your own operations that link to and build on top of :mod:`mlx.core`.
-We will introduce the inner-workings of MLX and go over a simple example to 
-learn the steps involved in adding new operations to MLX with your own CPU 
-and GPU implementations. 
+You can extend MLX with custom operations on the CPU or GPU. This guide
+explains how to do that with a simple example.
 
 Introducing the Example
 -----------------------
 
-Let's say that you would like an operation that takes in two arrays, 
-``x`` and ``y``, scales them both by some coefficients ``alpha`` and ``beta``
-respectively, and then adds them together to get the result 
-``z = alpha * x + beta * y``. Well, you can very easily do that by just 
-writing out a function as follows:
+Let's say you would like an operation that takes in two arrays, ``x`` and
+``y``, scales them both by coefficients ``alpha`` and ``beta`` respectively,
+and then adds them together to get the result ``z = alpha * x + beta * y``.
+You can do that in MLX directly:
 
 .. code-block:: python
 
@@ -27,44 +19,35 @@ writing out a function as follows:
     def simple_axpby(x: mx.array, y: mx.array, alpha: float, beta: float) -> mx.array:
         return alpha * x + beta * y
 
-This function performs that operation while leaving the implementations and 
-differentiation to MLX. 
+This function performs that operation while leaving the implementation and
+function transformations to MLX.
 
-However, you work with vector math libraries often and realize that the 
-``axpby`` routine defines the same operation ``Y = (alpha * X) + (beta * Y)``. 
-You would really like the part of your applications that does this operation 
-on the CPU to be very fast - so you decide that you want it to rely on the 
-``axpby`` routine provided by the Accelerate_ framework. Continuing to impose 
-our assumptions on to you, let's also assume that you want to learn how to add 
-your own implementation for the gradients of your new operation while going 
-over the ins-and-outs of the MLX framework. 
+However you may need to customize the underlying implementation, perhaps to
+make it faster or for custom differentiation. In this tutorial we will go
+through adding custom extensions. It will cover:
 
-Well, what a coincidence! You are in the right place. Over the course of this 
-example, we will learn:
-
-* The structure of the MLX library from the frontend API to the backend implementations.
-* How to implement your own CPU backend that redirects to Accelerate_ when appropriate (and a fallback if needed).
-* How to implement your own GPU implementation using metal.
-* How to add your own ``vjp`` and ``jvp``.
-* How to build your implementations, link them to MLX, and bind them to python.
+* The structure of the MLX library.
+* Implementing a CPU operation that redirects to Accelerate_ when appropriate.
+* Implementing a GPU operation using metal.
+* Adding the ``vjp`` and ``jvp`` function transformation.
+* Building a custom extension and binding it to python.
 
 Operations and Primitives
 -------------------------
 
-In one sentence, operations in MLX build the computation graph, and primitives 
-provide the rules for evaluation and transformations of said graph. Let's start 
-by discussing operations in more detail. 
+Operations in MLX build the computation graph. Primitives provide the rules for
+evaluating and transforming the graph. Let's start by discussing operations in
+more detail.
 
 Operations
 ^^^^^^^^^^^
 
-Operations are the frontend functions that operate on arrays. They are defined 
-in the C++ API (:ref:`cpp_ops`) and then we provide bindings to these 
-operations in the Python API (:ref:`ops`). 
+Operations are the front-end functions that operate on arrays. They are defined
+in the C++ API (:ref:`cpp_ops`), and the Python API (:ref:`ops`) binds them.
 
-We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and ``y``,
-and two scalars, ``alpha`` and ``beta``. This is how we would define it in the 
-C++ API:
+We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and
+``y``, and two scalars, ``alpha`` and ``beta``. This is how to define it in
+C++:
 
 .. code-block:: C++
 
@@ -83,10 +66,7 @@ C++ API:
         StreamOrDevice s = {} // Stream on which to schedule the operation
     );
 
-
-This operation itself can call other operations within it if needed. So, the 
-simplest way to go about implementing this operation would be do so in terms 
-of existing operations. 
+The simplest way to this operation is in terms of existing operations:
 
 .. code-block:: C++
 
@@ -100,25 +80,23 @@ of existing operations.
         // Scale x and y on the provided stream
         auto ax = multiply(array(alpha), x, s);
         auto by = multiply(array(beta), y, s);
-        
+
         // Add and return
         return add(ax, by, s);
     }
 
-However, as we discussed earlier, this is not our goal. The operations themselves 
-do not contain the implementations that act on the data, nor do they contain the
-rules of transformations. Rather, they are an easy to use interface that build 
-on top of the building blocks we call :class:`Primitive`. 
+The operations themselves do not contain the implementations that act on the
+data, nor do they contain the rules of transformations. Rather, they are an
+easy to use interface that use :class:`Primitive` building blocks.
 
 Primitives
 ^^^^^^^^^^^
 
-A :class:`Primitive` is part of the computation graph of an :class:`array`. It 
-defines how to create an output given a set of input :class:`array` . Further,
-a :class:`Primitive` is a class that contains rules on how it is evaluated 
-on the CPU or GPU, and how it acts under transformations such as ``vjp`` and 
-``jvp``. These words on their own can be a bit abstract, so lets take a step 
-back and go to our example to give ourselves a more concrete image. 
+A :class:`Primitive` is part of the computation graph of an :class:`array`. It
+defines how to create outputs arrays given a input arrays. Further, a
+:class:`Primitive` has methods to run on the CPU or GPU and for function
+transformations such as ``vjp`` and ``jvp``.  Lets go back to our example to be
+more concrete:
 
 .. code-block:: C++
 
@@ -134,11 +112,15 @@ back and go to our example to give ourselves a more concrete image.
         * To avoid unnecessary allocations, the evaluation function
         * is responsible for allocating space for the array.
         */
-        void eval_cpu(const std::vector<array>& inputs, array& out) override;
-        void eval_gpu(const std::vector<array>& inputs, array& out) override;
+        void eval_cpu(
+            const std::vector<array>& inputs,
+            std::vector<array>& outputs) override;
+        void eval_gpu(
+            const std::vector<array>& inputs,
+            std::vector<array>& outputs) override;
 
         /** The Jacobian-vector product. */
-        array jvp(
+        std::vector<array> jvp(
             const std::vector<array>& primals,
             const std::vector<array>& tangents,
             const std::vector<int>& argnums) override;
@@ -147,7 +129,8 @@ back and go to our example to give ourselves a more concrete image.
         std::vector<array> vjp(
             const std::vector<array>& primals,
             const array& cotan,
-            const std::vector<int>& argnums) override;
+            const std::vector<int>& argnums,
+            const std::vector<array>& outputs) override;
 
         /**
         * The primitive must know how to vectorize itself across
@@ -155,7 +138,7 @@ back and go to our example to give ourselves a more concrete image.
         * representing the vectorized computation and the axis which
         * corresponds to the output vectorized dimension.
         */
-        std::pair<array, int> vmap(
+        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) override;
 
@@ -175,22 +158,22 @@ back and go to our example to give ourselves a more concrete image.
         void eval(const std::vector<array>& inputs, array& out);
     };
 
-The :class:`Axpby` class derives from the base :class:`Primitive` class and 
-follows the above demonstrated interface. :class:`Axpby` treats ``alpha`` and 
-``beta`` as parameters. It then provides implementations of how the array ``out`` 
-is produced given ``inputs`` through :meth:`Axpby::eval_cpu` and 
-:meth:`Axpby::eval_gpu`. Further, it provides rules of transformations in 
-:meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and :meth:`Axpby::vmap`. 
+The :class:`Axpby` class derives from the base :class:`Primitive` class. The
+:class:`Axpby` treats ``alpha`` and ``beta`` as parameters. It then provides
+implementations of how the output array is produced given the inputs through
+:meth:`Axpby::eval_cpu` and :meth:`Axpby::eval_gpu`. It also provides rules
+of transformations in :meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and
+:meth:`Axpby::vmap`.
 
-Using the Primitives
-^^^^^^^^^^^^^^^^^^^^^
+Using the Primitive
+^^^^^^^^^^^^^^^^^^^
 
-Operations can use this :class:`Primitive` to add a new :class:`array` to 
-the computation graph. An :class:`array` can be constructed by providing its 
-data type, shape, the :class:`Primitive` that computes it, and the 
-:class:`array` inputs that are passed to the primitive.
+Operations can use this :class:`Primitive` to add a new :class:`array` to the
+computation graph. An :class:`array` can be constructed by providing its data
+type, shape, the :class:`Primitive` that computes it, and the :class:`array`
+inputs that are passed to the primitive.
 
-Let's re-implement our operation now in terms of our :class:`Axpby` primitive.
+Let's reimplement our operation now in terms of our :class:`Axpby` primitive.
 
 .. code-block:: C++
 
@@ -223,7 +206,7 @@ Let's re-implement our operation now in terms of our :class:`Axpby` primitive.
             /* const std::vector<int>& shape = */ out_shape,
             /* Dtype dtype = */ out_dtype,
             /* std::unique_ptr<Primitive> primitive = */
-            std::make_unique<Axpby>(to_stream(s), alpha, beta),
+            std::make_shared<Axpby>(to_stream(s), alpha, beta),
             /* const std::vector<array>& inputs = */ broadcasted_inputs);
     }
 
@@ -238,27 +221,26 @@ This operation now handles the following:
 Implementing the Primitive
 --------------------------
 
-No computation happens when we call the operation alone. In effect, the 
-operation only builds the computation graph. When we evaluate the output 
-array, MLX schedules the execution of the computation graph, and calls
-:meth:`Axpby::eval_cpu` or :meth:`Axpby::eval_gpu` depending on the 
-stream/device specified by the user. 
+No computation happens when we call the operation alone. The operation only
+builds the computation graph. When we evaluate the output array, MLX schedules
+the execution of the computation graph, and calls :meth:`Axpby::eval_cpu` or
+:meth:`Axpby::eval_gpu` depending on the stream/device specified by the user.
 
 .. warning::
     When :meth:`Primitive::eval_cpu` or :meth:`Primitive::eval_gpu` are called,
     no memory has been allocated for the output array. It falls on the implementation
-    of these functions to allocate memory as needed
+    of these functions to allocate memory as needed.
 
-Implementing the CPU Backend
+Implementing the CPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Let's start by trying to implement a naive and generic version of 
-:meth:`Axpby::eval_cpu`. We declared this as a private member function of 
-:class:`Axpby` earlier called :meth:`Axpby::eval`. 
+Let's start by implementing a naive and generic version of
+:meth:`Axpby::eval_cpu`. We declared this as a private member function of
+:class:`Axpby` earlier called :meth:`Axpby::eval`.
 
-Our naive method will go over each element of the output array, find the 
-corresponding input elements of ``x`` and ``y`` and perform the operation 
-pointwise. This is captured in the templated function :meth:`axpby_impl`. 
+Our naive method will go over each element of the output array, find the
+corresponding input elements of ``x`` and ``y`` and perform the operation
+point-wise. This is captured in the templated function :meth:`axpby_impl`.
 
 .. code-block:: C++
 
@@ -296,19 +278,19 @@ pointwise. This is captured in the templated function :meth:`axpby_impl`.
         }
     }
 
-Now, we would like our implementation to be able to do this pointwise operation 
-for all incoming floating point arrays. Accordingly, we add dispatches for 
-``float32``, ``float16``, ``bfloat16`` and ``complex64``. We throw an error 
-if we encounter an unexpected type.
+Our implementation should work for all incoming floating point arrays.
+Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
+``complex64``. We throw an error if we encounter an unexpected type.
 
 .. code-block:: C++
 
     /** Fall back implementation for evaluation on CPU */
-    void Axpby::eval(const std::vector<array>& inputs, array& out) {
-        // Check the inputs (registered in the op while constructing the out array)
-        assert(inputs.size() == 2);
+    void Axpby::eval(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Dispatch to the correct dtype
         if (out.dtype() == float32) {
@@ -321,28 +303,26 @@ if we encounter an unexpected type.
             return axpby_impl<complex64_t>(x, y, out, alpha_, beta_);
         } else {
             throw std::runtime_error(
-                "Axpby is only supported for floating point types.");
+                "[Axpby] Only supports floating point types.");
         }
     }
 
-We have a fallback implementation! Now, to do what we are really here to do. 
-Remember we wanted to use the ``axpby`` routine provided by the Accelerate_
-framework? Well, there are 3 complications to keep in mind:
+This is good as a fallback implementation. We can use the ``axpby`` routine
+provided by the Accelerate_ framework for a faster implementation in certain
+cases:
 
 #.  Accelerate does not provide implementations of ``axpby`` for half precision
-    floats. We can only direct to it for ``float32`` types 
-#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all elements
-    have fixed strides between them. Possibly due to broadcasts and transposes, 
-    we aren't guaranteed that the inputs fit this requirement. We can 
-    only direct to Accelerate if both ``x`` and ``y`` are row contiguous or 
-    column contiguous. 
-#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` inplace. 
-    MLX expects to write out the answer to a new array. We must copy the elements 
-    of ``y`` into the output array and use that as an input to ``axpby``
+    floats. We can only use it for ``float32`` types.
+#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all
+    elements have fixed strides between them. We only direct to Accelerate
+    if both ``x`` and ``y`` are row contiguous or column contiguous.
+#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` in-place.
+    MLX expects to write the output to a new array. We must copy the elements
+    of ``y`` into the output and use that as an input to ``axpby``.
 
-Let's write out an implementation that uses Accelerate in the right conditions. 
-It must simply allocate data for the output, copy elements of ``y`` into it, 
-and then call the :meth:`catlas_saxpby` from accelerate. 
+Let's write an implementation that uses Accelerate in the right conditions.
+It allocates data for the output, copies ``y`` into it, and then calls the
+:func:`catlas_saxpby` from accelerate.
 
 .. code-block:: C++
 
@@ -356,17 +336,7 @@ and then call the :meth:`catlas_saxpby` from accelerate.
         // Accelerate library provides catlas_saxpby which does
         // Y = (alpha * X) + (beta * Y) in place
         // To use it, we first copy the data in y over to the output array
-
-        // This specialization requires both x and y be contiguous in the same mode
-        // i.e: corresponding linear indices in both point to corresponding elements
-        // The data in the output array is allocated to match the strides in y
-        // such that x, y, and out are contiguous in the same mode and
-        // no transposition is needed
-        out.set_data(
-            allocator::malloc_or_wait(y.data_size() * out.itemsize()),
-            y.data_size(),
-            y.strides(),
-            y.flags());
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
         // We then copy over the elements using the contiguous vector specialization
         copy_inplace(y, out, CopyType::Vector);
@@ -389,18 +359,20 @@ and then call the :meth:`catlas_saxpby` from accelerate.
             /* INCY = */ 1);
     }
 
-Great! But what about the inputs that do not fit the criteria for accelerate?
-Luckily, we can always just direct back to :meth:`Axpby::eval`.
-
-With this in mind, lets finally implement our :meth:`Axpby::eval_cpu`.
+For inputs that do not fit the criteria for accelerate, we fall back to
+:meth:`Axpby::eval`. With this in mind, let's finish our
+:meth:`Axpby::eval_cpu`.
 
 .. code-block:: C++
 
     /** Evaluate primitive on CPU using accelerate specializations */
-    void Axpby::eval_cpu(const std::vector<array>& inputs, array& out) {
+    void Axpby::eval_cpu(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
         assert(inputs.size() == 2);
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Accelerate specialization for contiguous single precision float arrays
         if (out.dtype() == float32 &&
@@ -410,35 +382,33 @@ With this in mind, lets finally implement our :meth:`Axpby::eval_cpu`.
             return;
         }
 
-        // Fall back to common backend if specializations are not available
-        eval(inputs, out);
+        // Fall back to common back-end if specializations are not available
+        eval(inputs, outputs);
     }
 
-We have now hit a milestone! Just this much is enough to run the operation 
-:meth:`axpby` on a CPU stream! 
+Just this much is enough to run the operation :meth:`axpby` on a CPU stream! If
+you do not plan on running the operation on the GPU or using transforms on
+computation graphs that contain :class:`Axpby`, you can stop implementing the
+primitive here and enjoy the speed-ups you get from the Accelerate library.
 
-If you do not plan on running the operation on the GPU or using transforms on 
-computation graphs that contain :class:`Axpby`, you can stop implementing the 
-primitive here and enjoy the speed-ups you get from the Accelerate library. 
-
-Implementing the GPU Backend
+Implementing the GPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Apple silicon devices address their GPUs using the Metal_ shading language, and 
-all GPU kernels in MLX are written using metal. 
+Apple silicon devices address their GPUs using the Metal_ shading language, and
+GPU kernels in MLX are written using Metal.
 
 .. note::
 
-    Here are some helpful resources if you are new to metal!
+    Here are some helpful resources if you are new to Metal:
 
     * A walkthrough of the metal compute pipeline: `Metal Example`_
     * Documentation for metal shading language: `Metal Specification`_
     * Using metal from C++: `Metal-cpp`_
 
-Let's keep the GPU algorithm simple. We will launch exactly as many threads 
-as there are elements in the output. Each thread will pick the element it needs 
-from ``x`` and ``y``, do the pointwise operation, and then update its assigned 
-element in the output. 
+Let's keep the GPU kernel simple. We will launch exactly as many threads as
+there are elements in the output. Each thread will pick the element it needs
+from ``x`` and ``y``, do the point-wise operation, and update its assigned
+element in the output.
 
 .. code-block:: C++
 
@@ -457,15 +427,14 @@ element in the output.
         // Convert linear indices to offsets in array
         auto x_offset = elem_to_loc(index, shape, x_strides, ndim);
         auto y_offset = elem_to_loc(index, shape, y_strides, ndim);
-        
+
         // Do the operation and update the output
-        out[index] = 
+        out[index] =
             static_cast<T>(alpha) * x[x_offset] + static_cast<T>(beta) * y[y_offset];
     }
 
 We then need to instantiate this template for all floating point types and give
-each instantiation a unique host name so we can identify the right kernel for 
-each data type. 
+each instantiation a unique host name so we can identify it.
 
 .. code-block:: C++
 
@@ -488,29 +457,21 @@ each data type.
     instantiate_axpby(bfloat16, bfloat16_t);
     instantiate_axpby(complex64, complex64_t);
 
-This kernel will be compiled into a metal library ``mlx_ext.metallib`` as we 
-will see later in :ref:`Building with CMake`. In the following example, we 
-assume that the library ``mlx_ext.metallib`` will always be co-located with 
-the executable/ shared-library calling the :meth:`register_library` function. 
-The :meth:`register_library` function takes the library's name and potential 
-path (or in this case, a function that can produce the path of the metal 
-library) and tries to load that library if it hasn't already been registered 
-by the relevant static :class:`mlx::core::metal::Device` object. This is why, 
-it is important to package your C++ library with the metal library. We will 
-go over this process in more detail later. 
-
-The logic to determine the kernel, set the inputs, resolve the grid dimensions 
-and dispatch it to the GPU are contained in :meth:`Axpby::eval_gpu` as shown 
+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
 below.
 
 .. code-block:: C++
 
     /** Evaluate primitive on GPU */
-    void Axpby::eval_gpu(const std::vector<array>& inputs, array& out) {
+    void Axpby::eval_gpu(
+      const std::vector<array>& inputs,
+      std::vector<array>& outputs) {
         // Prepare inputs
         assert(inputs.size() == 2);
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Each primitive carries the stream it should execute on
         // and each stream carries its device identifiers
@@ -518,22 +479,21 @@ below.
         // We get the needed metal device using the stream
         auto& d = metal::device(s.device);
 
-        // Allocate output memory 
+        // Allocate output memory
         out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
-        // Resolve name of kernel (corresponds to axpby.metal)
+        // Resolve name of kernel
         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);
+        auto& compute_encoder = d.get_command_encoder(s.index);
         compute_encoder->setComputePipelineState(kernel);
 
         // Kernel parameters are registered with buffer indices corresponding to
@@ -542,17 +502,17 @@ 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);
 
-        // Encode shape, strides and ndim 
+        // 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);
@@ -570,33 +530,30 @@ below.
 
         // Launch the grid with the given number of threads divided among
         // the given threadgroups
-        compute_encoder->dispatchThreads(grid_dims, group_dims);
+        compute_encoder.dispatchThreads(grid_dims, group_dims);
     }
 
 We can now call the :meth:`axpby` operation on both the CPU and the GPU!
 
-A few things to note about MLX and metal before moving on. MLX keeps track 
-of the active ``compute_encoder``. We rely on :meth:`d.get_command_encoder` 
-to give us the active metal compute command encoder instead of building a 
-new one and calling :meth:`compute_encoder->end_encoding` at the end. 
-MLX keeps adding kernels (compute pipelines) to the active command encoder 
-until some specified limit is hit or the compute encoder needs to be flushed 
-for synchronization. MLX also handles enqueuing and committing the associated 
-command buffers as needed. We suggest taking a deeper dive into 
-:class:`metal::Device` if you would like to study this routine further.
+A few things to note about MLX and Metal before moving on. MLX keeps track of
+the active ``command_buffer`` and the ``MTLCommandBuffer`` to which it is
+associated. We rely on :meth:`d.get_command_encoder` to give us the active
+metal compute command encoder instead of building a new one and calling
+:meth:`compute_encoder->end_encoding` at the end. MLX adds kernels (compute
+pipelines) to the active command buffer until some specified limit is hit or
+the command buffer needs to be flushed for synchronization.
 
 Primitive Transforms
 ^^^^^^^^^^^^^^^^^^^^^
 
-Now that we have come this far, let's also learn how to add implementations to 
-transformations in a :class:`Primitive`. These transformations can be built on 
-top of our operations, including the one we just defined now. Which then gives 
-us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
+Next, let's add implementations for transformations in a :class:`Primitive`.
+These transformations can be built on top of other operations, including the
+one we just defined:
 
 .. code-block:: C++
 
     /** The Jacobian-vector product. */
-    array Axpby::jvp(
+    std::vector<array> Axpby::jvp(
             const std::vector<array>& primals,
             const std::vector<array>& tangents,
             const std::vector<int>& argnums) {
@@ -611,12 +568,12 @@ us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
         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());
+            return {multiply(scale_arr, tangents[0], stream())};
         }
         // If, argnums = {0, 1}, we take contributions from both
         // which gives us jvp = tangent_x * alpha + tangent_y * beta
         else {
-            return axpby(tangents[0], tangents[1], alpha_, beta_, stream());
+            return {axpby(tangents[0], tangents[1], alpha_, beta_, stream())};
         }
     }
 
@@ -625,34 +582,35 @@ us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
     /** The vector-Jacobian product. */
     std::vector<array> Axpby::vjp(
             const std::vector<array>& primals,
-            const array& cotan,
-            const std::vector<int>& argnums) {
+            const std::vector<array>& cotangents,
+            const std::vector<int>& argnums,
+            const std::vector<int>& /* unused */) {
         // Reverse mode diff
         std::vector<array> vjps;
         for (auto arg : argnums) {
             auto scale = arg == 0 ? alpha_ : beta_;
-            auto scale_arr = array(scale, cotan.dtype());
-            vjps.push_back(multiply(scale_arr, cotan, stream()));
+            auto scale_arr = array(scale, cotangents[0].dtype());
+            vjps.push_back(multiply(scale_arr, cotangents[0], stream()));
         }
         return vjps;
     }
 
-Finally, you need not have a transformation fully defined to start using your 
-own :class:`Primitive`.
+Note, a transformation does not need to be fully defined to start using
+the :class:`Primitive`.
 
 .. code-block:: C++
 
     /** Vectorize primitive along given axis */
-    std::pair<array, int> Axpby::vmap(
+    std::pair<std::vector<array>, std::vector<int>> Axpby::vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) {
-        throw std::runtime_error("Axpby has no vmap implementation.");
+        throw std::runtime_error("[Axpby] vmap not implemented.");
     }
 
 Building and Binding
 --------------------
 
-Let's look at the overall directory structure first. 
+Let's look at the overall directory structure first.
 
 | extensions
 | ├── axpby
@@ -666,40 +624,39 @@ Let's look at the overall directory structure first.
 | └── setup.py
 
 * ``extensions/axpby/`` defines the C++ extension library
-* ``extensions/mlx_sample_extensions`` sets out the structure for the 
-  associated python package
-* ``extensions/bindings.cpp`` provides python bindings for our operation
-* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and 
-  python bindings
+* ``extensions/mlx_sample_extensions`` sets out the structure for the
+  associated Python package
+* ``extensions/bindings.cpp`` provides Python bindings for our operation
+* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and
+  Python bindings
 * ``extensions/setup.py`` holds the ``setuptools`` rules to build and install
-  the python package
+  the Python package
 
 Binding to Python
 ^^^^^^^^^^^^^^^^^^
 
-We use PyBind11_ to build a Python API for the C++ library. Since bindings for
+We use nanobind_ to build a Python API for the C++ library. Since bindings for
 components such as :class:`mlx.core.array`, :class:`mlx.core.stream`, etc. are
-already provided, adding our :meth:`axpby` is simple!
+already provided, adding our :meth:`axpby` is simple.
 
 .. code-block:: C++
 
-    PYBIND11_MODULE(mlx_sample_extensions, m) {
-        m.doc() = "Sample C++ and metal extensions for MLX";
+   NB_MODULE(_ext, m) {
+        m.doc() = "Sample extension for MLX";
 
         m.def(
             "axpby",
             &axpby,
             "x"_a,
             "y"_a,
-            py::pos_only(),
             "alpha"_a,
             "beta"_a,
-            py::kw_only(),
-            "stream"_a = py::none(),
-            R"pbdoc(
+            nb::kw_only(),
+            "stream"_a = nb::none(),
+            R"(
                 Scale and sum two vectors element-wise
                 ``z = alpha * x + beta * y``
-                
+
                 Follows numpy style broadcasting between ``x`` and ``y``
                 Inputs are upcasted to floats if needed
 
@@ -711,17 +668,17 @@ already provided, adding our :meth:`axpby` is simple!
 
                 Returns:
                     array: ``alpha * x + beta * y``
-            )pbdoc");
+            )");
     }
 
-Most of the complexity in the above example comes from additional bells and 
+Most of the complexity in the above example comes from additional bells and
 whistles such as the literal names and doc-strings.
 
 .. warning::
 
-    :mod:`mlx.core` needs to be imported before importing 
-    :mod:`mlx_sample_extensions` as defined by the pybind11 module above to 
-    ensure that the casters for :mod:`mlx.core` components like 
+    :mod:`mlx.core` must be imported before importing
+    :mod:`mlx_sample_extensions` as defined by the nanobind module above to
+    ensure that the casters for :mod:`mlx.core` components like
     :class:`mlx.core.array` are available.
 
 .. _Building with CMake:
@@ -729,8 +686,8 @@ whistles such as the literal names and doc-strings.
 Building with CMake
 ^^^^^^^^^^^^^^^^^^^^
 
-Building the C++ extension library itself is simple, it only requires that you 
-``find_package(MLX CONFIG)`` and then link it to your library. 
+Building the C++ extension library only requires that you ``find_package(MLX
+CONFIG)`` and then link it to your library.
 
 .. code-block:: cmake
 
@@ -752,12 +709,12 @@ Building the C++ extension library itself is simple, it only requires that you
     # Link to mlx
     target_link_libraries(mlx_ext PUBLIC mlx)
 
-We also need to build the attached metal library. For convenience, we provide a 
-:meth:`mlx_build_metallib` function that builds a ``.metallib`` target given 
-sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and 
-automatically imported with MLX package). 
+We also need to build the attached Metal library. For convenience, we provide a
+:meth:`mlx_build_metallib` function that builds a ``.metallib`` target given
+sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and
+automatically imported with MLX package).
 
-Here is what that looks like in practice!
+Here is what that looks like in practice:
 
 .. code-block:: cmake
 
@@ -779,27 +736,29 @@ Here is what that looks like in practice!
 
     endif()
 
-Finally, we build the Pybind11_ bindings
+Finally, we build the nanobind_ bindings
 
 .. code-block:: cmake
 
-    pybind11_add_module(
-        mlx_sample_extensions
-        ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
+    nanobind_add_module(
+      _ext
+      NB_STATIC STABLE_ABI LTO NOMINSIZE
+      NB_DOMAIN mlx
+      ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
     )
-    target_link_libraries(mlx_sample_extensions PRIVATE mlx_ext)
+    target_link_libraries(_ext PRIVATE mlx_ext)
 
     if(BUILD_SHARED_LIBS)
-        target_link_options(mlx_sample_extensions PRIVATE -Wl,-rpath,@loader_path)
+      target_link_options(_ext PRIVATE -Wl,-rpath,@loader_path)
     endif()
 
 Building with ``setuptools``
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Once we have set out the CMake build rules as described above, we can use the
-build utilities defined in :mod:`mlx.extension` for a simple build process. 
+build utilities defined in :mod:`mlx.extension`:
 
-.. code-block:: python 
+.. code-block:: python
 
     from mlx import extension
     from setuptools import setup
@@ -809,48 +768,50 @@ build utilities defined in :mod:`mlx.extension` for a simple build process.
             name="mlx_sample_extensions",
             version="0.0.0",
             description="Sample C++ and Metal extensions for MLX primitives.",
-            ext_modules=[extension.CMakeExtension("mlx_sample_extensions")],
+            ext_modules=[extension.CMakeExtension("mlx_sample_extensions._ext")],
             cmdclass={"build_ext": extension.CMakeBuild},
-            packages = ["mlx_sample_extensions"],
-            package_dir = {"": "mlx_sample_extensions"},
-            package_data = {"mlx_sample_extensions" : ["*.so", "*.dylib", "*.metallib"]},
+            packages=["mlx_sample_extensions"],
+            package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
+            extras_require={"dev":[]},
             zip_safe=False,
-            python_requires=">=3.7",
+            python_requires=">=3.8",
         )
 
 .. note::
     We treat ``extensions/mlx_sample_extensions`` as the package directory
     even though it only contains a ``__init__.py`` to ensure the following:
-    
-    * :mod:`mlx.core` is always imported before importing  :mod:`mlx_sample_extensions`
-    * The C++ extension library and the metal library are co-located with the python 
-      bindings and copied together if the package is installed 
 
-You can build inplace for development using
+    * :mod:`mlx.core` must be imported before importing :mod:`_ext`
+    * The C++ extension library and the metal library are co-located with the python
+      bindings and copied together if the package is installed
+
+To build the package, first install the build dependencies with ``pip install
+-r requirements.txt``.  You can then build inplace for development using
 ``python setup.py build_ext -j8 --inplace`` (in ``extensions/``)
 
-This will result in a directory structure as follows:
+This results in the directory structure:
 
 | extensions
 | ├── mlx_sample_extensions
 | │   ├── __init__.py
 | │   ├── libmlx_ext.dylib # C++ extension library
 | │   ├── mlx_ext.metallib # Metal library
-| │   └── mlx_sample_extensions.cpython-3x-darwin.so # Python Binding
+| │   └── _ext.cpython-3x-darwin.so # Python Binding
 | ...
 
-When you try to install using the command ``python -m pip install .`` 
-(in ``extensions/``), the package will be installed with the same structure as 
-``extensions/mlx_sample_extensions`` and the C++ and metal library will be 
-copied along with the python binding since they are specified as ``package_data``.
+When you try to install using the command ``python -m pip install .`` (in
+``extensions/``), the package will be installed with the same structure as
+``extensions/mlx_sample_extensions`` and the C++ and Metal library will be
+copied along with the Python binding since they are specified as
+``package_data``.
 
 Usage
 -----
 
-After installing the extension as described above, you should be able to simply 
-import the python package and play with it as you would any other MLX operation!
+After installing the extension as described above, you should be able to simply
+import the Python package and play with it as you would any other MLX operation.
 
-Let's looks at a simple script and it's results!
+Let's look at a simple script and its results:
 
 .. code-block:: python
 
@@ -863,7 +824,7 @@ Let's looks at a simple script and it's 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:
 
@@ -874,12 +835,12 @@ Output:
     c correctness: True
 
 Results
-^^^^^^^^^^^^^^^^
+^^^^^^^
 
-Let's run a quick benchmark and see how our new ``axpby`` operation compares 
-with the naive :meth:`simple_axpby` we defined at first on the CPU. 
+Let's run a quick benchmark and see how our new ``axpby`` operation compares
+with the naive :meth:`simple_axpby` we first defined on the CPU.
 
-.. code-block:: python 
+.. code-block:: python
 
     import mlx.core as mx
     from mlx_sample_extensions import axpby
@@ -898,7 +859,7 @@ with the naive :meth:`simple_axpby` we defined at first on the CPU.
     alpha = 4.0
     beta = 2.0
 
-    mx.eval((x, y))
+    mx.eval(x, y)
 
     def bench(f):
         # Warm up
@@ -919,30 +880,23 @@ with the naive :meth:`simple_axpby` we defined at first on the CPU.
 
     print(f"Simple axpby: {simple_time:.3f} s | Custom axpby: {custom_time:.3f} s")
 
-Results:
-
-.. code-block::
-
-    Simple axpby: 0.114 s | Custom axpby: 0.109 s
-
-We see some modest improvements right away! 
+The results are ``Simple axpby: 0.114 s | Custom axpby: 0.109 s``. We see
+modest improvements right away!
 
 This operation is now good to be used to build other operations, in
 :class:`mlx.nn.Module` calls, and also as a part of graph transformations like
-:meth:`grad`!
+:meth:`grad`.
 
 Scripts
 -------
 
 .. admonition:: Download the code
 
-   The full example code is available in `mlx <code>`_.
-
-.. code: `https://github.com/ml-explore/mlx/tree/main/examples/extensions/`_
+   The full example code is available in `mlx <https://github.com/ml-explore/mlx/tree/main/examples/extensions/>`_.
 
 .. _Accelerate: https://developer.apple.com/documentation/accelerate/blas?language=objc
 .. _Metal: https://developer.apple.com/documentation/metal?language=objc
 .. _Metal-cpp: https://developer.apple.com/metal/cpp/
 .. _`Metal Specification`: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
 .. _`Metal Example`: https://developer.apple.com/documentation/metal/performing_calculations_on_a_gpu?language=objc
-.. _PyBind11: https://pybind11.readthedocs.io/en/stable/
+.. _nanobind: https://nanobind.readthedocs.io/en/latest/

docs/src/dev/metal_debugger.rst

@@ -0,0 +1,68 @@
+Metal Debugger
+==============
+
+.. currentmodule:: mlx.core
+
+Profiling is a key step for performance optimization. You can build MLX with
+the ``MLX_METAL_DEBUG`` option to improve the Metal debugging and
+optimization workflow. The ``MLX_METAL_DEBUG`` debug option:
+
+* Records source during Metal compilation, for later inspection while
+  debugging.
+* Labels Metal objects such as command queues, improving capture readability.
+
+To build with debugging enabled in Python prepend
+``CMAKE_ARGS="-DMLX_METAL_DEBUG=ON"`` to the build call.
+
+The :func:`metal.start_capture` function initiates a capture of all MLX GPU
+work.
+
+.. note::
+
+   To capture a GPU trace you must run the application with
+   ``MTL_CAPTURE_ENABLED=1``.
+
+.. code-block:: python
+
+    import mlx.core as mx
+
+    a = mx.random.uniform(shape=(512, 512))
+    b = mx.random.uniform(shape=(512, 512))
+    mx.eval(a, b)
+
+    trace_file = "mlx_trace.gputrace"
+
+    # Make sure to run with MTL_CAPTURE_ENABLED=1 and
+    # that the path trace_file does not already exist.
+    mx.metal.start_capture(trace_file)
+
+    for _ in range(10):
+      mx.eval(mx.add(a, b))
+
+    mx.metal.stop_capture()
+
+You can open and replay the GPU trace in Xcode. The ``Dependencies`` view
+has a great overview of all operations. Checkout the `Metal debugger
+documentation`_ for more information.
+
+.. image:: ../_static/metal_debugger/capture.png
+    :class: dark-light
+
+Xcode Workflow
+--------------
+
+You can skip saving to a path by running within Xcode. First, generate an
+Xcode project using CMake.
+
+.. code-block::
+
+    mkdir build && cd build
+    cmake .. -DMLX_METAL_DEBUG=ON -G Xcode
+    open mlx.xcodeproj
+
+Select the ``metal_capture`` example schema and run.
+
+.. image:: ../_static/metal_debugger/schema.png
+    :class: dark-light
+
+.. _`Metal debugger documentation`: https://developer.apple.com/documentation/xcode/metal-debugger

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

@@ -43,6 +43,7 @@ are the CPU and GPU.
    usage/function_transforms
    usage/compile
    usage/numpy
+   usage/distributed
    usage/using_streams
 
 .. toctree::
@@ -58,15 +59,18 @@ are the CPU and GPU.
    :maxdepth: 1
 
    python/array
+   python/data_types
    python/devices_and_streams
    python/ops
    python/random
    python/transforms
+   python/fast
    python/fft
    python/linalg
    python/metal
    python/nn
    python/optimizers
+   python/distributed
    python/tree_utils
 
 .. toctree::
@@ -80,3 +84,5 @@ are the CPU and GPU.
    :maxdepth: 1
 
    dev/extensions
+   dev/metal_debugger
+   dev/custom_metal_kernels

docs/src/install.rst

@@ -14,11 +14,11 @@ 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
-- macOS >= 13.3
+- Using a native Python >= 3.9
+- macOS >= 13.5
 
 .. note::
-    MLX is only available on devices running macOS >= 13.3 
+    MLX is only available on devices running macOS >= 13.5
     It is highly recommended to use macOS 14 (Sonoma)
 
 
@@ -54,7 +54,7 @@ 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``
-- Xcode >= 14.3 (Xcode >= 15.0 for macOS 14 and above)
+- Xcode >= 15.0 and macOS SDK >= 14.0
 
 .. note::
    Ensure your shell environment is native ``arm``, not ``x86`` via Rosetta. If
@@ -70,39 +70,36 @@ 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
 
-Make sure that you have `pybind11 <https://pybind11.readthedocs.io/en/stable/index.html>`_
-installed. You can install ``pybind11`` with ``pip``, ``brew`` or ``conda`` as follows:
+Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-    pip install "pybind11[global]"
-    conda install pybind11
-    brew install pybind11
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
 
-Then simply build and install it using pip:
+For developing, install the package with development dependencies, and use an
+editable install:
 
 .. code-block:: shell
 
-   env CMAKE_BUILD_PARALLEL_LEVEL="" pip install .
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
 
-For developing use an editable install:
+Once the development dependencies are installed, you can build faster with:
 
 .. code-block:: shell
 
-  env CMAKE_BUILD_PARALLEL_LEVEL="" pip install -e .
+ CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
 
-To make sure the install is working run the tests with:
+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
@@ -123,7 +120,7 @@ Create a build directory and run CMake and make:
 .. code-block:: shell
 
    mkdir -p build && cd build
-   cmake .. && make -j 
+   cmake .. && make -j
 
 Run tests with:
 
@@ -142,7 +139,7 @@ directory as the executable statically linked to ``libmlx.a`` or the
 preprocessor constant ``METAL_PATH`` should be defined at build time and it
 should point to the path to the built metal library.
 
-.. list-table:: Build Options 
+.. list-table:: Build Options
    :widths: 25 8
    :header-rows: 1
 
@@ -156,31 +153,67 @@ 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::
 
-    If you have multiple Xcode installations and wish to use 
-    a specific one while building, you can do so by adding the 
-    following environment variable before building 
+    If you have multiple Xcode installations and wish to use
+    a specific one while building, you can do so by adding the
+    following environment variable before building
 
     .. code-block:: shell
 
       export DEVELOPER_DIR="/path/to/Xcode.app/Contents/Developer/"
 
-    Further, you can use the following command to find out which 
+    Further, you can use the following command to find out which
     macOS SDK will be used
 
     .. code-block:: shell
 
       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 accross reboots.
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 
-
 Metal not found
 ~~~~~~~~~~~~~~~
 
@@ -202,12 +235,12 @@ Then set the active developer directory:
 
   sudo xcode-select --switch /Applications/Xcode.app/Contents/Developer
 
-x86 Shell 
+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,
@@ -231,4 +264,4 @@ 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

@@ -10,27 +10,39 @@ Array
 
     array
     array.astype
+    array.at
     array.item
     array.tolist
     array.dtype
+    array.itemsize
+    array.nbytes
     array.ndim
     array.shape
     array.size
-    Dtype
     array.abs
     array.all
     array.any
     array.argmax
     array.argmin
+    array.conj
     array.cos
-    array.dtype
+    array.cummax
+    array.cummin
+    array.cumprod
+    array.cumsum
+    array.diag
+    array.diagonal
     array.exp
+    array.flatten
     array.log
+    array.log10
     array.log1p
+    array.log2
     array.logsumexp
     array.max
     array.mean
     array.min
+    array.moveaxis
     array.prod
     array.reciprocal
     array.reshape
@@ -40,7 +52,11 @@ Array
     array.split
     array.sqrt
     array.square
+    array.squeeze
+    array.std
     array.sum
+    array.swapaxes
     array.transpose
     array.T
     array.var
+    array.view

docs/src/python/data_types.rst

@@ -1,7 +1,5 @@
 .. _data_types:
 
-:orphan:
-
 Data Types
 ==========
 
@@ -44,9 +42,27 @@ The default floating point type is ``float32`` and the default integer type is
    * - ``int64``
      - 8 
      - 64-bit signed integer 
+   * - ``bfloat16``
+     - 2 
+     - 16-bit brain float (e8, m7)
    * - ``float16``
      - 2 
-     - 16-bit float, only available with `ARM C language extensions <https://developer.arm.com/documentation/101028/0012/3--C-language-extensions?lang=en>`_
+     - 16-bit IEEE float (e5, m10)
    * - ``float32``
      - 4 
      - 32-bit float
+   * - ``complex64``
+     - 8 
+     - 64-bit complex float
+
+
+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.
+
+.. autosummary::
+   :toctree: _autosummary
+
+   Dtype
+   DtypeCategory
+   issubdtype

docs/src/python/devices_and_streams.rst

@@ -16,3 +16,4 @@ Devices and Streams
    new_stream
    set_default_stream
    stream
+   synchronize

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/fast.rst

@@ -0,0 +1,16 @@
+.. _fast:
+
+Fast
+====
+
+.. currentmodule:: mlx.core.fast
+
+.. autosummary:: 
+  :toctree: _autosummary
+
+  rms_norm
+  layer_norm
+  rope
+  scaled_dot_product_attention
+  affine_quantize
+  metal_kernel

docs/src/python/linalg.rst

@@ -8,5 +8,13 @@ Linear Algebra
 .. autosummary:: 
    :toctree: _autosummary 
 
+    inv
+    tri_inv
     norm
+    cholesky
+    cholesky_inv
+    cross
     qr
+    svd
+    eigvalsh
+    eigh

docs/src/python/metal.rst

@@ -3,12 +3,18 @@ Metal
 
 .. currentmodule:: mlx.core.metal
 
-.. autosummary:: 
+.. autosummary::
   :toctree: _autosummary
 
   is_available
+  device_info
   get_active_memory
   get_peak_memory
+  reset_peak_memory
   get_cache_memory
   set_memory_limit
   set_cache_limit
+  set_wired_limit
+  clear_cache
+  start_capture
+  stop_capture

docs/src/python/nn.rst

@@ -173,6 +173,7 @@ In detail:
    :toctree: _autosummary
 
    value_and_grad
+   quantize
 
 .. toctree::
 

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

@@ -13,31 +13,55 @@ Layers
    AvgPool1d
    AvgPool2d
    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
    Mish
    MultiHeadAttention
    PReLU
+   QuantizedEmbedding
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU6
+   RNN
    RoPE
    SELU
    Sequential
+   Sigmoid
    SiLU
    SinusoidalPositionalEncoding
+   Softmin
    Softshrink
+   Softsign
+   Softmax
+   Softplus
    Step
+   Tanh
    Transformer
-   Upsample
+   Upsample

docs/src/python/nn/module.rst

@@ -30,6 +30,7 @@ Module
       Module.named_modules
       Module.parameters
       Module.save_weights
+      Module.set_dtype
       Module.train
       Module.trainable_parameters
       Module.unfreeze

docs/src/python/ops.rst

@@ -5,13 +5,14 @@ Operations
 
 .. currentmodule:: mlx.core
 
-.. autosummary:: 
+.. autosummary::
   :toctree: _autosummary
 
    abs
    add
+   addmm
    all
-   allclose 
+   allclose
    any
    arange
    arccos
@@ -19,49 +20,76 @@ Operations
    arcsin
    arcsinh
    arctan
+   arctan2
    arctanh
    argmax
    argmin
    argpartition
    argsort
    array_equal
+   as_strided
    atleast_1d
    atleast_2d
    atleast_3d
+   bitwise_and
+   bitwise_or
+   bitwise_xor
+   block_masked_mm
    broadcast_to
    ceil
    clip
    concatenate
+   conj
+   conjugate
    convolve
    conv1d
    conv2d
+   conv3d
+   conv_transpose1d
+   conv_transpose2d
+   conv_transpose3d
    conv_general
    cos
    cosh
+   cummax
+   cummin
+   cumprod
+   cumsum
+   degrees
    dequantize
    diag
    diagonal
    divide
    divmod
+   einsum
+   einsum_path
    equal
    erf
    erfinv
    exp
+   expm1
    expand_dims
    eye
    flatten
    floor
    floor_divide
    full
+   gather_mm
+   gather_qmm
    greater
    greater_equal
+   hadamard_transform
    identity
+   imag
    inner
+   isfinite
    isclose
-   isnan
-   isposinf
-   isneginf
    isinf
+   isnan
+   isneginf
+   isposinf
+   issubdtype
+   left_shift
    less
    less_equal
    linspace
@@ -79,22 +107,32 @@ Operations
    max
    maximum
    mean
+   meshgrid
    min
    minimum
    moveaxis
    multiply
+   nan_to_num
    negative
+   not_equal
    ones
    ones_like
    outer
    partition
    pad
+   power
    prod
+   put_along_axis
    quantize
    quantized_matmul
+   radians
+   real
    reciprocal
+   remainder
    repeat
    reshape
+   right_shift
+   roll
    round
    rsqrt
    save
@@ -113,6 +151,7 @@ Operations
    square
    squeeze
    stack
+   std
    stop_gradient
    subtract
    sum
@@ -124,11 +163,13 @@ Operations
    tensordot
    tile
    topk
+   trace
    transpose
    tri
    tril
    triu
    var
+   view
    where
    zeros
    zeros_like

docs/src/python/optimizers.rst

@@ -1,5 +1,7 @@
 .. _optimizers:
 
+.. currentmodule:: mlx.optimizers
+
 Optimizers
 ==========
 
@@ -29,8 +31,48 @@ 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
    optimizers/common_optimizers
    optimizers/schedulers
+
+.. autosummary::
+   :toctree: _autosummary
+
+   clip_grad_norm

docs/src/python/random.rst

@@ -38,8 +38,11 @@ we use a splittable version of Threefry, which is a counter-based PRNG.
    gumbel
    key
    normal
+   multivariate_normal
    randint
    seed
    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/python/tree_utils.rst

@@ -19,3 +19,5 @@ return python trees will be using the default python ``dict``, ``list`` and
    tree_flatten
    tree_unflatten
    tree_map
+   tree_map_with_path
+   tree_reduce

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`

docs/src/usage/distributed.rst

@@ -0,0 +1,166 @@
+.. _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 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()
+        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/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
 
@@ -40,7 +40,7 @@ getting higher order derivatives.
 
 Any of the MLX function transformations can be composed in any order to any
 depth. See the following sections for more information on :ref:`automatic
-differentiaion <auto diff>` and :ref:`automatic vectorization <vmap>`.
+differentiation <auto diff>` and :ref:`automatic vectorization <vmap>`.
 For more information on :func:`compile` see the :ref:`compile documentation <compile>`.
 
 
@@ -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.
 
 
@@ -166,14 +166,14 @@ A naive way to add the elements from two sets of vectors is with a loop:
 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))
 
 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:
 

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)
 
@@ -82,7 +82,7 @@ general, MLX has limited support for operations for which outputs
 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

@@ -13,12 +13,12 @@ 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
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Lazy evaluation let's us record a compute graph without actually doing any
+Lazy evaluation lets us record a compute graph without actually doing any
 computations. This is useful for function transformations like :func:`grad` and
 :func:`vmap` and graph optimizations.
 
@@ -116,7 +116,7 @@ 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
@@ -62,7 +66,7 @@ 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.

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:
 
@@ -49,7 +49,7 @@ it will be added. You can load the array with:
 
 .. code-block:: shell
 
-   >>> mx.load("array.npy", a)
+   >>> mx.load("array.npy")
    array([1], dtype=float32)
 
 Here's an example of saving several arrays to a single 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/cpp/CMakeLists.txt

@@ -8,3 +8,5 @@ endfunction(build_example)
 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"
+
+using namespace mlx::core;
+
+int main() {
+  if (!distributed::is_available()) {
+    std::cout << "No communication backend found" << std::endl;
+    return 1;
+  }
+
+  auto global_group = distributed::init();
+  std::cout << global_group.rank() << " / " << global_group.size() << std::endl;
+
+  array x = ones({10});
+  array out = distributed::all_sum(x, global_group);
+
+  std::cout << out << std::endl;
+}

examples/cpp/metal_capture.cpp

@@ -0,0 +1,31 @@
+// Copyright © 2024 Apple Inc.
+
+#include <cassert>
+#include <iostream>
+
+#include "mlx/mlx.h"
+
+using namespace mlx::core;
+
+int main() {
+  // To use Metal debugging and profiling:
+  // 1. Build with the MLX_METAL_DEBUG CMake option (i.e. -DMLX_METAL_DEBUG=ON).
+  // 2. Run with MTL_CAPTURE_ENABLED=1.
+  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 a = arange(1.f, 10.f, 1.f, float32, s2);
+  auto b = arange(1.f, 10.f, 1.f, float32, s3);
+  auto x = add(a, a, s2);
+  auto y = add(b, b, s3);
+
+  // The multiply will happen on the default stream.
+  std::cout << multiply(x, y) << std::endl;
+
+  metal::stop_capture();
+}

examples/cpp/tutorial.cpp

@@ -89,8 +89,8 @@ void automatic_differentiation() {
   // dfdx is 2 * x
 
   // Get the second derivative by composing grad with grad
-  auto df2dx2 = grad(grad(fn))(x);
-  // df2dx2 is 2
+  auto d2fdx2 = grad(grad(fn))(x);
+  // d2fdx2 is 2
 }
 
 int main() {

examples/extensions/CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.27)
 
-project(mlx_sample_extensions LANGUAGES CXX)
+project(_ext LANGUAGES CXX)
 
 # ----------------------------- Setup -----------------------------
 set(CMAKE_CXX_STANDARD 17)
@@ -11,8 +11,16 @@ option(BUILD_SHARED_LIBS "Build extensions as a shared library" ON)
 
 # ----------------------------- Dependencies -----------------------------
 find_package(MLX CONFIG REQUIRED)
-find_package(Python COMPONENTS Interpreter Development)
-find_package(pybind11 CONFIG 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}")
+find_package(nanobind CONFIG REQUIRED)
 
 # ----------------------------- Extensions -----------------------------
 
@@ -20,16 +28,10 @@ find_package(pybind11 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)
@@ -38,29 +40,35 @@ 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()
 
-# ----------------------------- Pybind -----------------------------
-pybind11_add_module(
-  mlx_sample_extensions
-  ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
-)
-target_link_libraries(mlx_sample_extensions PRIVATE mlx_ext)
+# ----------------------------- Python Bindings -----------------------------
+nanobind_add_module(
+  _ext
+  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)
-  target_link_options(mlx_sample_extensions PRIVATE -Wl,-rpath,@loader_path)
+  target_link_options(_ext PRIVATE -Wl,-rpath,@loader_path)
 endif()

examples/extensions/README.md

@@ -0,0 +1,24 @@
+
+## Build
+
+```
+pip install -e .
+```
+
+For faster builds during development, you can also pre-install the requirements:
+
+```
+pip install -r requirements.txt
+```
+
+And then run:
+
+```
+python setup.py build_ext -j8 --inplace
+```
+
+## Test
+
+```
+python test.py
+```

examples/extensions/axpby/axpby.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <iostream>
@@ -43,7 +43,7 @@ array axpby(
   auto promoted_dtype = promote_types(x.dtype(), y.dtype());
 
   // Upcast to float32 for non-floating point inputs x and y
-  auto out_dtype = is_floating_point(promoted_dtype)
+  auto out_dtype = issubdtype(promoted_dtype, float32)
       ? promoted_dtype
       : promote_types(promoted_dtype, float32);
 
@@ -61,7 +61,7 @@ array axpby(
       /* const std::vector<int>& shape = */ out_shape,
       /* Dtype dtype = */ out_dtype,
       /* std::unique_ptr<Primitive> primitive = */
-      std::make_unique<Axpby>(to_stream(s), alpha, beta),
+      std::make_shared<Axpby>(to_stream(s), alpha, beta),
       /* const std::vector<array>& inputs = */ broadcasted_inputs);
 }
 
@@ -106,12 +106,12 @@ void axpby_impl(
 /** Fall back implementation for evaluation on CPU */
 void Axpby::eval(
     const std::vector<array>& inputs,
-    std::vector<array>& out_arr) {
-  auto out = out_arr[0];
+    std::vector<array>& outputs) {
   // Check the inputs (registered in the op while constructing the out array)
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Dispatch to the correct dtype
   if (out.dtype() == float32) {
@@ -150,11 +150,7 @@ 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(y.data_size() * out.itemsize()),
-      y.data_size(),
-      y.strides(),
-      y.flags());
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
   // We then copy over the elements using the contiguous vector specialization
   copy_inplace(y, out, CopyType::Vector);
@@ -180,11 +176,11 @@ void axpby_impl_accelerate(
 /** Evaluate primitive on CPU using accelerate specializations */
 void Axpby::eval_cpu(
     const std::vector<array>& inputs,
-    std::vector<array>& outarr) {
-  auto out = outarr[0];
+    std::vector<array>& outputs) {
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Accelerate specialization for contiguous single precision float arrays
   if (out.dtype() == float32 &&
@@ -195,7 +191,7 @@ void Axpby::eval_cpu(
   }
 
   // Fall back to common backend if specializations are not available
-  eval(inputs, outarr);
+  eval(inputs, outputs);
 }
 
 #else // Accelerate not available
@@ -203,8 +199,8 @@ void Axpby::eval_cpu(
 /** Evaluate primitive on CPU falling back to common backend */
 void Axpby::eval_cpu(
     const std::vector<array>& inputs,
-    std::vector<array>& out) {
-  eval(inputs, out);
+    const std::vector<array>& outputs) {
+  eval(inputs, outputs);
 }
 
 #endif
@@ -218,12 +214,12 @@ void Axpby::eval_cpu(
 /** Evaluate primitive on GPU */
 void Axpby::eval_gpu(
     const std::vector<array>& inputs,
-    std::vector<array>& outarr) {
+    std::vector<array>& outputs) {
   // Prepare inputs
-  auto out = outarr[0];
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Each primitive carries the stream it should execute on
   // and each stream carries its device identifiers
@@ -253,15 +249,14 @@ 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);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
 
   // Kernel parameters are registered with buffer indices corresponding to
@@ -270,11 +265,11 @@ 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);
@@ -300,7 +295,7 @@ 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.dispatchThreads(grid_dims, group_dims);
 }
 
 #else // Metal is not available
@@ -372,4 +367,4 @@ bool Axpby::is_equivalent(const Primitive& other) const {
   return alpha_ == r_other.alpha_ && beta_ == r_other.beta_;
 }
 
-} // namespace mlx::core
+} // namespace mlx::core

examples/extensions/axpby/axpby.h

@@ -33,7 +33,7 @@ array axpby(
 class Axpby : public Primitive {
  public:
   explicit Axpby(Stream stream, float alpha, float beta)
-      : Primitive(stream), alpha_(alpha), beta_(beta){};
+      : Primitive(stream), alpha_(alpha), beta_(beta) {};
 
   /**
    * A primitive must know how to evaluate itself on the CPU/GPU
@@ -42,9 +42,9 @@ 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>& out)
+  void eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs)
       override;
-  void eval_gpu(const std::vector<array>& inputs, std::vector<array>& out)
+  void eval_gpu(const std::vector<array>& inputs, std::vector<array>& outputs)
       override;
 
   /** The Jacobian-vector product. */
@@ -83,7 +83,7 @@ class Axpby : public Primitive {
   float beta_;
 
   /** Fall back implementation for evaluation on CPU */
-  void eval(const std::vector<array>& inputs, std::vector<array>& out);
+  void eval(const std::vector<array>& inputs, std::vector<array>& outputs);
 };
 
-} // namespace mlx::core
+} // namespace mlx::core

examples/extensions/axpby/axpby.metal

@@ -19,7 +19,7 @@ template <typename T>
     uint index [[thread_position_in_grid]]) {
   auto x_offset = elem_to_loc(index, shape, x_strides, ndim);
   auto y_offset = elem_to_loc(index, shape, y_strides, ndim);
-  out[index] = 
+  out[index] =
       static_cast<T>(alpha) * x[x_offset] + static_cast<T>(beta) * y[y_offset];
 }
 
@@ -31,30 +31,30 @@ template <typename T>
     constant const float& alpha [[buffer(3)]],
     constant const float& beta [[buffer(4)]],
     uint index [[thread_position_in_grid]]) {
-  out[index] = 
+  out[index] =
       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)]],          \
+#define instantiate_axpby(type_name, type)                               \
+  template [[host_name("axpby_general_" #type_name)]] [[kernel]] void    \
+  axpby_general<type>(                                                   \
+      device const type* x [[buffer(0)]],                                \
+      device const type* y [[buffer(1)]],                                \
+      device type* out [[buffer(2)]],                                    \
+      constant const float& alpha [[buffer(3)]],                         \
+      constant const float& beta [[buffer(4)]],                          \
+      constant const int* shape [[buffer(5)]],                           \
+      constant const size_t* x_strides [[buffer(6)]],                    \
+      constant const size_t* y_strides [[buffer(7)]],                    \
+      constant const int& ndim [[buffer(8)]],                            \
+      uint index [[thread_position_in_grid]]);                           \
+  template [[host_name("axpby_contiguous_" #type_name)]] [[kernel]] void \
+  axpby_contiguous<type>(                                                \
+      device const type* x [[buffer(0)]],                                \
+      device const type* y [[buffer(1)]],                                \
+      device type* out [[buffer(2)]],                                    \
+      constant const float& alpha [[buffer(3)]],                         \
+      constant const float& beta [[buffer(4)]],                          \
       uint index [[thread_position_in_grid]]);
 
 instantiate_axpby(float32, float);

examples/extensions/bindings.cpp

@@ -1,31 +1,31 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
-#include <pybind11/pybind11.h>
-#include <pybind11/stl.h>
+#include <nanobind/nanobind.h>
+#include <nanobind/stl/variant.h>
 
 #include "axpby/axpby.h"
 
-namespace py = pybind11;
-using namespace py::literals;
+namespace nb = nanobind;
+using namespace nb::literals;
+
 using namespace mlx::core;
 
-PYBIND11_MODULE(mlx_sample_extensions, m) {
-  m.doc() = "Sample C++ and metal extensions for MLX";
+NB_MODULE(_ext, m) {
+  m.doc() = "Sample extension for MLX";
 
   m.def(
       "axpby",
       &axpby,
       "x"_a,
       "y"_a,
-      py::pos_only(),
       "alpha"_a,
       "beta"_a,
-      py::kw_only(),
-      "stream"_a = py::none(),
-      R"pbdoc(
+      nb::kw_only(),
+      "stream"_a = nb::none(),
+      R"(
         Scale and sum two vectors element-wise
         ``z = alpha * x + beta * y``
-        
+
         Follows numpy style broadcasting between ``x`` and ``y``
         Inputs are upcasted to floats if needed
 
@@ -37,5 +37,5 @@ PYBIND11_MODULE(mlx_sample_extensions, m) {
 
         Returns:
             array: ``alpha * x + beta * y``
-      )pbdoc");
-}
+      )");
+}

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

@@ -1,3 +1,8 @@
 [build-system]
-requires = ["setuptools>=42", "pybind11>=2.10", "cmake>=3.24", "mlx @ git+https://github.com/mlx-explore/mlx@main"]
-build-backend = "setuptools.build_meta"
+requires = [
+  "setuptools>=42",
+  "cmake>=3.24",
+  "mlx>=0.18.0",
+  "nanobind==2.2.0",
+]
+build-backend = "setuptools.build_meta"

examples/extensions/requirements.txt

@@ -0,0 +1,4 @@
+setuptools>=42
+cmake>=3.24
+mlx>=0.18.1
+nanobind==2.2.0

examples/extensions/setup.py

@@ -1,4 +1,4 @@
-# Copyright © 2023 Apple Inc.
+# Copyright © 2023-2024 Apple Inc.
 
 from setuptools import setup
 
@@ -9,10 +9,9 @@ if __name__ == "__main__":
         name="mlx_sample_extensions",
         version="0.0.0",
         description="Sample C++ and Metal extensions for MLX primitives.",
-        ext_modules=[extension.CMakeExtension("mlx_sample_extensions")],
+        ext_modules=[extension.CMakeExtension("mlx_sample_extensions._ext")],
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
-        package_dir={"": "."},
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
         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/CMakeLists.txt

@@ -1,37 +1,40 @@
 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}/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(${CMAKE_CURRENT_SOURCE_DIR}/backend/accelerate)
+if(MLX_BUILD_CPU)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
 else()
-  target_sources(
-    mlx
-    PRIVATE
-    ${CMAKE_CURRENT_SOURCE_DIR}/backend/common/default_primitives.cpp
-  )
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_cpu)
 endif()
 
-if (MLX_BUILD_METAL)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/distributed)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
+if(MLX_BUILD_ACCELERATE)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/accelerate)
+elseif(MLX_BUILD_CPU)
+  target_sources(
+    mlx
+    PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/common/default_primitives.cpp)
+endif()
+
+if(MLX_BUILD_METAL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/metal)
 else()
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_metal)

mlx/allocator.cpp

@@ -23,11 +23,22 @@ void free(Buffer 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

@@ -14,7 +14,7 @@ class Buffer {
   void* ptr_;
 
  public:
-  Buffer(void* ptr) : ptr_(ptr){};
+  Buffer(void* ptr) : ptr_(ptr) {};
 
   // Get the raw data pointer from the buffer
   void* raw_ptr();
@@ -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 +1,4 @@
 // Copyright © 2023-2024 Apple Inc.
-
 #include <functional>
 
 #include "mlx/array.h"
@@ -12,22 +11,16 @@ namespace mlx::core {
 
 namespace {
 
-std::pair<size_t, std::vector<size_t>> cum_prod(const std::vector<int>& shape) {
-  std::vector<size_t> strides(shape.size());
-  size_t cum_prod = 1;
-  for (int i = shape.size() - 1; i >= 0; --i) {
-    strides[i] = cum_prod;
-    cum_prod *= shape[i];
-  }
-  return {cum_prod, strides};
-}
-
 /** Return true if we are currently performing a function transformation in
  * order to keep the graph when evaluating tracer arrays. */
 bool in_tracing() {
   return detail::InTracing::in_tracing();
 }
 
+bool retain_graph() {
+  return detail::RetainGraph::retain_graph();
+}
+
 } // namespace
 
 array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
@@ -36,22 +29,11 @@ array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
   init(&cval);
 }
 
-array::array(
-    const std::vector<int>& shape,
-    Dtype dtype,
-    std::shared_ptr<Primitive> primitive,
-    const std::vector<array>& inputs)
-    : array_desc_(std::make_shared<ArrayDesc>(
-          shape,
-          dtype,
-          std::move(primitive),
-          inputs)) {}
-
 array::array(
     std::vector<int> shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
-    std::vector<array>&& inputs)
+    std::vector<array> inputs)
     : array_desc_(std::make_shared<ArrayDesc>(
           std::move(shape),
           dtype,
@@ -59,15 +41,16 @@ array::array(
           std::move(inputs))) {}
 
 std::vector<array> array::make_arrays(
-    const std::vector<std::vector<int>>& shapes,
+    std::vector<std::vector<int>> shapes,
     const std::vector<Dtype>& dtypes,
-    std::shared_ptr<Primitive> primitive,
+    const std::shared_ptr<Primitive>& primitive,
     const std::vector<array>& inputs) {
   std::vector<array> outputs;
-  for (int i = 0; i < shapes.size(); ++i) {
-    outputs.push_back(array(shapes[i], dtypes[i], primitive, inputs));
+  for (size_t i = 0; i < shapes.size(); ++i) {
+    outputs.emplace_back(std::move(shapes[i]), dtypes[i], primitive, inputs);
   }
-  for (int i = 0; i < outputs.size(); ++i) {
+  // For each node in |outputs|, its siblings are the other nodes.
+  for (size_t i = 0; i < outputs.size(); ++i) {
     auto siblings = outputs;
     siblings.erase(siblings.begin() + i);
     outputs[i].set_siblings(std::move(siblings), i);
@@ -92,10 +75,10 @@ array::array(std::initializer_list<int> data, Dtype dtype)
 /* Build an array from a shared buffer */
 array::array(
     allocator::Buffer data,
-    const std::vector<int>& shape,
+    std::vector<int> shape,
     Dtype dtype,
     deleter_t deleter)
-    : array_desc_(std::make_shared<ArrayDesc>(shape, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   set_data(data, deleter);
 }
 
@@ -104,22 +87,42 @@ void array::detach() {
     s.array_desc_->inputs.clear();
     s.array_desc_->siblings.clear();
     s.array_desc_->position = 0;
-    s.array_desc_->depth = 0;
     s.array_desc_->primitive = nullptr;
   }
   array_desc_->inputs.clear();
   array_desc_->siblings.clear();
   array_desc_->position = 0;
-  array_desc_->depth = 0;
   array_desc_->primitive = nullptr;
 }
 
+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);
+  }
+}
+
 void array::eval() {
-  mlx::core::eval({*this});
+  // 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 && in_tracing() || retain_graph();
 }
 
 void array::set_data(allocator::Buffer buffer, deleter_t d) {
@@ -164,51 +167,129 @@ void array::copy_shared_buffer(const array& other) {
   copy_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
-void array::move_shared_buffer(array other) {
+void array::move_shared_buffer(
+    array other,
+    const std::vector<size_t>& strides,
+    Flags flags,
+    size_t data_size,
+    size_t offset /* = 0 */) {
   array_desc_->data = std::move(other.array_desc_->data);
-  array_desc_->strides = other.strides();
-  array_desc_->flags = other.flags();
-  array_desc_->data_size = other.data_size();
-  array_desc_->data_ptr = other.array_desc_->data_ptr;
+  array_desc_->strides = strides;
+  array_desc_->flags = flags;
+  array_desc_->data_size = data_size;
+  auto char_offset = sizeof(char) * itemsize() * offset;
+  auto data_ptr = other.array_desc_->data_ptr;
+  other.array_desc_->data_ptr = nullptr;
+  array_desc_->data_ptr =
+      static_cast<void*>(static_cast<char*>(data_ptr) + char_offset);
 }
 
-array::ArrayDesc::ArrayDesc(const std::vector<int>& shape, Dtype dtype)
-    : shape(shape), dtype(dtype) {
-  std::tie(size, strides) = cum_prod(shape);
+void array::move_shared_buffer(array other) {
+  move_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
-array::ArrayDesc::ArrayDesc(
-    const std::vector<int>& shape,
-    Dtype dtype,
-    std::shared_ptr<Primitive> primitive,
-    const std::vector<array>& inputs)
-    : shape(shape),
-      dtype(dtype),
-      primitive(std::move(primitive)),
-      inputs(inputs) {
-  std::tie(size, strides) = cum_prod(this->shape);
-  for (auto& in : this->inputs) {
-    is_tracer |= in.is_tracer();
-    depth = std::max(in.graph_depth(), depth);
+array::~array() {
+  if (array_desc_ == nullptr) {
+    return;
   }
-  depth++;
+
+  // 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;
+    // If all siblings have siblings.size() references except
+    // the one we are currently destroying (which has siblings.size() + 1)
+    // then there are no more external references
+    do_detach &= (array_desc_.use_count() == (n + 1));
+    for (auto& s : siblings()) {
+      do_detach &= (s.array_desc_.use_count() == n);
+      if (!do_detach) {
+        break;
+      }
+    }
+    if (do_detach) {
+      for (auto& s : siblings()) {
+        for (auto& ss : s.siblings()) {
+          ss.array_desc_ = nullptr;
+        }
+        s.array_desc_->siblings.clear();
+      }
+    }
+  }
+}
+
+void array::ArrayDesc::init() {
+  strides.resize(shape.size());
+  size = 1;
+  for (int i = shape.size() - 1; i >= 0; --i) {
+    strides[i] = size;
+    size *= shape[i];
+  }
+  for (const auto& in : inputs) {
+    is_tracer |= in.is_tracer();
+  }
+}
+
+array::ArrayDesc::ArrayDesc(std::vector<int> shape, Dtype dtype)
+    : shape(std::move(shape)), dtype(dtype), status(Status::available) {
+  init();
 }
 
 array::ArrayDesc::ArrayDesc(
-    std::vector<int>&& shape,
+    std::vector<int> shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
-    std::vector<array>&& inputs)
+    std::vector<array> inputs)
     : shape(std::move(shape)),
       dtype(dtype),
+      status(Status::unscheduled),
       primitive(std::move(primitive)),
       inputs(std::move(inputs)) {
-  std::tie(size, strides) = cum_prod(this->shape);
-  for (auto& in : this->inputs) {
-    is_tracer |= in.is_tracer();
-    depth = std::max(in.graph_depth(), depth);
+  init();
+}
+
+array::ArrayDesc::~ArrayDesc() {
+  // When an array description is destroyed it will delete a bunch of arrays
+  // 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;
+
+  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});
+      }
+    }
+    ad.inputs.clear();
+    for (auto& [_, a] : input_map) {
+      if (a.array_desc_.use_count() <= a.siblings().size() + 1) {
+        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);
   }
-  depth++;
 }
 
 array::ArrayIterator::ArrayIterator(const array& arr, int idx)

mlx/array.h

@@ -1,5 +1,6 @@
 // Copyright © 2023 Apple Inc.
 #pragma once
+
 #include <algorithm>
 #include <cstdint>
 #include <functional>
@@ -8,6 +9,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/dtype.h"
+#include "mlx/event.h"
 
 namespace mlx::core {
 
@@ -31,7 +33,7 @@ class array {
   template <typename It>
   array(
       It data,
-      const std::vector<int>& shape,
+      std::vector<int> shape,
       Dtype dtype =
           TypeToDtype<typename std::iterator_traits<It>::value_type>());
 
@@ -47,13 +49,13 @@ class array {
   template <typename T>
   array(
       std::initializer_list<T> data,
-      const std::vector<int>& shape,
+      std::vector<int> shape,
       Dtype dtype = TypeToDtype<T>());
 
   /* Build an array from a buffer */
   array(
       allocator::Buffer data,
-      const std::vector<int>& shape,
+      std::vector<int> shape,
       Dtype dtype,
       deleter_t deleter = allocator::free);
 
@@ -71,32 +73,32 @@ 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 {
     return array_desc_->shape;
-  };
+  }
 
   /**
    *  Get the size of the corresponding dimension.
@@ -105,17 +107,26 @@ class array {
    *  bounds checking. */
   int shape(int dim) const {
     return shape().at(dim < 0 ? dim + ndim() : dim);
-  };
+  }
 
   /** The strides of the array. */
   const std::vector<size_t>& 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 {
+    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();
@@ -149,10 +160,10 @@ 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;
@@ -172,22 +183,16 @@ class array {
    * API may change.
    */
 
-  array(
-      const std::vector<int>& shape,
-      Dtype dtype,
-      std::shared_ptr<Primitive> primitive,
-      const std::vector<array>& inputs);
-
   array(
       std::vector<int> shape,
       Dtype dtype,
       std::shared_ptr<Primitive> primitive,
-      std::vector<array>&& inputs);
+      std::vector<array> inputs);
 
   static std::vector<array> make_arrays(
-      const std::vector<std::vector<int>>& shapes,
+      std::vector<std::vector<int>> shapes,
       const std::vector<Dtype>& dtypes,
-      std::shared_ptr<Primitive> primitive,
+      const std::shared_ptr<Primitive>& primitive,
       const std::vector<array>& inputs);
 
   /** A unique identifier for an array. */
@@ -204,7 +209,7 @@ class array {
     allocator::Buffer buffer;
     deleter_t d;
     Data(allocator::Buffer buffer, deleter_t d = allocator::free)
-        : buffer(buffer), d(d){};
+        : buffer(buffer), d(d) {}
     // Not copyable
     Data(const Data& d) = delete;
     Data& operator=(const Data& d) = delete;
@@ -214,33 +219,45 @@ 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;
   };
 
   /** 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;
@@ -254,7 +271,12 @@ 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);
@@ -271,11 +293,6 @@ class array {
     outputs.push_back(*this);
     outputs.insert(outputs.end(), siblings().begin() + idx, siblings().end());
     return outputs;
-  };
-
-  /** The depth of the array in the graph. Evaluated arrays have depth 0. */
-  uint16_t graph_depth() const {
-    return array_desc_->depth;
   }
 
   /** Detach the array from the graph. */
@@ -284,19 +301,32 @@ 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
@@ -307,16 +337,57 @@ 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
   };
 
-  // Check if the array has been evaluated
-  bool is_evaled() const {
-    return array_desc_->data != nullptr;
+  // Check if the array is safe to read.
+  bool is_available() const;
+
+  // Wait on the array to be available. After this `is_available` returns
+  // `true`.
+  void wait();
+
+  Status status() const {
+    return array_desc_->status;
+  }
+
+  void set_status(Status s) const {
+    array_desc_->status = s;
+  }
+
+  // Get the array's shared event
+  Event& event() const {
+    return array_desc_->event;
+  }
+
+  // Attach an event to a not yet evaluated array
+  void attach_event(Event e) const {
+    array_desc_->event = std::move(e);
   }
 
   // Mark the array as a tracer array (true) or not.
@@ -344,12 +415,21 @@ class array {
 
   void copy_shared_buffer(const array& other);
 
+  void move_shared_buffer(
+      array other,
+      const std::vector<size_t>& strides,
+      Flags flags,
+      size_t data_size,
+      size_t offset = 0);
+
   void move_shared_buffer(array other);
 
   void overwrite_descriptor(const array& other) {
     array_desc_ = other.array_desc_;
   }
 
+  ~array();
+
  private:
   // Initialize the arrays data
   template <typename It>
@@ -360,7 +440,12 @@ class array {
     std::vector<size_t> strides;
     size_t size;
     Dtype dtype;
-    std::shared_ptr<Primitive> primitive{nullptr};
+    std::shared_ptr<Primitive> primitive;
+
+    Status status;
+
+    // An event on the array used for synchronization
+    Event event;
 
     // Indicates an array is being used in a graph transform
     // and should not be detached from the graph
@@ -368,14 +453,12 @@ class array {
 
     // This is a shared pointer so that *different* arrays
     // can share the underlying data buffer.
-    std::shared_ptr<Data> data{nullptr};
+    std::shared_ptr<Data> data;
 
     // Properly offset data pointer
     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
@@ -388,29 +471,26 @@ class array {
     // The arrays position in the output list
     uint32_t position{0};
 
-    // The depth of the array in the graph.
-    uint16_t depth{0};
-
-    explicit ArrayDesc(const std::vector<int>& shape, Dtype dtype);
+    explicit ArrayDesc(std::vector<int> shape, Dtype dtype);
 
     explicit ArrayDesc(
-        const std::vector<int>& shape,
+        std::vector<int> shape,
         Dtype dtype,
         std::shared_ptr<Primitive> primitive,
-        const std::vector<array>& inputs);
+        std::vector<array> inputs);
 
-    explicit ArrayDesc(
-        std::vector<int>&& shape,
-        Dtype dtype,
-        std::shared_ptr<Primitive> primitive,
-        std::vector<array>&& inputs);
+    ~ArrayDesc();
+
+   private:
+    // Initialize size, strides, and other metadata
+    void init();
   };
 
   // The ArrayDesc contains the details of the materialized array including the
   // shape, strides, the data type. It also includes
   // the primitive which knows how to compute the array's data from its inputs
   // and the list of array's inputs for the primitive.
-  std::shared_ptr<ArrayDesc> array_desc_{nullptr};
+  std::shared_ptr<ArrayDesc> array_desc_;
 };
 
 template <typename T>
@@ -422,9 +502,9 @@ array::array(T val, Dtype dtype /* = TypeToDtype<T>() */)
 template <typename It>
 array::array(
   It data,
-  const std::vector<int>& shape,
+  std::vector<int> shape,
   Dtype dtype /* = TypeToDtype<typename std::iterator_traits<It>::value_type>() */) :
-    array_desc_(std::make_shared<ArrayDesc>(shape, dtype)) {
+    array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   init(data);
 }
 
@@ -441,9 +521,9 @@ array::array(
 template <typename T>
 array::array(
     std::initializer_list<T> data,
-    const std::vector<int>& shape,
+    std::vector<int> shape,
     Dtype dtype /* = TypeToDtype<T>() */)
-    : array_desc_(std::make_shared<ArrayDesc>(shape, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   if (data.size() != size()) {
     throw std::invalid_argument(
         "Data size and provided shape mismatch in array construction.");
@@ -465,10 +545,11 @@ T array::item() const {
   if (size() != 1) {
     throw std::invalid_argument("item can only be called on arrays of size 1.");
   }
-  if (!is_evaled()) {
+  if (status() == Status::unscheduled) {
     throw std::invalid_argument(
         "item() const can only be called on evaled arrays");
   }
+  const_cast<array*>(this)->eval();
   return *data<T>();
 }
 
@@ -518,4 +599,15 @@ void array::init(It src) {
   }
 }
 
+/* Utilities for determining whether a template parameter is array. */
+template <typename T>
+inline constexpr bool is_array_v =
+    std::is_same_v<std::remove_cv_t<std::remove_reference_t<T>>, array>;
+
+template <typename... T>
+inline constexpr bool is_arrays_v = (is_array_v<T> && ...);
+
+template <typename... T>
+using enable_for_arrays_t = typename std::enable_if_t<is_arrays_v<T...>>;
+
 } // namespace mlx::core

mlx/backend/accelerate/CMakeLists.txt

@@ -1,10 +1,8 @@
 target_sources(
   mlx
-  PRIVATE
-  ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
-)
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp)

mlx/backend/accelerate/conv.cpp

@@ -1,9 +1,9 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 
+#include <Accelerate/Accelerate.h>
 #include <simd/vector.h>
-#include <vecLib/vDSP.h>
 
 #include "mlx/backend/common/copy.h"
 #include "mlx/primitives.h"

mlx/backend/accelerate/matmul.cpp

@@ -1,9 +1,8 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 
-#include <vecLib/BNNS/bnns.h>
-#include <vecLib/cblas_new.h>
+#include <Accelerate/Accelerate.h>
 
 #include "mlx/backend/accelerate/utils.h"
 #include "mlx/backend/common/copy.h"
@@ -196,6 +195,40 @@ inline void matmul_bnns(const array& a_pre, const array& b_pre, array& out) {
   return matmul_bnns_general(a_pre, b_pre, out);
 }
 
+template <typename T>
+inline void mask_matrix(
+    T* data,
+    const bool* mask,
+    int tile_size,
+    const int X,
+    const int Y,
+    const size_t X_data_str,
+    const size_t Y_data_str,
+    const size_t X_mask_str,
+    const size_t Y_mask_str) {
+  int tX = (X + tile_size - 1) / tile_size;
+  int tY = (Y + tile_size - 1) / tile_size;
+
+  for (int i = 0; i < tX; i++) {
+    for (int j = 0; j < tY; j++) {
+      bool do_mask = mask[i * X_mask_str + j * Y_mask_str];
+      if (!do_mask) {
+        int loc_x = i * tile_size;
+        int loc_y = j * tile_size;
+        T* data_block = data + loc_x * X_data_str + loc_y * Y_data_str;
+
+        int size_x = std::min(tile_size, X - loc_x);
+        int size_y = std::min(tile_size, Y - loc_y);
+        for (int ii = 0; ii < size_x; ii++) {
+          for (int jj = 0; jj < size_y; jj++) {
+            data_block[ii * X_data_str + jj * Y_data_str] = T(0.);
+          }
+        }
+      }
+    }
+  }
+}
+
 } // namespace
 
 void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {

mlx/backend/accelerate/primitives.cpp

@@ -3,8 +3,7 @@
 #include <cassert>
 #include <cmath>
 
-#include <vecLib/vDSP.h>
-#include <vecLib/vForce.h>
+#include <Accelerate/Accelerate.h>
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/binary.h"
@@ -31,21 +30,27 @@ DEFAULT(ArgPartition)
 DEFAULT(ArgReduce)
 DEFAULT(ArgSort)
 DEFAULT(AsStrided)
+DEFAULT(BlockMaskedMM)
 DEFAULT(Broadcast)
 DEFAULT(Ceil)
 DEFAULT(Concatenate)
+DEFAULT(Conjugate)
 DEFAULT(Copy)
-DEFAULT_MULTI(CustomVJP)
+DEFAULT_MULTI(CustomTransforms)
 DEFAULT_MULTI(Depends)
 DEFAULT_MULTI(DivMod)
+DEFAULT(NumberOfElements)
 DEFAULT(Equal)
 DEFAULT(Erf)
 DEFAULT(ErfInv)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Gather)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
 DEFAULT(Less)
 DEFAULT(LessEqual)
 DEFAULT(Load)
@@ -68,10 +73,15 @@ DEFAULT(Select)
 DEFAULT(Sigmoid)
 DEFAULT(Sign)
 DEFAULT(Slice)
+DEFAULT(SliceUpdate)
 DEFAULT_MULTI(Split)
 DEFAULT(Sort)
 DEFAULT(StopGradient)
+DEFAULT_MULTI(SVD)
 DEFAULT(Transpose)
+DEFAULT(Inverse)
+DEFAULT(Cholesky)
+DEFAULT_MULTI(Eigh)
 
 void Abs::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
@@ -93,7 +103,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -108,7 +118,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vadd((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
         });
   } else if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -123,7 +133,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vaddi((const int*)a, 1, (const int*)b, 1, (int*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x + y; });
+    eval(inputs, out);
   }
 }
 
@@ -187,6 +197,26 @@ void ArcTan::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  if (out.dtype() == float32 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    if (a.is_donatable()) {
+      out.copy_shared_buffer(a);
+    } else if (b.is_donatable()) {
+      out.copy_shared_buffer(b);
+    } else {
+      out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    }
+    int size = a.data_size();
+    vvatan2f(out.data<float>(), a.data<float>(), b.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
 void ArcTanh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -258,7 +288,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -271,7 +301,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdivi((const int*)b, 1, (const int*)a, 1, (int*)o, 1, n);
         });
   } else if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -286,7 +316,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdiv((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x / y; });
+    eval(inputs, out);
   }
 }
 
@@ -297,12 +327,21 @@ void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
     set_unary_output_data(in, out);
     auto size = in.data_size();
     vvexpf(out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else if (is_floating_point(out.dtype())) {
-    unary_fp(in, out, [](auto x) { return std::exp(x); });
   } else {
-    throw std::invalid_argument(
-        "[exp] Cannot exponentiate elements in array"
-        " with non floating point type.");
+    eval(inputs, out);
+  }
+}
+
+void Expm1::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vvexpm1f(
+        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+  } else {
+    eval(inputs, out);
   }
 }
 
@@ -351,12 +390,8 @@ void Log1p::eval_cpu(const std::vector<array>& inputs, array& out) {
     auto size = in.data_size();
     vvlog1pf(
         out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else if (is_floating_point(out.dtype())) {
-    unary_fp(in, out, [](auto x) { return std::log1p(x); });
   } else {
-    throw std::invalid_argument(
-        "[log1p] Cannot compute log of elements in array with"
-        " non floating point type.");
+    eval(inputs, out);
   }
 }
 
@@ -366,7 +401,7 @@ void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -381,7 +416,7 @@ void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vmul((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x * y; });
+    eval(inputs, out);
   }
 }
 
@@ -392,7 +427,7 @@ void Negative::eval_cpu(const std::vector<array>& inputs, array& out) {
     set_unary_output_data(in, out);
     vDSP_vneg(in.data<float>(), 1, out.data<float>(), 1, in.data_size());
   } else {
-    unary(in, out, [](auto x) { return -x; });
+    eval(inputs, out);
   }
 }
 
@@ -479,7 +514,7 @@ void Square::eval_cpu(const std::vector<array>& inputs, array& out) {
     auto size = in.data_size();
     vDSP_vsq(in.data<float>(), 1, out.data<float>(), 1, size);
   } else {
-    unary(in, out, [](auto x) { return x * x; });
+    eval(inputs, out);
   }
 }
 
@@ -505,7 +540,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -523,7 +558,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vsub((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
         });
   } else if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -535,7 +570,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
         },
         UseDefaultBinaryOp());
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x - y; });
+    eval(inputs, out);
   }
 }
 

mlx/backend/accelerate/quantized.cpp

@@ -18,49 +18,61 @@ void _qmm_t_4_64(
     const float* biases,
     int M,
     int N,
-    int K) {
+    int K,
+    int B,
+    bool batched_w) {
   constexpr int bits = 4;
   constexpr int group_size = 64;
   constexpr int bitmask = (1 << bits) - 1;
   constexpr int pack_factor = 32 / bits;
   constexpr int packs_in_group = group_size / pack_factor;
 
-  for (int m = 0; m < M; m++) {
-    const uint32_t* w_local = w;
-    const float* scales_local = scales;
-    const float* biases_local = biases;
+  int w_els = N * K / pack_factor;
+  int g_els = w_els * pack_factor / group_size;
 
-    for (int n = 0; n < N; n++) {
-      const simd_float16* x_local = (simd_float16*)x;
-      simd_float16 sum = 0;
-      for (int k = 0; k < K; k += group_size) {
-        float scale = *scales_local++;
-        float bias = *biases_local++;
+  for (int i = 0; i < B; i++) {
+    for (int m = 0; m < M; m++) {
+      const uint32_t* w_local = w;
+      const float* scales_local = scales;
+      const float* biases_local = biases;
 
-        for (int kw = 0; kw < packs_in_group; kw += 2) {
-          // TODO: vectorize this properly
-          simd_uint16 wi;
-          for (int e = 0; e < 2; e++) {
-            uint32_t wii = *w_local++;
-            for (int p = 0; p < 8; p++) {
-              wi[e * 8 + p] = wii & bitmask;
-              wii >>= bits;
+      for (int n = 0; n < N; n++) {
+        const simd_float16* x_local = (simd_float16*)x;
+        simd_float16 sum = 0;
+        for (int k = 0; k < K; k += group_size) {
+          float scale = *scales_local++;
+          float bias = *biases_local++;
+
+          for (int kw = 0; kw < packs_in_group; kw += 2) {
+            // TODO: vectorize this properly
+            simd_uint16 wi;
+            for (int e = 0; e < 2; e++) {
+              uint32_t wii = *w_local++;
+              for (int p = 0; p < 8; p++) {
+                wi[e * 8 + p] = wii & bitmask;
+                wii >>= bits;
+              }
             }
-          }
-          simd_float16 wf = simd_float(wi);
-          wf *= scale;
-          wf += bias;
+            simd_float16 wf = simd_float(wi);
+            wf *= scale;
+            wf += bias;
 
-          sum += (*x_local) * wf;
-          x_local++;
+            sum += (*x_local) * wf;
+            x_local++;
+          }
         }
+
+        *result = simd_reduce_add(sum);
+        result++;
       }
 
-      *result = simd_reduce_add(sum);
-      result++;
+      x += K;
+    }
+    if (batched_w) {
+      w += w_els;
+      scales += g_els;
+      biases += g_els;
     }
-
-    x += K;
   }
 }
 
@@ -82,8 +94,10 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
   if (condition) {
     out.set_data(allocator::malloc_or_wait(out.nbytes()));
     int K = x.shape(-1);
-    int M = x.size() / K;
+    int M = x.shape(-2);
     int N = out.shape(-1);
+    int B = x.size() / K / M;
+    bool batched_w = w.ndim() > 2;
     _qmm_t_4_64(
         out.data<float>(),
         x.data<float>(),
@@ -92,7 +106,9 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
         biases.data<float>(),
         M,
         N,
-        K);
+        K,
+        B,
+        batched_w);
   } else {
     eval(inputs, out);
   }

mlx/backend/accelerate/reduce.cpp

@@ -2,86 +2,73 @@
 
 #include <cassert>
 
+#include <Accelerate/Accelerate.h>
 #include <simd/vector.h>
-#include <vecLib/vDSP.h>
 
 #include "mlx/backend/common/reduce.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
-template <typename T, typename VT, int N>
-void _vectorized_strided_sum(const T* x, T* accum, int size, size_t stride) {
-  for (int i = 0; i < size; i++) {
-    size_t s = stride;
-    T* a = accum;
-    while (s >= N) {
-      VT val = (*(VT*)x);
-      *(VT*)a += val;
-      x += N;
-      a += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      *a++ += *x++;
-    }
-  }
-}
+namespace {
 
-// TODO: Add proper templates for the strided reduce algorithm so we don't have
-// to write max/min/sum etc.
-template <typename T, typename VT, int N>
-void _vectorized_strided_max(const T* x, T* accum, int size, size_t stride) {
-  for (int i = 0; i < size; i++) {
-    size_t s = stride;
-    T* a = accum;
-    while (s >= N) {
-      *(VT*)a = simd_max((*(VT*)x), (*(VT*)a));
-      x += N;
-      a += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      *a = std::max(*a, *x);
-      a++;
-      x++;
-    }
+template <typename T, typename VT>
+struct MinReduction {
+  T operator()(const T& a, const T& b) {
+    return std::min(a, b);
   }
-}
 
-template <typename T, typename VT, int N>
-void _vectorized_strided_min(const T* x, T* accum, int size, size_t stride) {
-  for (int i = 0; i < size; i++) {
-    size_t s = stride;
-    T* a = accum;
-    while (s >= N) {
-      *(VT*)a = simd_min((*(VT*)x), (*(VT*)a));
-      x += N;
-      a += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      *a = std::min(*a, *x);
-      a++;
-      x++;
-    }
+  VT operator()(VT a, VT b) {
+    return simd_min(a, b);
   }
-}
+};
 
-template <typename T, typename VT, int N>
-void _vectorized_sum(const T* x, T* accum, int size) {
-  VT _sum = {0};
-  while (size >= N) {
-    _sum += (*(VT*)x);
-    x += N;
-    size -= N;
+template <typename T, typename VT>
+struct MaxReduction {
+  T operator()(const T& a, const T& b) {
+    return std::max(a, b);
   }
-  T sum = _sum[0];
-  for (int i = 1; i < N; i++) {
-    sum += _sum[i];
+
+  VT operator()(VT a, VT b) {
+    return simd_max(a, b);
   }
-  *accum += sum;
-}
+};
+
+template <typename T, typename VT>
+struct SumReduction {
+  T operator()(const T& a, const T& b) {
+    return a + b;
+  }
+
+  VT operator()(VT a, VT b) {
+    return a + b;
+  }
+};
+
+template <typename T, typename VT, int N, typename Reduction>
+struct StridedReduce {
+  void operator()(const T* x, T* accum, int size, size_t stride) {
+    Reduction op;
+
+    for (int i = 0; i < size; i++) {
+      size_t s = stride;
+      T* a = accum;
+      while (s >= N) {
+        *(VT*)a = op((*(VT*)x), (*(VT*)a));
+        x += N;
+        a += N;
+        s -= N;
+      }
+      while (s-- > 0) {
+        *a = op(*a, *x);
+        a++;
+        x++;
+      }
+    }
+  }
+};
+
+} // namespace
 
 void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
@@ -94,10 +81,11 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
           out,
           axes_,
           0,
-          [](const auto* x, auto* accum, int size, size_t stride) {
-            _vectorized_strided_sum<float, simd_float16, 16>(
-                (const float*)x, (float*)accum, size, stride);
-          },
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              SumReduction<float, simd_float16>>(),
           [](const auto* x, auto* accum, int size) {
             float acc;
             vDSP_sve((const float*)x, 1, &acc, size);
@@ -111,10 +99,11 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
           out,
           axes_,
           -std::numeric_limits<float>::infinity(),
-          [](const auto* x, auto* accum, int size, size_t stride) {
-            _vectorized_strided_max<float, simd_float16, 16>(
-                (const float*)x, (float*)accum, size, stride);
-          },
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              MaxReduction<float, simd_float16>>(),
           [](const auto* x, auto* accum, int size) {
             float max;
             vDSP_maxv((const float*)x, 1, &max, size);
@@ -128,10 +117,11 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
           out,
           axes_,
           std::numeric_limits<float>::infinity(),
-          [](const auto* x, auto* accum, int size, size_t stride) {
-            _vectorized_strided_min<float, simd_float16, 16>(
-                (const float*)x, (float*)accum, size, stride);
-          },
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              MinReduction<float, simd_float16>>(),
           [](const auto* x, auto* accum, int size) {
             float min;
             vDSP_minv((const float*)x, 1, &min, size);

mlx/backend/accelerate/softmax.cpp

@@ -1,9 +1,12 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <limits>
 
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 #include <arm_neon.h>
+#endif
+
 #include <simd/math.h>
 #include <simd/vector.h>
 
@@ -30,8 +33,8 @@ namespace {
  * Note: The implementation below is a general fast exp. There could be faster
  *       implementations for numbers strictly < 0.
  */
-inline simd_float16 simd_fast_exp(simd_float16 x) {
-  x *= 1.442695; // multiply with log_2(e)
+inline simd_float16 simd_fast_exp(simd_float16 x_init) {
+  auto x = x_init * 1.442695; // multiply with log_2(e)
   simd_float16 ipart, fpart;
   simd_int16 epart;
   x = simd_clamp(x, -80, 80);
@@ -50,28 +53,30 @@ inline simd_float16 simd_fast_exp(simd_float16 x) {
   // bitshifting
   epart = (simd_int(ipart) + 127) << 23;
 
-  return (*(simd_float16*)&epart) * x;
+  // Avoid supressing NaNs
+  simd_int16 eq = (x_init == x_init);
+  return simd_bitselect(x_init, (*(simd_float16*)&epart) * x, eq);
 }
 
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 /**
  * The ARM neon equivalent of the fast exp above.
  */
 inline float16x8_t neon_fast_exp(float16x8_t x) {
-  x = vmulq_f16(x, vdupq_n_f16(1.442695)); // multiply with log_2(e)
-  x = vmaxq_f16(x, vdupq_n_f16(-14)); // clamp under with -14
-  x = vminq_f16(x, vdupq_n_f16(14)); // clamp over with 14
+  x = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
+  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f))); // clamp under with -14
+  x = vminq_f16(x, vdupq_n_f16(float16_t(14.f))); // clamp over with 14
 
-  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(0.5)));
+  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(float16_t(0.5f))));
   float16x8_t fpart = vsubq_f16(x, ipart);
 
-  x = vdupq_n_f16(1.535336188319500e-4f);
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(9.618437357674640e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(5.550332471162809e-2f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(2.402264791363012e-1f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(6.931472028550421e-1f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.000000000000000f), x, fpart);
+  x = vdupq_n_f16(float16_t(1.535336188319500e-4f));
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.000000000000000f)), x, fpart);
 
   // generate 2**ipart in the floating point representation using integer
   // bitshifting
@@ -107,53 +112,6 @@ inline float16_t neon_reduce_add(float16x8_t x) {
   return vget_lane_f16(y, 0);
 }
 
-template <typename T, typename VT>
-struct AccelerateSimdOps {
-  VT init(T a) {
-    return a;
-  }
-
-  VT load(const T* a) {
-    return *(VT*)a;
-  }
-
-  void store(T* dst, VT x) {
-    *(VT*)dst = x;
-  }
-
-  VT max(VT a, VT b) {
-    return simd_max(a, b);
-  };
-
-  VT exp(VT x) {
-    return simd_fast_exp(x);
-  }
-
-  VT add(VT a, VT b) {
-    return a + b;
-  }
-
-  VT sub(VT a, T b) {
-    return a - b;
-  }
-
-  VT mul(VT a, VT b) {
-    return a * b;
-  }
-
-  VT mul(VT a, T b) {
-    return a * b;
-  }
-
-  T reduce_max(VT x) {
-    return simd_reduce_max(x);
-  }
-
-  T reduce_add(VT x) {
-    return simd_reduce_add(x);
-  }
-};
-
 template <typename T, typename VT>
 struct NeonFp16SimdOps {
   VT init(T a) {
@@ -170,7 +128,7 @@ struct NeonFp16SimdOps {
 
   VT max(VT a, VT b) {
     return vmaxq_f16(a, b);
-  };
+  }
 
   VT exp(VT x) {
     return neon_fast_exp(x);
@@ -201,7 +159,56 @@ struct NeonFp16SimdOps {
   }
 };
 
-template <typename T, typename VT, typename Ops, int N>
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+template <typename T, typename VT>
+struct AccelerateSimdOps {
+  VT init(T a) {
+    return a;
+  }
+
+  VT load(const T* a) {
+    return *(VT*)a;
+  }
+
+  void store(T* dst, VT x) {
+    *(VT*)dst = x;
+  }
+
+  VT max(VT a, VT b) {
+    return simd_max(a, b);
+  }
+
+  VT exp(VT x) {
+    return simd_fast_exp(x);
+  }
+
+  VT add(VT a, VT b) {
+    return a + b;
+  }
+
+  VT sub(VT a, T b) {
+    return a - b;
+  }
+
+  VT mul(VT a, VT b) {
+    return a * b;
+  }
+
+  VT mul(VT a, T b) {
+    return a * b;
+  }
+
+  T reduce_max(VT x) {
+    return simd_reduce_max(x);
+  }
+
+  T reduce_add(VT x) {
+    return simd_reduce_add(x);
+  }
+};
+
+template <typename T, typename AccT, typename VT, typename Ops, int N>
 void softmax(const array& in, array& out) {
   Ops ops;
 
@@ -218,13 +225,21 @@ void softmax(const array& in, array& out) {
     VT vmaximum = ops.init(-std::numeric_limits<float>::infinity());
     size_t s = M;
     while (s >= N) {
-      vmaximum = ops.max(ops.load(current_in_ptr), vmaximum);
+      VT vals;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vals = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vals[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vmaximum = ops.max(vals, vmaximum);
       current_in_ptr += N;
       s -= N;
     }
-    T maximum = ops.reduce_max(vmaximum);
+    AccT maximum = ops.reduce_max(vmaximum);
     while (s-- > 0) {
-      maximum = std::max(maximum, *current_in_ptr);
+      maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
       current_in_ptr++;
     }
 
@@ -234,18 +249,29 @@ void softmax(const array& in, array& out) {
     current_in_ptr = in_ptr;
     s = M;
     while (s >= N) {
-      VT vexp = ops.exp(ops.sub(*(VT*)current_in_ptr, maximum));
-      ops.store(current_out_ptr, vexp);
-      *(VT*)current_out_ptr = vexp;
+      VT vexp;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vexp = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vexp = ops.exp(ops.sub(vexp, maximum));
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, vexp);
+      }
       vnormalizer = ops.add(vnormalizer, vexp);
       current_in_ptr += N;
       current_out_ptr += N;
       s -= N;
     }
-    T normalizer = ops.reduce_add(vnormalizer);
+    AccT normalizer = ops.reduce_add(vnormalizer);
     while (s-- > 0) {
-      T _exp = std::exp(*current_in_ptr - maximum);
-      *current_out_ptr = _exp;
+      AccT _exp = std::exp(*current_in_ptr - maximum);
+      if (std::is_same<T, AccT>::value) {
+        *current_out_ptr = _exp;
+      }
       normalizer += _exp;
       current_in_ptr++;
       current_out_ptr++;
@@ -254,14 +280,33 @@ void softmax(const array& in, array& out) {
 
     // Normalize
     current_out_ptr = out_ptr;
+    current_in_ptr = in_ptr;
     s = M;
     while (s >= N) {
-      ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
+      } else {
+        VT vexp;
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+        vexp = ops.mul(ops.exp(ops.sub(vexp, maximum)), normalizer);
+        for (int i = 0; i < N; ++i) {
+          current_out_ptr[i] = vexp[i];
+        }
+        current_in_ptr += N;
+      }
       current_out_ptr += N;
       s -= N;
     }
     while (s-- > 0) {
-      *current_out_ptr *= normalizer;
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr *= normalizer;
+      } else {
+        AccT _exp = std::exp(*current_in_ptr - maximum);
+        *current_out_ptr = static_cast<T>(_exp * normalizer);
+        current_in_ptr++;
+      }
       current_out_ptr++;
     }
   }
@@ -308,15 +353,33 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
           "Softmax is defined only for floating point types");
       break;
     case float32:
-      softmax<float, simd_float16, AccelerateSimdOps<float, simd_float16>, 16>(
-          in, out);
+      softmax<
+          float,
+          float,
+          simd_float16,
+          AccelerateSimdOps<float, simd_float16>,
+          16>(in, out);
       break;
     case float16:
-      softmax<
-          float16_t,
-          float16x8_t,
-          NeonFp16SimdOps<float16_t, float16x8_t>,
-          8>(in, out);
+      if (precise_) {
+        softmax<
+            float16_t,
+            float,
+            simd_float16,
+            AccelerateSimdOps<float, simd_float16>,
+            16>(in, out);
+      } else {
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        softmax<
+            float16_t,
+            float16_t,
+            float16x8_t,
+            NeonFp16SimdOps<float16_t, float16x8_t>,
+            8>(in, out);
+#else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        eval(inputs, out); // Redirect to common backend for consistency
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+      }
       break;
     case bfloat16:
       eval(inputs, out);

mlx/backend/accelerate/utils.h

@@ -1,8 +1,8 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #pragma once
 
-#include <vecLib/BNNS/bnns.h>
+#include <Accelerate/Accelerate.h>
 #include "mlx/dtype.h"
 
 namespace mlx::core {

mlx/backend/common/CMakeLists.txt

@@ -1,5 +1,4 @@
-
-if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
   set(COMPILER ${CMAKE_C_COMPILER})
   set(CLANG TRUE)
 else()
@@ -7,65 +6,57 @@ else()
 endif()
 
 add_custom_command(
-    OUTPUT  compiled_preamble.cpp
-    COMMAND /bin/bash
-              ${CMAKE_CURRENT_SOURCE_DIR}/make_compiled_preamble.sh
-              ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp
-              ${COMPILER}
-              ${PROJECT_SOURCE_DIR}
-              ${CLANG}
+  OUTPUT compiled_preamble.cpp
+  COMMAND
+    /bin/bash ${CMAKE_CURRENT_SOURCE_DIR}/make_compiled_preamble.sh
+    ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp ${COMPILER}
+    ${PROJECT_SOURCE_DIR} ${CLANG}
+  DEPENDS make_compiled_preamble.sh
+          compiled_preamble.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/half_types.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/fp16.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/bf16.h
+          ${PROJECT_SOURCE_DIR}/mlx/types/complex.h
+          ops.h)
 
-    DEPENDS make_compiled_preamble.sh
-            compiled_preamble.h
-            ${PROJECT_SOURCE_DIR}/mlx/types/half_types.h
-            ${PROJECT_SOURCE_DIR}/mlx/types/fp16.h
-            ${PROJECT_SOURCE_DIR}/mlx/types/bf16.h
-            ${PROJECT_SOURCE_DIR}/mlx/types/complex.h
-            ops.h
-)
-
-add_custom_target(
-  cpu_compiled_preamble
-  DEPENDS compiled_preamble.cpp
-)
+add_custom_target(cpu_compiled_preamble DEPENDS compiled_preamble.cpp)
 
 add_dependencies(mlx cpu_compiled_preamble)
 
 target_sources(
   mlx
-  PRIVATE
-  ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/erf.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/rope.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp
-)
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/common.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eigh.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/erf.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/masked_mm.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce_utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/svd.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/cholesky.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
+          ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp)
 
-if (IOS)
-  target_sources(
-    mlx
-    PRIVATE
-    ${CMAKE_CURRENT_SOURCE_DIR}/compiled_nocpu.cpp
-  )
+if(IOS)
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled_nocpu.cpp)
 else()
-  target_sources(
-    mlx
-    PRIVATE
-    ${CMAKE_CURRENT_SOURCE_DIR}/compiled_cpu.cpp
-  )
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/compiled_cpu.cpp)
 endif()

mlx/backend/common/binary.cpp

@@ -179,18 +179,16 @@ void LogAddExp::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  if (is_floating_point(out.dtype())) {
-    if (out.dtype() == float32) {
-      binary_op<float>(a, b, out, detail::LogAddExp());
-    } else if (out.dtype() == float16) {
-      binary_op<float16_t>(a, b, out, detail::LogAddExp());
-    } else if (out.dtype() == bfloat16) {
-      binary_op<bfloat16_t>(a, b, out, detail::LogAddExp());
-    } else {
-      std::ostringstream err;
-      err << "[logaddexp] Does not support " << out.dtype();
-      throw std::invalid_argument(err.str());
-    }
+  if (out.dtype() == float32) {
+    binary_op<float>(a, b, out, detail::LogAddExp());
+  } else if (out.dtype() == float16) {
+    binary_op<float16_t>(a, b, out, detail::LogAddExp());
+  } else if (out.dtype() == bfloat16) {
+    binary_op<bfloat16_t>(a, b, out, detail::LogAddExp());
+  } else if (issubdtype(out.dtype(), inexact)) {
+    std::ostringstream err;
+    err << "[logaddexp] Does not support " << out.dtype();
+    throw std::invalid_argument(err.str());
   } else {
     throw std::invalid_argument(
         "[logaddexp] Cannot compute logaddexp for arrays with"
@@ -198,6 +196,20 @@ void LogAddExp::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void LogicalAnd::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalAnd());
+}
+
+void LogicalOr::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalOr requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalOr());
+}
+
 void Maximum::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
@@ -238,4 +250,82 @@ void Subtract::eval(const std::vector<array>& inputs, array& out) {
   binary(a, b, out, detail::Subtract());
 }
 
+void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto dispatch_type = [&a, &b, &out](auto op) {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool>(a, b, out, op);
+      case uint8:
+        binary_op<uint8_t>(a, b, out, op);
+        break;
+      case uint16:
+        binary_op<uint16_t>(a, b, out, op);
+        break;
+      case uint32:
+        binary_op<uint32_t>(a, b, out, op);
+        break;
+      case uint64:
+        binary_op<uint64_t>(a, b, out, op);
+        break;
+      case int8:
+        binary_op<int8_t>(a, b, out, op);
+        break;
+      case int16:
+        binary_op<int16_t>(a, b, out, op);
+        break;
+      case int32:
+        binary_op<int32_t>(a, b, out, op);
+        break;
+      case int64:
+        binary_op<int64_t>(a, b, out, op);
+        break;
+      default:
+        throw std::runtime_error(
+            "[BitwiseBinary::eval_cpu] Type not supported");
+        break;
+    }
+  };
+  switch (op_) {
+    case BitwiseBinary::And:
+      dispatch_type(detail::BitwiseAnd());
+      break;
+    case BitwiseBinary::Or:
+      dispatch_type(detail::BitwiseOr());
+      break;
+    case BitwiseBinary::Xor:
+      dispatch_type(detail::BitwiseXor());
+      break;
+    case BitwiseBinary::LeftShift:
+      dispatch_type(detail::LeftShift());
+      break;
+    case BitwiseBinary::RightShift:
+      dispatch_type(detail::RightShift());
+      break;
+  }
+}
+
+void ArcTan2::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  if (out.dtype() == float32) {
+    binary_op<float>(a, b, out, detail::ArcTan2());
+  } else if (out.dtype() == float16) {
+    binary_op<float16_t>(a, b, out, detail::ArcTan2());
+  } else if (out.dtype() == bfloat16) {
+    binary_op<bfloat16_t>(a, b, out, detail::ArcTan2());
+  } else if (issubdtype(out.dtype(), inexact)) {
+    std::ostringstream err;
+    err << "[arctan2] Does not support " << out.dtype();
+    throw std::invalid_argument(err.str());
+  } else {
+    throw std::invalid_argument(
+        "[arctan2] Cannot compute inverse tangent for arrays"
+        " with non floating point type.");
+  }
+}
+
 } // namespace mlx::core