patch bump (#1656)

* allow compiling lambdas in C++

* fix test

* more tests

* auto detect capture-less lambda

.circleci/config.yml

@@ -1,5 +1,8 @@
 version: 2.1
 
+orbs:
+  apple: ml-explore/pr-approval@0.1.0
+
 parameters:
   nightly_build:
     type: boolean
@@ -7,8 +10,65 @@ parameters:
   weekly_build:
     type: boolean
     default: false
+  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
@@ -25,196 +85,262 @@ jobs:
           name: Install dependencies
           command: |
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
+            pip install nanobind==2.2.0
             pip install numpy
             sudo apt-get update
-            sudo apt-get install libblas-dev
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
       - run:
-          name: Build python package
+          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_ARGS="-DMLX_BUILD_METAL=OFF" CMAKE_BUILD_PARALLEL_LEVEL="" python3 setup.py develop
+              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: Run the python tests
+          name: Generate package stubs
           command: |
-            python3 -m unittest discover python/tests
+            echo "stubs"
-      # TODO: Reenable when extension api becomes stable
+            pip install typing_extensions
-      # - run:
+            python setup.py generate_stubs 
-      #     name: Build example extension
+      - run:
-      #     command: |
+          name: Run Python tests
-      #       cd examples/extensions && python3 -m pip install . 
+          command: |
+            python3 -m unittest discover python/tests -v
       - 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:
-    machine: true
+    parameters:
-    resource_class: ml-explore/m-builder
+      xcode_version:
+        type: string
+        default: "15.2.0"
+    macos:
+      xcode: << parameters.xcode_version >>
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
-            eval "$(conda shell.bash hook)"
+            brew install python@3.9
-            rm -r $CONDA_PREFIX/envs/runner-env
+            brew install openmpi
-            conda create -y -n runner-env python=3.9
+            python3.9 -m venv env
-            conda activate runner-env
+            source env/bin/activate
+            pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
+            pip install nanobind==2.2.0
             pip install numpy
             pip install torch
             pip install tensorflow
             pip install unittest-xml-reporting
       - run:
-          name: Build python package
+          name: Install Python package
           command: |
-            eval "$(conda shell.bash hook)"
+            source env/bin/activate
-            conda activate runner-env
+            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install -e . -v
-            CMAKE_BUILD_PARALLEL_LEVEL="" python setup.py build_ext --inplace
-            CMAKE_BUILD_PARALLEL_LEVEL="" python setup.py develop
       - run:
-          name: Run the python tests
+          name: Generate package stubs
           command: |
-            eval "$(conda shell.bash hook)"
+            source env/bin/activate
-            conda activate runner-env
+            pip install typing_extensions
-            DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
+            python setup.py generate_stubs 
-            DEVICE=gpu python -m xmlrunner discover -v python/tests -o test-results/gpu
+      - run:
-      # TODO: Reenable when extension api becomes stable
+          name: Run Python tests
-      # - run:
+          command: |
-      #     name: Build example extension
+            source env/bin/activate
-      #     command: |
+            LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
-      #       eval "$(conda shell.bash hook)"
+            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
-      #       conda activate runner-env
+            mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-      #       cd examples/extensions && python -m pip install . 
+      - 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: |
-            mkdir -p build && cd build && cmake .. && make -j
+            source env/bin/activate
+            mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
       - run:
           name: Run CPP tests
-          command: METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 ./build/tests/tests
+          command: |
+            DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 ./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:
-    machine: true
-    resource_class: ml-explore/m-builder
     parameters:
       python_version:
         type: string
         default: "3.9"
-      macos_version:
+      xcode_version:
         type: string
-        default: "14"
+        default: "15.2.0"
+      build_env:
+        type: string
+        default: ""
+    macos:
+      xcode: << parameters.xcode_version >>
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
-            eval "$(conda shell.bash hook)"
+            brew install python@<< parameters.python_version >>
-            rm -r $CONDA_PREFIX/envs/runner-env
+            brew install openmpi
-            conda create -y -n runner-env python=<< parameters.python_version >>
+            python<< parameters.python_version >> -m venv env
-            conda activate runner-env
+            source env/bin/activate
+            pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
+            pip install nanobind==2.2.0
+            pip install --upgrade setuptools
             pip install numpy
             pip install twine
+            pip install build
       - run:
-          name: Build package
+          name: Install Python package
           command: |
-            eval "$(conda shell.bash hook)"
+            source env/bin/activate
-            conda activate runner-env
+            DEV_RELEASE=1 \
-            DEVELOPER_DIR=$(developer_dir_macos_<< parameters.macos_version >>) \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              PYPI_RELEASE=1 \
+              pip install . -v
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+      - run:
-              python setup.py bdist_wheel
+          name: Generate package stubs
-            twine upload dist/* --repository mlx
+          command: |
+            source env/bin/activate
+            pip install typing_extensions
+            python setup.py generate_stubs 
+      - run:
+          name: Build Python package
+          command: |
+            source env/bin/activate
+            << parameters.build_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
+              python -m build -w
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload package
+                command: |
+                  source env/bin/activate
+                  twine upload dist/*
       - store_artifacts:
           path: dist/
 
-  build_dev_release:
+  build_linux_release:
-    machine: true
-    resource_class: ml-explore/m-builder
     parameters:
       python_version:
         type: string
         default: "3.9"
-      macos_version:
+      extra_env:
         type: string
-        default: "14"
+        default: "DEV_RELEASE=1"
+    docker:
+      - image: ubuntu:20.04
     steps:
       - checkout
       - run:
-          name: Install dependencies
+          name: Build wheel
           command: |
-            eval "$(conda shell.bash hook)"
+            PYTHON=python<< parameters.python_version >>
-            rm -r $CONDA_PREFIX/envs/runner-env
+            apt-get update
-            conda create -y -n runner-env python=<< parameters.python_version >>
+            apt-get upgrade -y
-            conda activate runner-env
+            DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata
+            apt-get install -y apt-utils
+            apt-get install -y software-properties-common
+            add-apt-repository -y ppa:deadsnakes/ppa
+            apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
+            apt-get install -y libblas-dev liblapack-dev liblapacke-dev
+            apt-get install -y build-essential git
+            $PYTHON -m venv env
+            source env/bin/activate
+            pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
+            pip install nanobind==2.2.0
+            pip install --upgrade setuptools
             pip install numpy
+            pip install auditwheel
+            pip install patchelf
+            pip install build
             pip install twine
+            << parameters.extra_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              pip install . -v
+            pip install typing_extensions
+            python setup.py generate_stubs 
+            << parameters.extra_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              python -m build --wheel
+            auditwheel show dist/*
+            auditwheel repair dist/* --plat manylinux_2_31_x86_64
       - run:
-          name: Build package
+          name: Upload package
           command: |
-            eval "$(conda shell.bash hook)"
+            source env/bin/activate
-            conda activate runner-env
+            twine upload wheelhouse/*
-            DEVELOPER_DIR=$(developer_dir_macos_<< parameters.macos_version >>) \
-              DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
-              python setup.py bdist_wheel
-            twine upload dist/* --repository mlx
       - store_artifacts:
-          path: dist/
+          path: wheelhouse/
-
-  build_package:
-    machine: true
-    resource_class: ml-explore/m-builder
-    parameters:
-      python_version:
-        type: string
-        default: "3.9"
-      macos_version:
-        type: string
-        default: "14"
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            eval "$(conda shell.bash hook)"
-            rm -r $CONDA_PREFIX/envs/runner-env
-            conda create -y -n runner-env python=<< parameters.python_version >>
-            conda activate runner-env
-            pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
-            pip install numpy
-            pip install twine
-      - run:
-          name: Build package
-          command: |
-            eval "$(conda shell.bash hook)"
-            conda activate runner-env
-            DEVELOPER_DIR=$(developer_dir_macos_<< parameters.macos_version >>) \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
-              python setup.py bdist_wheel
-      - store_artifacts:
-          path: dist/
 
 workflows:
   build_and_test:
     when:
       and:
+        - matches:
+            pattern: "^(?!pull/)[-\\w]+$"
+            value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
         - not: << pipeline.parameters.weekly_build >>
+        - not: << pipeline.parameters.test_release >>
     jobs:
+      - mac_build_and_test:
+          matrix:
+            parameters:
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test
-      - mac_build_and_test
+      - build_documentation 
+
+  build_pypi_release:
+    when:
+      and:
+        - not: << pipeline.parameters.nightly_build >>
+        - not: << pipeline.parameters.weekly_build >>
+        - not: << pipeline.parameters.test_release >>
+    jobs:
       - build_release:
           filters:
             tags:
@@ -223,21 +349,65 @@ workflows:
               ignore: /.*/
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macos_version: ["13", "14"]
+              xcode_version: ["15.0.0", "15.2.0"]
+              build_env: ["PYPI_RELEASE=1"]
+      - build_documentation:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          upload-docs: true
+
+  prb:
+    when:
+      matches:
+        pattern: "^pull/\\d+(/head)?$"
+        value: << pipeline.git.branch >>
+    jobs:
+      - hold:
+          type: approval
+      - apple/authenticate:
+          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:
-    when: << pipeline.parameters.nightly_build >>
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.nightly_build >>
     jobs:
-      - build_package:
+      - build_release:
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macos_version: ["13", "14"]
+              xcode_version: ["15.0.0", "15.2.0"]
   weekly_build:
-    when: << pipeline.parameters.weekly_build >>
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.weekly_build >>
     jobs:
-      - build_dev_release:
+      - build_release:
           matrix:
             parameters:
-              python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macos_version: ["13", "14"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
+              build_env: ["DEV_RELEASE=1"]
+  linux_test_release:
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.linux_release >>
+    jobs:
+      - build_linux_release:
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              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

@@ -1,16 +1,21 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v17.0.6
+    rev: v19.1.4
     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: 23.12.1
+    rev: 24.10.0
     hooks:
     -   id: black
+    
 -   repo: https://github.com/pycqa/isort
-    rev: 5.12.0
+    rev: 5.13.2
     hooks:
     -   id: isort
         args:
             - --profile=black
+- repo: https://github.com/cheshirekow/cmake-format-precommit
+  rev: v0.6.13
+  hooks:
+    - id: cmake-format

ACKNOWLEDGMENTS.md

@@ -7,11 +7,18 @@ with a short description of your contribution(s) below. For example:
 
 MLX was developed with contributions from the following individuals:
 
-- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops.
+- 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` 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.
+- 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" />
@@ -252,4 +259,4 @@ Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
-limitations under the License.
+limitations under the License.

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

@@ -15,36 +15,43 @@ 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 0.0.10)
+  set(MLX_VERSION 0.21.1)
 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_SYSTEM_PROCESSOR} MATCHES "x86_64")
-if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+    if(NOT MLX_ENABLE_X64_MAC)
-
+      message(
-  if (${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "x86_64" AND ${CMAKE_HOST_APPLE})
+        FATAL_ERROR
-    message(FATAL_ERROR
+          "Building for x86_64 on macOS is not supported."
-      "Building for x86_64 on macOS is not supported." 
+          " If you are on an Apple silicon system, check the build"
-      " If you are on an Apple silicon system, check the build"
+          " documentation for possible fixes: "
-      " documentation for possible fixes: "
+          "https://ml-explore.github.io/mlx/build/html/install.html#build-from-source"
-      "https://ml-explore.github.io/mlx/build/html/install.html#build-from-source")
+      )
-  elseif (${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "x86_64")
+    else()
-    message(WARNING 
+      set(MLX_BUILD_METAL OFF)
-      "Building for x86_64 on macOS is not supported." 
+      message(WARNING "Building for x86_64 arch is not officially supported.")
-      " If you are on an Apple silicon system, "
+    endif()
-      " make sure you are building for arm64.")
-  elseif(${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "arm64")
-    set(MLX_BUILD_ARM ON)
   endif()
 
 else()
+  set(MLX_BUILD_METAL OFF)
   message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
 endif()
 
@@ -56,141 +63,199 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-if (MLX_BUILD_METAL)
+if(MLX_BUILD_METAL)
-  find_library(METAL_LIB Metal)
+  set(METAL_LIB "-framework Metal")
-  find_library(FOUNDATION_LIB Foundation)
+  set(FOUNDATION_LIB "-framework Foundation")
-  find_library(QUARTZ_LIB QuartzCore)
+  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")
-  add_compile_definitions(_METAL_)
 
-  # Throw an error if xcrun not found
+  if(MLX_METAL_DEBUG)
-  execute_process(COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
+    add_compile_definitions(MLX_METAL_DEBUG)
-                  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" )
   endif()
 
-  FetchContent_Declare(
+  # Throw an error if xcrun not found
-    metal_cpp
+  execute_process(
-    URL ${METAL_CPP_URL}
+    COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
+    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
+
+  if(${MACOS_SDK_VERSION} LESS 14.0)
+    message(
+      FATAL_ERROR
+        "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON")
+  endif()
+  message(STATUS "Building with macOS SDK version ${MACOS_SDK_VERSION}")
+
+  set(METAL_CPP_URL
+      https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18-beta.zip
   )
 
+  if(NOT CMAKE_OSX_DEPLOYMENT_TARGET STREQUAL "")
+    set(XCRUN_FLAGS "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
+  endif()
+  execute_process(
+    COMMAND
+      zsh "-c"
+      "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal ${XCRUN_FLAGS} -E -x metal -P - | tail -1 | tr -d '\n'"
+    OUTPUT_VARIABLE MLX_METAL_VERSION COMMAND_ERROR_IS_FATAL ANY)
+  FetchContent_Declare(metal_cpp URL ${METAL_CPP_URL})
+
   FetchContent_MakeAvailable(metal_cpp)
   target_include_directories(
-    mlx PUBLIC
+    mlx PUBLIC $<BUILD_INTERFACE:${metal_cpp_SOURCE_DIR}>
-    $<BUILD_INTERFACE:${metal_cpp_SOURCE_DIR}>
+               $<INSTALL_INTERFACE:include/metal_cpp>)
-    $<INSTALL_INTERFACE:include/metal_cpp>
+  target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
-  )
-  target_link_libraries(
-    mlx
-    ${METAL_LIB}
-    ${FOUNDATION_LIB}
-    ${QUARTZ_LIB})
 endif()
 
-find_library(ACCELERATE_LIBRARY Accelerate)
+if(MLX_BUILD_CPU)
-if (MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
+  find_library(ACCELERATE_LIBRARY Accelerate)
-  message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
+  if(ACCELERATE_LIBRARY)
-  set(MLX_BUILD_ACCELERATE ON)
+    message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
-  target_link_libraries(mlx ${ACCELERATE_LIBRARY})
+    set(MLX_BUILD_ACCELERATE ON)
-  add_compile_definitions(ACCELERATE_NEW_LAPACK)
+    target_link_libraries(mlx PUBLIC ${ACCELERATE_LIBRARY})
-else()
+    add_compile_definitions(ACCELERATE_NEW_LAPACK)
-  message(STATUS "Accelerate or arm neon not found, using default backend.")
+  else()
-  set(MLX_BUILD_ACCELERATE OFF)
+    message(STATUS "Accelerate or arm neon not found, using default backend.")
-  #set(BLA_VENDOR Generic)
+    set(MLX_BUILD_ACCELERATE OFF)
-  find_package(BLAS REQUIRED)
+    if(${CMAKE_HOST_APPLE})
-  if (NOT BLAS_FOUND)
+      # The blas shipped in macOS SDK is not supported, search homebrew for
-    message(FATAL_ERROR "Must have BLAS installed")
+      # 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})
+
+    if(WIN32)
+      find_package(dlfcn-win32 REQUIRED)
+      message(STATUS "dlfcn-win32 lib " ${dlfcn-win32_LIBRARIES})
+      message(STATUS "dlfcn-win32 include " ${dlfcn-win32_INCLUDE_DIRS})
+      target_link_libraries(mlx PUBLIC ${dlfcn-win32_LIBRARIES})
+    endif()
+  endif()
+else()
+  set(MLX_BUILD_ACCELERATE OFF)
+endif()
+
+find_package(MPI)
+if(MPI_FOUND)
+  execute_process(
+    COMMAND zsh "-c" "mpirun --version"
+    OUTPUT_VARIABLE MPI_VERSION
+    ERROR_QUIET)
+  if(${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()
-  # 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_LIBRARIES})
-  message(STATUS ${BLAS_INCLUDE_DIRS})
-  target_include_directories(mlx PRIVATE ${BLAS_INCLUDE_DIRS})
-  target_link_libraries(mlx ${BLAS_LIBRARIES})
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
 
 target_include_directories(
-  mlx 
+  mlx PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
-  PUBLIC
+             $<INSTALL_INTERFACE:include>)
-  $<BUILD_INTERFACE:${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 $<BUILD_INTERFACE:fmt::fmt-header-only>)
+
+if(MLX_BUILD_PYTHON_BINDINGS)
   message(STATUS "Building Python bindings.")
-  find_package(Python COMPONENTS Interpreter Development)
+  find_package(
-  find_package(pybind11 CONFIG REQUIRED)
+    Python 3.8
+    COMPONENTS Interpreter Development.Module
+    REQUIRED)
+  execute_process(
+    COMMAND "${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
+  TARGETS mlx
-    EXPORT MLXTargets
+  EXPORT MLXTargets
-    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+  LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
-    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+  ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
-    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+  RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
-    INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
+  INCLUDES
-)
+  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
-
 
 # Install headers
 install(
-    DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/mlx
+  DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/mlx
-    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
+  DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
-    COMPONENT headers
+  COMPONENT headers
-    FILES_MATCHING PATTERN "*.h"
+  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}/
+    DIRECTORY ${metal_cpp_SOURCE_DIR}/
-      DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/metal_cpp
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/metal_cpp
-      COMPONENT metal_cpp_source
+    COMPONENT metal_cpp_source)
-  )
 
 endif()
 
@@ -202,31 +267,24 @@ 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
+  ${CMAKE_CURRENT_LIST_DIR}/mlx.pc.in ${MLX_CMAKE_BUILD_CONFIG}
-  ${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(
+install(FILES ${MLX_CMAKE_BUILD_CONFIG} ${MLX_CMAKE_BUILD_VERSION_CONFIG}
-  FILES ${MLX_CMAKE_BUILD_CONFIG} ${MLX_CMAKE_BUILD_VERSION_CONFIG}
+        DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR})
-  DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR}
-)
 
-install(
+install(DIRECTORY ${CMAKE_MODULE_PATH}/
-  DIRECTORY ${CMAKE_MODULE_PATH}/
+        DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR})
-  DESTINATION ${MLX_CMAKE_INSTALL_MODULE_DIR}
-)

MANIFEST.in

@@ -1,3 +1,4 @@
 include CMakeLists.txt
 recursive-include mlx/ *
 include python/src/*
+include python/mlx/py.typed # support type hinting as in PEP-561

README.md

@@ -6,15 +6,17 @@
 
 [![CircleCI](https://circleci.com/gh/ml-explore/mlx.svg?style=svg)](https://circleci.com/gh/ml-explore/mlx)
 
-MLX is an array framework for machine learning on Apple silicon, brought to you
+MLX is an array framework for machine learning on Apple silicon,
-by Apple machine learning research.
+brought to you by Apple machine learning research.
 
 Some key features of MLX include:
 
- - **Familiar APIs**: MLX has a Python API that closely follows NumPy.
+ - **Familiar APIs**: MLX has a Python API that closely follows NumPy.  MLX
-   MLX also has a fully featured C++ API, which closely mirrors the Python API. 
+   also has fully featured C++, [C](https://github.com/ml-explore/mlx-c), and
-   MLX has higher-level packages like `mlx.nn` and `mlx.optimizers` with APIs
+   [Swift](https://github.com/ml-explore/mlx-swift/) APIs, which closely mirror
-   that closely follow PyTorch to simplify building more complex models.
+   the Python API.  MLX has higher-level packages like `mlx.nn` and
+   `mlx.optimizers` with APIs that closely follow PyTorch to simplify building
+   more complex models.
 
  - **Composable function transformations**: MLX supports composable function
    transformations for automatic differentiation, automatic vectorization,
@@ -68,23 +70,31 @@ in the documentation.
 
 MLX is available on [PyPI](https://pypi.org/project/mlx/). To install the Python API, run:
 
+**With `pip`**:
+
 ```
 pip install mlx
 ```
 
+**With `conda`**:
+
+```
+conda install -c conda-forge mlx
+```
+
 Checkout the
 [documentation](https://ml-explore.github.io/mlx/build/html/install.html#)
 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/single_ops.cpp

@@ -73,6 +73,7 @@ void time_unary_ops() {
 
 void time_binary_ops() {
   int M = 1000, N = 100, K = 10;
+  auto condition = random::randint(0, 2, {M, N, K});
   auto a = random::uniform({M, N, K});
   auto b = random::uniform({M, N, K});
   auto device = default_device();
@@ -84,7 +85,9 @@ void time_binary_ops() {
   TIME(divide, a, b, device);
   TIME(maximum, a, b, device);
   TIME(minimum, a, b, device);
+  TIME(where, condition, a, b, device);
 
+  condition = array({true});
   b = random::uniform({1});
   eval(b);
   TIMEM("scalar", add, a, b, device);
@@ -93,7 +96,9 @@ void time_binary_ops() {
   TIMEM("scalar", multiply, a, b, device);
   TIMEM("vector-scalar", divide, a, b, device);
   TIMEM("scalar-vector", divide, b, a, device);
+  TIMEM("scalar-vector", where, condition, a, b, device);
 
+  condition = broadcast_to(array({true}), {1000, 100});
   a = broadcast_to(random::uniform({1}), {1000, 100});
   b = broadcast_to(random::uniform({1}), {1000, 100});
   eval(a, b);
@@ -101,6 +106,7 @@ void time_binary_ops() {
   TIMEM("scalar-scalar broadcast", subtract, a, b, device);
   TIMEM("scalar-scalar broadcast", multiply, a, b, device);
   TIMEM("scalar-scalar broadcast", divide, a, b, device);
+  TIMEM("scalar-scalar broadcast", where, condition, a, b, device);
 }
 
 void time_strided_ops() {

benchmarks/cpp/time_utils.h

@@ -17,14 +17,13 @@
             << std::setprecision(5) << time_fn(FUNC, ##__VA_ARGS__) << " msec" \
             << std::endl;
 
-#define TIMEM(MSG, FUNC, ...)                                                  \
+#define TIMEM(MSG, FUNC, ...)                                      \
-  std::cout << "Timing "                                                       \
+  std::cout << "Timing " << "(" << MSG << ") " << #FUNC << " ... " \
-            << "(" << MSG << ") " << #FUNC << " ... " << std::flush            \
+            << std::flush << std::setprecision(5)                  \
-            << std::setprecision(5) << time_fn(FUNC, ##__VA_ARGS__) << " msec" \
+            << time_fn(FUNC, ##__VA_ARGS__) << " msec" << std::endl;
-            << 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_mlx.py

@@ -72,6 +72,9 @@ def _quant_matmul(x, w, s, b, transpose, group_size, bits):
 
 
 quant_matmul = {
+    "quant_matmul_32_2": partial(_quant_matmul, transpose=False, group_size=32, bits=2),
+    "quant_matmul_32_4": partial(_quant_matmul, transpose=False, group_size=32, bits=4),
+    "quant_matmul_32_8": partial(_quant_matmul, transpose=False, group_size=32, bits=8),
     "quant_matmul_64_2": partial(_quant_matmul, transpose=False, group_size=64, bits=2),
     "quant_matmul_64_4": partial(_quant_matmul, transpose=False, group_size=64, bits=4),
     "quant_matmul_64_8": partial(_quant_matmul, transpose=False, group_size=64, bits=8),
@@ -84,6 +87,15 @@ quant_matmul = {
     "quant_matmul_128_8": partial(
         _quant_matmul, transpose=False, group_size=128, bits=8
     ),
+    "quant_matmul_t_32_2": partial(
+        _quant_matmul, transpose=True, group_size=32, bits=2
+    ),
+    "quant_matmul_t_32_4": partial(
+        _quant_matmul, transpose=True, group_size=32, bits=4
+    ),
+    "quant_matmul_t_32_8": partial(
+        _quant_matmul, transpose=True, group_size=32, bits=8
+    ),
     "quant_matmul_t_64_2": partial(
         _quant_matmul, transpose=True, group_size=64, bits=2
     ),
@@ -132,6 +144,13 @@ def reduction(op, axis, x):
     mx.eval(ys)
 
 
+def sum_and_add(axis, x, y):
+    z = x.sum(axis=axis, keepdims=True)
+    for i in range(50):
+        z = (z + y).sum(axis=axis, keepdims=True)
+    mx.eval(z)
+
+
 def softmax(axis, x):
     ys = []
     for i in range(100):
@@ -368,10 +387,6 @@ if __name__ == "__main__":
     if len(args.axis) > 1:
         args.axis.pop(0)
 
-    if args.print_pid:
-        print(os.getpid())
-        input("Press enter to run")
-
     if args.cpu:
         mx.set_default_device(mx.cpu)
     else:
@@ -394,6 +409,10 @@ if __name__ == "__main__":
     x = xs[0]
     axis = args.axis[0]
 
+    if args.print_pid:
+        print(os.getpid())
+        input("Press enter to run")
+
     if args.benchmark == "matmul_square":
         print(bench(matmul_square, x))
 
@@ -493,5 +512,8 @@ if __name__ == "__main__":
     elif args.benchmark == "selu":
         print(bench(selu, x))
 
+    elif args.benchmark == "sum_and_add":
+        print(bench(sum_and_add, axis, *xs))
+
     else:
         raise ValueError("Unknown benchmark")

benchmarks/python/comparative/bench_torch.py

@@ -185,7 +185,7 @@ def prelu(x: torch.Tensor) -> torch.Tensor:
 def mish(x: torch.Tensor) -> torch.Tensor:
     y = x
     for _ in range(100):
-        return torch.nn.functional.mish(y)
+        y = torch.nn.functional.mish(y)
     sync_if_needed(x)
 
 
@@ -283,6 +283,14 @@ def topk(axis, x):
     sync_if_needed(x)
 
 
+@torch.no_grad()
+def step_function(x):
+    y = x
+    for i in range(100):
+        y = torch.where(y < 0, 0, 1)
+    sync_if_needed(x)
+
+
 @torch.no_grad()
 def selu(x):
     y = x
@@ -331,10 +339,6 @@ if __name__ == "__main__":
     if len(args.axis) > 1:
         args.axis.pop(0)
 
-    if args.print_pid:
-        print(os.getpid())
-        input("Press enter to run")
-
     torch.set_num_threads(1)
     device = "cpu" if args.cpu else "mps"
 
@@ -354,6 +358,10 @@ if __name__ == "__main__":
     x = xs[0]
     axis = args.axis[0]
 
+    if args.print_pid:
+        print(os.getpid())
+        input("Press enter to run")
+
     if args.benchmark == "matmul_square":
         print(bench(matmul_square, 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):
@@ -80,10 +82,8 @@ if __name__ == "__main__":
     _filter = make_predicate(args.filter, args.negative_filter)
 
     if args.mlx_dtypes:
-        compare_filtered = (
+        compare_filtered = lambda x: (
-            lambda x: compare_mlx_dtypes(
+            compare_mlx_dtypes(x.split() + rest, args.mlx_dtypes[0], args.mlx_dtypes[1])
-                x.split() + rest, args.mlx_dtypes[0], args.mlx_dtypes[1]
-            )
             if _filter(x)
             else None
         )

benchmarks/python/compile_bench.py

@@ -0,0 +1,107 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import argparse
+import math
+import random
+
+import mlx.core as mx
+from time_utils import time_fn
+
+
+def bench_gelu():
+    def gelu(x):
+        return x * (1 + mx.erf(x / math.sqrt(2))) / 2
+
+    x = mx.random.uniform(shape=(1000, 1024))
+
+    def gen_fun(fun):
+        def bench_fun(x):
+            for _ in range(10):
+                x = fun(x)
+            return x
+
+        return bench_fun
+
+    time_fn(gen_fun(gelu), x, msg="fixed gelu")
+    time_fn(gen_fun(mx.compile(gelu)), x, msg="compiled fixed gelu")
+
+    def randint():
+        return random.randint(1, x.shape[0])
+
+    def gen_fun(fun):
+        def bench_fun(x, y):
+            x = x[: randint()]
+            for _ in range(10):
+                x = fun(x)
+                y = fun(y)
+            return x, y
+
+        return bench_fun
+
+    y = mx.random.uniform(shape=(1000, 1024))
+    time_fn(gen_fun(gelu), x, y, msg="variable gelu")
+    time_fn(gen_fun(mx.compile(gelu)), x, y, msg="compiled variable gelu")
+    time_fn(
+        gen_fun(mx.compile(gelu, shapeless=True)),
+        x,
+        y,
+        msg="shapeless variable 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)
+
+    def layernorm(x):
+        x = x.astype(mx.float32)
+        means = mx.mean(x, axis=-1, keepdims=True)
+        var = mx.var(x, axis=-1, keepdims=True)
+        x = (x - means) * mx.rsqrt(var + 1e-4)
+        x = x.astype(mx.float16)
+        return weight * x + bias
+
+    x = mx.random.uniform(shape=(1000, 4096)).astype(mx.float16)
+
+    def gen_fun(fun):
+        def bench_fun(x):
+            for _ in range(10):
+                x = fun(x)
+            return x
+
+        return bench_fun
+
+    time_fn(gen_fun(layernorm), x, msg="fixed layernorm")
+    time_fn(gen_fun(mx.compile(layernorm)), x, msg="compiled fixed layernorm")
+
+    def randint():
+        return random.randint(1, x.shape[0])
+
+    def gen_fun(fun):
+        def bench_fun(x):
+            x = x[: randint()]
+            for _ in range(10):
+                x = fun(x)
+            return x
+
+        return bench_fun
+
+    random.seed(0)
+    time_fn(gen_fun(layernorm), x, msg="variable layernorm")
+    random.seed(0)
+    time_fn(gen_fun(mx.compile(layernorm)), x, msg="compiled variable layernorm")
+    random.seed(0)
+    time_fn(
+        gen_fun(mx.compile(layernorm, shapeless=True)),
+        x,
+        msg="shapeless variable layernorm",
+    )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("Compile benchmarks.")
+    args = parser.parse_args()
+
+    bench_gelu()
+    bench_layernorm()

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

@@ -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
+
+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_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    def mx_conv_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv2d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_2D
+
+
+def make_pt_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv2d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        torch.mps.synchronize()
+        return ys
+
+    return pt_conv_2D
+
+
+def bench_shape(N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype):
+    scale = 1.0 / math.sqrt(kH * kH * C)
+    a_np = np.random.uniform(0, 0.5, (N, H, W, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kH, kW, int(C / groups))).astype(
+        np_dtype
+    )
+
+    a_mx = mx.array(a_np)
+    b_mx = mx.array(b_np)
+
+    a_pt = torch.from_numpy(a_np.transpose((0, 3, 1, 2))).to("mps")
+    b_pt = torch.from_numpy(b_np.transpose((0, 3, 1, 2))).to("mps")
+
+    torch.mps.synchronize()
+
+    f_mx = make_mx_conv_2D(strides, padding, groups)
+    f_pt = make_pt_conv_2D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv2d(a_mx, b_mx, stride=strides, padding=padding, groups=groups)
+    out_pt = torch.conv2d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 3, 1))
+    out_pt = out_pt.numpy(force=True)
+
+    atol = 2e-5 if np_dtype == np.float32 else 1e-4
+
+    if not np.allclose(out_pt, out_mx, atol=atol):
+        print(
+            f"Failed at {(N, H, W, C)}, {(O, kH, kW, C)} [strides = {strides}, padding = {padding}, groups = {groups}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
+        )
+
+    return time_mlx, time_torch
+
+
+if __name__ == "__main__":
+    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/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/gather_bench.py

@@ -0,0 +1,53 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import argparse
+from time import time
+
+import mlx.core as mx
+import torch
+from time_utils import measure_runtime
+
+
+def benchmark_gather_mlx(x_shape, idx_shape):
+    def gather(x, idx):
+        mx.eval(x[idx])
+
+    idx = mx.random.randint(0, x_shape[0] - 1, idx_shape)
+    x = mx.random.normal(x_shape).astype(mx.float32)
+
+    runtime = measure_runtime(gather, x=x, idx=idx)
+    print(f"MLX: {runtime:.3f}ms")
+
+
+def benchmark_gather_torch(x_shape, idx_shape, device):
+    def gather(x, idx, device):
+        _ = x[idx]
+        if device == torch.device("mps"):
+            torch.mps.synchronize()
+
+    idx = torch.randint(0, x_shape[0] - 1, idx_shape).to(device)
+    x = torch.randn(x_shape, dtype=torch.float32).to(device)
+
+    runtime = measure_runtime(gather, x=x, idx=idx, device=device)
+    print(f"PyTorch: {runtime:.3f}ms")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("Gather benchmarks.")
+    parser.add_argument("--cpu", action="store_true", help="Use the CPU.")
+    args = parser.parse_args()
+
+    if args.cpu:
+        mx.set_default_device(mx.cpu)
+        device = torch.device("cpu")
+    else:
+        device = torch.device("mps")
+
+    idx_shapes = [(1_000_000,), (100_000,), ()]
+    x_shapes = [(100, 64), (100, 1024), (4, 1_000_000)]
+
+    for x_shape, idx_shape in zip(x_shapes, idx_shapes):
+        print("=" * 20)
+        print(f"X {x_shape}, Indices {idx_shape}")
+        benchmark_gather_mlx(x_shape, idx_shape)
+        benchmark_gather_torch(x_shape, idx_shape, device=device)

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/llama_jax_bench.py

@@ -1,198 +0,0 @@
-# Copyright © 2023 Apple Inc.
-
-import math
-import time
-
-import jax
-import jax.numpy as jnp
-from flax import linen as nn
-
-
-class RoPE(nn.Module):
-    dims: int
-    traditional: bool = False
-
-    def _compute_rope(self, costheta, sintheta, x):
-        x1 = x[..., : self.dims // 2]
-        x2 = x[..., self.dims // 2 : self.dims]
-        rx1 = x1 * costheta - x2 * sintheta
-        rx2 = x1 * sintheta + x2 * costheta
-
-        if self.dims < x.shape[-1]:
-            rx = jnp.concatenate([rx1, rx2, x[..., self.dims :]], axis=-1)
-        else:
-            rx = jnp.concatenate([rx1, rx2], axis=-1)
-
-        return rx
-
-    def _compute_traditional_rope(self, costheta, sintheta, x):
-        x1 = x[..., ::2]
-        x2 = x[..., 1::2]
-        rx1 = x1 * costheta - x2 * sintheta
-        rx2 = x1 * sintheta + x2 * costheta
-
-        if self.dims < x.shape[-1]:
-            raise NotImplementedError(
-                "RoPE doesn't implement partial traditional application"
-            )
-
-        rx = jnp.concatenate([rx1[..., None], rx2[..., None]], axis=-1)
-
-        return rx
-
-    @staticmethod
-    def create_cos_sin_theta(
-        N: int,
-        D: int,
-        offset: int = 0,
-        base: float = 10000,
-        dtype=jnp.float32,
-    ):
-        D = D // 2
-        positions = jnp.arange(offset, N, dtype=dtype)
-        freqs = jnp.exp(-jnp.arange(0, D, dtype=dtype) * (math.log(base) / D))
-        theta = positions.reshape((-1, 1)) * freqs.reshape((1, -1))
-        costheta = jnp.cos(theta)
-        sintheta = jnp.sin(theta)
-
-        return costheta, sintheta
-
-    @nn.compact
-    def __call__(self, x, offset: int = 0):
-        shape = x.shape
-        x = x.reshape((-1, shape[-2], shape[-1]))
-        N = x.shape[1] + offset
-        costheta, sintheta = RoPE.create_cos_sin_theta(
-            N, self.dims, offset=offset, dtype=x.dtype
-        )
-
-        rope = (
-            self._compute_traditional_rope if self.traditional else self._compute_rope
-        )
-        rx = rope(costheta, sintheta, x)
-
-        return rx.reshape(shape)
-
-
-class LlamaAttention(nn.Module):
-    dims: int
-    num_heads: int
-    dtype: jnp.dtype
-
-    def setup(self):
-        num_heads = self.num_heads
-        dims = self.dims
-
-        self.rope = RoPE(dims // num_heads, True)
-        self.query_proj = nn.Dense(dims, use_bias=False, param_dtype=self.dtype)
-        self.key_proj = nn.Dense(dims, use_bias=False, param_dtype=self.dtype)
-        self.value_proj = nn.Dense(dims, use_bias=False, param_dtype=self.dtype)
-        self.out_proj = nn.Dense(dims, use_bias=False, param_dtype=self.dtype)
-
-    def __call__(self, queries, keys, values, mask=None, cache=None):
-        queries = self.query_proj(queries)
-        keys = self.key_proj(keys)
-        values = self.value_proj(values)
-
-        num_heads = self.num_heads
-        B, L, D = queries.shape
-        queries = queries.reshape((B, L, num_heads, -1)).transpose((0, 2, 1, 3))
-        keys = keys.reshape((B, L, num_heads, -1)).transpose((0, 2, 1, 3))
-        values = values.reshape((B, L, num_heads, -1)).transpose((0, 2, 1, 3))
-
-        if cache is not None:
-            key_cache, value_cache = cache
-            queries = self.rope(queries, offset=key_cache.shape[2])
-            keys = self.rope(keys, offset=key_cache.shape[2])
-            keys = jnp.concatenate([key_cache, keys], axis=2)
-            values = jnp.concatenate([value_cache, values], axis=2)
-        else:
-            queries = self.rope(queries)
-            keys = self.rope(keys)
-
-        # Dimensions are [batch x num heads x sequence x hidden dim]
-        scale = math.sqrt(1 / queries.shape[-1])
-        scores = (queries * scale) @ keys.transpose((0, 1, 3, 2))
-        if mask is not None:
-            scores = scores + mask
-        scores = jax.nn.softmax(scores, axis=-1)
-        values_hat = (scores @ values).transpose((0, 2, 1, 3)).reshape((B, L, -1))
-
-        return self.out_proj(values_hat), (keys, values)
-
-
-class LlamaEncoderLayer(nn.Module):
-    dims: int
-    mlp_dims: int
-    num_heads: int
-    dtype: jnp.dtype
-
-    def setup(self):
-        dims = self.dims
-        mlp_dims = self.mlp_dims
-        num_heads = self.num_heads
-
-        self.attention = LlamaAttention(dims, num_heads, dtype)
-
-        self.norm1 = nn.RMSNorm(param_dtype=self.dtype)
-        self.norm2 = nn.RMSNorm(param_dtype=self.dtype)
-
-        self.linear1 = nn.Dense(mlp_dims, use_bias=False, param_dtype=self.dtype)
-        self.linear2 = nn.Dense(mlp_dims, use_bias=False, param_dtype=self.dtype)
-        self.linear3 = nn.Dense(dims, use_bias=False, param_dtype=self.dtype)
-
-    def __call__(self, x, mask=None, cache=None):
-        y = self.norm1(x)
-        y, cache = self.attention(y, y, y, mask, cache)
-        x = x + y
-
-        y = self.norm2(x)
-        a = self.linear1(y)
-        b = self.linear2(y)
-        y = jax.nn.silu(a) * b
-        y = self.linear3(y)
-        x = x + y
-
-        return x, cache
-
-
-def measure(model, x, cache):
-    for i in range(5):
-        y, c = model(x, mask=None, cache=cache)
-        jax.block_until_ready((y, c))
-
-    start = time.time()
-    for i in range(5):
-        y, c = model(x, mask=None, cache=cache)
-        jax.block_until_ready((y, c))
-
-    end = time.time()
-    return (end - start) * 1000 / 5
-
-
-if __name__ == "__main__":
-    H = 32
-    D = 4096
-    F = 43 * 256
-    C = 1000
-    dtype = jnp.float16
-
-    k1, k2, k3, k4 = jax.random.split(jax.random.PRNGKey(0), 4)
-
-    x = jax.random.normal(k1, (1, 1, D), dtype)
-    cache = [
-        jax.random.normal(k2, [1, H, C, D // H], dtype),
-        jax.random.normal(k3, [1, H, C, D // H], dtype),
-    ]
-
-    layer = LlamaEncoderLayer(D, F, H, dtype=dtype)
-    params = layer.init(k4, x, mask=None, cache=cache)["params"]
-
-    @jax.jit
-    def model_fn(x, mask, cache):
-        return layer.apply({"params": params}, x, mask=mask, cache=cache)
-
-    T = measure(model_fn, x, cache)
-
-    print("Time per layer per token:", T, "ms")
-    print("Lower bound total time per token:", T * 32, "ms")

benchmarks/python/llama_mlx_bench.py

@@ -1,118 +0,0 @@
-# Copyright © 2023 Apple Inc.
-
-import math
-import time
-
-import mlx.core as mx
-import mlx.nn as nn
-import mlx.utils
-
-
-class LlamaAttention(nn.Module):
-    def __init__(self, dims: int, num_heads: int):
-        super().__init__()
-        self.num_heads = num_heads
-        self.rope = nn.RoPE(dims // num_heads, True)
-        self.query_proj = nn.Linear(dims, dims, False)
-        self.key_proj = nn.Linear(dims, dims, False)
-        self.value_proj = nn.Linear(dims, dims, False)
-        self.out_proj = nn.Linear(dims, dims, False)
-
-    def __call__(self, queries, keys, values, mask=None, cache=None):
-        queries = self.query_proj(queries)
-        keys = self.key_proj(keys)
-        values = self.value_proj(values)
-
-        num_heads = self.num_heads
-        B, L, D = queries.shape
-        queries = mx.transpose(mx.reshape(queries, (B, L, num_heads, -1)), (0, 2, 1, 3))
-        keys = mx.transpose(mx.reshape(keys, (B, L, num_heads, -1)), (0, 2, 1, 3))
-        values = mx.transpose(mx.reshape(values, (B, L, num_heads, -1)), (0, 2, 1, 3))
-
-        if cache is not None:
-            key_cache, value_cache = cache
-            queries = self.rope(queries, offset=key_cache.shape[2])
-            keys = self.rope(keys, offset=key_cache.shape[2])
-            keys = mx.concatenate([key_cache, keys], axis=2)
-            values = mx.concatenate([value_cache, values], axis=2)
-        else:
-            queries = self.rope(queries)
-            keys = self.rope(keys)
-
-        # Dimensions are [batch x num heads x sequence x hidden dim]
-        scale = mx.array(math.sqrt(1 / queries.shape[-1]), dtype=queries.dtype)
-        scores = (queries * scale) @ mx.transpose(keys, (0, 1, 3, 2))
-        if mask is not None:
-            scores = scores + mask
-        scores = mx.softmax(scores, axis=-1)
-        values_hat = mx.reshape(mx.transpose(scores @ values, (0, 2, 1, 3)), (B, L, -1))
-
-        return self.out_proj(values_hat), (keys, values)
-
-
-class LlamaEncoderLayer(nn.Module):
-    def __init__(self, dims: int, mlp_dims: int, num_heads: int):
-        super().__init__()
-
-        self.attention = LlamaAttention(dims, num_heads)
-
-        self.norm1 = nn.RMSNorm(dims)
-        self.norm2 = nn.RMSNorm(dims)
-
-        self.linear1 = nn.Linear(dims, mlp_dims, False)
-        self.linear2 = nn.Linear(dims, mlp_dims, False)
-        self.linear3 = nn.Linear(mlp_dims, dims, False)
-
-    def __call__(self, x, mask=None, cache=None):
-        y = self.norm1(x)
-        y, cache = self.attention(y, y, y, mask, cache)
-        x = x + y
-
-        y = self.norm2(x)
-        a = self.linear1(y)
-        b = self.linear2(y)
-        y = a * mx.sigmoid(a) * b
-        y = self.linear3(y)
-        x = x + y
-
-        return x, cache
-
-
-def measure(model, x, cache):
-    for i in range(5):
-        y, c = model(x, mask=None, cache=cache)
-        mx.eval(y, c)
-
-    start = time.time()
-    rs = []
-    for i in range(5):
-        y, c = model(x, mask=None, cache=cache)
-        rs.append((y, c))
-    mx.eval(rs)
-    end = time.time()
-
-    return (end - start) * 1000 / 5
-
-
-if __name__ == "__main__":
-    H = 32
-    D = 4096
-    F = 43 * 256
-    C = 1000
-    mx.set_default_device(mx.gpu)
-    dtype = mx.float16
-
-    layer = LlamaEncoderLayer(D, F, H)
-    layer.update(mlx.utils.tree_map(lambda x: x.astype(dtype), layer.parameters()))
-    k1, k2, k3 = mx.random.split(mx.random.key(0), 3)
-    x = mx.random.normal([1, 1, D], dtype=dtype)
-    cache = [
-        mx.random.normal([1, H, C, D // H], dtype=dtype),
-        mx.random.normal([1, H, C, D // H], dtype=dtype),
-    ]
-    mx.eval(x, cache)
-
-    T = measure(layer, x, cache)
-
-    print("Time per layer per token:", T, "ms")
-    print("Lower bound total time per token:", T * 32, "ms")

benchmarks/python/llama_torch_bench.py

@@ -1,199 +0,0 @@
-# Copyright © 2023 Apple Inc.
-
-import math
-import time
-
-import torch
-import torch.mps
-import torch.nn as nn
-
-
-def sync_if_needed(x):
-    if x.device != torch.device("cpu"):
-        torch.mps.synchronize()
-
-
-class RoPE(nn.Module):
-    def __init__(self, dims: int, traditional: bool = False):
-        super().__init__()
-        self.dims = dims
-        self.traditional = traditional
-
-    def _compute_rope(self, costheta, sintheta, x):
-        x1 = x[..., : self.dims // 2]
-        x2 = x[..., self.dims // 2 : self.dims]
-        rx1 = x1 * costheta - x2 * sintheta
-        rx2 = x1 * sintheta + x2 * costheta
-
-        if self.dims < x.shape[-1]:
-            rx = torch.cat([rx1, rx2, x[..., self.dims :]], dim=-1)
-        else:
-            rx = torch.cat([rx1, rx2], dim=-1)
-
-        return rx
-
-    def _compute_traditional_rope(self, costheta, sintheta, x):
-        x1 = x[..., ::2]
-        x2 = x[..., 1::2]
-        rx1 = x1 * costheta - x2 * sintheta
-        rx2 = x1 * sintheta + x2 * costheta
-
-        if self.dims < x.shape[-1]:
-            raise NotImplementedError(
-                "RoPE doesn't implement partial traditional application"
-            )
-
-        rx = torch.cat([rx1[..., None], rx2[..., None]], dim=-1)
-
-        return rx
-
-    def forward(self, x, offset: int = 0):
-        shape = x.shape
-        x = x.view(-1, shape[-2], shape[-1])
-        N = x.shape[1] + offset
-        costheta, sintheta = RoPE.create_cos_sin_theta(
-            N, self.dims, offset=offset, device=x.device, dtype=x.dtype
-        )
-
-        rope = (
-            self._compute_traditional_rope if self.traditional else self._compute_rope
-        )
-        rx = rope(costheta, sintheta, x)
-
-        return rx.view(*shape)
-
-    @staticmethod
-    def create_cos_sin_theta(
-        N: int,
-        D: int,
-        offset: int = 0,
-        base: float = 10000,
-        device="cpu",
-        dtype=torch.float32,
-    ):
-        D = D // 2
-        positions = torch.arange(offset, N, dtype=dtype, device=device)
-        freqs = torch.exp(
-            -torch.arange(0, D, dtype=dtype, device=device) * (math.log(base) / D)
-        )
-        theta = positions.view(-1, 1) * freqs.view(1, -1)
-        costheta = torch.cos(theta)
-        sintheta = torch.sin(theta)
-
-        return costheta, sintheta
-
-
-class RMSNorm(nn.Module):
-    def __init__(self, dims: int, epsilon: float = 1e-6):
-        super().__init__()
-        self.gamma = nn.Parameter(torch.ones((dims,)))
-        self.epsilon = epsilon
-
-    def forward(self, x):
-        n = torch.rsqrt(x.square().mean(dim=-1, keepdims=True) + self.epsilon)
-        return self.gamma * x * n
-
-
-class LlamaAttention(nn.Module):
-    def __init__(self, dims: int, num_heads: int):
-        super().__init__()
-        self.num_heads = num_heads
-        self.rope = RoPE(dims // num_heads, True)
-        self.query_proj = nn.Linear(dims, dims, bias=False)
-        self.key_proj = nn.Linear(dims, dims, bias=False)
-        self.value_proj = nn.Linear(dims, dims, bias=False)
-        self.out_proj = nn.Linear(dims, dims, bias=False)
-
-    def forward(self, queries, keys, values, mask=None, cache=None):
-        queries = self.query_proj(queries)
-        keys = self.key_proj(keys)
-        values = self.value_proj(values)
-
-        num_heads = self.num_heads
-        B, L, D = queries.shape
-        queries = queries.view(B, L, num_heads, -1).permute(0, 2, 1, 3)
-        keys = keys.view(B, L, num_heads, -1).permute(0, 2, 1, 3)
-        values = values.view(B, L, num_heads, -1).permute(0, 2, 1, 3)
-
-        if cache is not None:
-            key_cache, value_cache = cache
-            queries = self.rope(queries, offset=key_cache.shape[2])
-            keys = self.rope(keys, offset=key_cache.shape[2])
-            keys = torch.cat([key_cache, keys], dim=2)
-            values = torch.cat([value_cache, values], dim=2)
-        else:
-            queries = self.rope(queries)
-            keys = self.rope(keys)
-
-        # Dimensions are [batch x num heads x sequence x hidden dim]
-        scale = math.sqrt(1 / queries.shape[-1])
-        scores = (queries * scale) @ keys.permute(0, 1, 3, 2)
-        if mask is not None:
-            scores = scores + mask
-        scores = torch.softmax(scores, dim=-1)
-        values_hat = (scores @ values).permute(0, 2, 1, 3).reshape(B, L, -1)
-
-        return self.out_proj(values_hat), (keys, values)
-
-
-class LlamaEncoderLayer(nn.Module):
-    def __init__(self, dims: int, mlp_dims: int, num_heads: int):
-        super().__init__()
-
-        self.attention = LlamaAttention(dims, num_heads)
-
-        self.norm1 = RMSNorm(dims)
-        self.norm2 = RMSNorm(dims)
-
-        self.linear1 = nn.Linear(dims, mlp_dims, bias=False)
-        self.linear2 = nn.Linear(dims, mlp_dims, bias=False)
-        self.linear3 = nn.Linear(mlp_dims, dims, bias=False)
-
-    def forward(self, x, mask=None, cache=None):
-        y = self.norm1(x)
-        y, cache = self.attention(y, y, y, mask, cache)
-        x = x + y
-
-        y = self.norm2(x)
-        a = self.linear1(y)
-        b = self.linear2(y)
-        y = torch.nn.functional.silu(a) * b
-        y = self.linear3(y)
-        x = x + y
-
-        return x, cache
-
-
-@torch.no_grad()
-def measure(model, x, cache):
-    for i in range(5):
-        y, c = model(x, mask=None, cache=cache)
-    sync_if_needed(x)
-
-    start = time.time()
-    for i in range(5):
-        y, c = model(x, mask=None, cache=cache)
-    sync_if_needed(x)
-    end = time.time()
-    return (end - start) * 1000 / 5
-
-
-if __name__ == "__main__":
-    H = 32
-    D = 4096
-    F = 43 * 256
-    C = 1000
-    device = torch.device("mps")
-    dtype = torch.float16
-
-    layer = LlamaEncoderLayer(D, F, H).to(device).to(dtype)
-    x = torch.randn(1, 1, D).to(device).to(dtype)
-    cache = [
-        torch.randn(1, H, C, D // H).to(device).to(dtype),
-        torch.randn(1, H, C, D // H).to(device).to(dtype),
-    ]
-
-    T = measure(layer, x, cache)
-
-    print("Time per layer per token:", T, "ms")
-    print("Lower bound total time per token:", T * 32, "ms")

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

@@ -0,0 +1,35 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import mlx.core as mx
+import mlx.nn as nn
+from time_utils import time_fn
+
+
+def time_rope():
+    rope = nn.RoPE(64)
+
+    # vec
+    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, offset=100)
+        return x
+
+    time_fn(rope_vec, x)
+
+    # matrix
+    x = mx.random.uniform(shape=(1, 32, 1024, 128)).astype(mx.float16)
+    mx.eval(x)
+
+    def rope_mat(x):
+        for _ in range(32):
+            x = rope(x)
+        return x
+
+    time_fn(rope_mat, x)
+
+
+if __name__ == "__main__":
+    time_rope()

benchmarks/python/scatter_bench.py

@@ -0,0 +1,96 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import argparse
+
+import mlx.core as mx
+import torch
+from time_utils import measure_runtime
+
+
+def benchmark_scatter_mlx(dst_shape, x_shape, idx_shapes):
+    def scatter(dst, x, idx):
+        dst[tuple(idx)] = x
+        mx.eval(dst)
+
+    idx = []
+    for idx_shape in idx_shapes:
+        idx.append(mx.random.randint(0, dst_shape[0] - 1, idx_shape))
+    x = mx.random.normal(x_shape).astype(mx.float32)
+    dst = mx.random.normal(dst_shape).astype(mx.float32)
+
+    runtime = measure_runtime(scatter, dst=dst, x=x, idx=idx)
+    print(f"MLX: {runtime:.3f}ms")
+
+
+def benchmark_scatter_torch(dst_shape, x_shape, idx_shapes, device):
+    def scatter(dst, x, idx, device):
+        dst[tuple(idx)] = x
+        if device == torch.device("mps"):
+            torch.mps.synchronize()
+
+    idx = []
+    for idx_shape in idx_shapes:
+        idx.append(torch.randint(0, dst_shape[0] - 1, idx_shape).to(device))
+    x = torch.randn(x_shape, dtype=torch.float32).to(device)
+    dst = torch.randn(dst_shape, dtype=torch.float32).to(device)
+
+    runtime = measure_runtime(scatter, dst=dst, x=x, idx=idx, device=device)
+    print(f"PyTorch: {runtime:.3f}ms")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("Gather benchmarks.")
+    parser.add_argument("--cpu", action="store_true", help="Use the CPU.")
+    args = parser.parse_args()
+
+    if args.cpu:
+        mx.set_default_device(mx.cpu)
+        device = torch.device("cpu")
+    else:
+        device = torch.device("mps")
+
+    dst_shapes = [
+        (10, 64),
+        (100_000, 64),
+        (1_000_000, 64),
+        (100_000,),
+        (200_000,),
+        (20_000_000,),
+        (10000, 64),
+        (100, 64),
+        (100, 10_000, 64),
+        (10, 100, 100, 21),
+        (1_000, 1_000, 10),
+    ]
+    idx_shapes = [
+        [(1_000_000,)],
+        [(1_000_000,)],
+        [(100_000,)],
+        [(1_000_000,)],
+        [(20_000_000,)],
+        [(20_000_000,)],
+        [(1000000,)],
+        [(10000000,)],
+        [(1_000,)],
+        [(10_000,)],
+        [(1_000,), (1_000,)],
+    ]
+    x_shapes = [
+        (1_000_000, 64),
+        (1_000_000, 64),
+        (100_000, 64),
+        (1_000_000,),
+        (20_000_000,),
+        (20_000_000,),
+        (1000000, 64),
+        (10000000, 64),
+        (1_000, 10_000, 64),
+        (10_000, 100, 100, 21),
+        (1_000, 10),
+    ]
+
+    for dst_shape, x_shape, idx_shape in zip(dst_shapes, x_shapes, idx_shapes):
+        print("=" * 20)
+        print(f"Dst: {dst_shape}, X {x_shape}, Indices {idx_shape}")
+        benchmark_scatter_mlx(dst_shape, x_shape, idx_shape)
+        benchmark_scatter_torch(dst_shape, x_shape, idx_shape, device=device)

benchmarks/python/sdpa_bench.py

@@ -0,0 +1,189 @@
+# Copyright © 2024 Apple Inc.
+
+import argparse
+import math
+import os
+import subprocess
+import time
+
+import mlx.core as mx
+import numpy as np
+
+device_name = subprocess.check_output(["sysctl", "-n", "machdep.cpu.brand_string"])
+device_name = device_name.decode("utf-8").strip("\n")
+
+N_warmup = 5
+N_iter_bench = 40
+N_iter_func = 8
+
+
+def bench(f, *args):
+    for i in range(N_warmup):
+        f(*args)
+
+    s = time.perf_counter_ns()
+    for i in range(N_iter_bench):
+        f(*args)
+    e = time.perf_counter_ns()
+    return (e - s) * 1e-9
+
+
+def mlx_sdpa_fused_inner(q, k, v, scale):
+    return mx.fast.scaled_dot_product_attention(q, k, v, scale=scale, mask=None)
+
+
+def mlx_sdpa_unfused_inner(q, k, v, scale, f32softmax=False):
+    q_dtype = q.dtype
+    q = q * mx.array(scale, q_dtype)
+    n_q_heads = q.shape[-3]
+    n_kv_heads = k.shape[-3]
+    n_repeats = n_q_heads // n_kv_heads
+
+    B = q.shape[0]
+    L = q.shape[2]
+
+    if n_repeats > 1:
+        q = mx.reshape(q, [B, n_kv_heads, n_repeats, L, -1])
+        k = mx.expand_dims(k, 2)
+        v = mx.expand_dims(v, 2)
+
+    scores = q @ mx.swapaxes(k, -1, -2)
+    if f32softmax:
+        scores = mx.softmax(scores.astype(mx.float32), axis=-1).astype(q_dtype)
+    else:
+        scores = mx.softmax(scores, axis=-1)
+
+    out = scores @ v
+    if n_repeats > 1:
+        out = mx.reshape(out, [B, n_q_heads, L, -1])
+
+    return out
+
+
+def mlx_spda_unfused(q, k, v, scale, transpose):
+    q_out = q
+    if transpose:
+        k = mx.transpose(k, (0, 2, 1, 3))
+        v = mx.transpose(v, (0, 2, 1, 3))
+
+    for i in range(N_iter_func):
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+        q_out = mlx_sdpa_unfused_inner(q_out, k, v, scale)
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+
+    mx.eval(q_out)
+    return q_out
+
+
+def mlx_spda_fused(q, k, v, scale, transpose):
+    q_out = q
+    if transpose:
+        k = mx.transpose(k, (0, 2, 1, 3))
+        v = mx.transpose(v, (0, 2, 1, 3))
+
+    for i in range(N_iter_func):
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+        q_out = mlx_sdpa_fused_inner(q_out, k, v, scale)
+        if transpose:
+            q_out = mx.transpose(q_out, (0, 2, 1, 3))
+
+    mx.eval(q_out)
+    return q_out
+
+
+def bench_shape(B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, np_dtype, transpose=True):
+    shape_q = (
+        (B, qsl, n_q_heads, head_dim) if transpose else (B, n_q_heads, qsl, head_dim)
+    )
+    shape_kv = (
+        (B, ksl, n_kv_heads, head_dim) if transpose else (B, n_kv_heads, ksl, head_dim)
+    )
+
+    q_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_q).astype(np_dtype)
+    k_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_kv).astype(np_dtype)
+    v_np = np.random.normal(0.0, 1.0 / math.sqrt(head_dim), shape_kv).astype(np_dtype)
+
+    scale = math.sqrt(1.0 / head_dim)
+
+    q_mx = mx.array(q_np)
+    k_mx = mx.array(k_np)
+    v_mx = mx.array(v_np)
+
+    time_mlx_unfused = bench(mlx_spda_unfused, q_mx, k_mx, v_mx, scale, transpose)
+    time_mlx_fused = bench(mlx_spda_fused, q_mx, k_mx, v_mx, scale, transpose)
+
+    if transpose:
+        q_mx = mx.transpose(q_mx, (0, 2, 1, 3))
+        k_mx = mx.transpose(k_mx, (0, 2, 1, 3))
+        v_mx = mx.transpose(v_mx, (0, 2, 1, 3))
+
+    o_mlx_fused = mlx_sdpa_fused_inner(q_mx, k_mx, v_mx, scale)
+    o_mlx_unfused = mlx_sdpa_unfused_inner(q_mx, k_mx, v_mx, scale, f32softmax=True)
+
+    atol = 1e-5 if np_dtype == np.float32 else 1e-4
+
+    if not mx.allclose(o_mlx_fused, o_mlx_unfused, atol=atol):
+        print(
+            f"Failed at (B: {B}, qsl: {qsl}, ksl: {ksl}, head_dim: {head_dim}, n_qh: {n_q_heads}, n_kvh: {n_kv_heads}) [tpose = {transpose}] with max(|a - b|) = {mx.max(mx.abs(o_mlx_unfused - o_mlx_fused)):3.2e}"
+        )
+
+    return time_mlx_fused, time_mlx_unfused
+
+
+def get_gflop_count(B, M, N, K):
+    return float(2.0 * N_iter_bench * N_iter_func * B * M * N * K) / float(1024.0**3)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run gemm benchmarks")
+
+    dtypes = ("float16", "float32")[:1]
+    transposes = (False,)
+
+    # fmt: off
+    shapes_64 = (
+        # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
+          (  1,    32,    32,       64,   32,    32),
+          (  1,    64,    64,       64,   32,    32),
+          (  1,   128,   128,       64,   32,    32),
+          (  1,   256,   256,       64,   32,    32),
+          (  1,   512,   512,       64,   32,    32),
+          (  1,  1024,  1024,       64,   32,    32),
+          (  1,  2048,  2048,       64,   32,    32),
+          (  1,  4096,  4096,       64,   32,    32),
+    )
+
+    shapes_80 = (
+        # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
+          (  1,  1024,  1024,       80,   32,    32),
+          (  1,  2048,  2048,       80,   32,    32),
+          (  1,  4096,  4096,       80,   32,    32),
+    )
+
+    shapes_128 = (
+        # (  B,   qsl,   ksl, head_dim, n_qh, n_kvh)
+          (  1,  1024,  1024,      128,   32,    32),
+          (  1,  2048,  2048,      128,   32,    32),
+          (  1,  4096,  4096,      128,   32,    32),
+    )
+    # fmt: on
+
+    shapes = shapes_64 + shapes_80 + shapes_128
+
+    print("  B,   qsl,   ksl, hdim, n_qh, n_kvh, tpose,   dtype, t_unfs, t_fuse, diff%")
+
+    for dtype in dtypes:
+        for transpose in transposes:
+            for B, qsl, ksl, head_dim, n_q_heads, n_kv_heads in shapes:
+                np_dtype = getattr(np, dtype)
+                time_mlx_fused, time_mlx_unfused = bench_shape(
+                    B, qsl, ksl, head_dim, n_q_heads, n_kv_heads, np_dtype, transpose
+                )
+                diff = time_mlx_unfused / time_mlx_fused - 1.0
+                t_str = 1 if transpose else 0
+                print(
+                    f"{B:3d}, {qsl:5d}, {ksl:5d}, {head_dim:4d}, {n_q_heads:4d}, {n_kv_heads:5d}, {t_str:5d}, {dtype}, {time_mlx_unfused: 2.3f}, {time_mlx_fused: 2.3f}, {100. * diff:+5.2f}%"
+                )

benchmarks/python/sdpa_vector_bench.py

@@ -0,0 +1,58 @@
+import argparse
+import math
+
+import mlx.core as mx
+from time_utils import time_fn
+
+L = 16384
+H = 32
+H_k = H // 4
+D = 128
+dtype = mx.float16
+loops = 10
+
+
+def attention(q, k, v):
+    def _sdpa(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)
+
+    for i in range(loops):
+        q = _sdpa(q, k, v)
+    return q
+
+
+def sdpa(q, k, v):
+    for i in range(loops):
+        q = mx.fast.scaled_dot_product_attention(q, k, v, scale=1.0)
+    return q
+
+
+def time_self_attention_primitives():
+    mx.random.seed(3)
+    q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
+    k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
+    v = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
+    mx.eval(q, k, v)
+    time_fn(attention, q, k, v)
+
+
+def time_self_attention_sdpa():
+    mx.random.seed(3)
+    q = mx.random.uniform(shape=(1, H, 1, D)).astype(dtype)
+    k = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
+    v = mx.random.uniform(shape=(1, H_k, L, D)).astype(dtype)
+    mx.eval(q, k, v)
+    time_fn(sdpa, q, k, v)
+
+
+if __name__ == "__main__":
+    time_self_attention_sdpa()
+    time_self_attention_primitives()

benchmarks/python/single_ops.py

@@ -44,6 +44,13 @@ def time_matmul():
     time_fn(mx.matmul, a, b)
 
 
+def time_maximum():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    b = mx.random.uniform(shape=(32, 1024, 1024))
+    mx.eval(a, b)
+    time_fn(mx.maximum, a, b)
+
+
 def time_negative():
     a = mx.random.uniform(shape=(10000, 1000))
     mx.eval(a)
@@ -101,6 +108,7 @@ if __name__ == "__main__":
 
     time_add()
     time_matmul()
+    time_maximum()
     time_exp()
     time_negative()
     time_logsumexp()

benchmarks/python/time_utils.py

@@ -1,4 +1,4 @@
-# Copyright © 2023 Apple Inc.
+# Copyright © 2023-2024 Apple Inc.
 
 import time
 
@@ -6,7 +6,11 @@ import mlx.core as mx
 
 
 def time_fn(fn, *args, **kwargs):
-    print(f"Timing {fn.__name__} ...", end=" ")
+    msg = kwargs.pop("msg", None)
+    if msg:
+        print(f"Timing {msg} ...", end=" ")
+    else:
+        print(f"Timing {fn.__name__} ...", end=" ")
 
     # warmup
     for _ in range(5):
@@ -20,3 +24,15 @@ def time_fn(fn, *args, **kwargs):
 
     msec = 1e3 * (toc - tic) / num_iters
     print(f"{msec:.5f} msec")
+
+
+def measure_runtime(fn, **kwargs):
+    # Warmup
+    for _ in range(5):
+        fn(**kwargs)
+
+    tic = time.time()
+    iters = 100
+    for _ in range(iters):
+        fn(**kwargs)
+    return (time.time() - tic) * 1000 / iters

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:
+# Args: TARGET: Custom target to be added for the metal library TITLE: Name of
-#     TARGET: Custom target to be added for the metal library 
+# the .metallib OUTPUT_DIRECTORY: Where to place ${TITLE}.metallib SOURCES: List
-#     TITLE: Name of the .metallib
+# of source files INCLUDE_DIRS: List of include dirs DEPS: List of dependency
-#     OUTPUT_DIRECTORY: Where to place ${TITLE}.metallib
+# files (like headers)
-#     SOURCES: List of source files
-#     INCLUDE_DIRS: List of include dirs
-#     DEPS: List of dependency files (like headers)
 #
 macro(mlx_build_metallib)
   # Parse args
   set(oneValueArgs TARGET TITLE OUTPUT_DIRECTORY)
   set(multiValueArgs SOURCES INCLUDE_DIRS DEPS)
-  cmake_parse_arguments(
+  cmake_parse_arguments(MTLLIB "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-      MTLLIB 
-      ""
-      "${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 ${MTLLIB_BUILD_TARGET}
-    COMMAND xcrun -sdk macosx metal 
+    COMMAND
-                  "$<LIST:TRANSFORM,${MTLLIB_INCLUDE_DIRS},PREPEND,-I>"
+      xcrun -sdk macosx metal
-                  ${MTLLIB_COMPILE_OPTIONS}
+      "$<LIST:TRANSFORM,${MTLLIB_INCLUDE_DIRS},PREPEND,-I>"
-                  ${MTLLIB_SOURCES}
+      ${MTLLIB_COMPILE_OPTIONS} ${MTLLIB_SOURCES} -o ${MTLLIB_BUILD_TARGET}
-                  -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(
+  add_custom_target(${MTLLIB_TARGET} DEPENDS ${MTLLIB_BUILD_TARGET})
-    ${MTLLIB_TARGET}
-    DEPENDS
-    ${MTLLIB_BUILD_TARGET}
-  )
 
-endmacro(mlx_build_metallib)
+endmacro(mlx_build_metallib)

docs/.gitignore

@@ -1,2 +1,3 @@
 src/python/_autosummary*/
 src/python/nn/_autosummary*/
+src/python/optimizers/_autosummary*/

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)
+Install Doxygen:
-for example with `conda`:
 
 ```
-conda install sphinx
+brew install doxygen
-pip install sphinx-book-theme
+```
+
+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

@@ -4,16 +4,17 @@
 
 .. autoclass:: {{ objname }}
 
-   {#{% block methods %}
+   {% block methods %}
 
    {% if methods %}
    .. rubric:: {{ _('Methods') }}
 
    .. autosummary::
    {% for item in methods %}
-      {%- if item not in inherited_members and item != '__init__' %}
+      {%- if item not in inherited_members and item != "__init__" %}
          ~{{ name }}.{{ item }}
       {%- endif %}
    {%- endfor %}
    {% endif %}
-   {% endblock %}#}
+   {% endblock %}
+

docs/src/conf.py

@@ -12,7 +12,7 @@ import mlx.core as mx
 project = "MLX"
 copyright = "2023, MLX Contributors"
 author = "MLX Contributors"
-version = ".".join(mx.__version__.split()[:-1])
+version = ".".join(mx.__version__.split(".")[:3])
 release = version
 
 # -- General configuration ---------------------------------------------------
@@ -22,22 +22,28 @@ extensions = [
     "sphinx.ext.autosummary",
     "sphinx.ext.intersphinx",
     "sphinx.ext.napoleon",
+    "breathe",
 ]
 
 python_use_unqualified_type_names = True
 autosummary_generate = True
+autosummary_filename_map = {"mlx.core.Stream": "stream_class"}
 
 intersphinx_mapping = {
-    "https://docs.python.org/3": None,
+    "python": ("https://docs.python.org/3", None),
-    "https://numpy.org/doc/stable/": 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 -------------------------------------------------
 
@@ -48,11 +54,45 @@ html_theme_options = {
     "repository_url": "https://github.com/ml-explore/mlx",
     "use_repository_button": True,
     "navigation_with_keys": False,
+    "logo": {
+        "image_light": "_static/mlx_logo.png",
+        "image_dark": "_static/mlx_logo_dark.png",
+    },
 }
 
-html_logo = "_static/mlx_logo.png"
+html_favicon = html_theme_options["logo"]["image_light"]
-
 
 # -- 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 +1,427 @@
+.. _custom_metal_kernels:
+
+Custom Metal Kernels
+====================
+
+MLX supports writing custom Metal kernels through the Python and C++ APIs.
+
+Simple Example
+--------------
+
+Let's write a custom kernel that computes ``exp`` elementwise:
+
+.. code-block:: python
+
+  def exp_elementwise(a: mx.array):
+      source = """
+          uint elem = thread_position_in_grid.x;
+          T tmp = inp[elem];
+          out[elem] = metal::exp(tmp);
+      """
+
+      kernel = mx.fast.metal_kernel(
+          name="myexp",
+          input_names=["inp"],
+          output_names=["out"],
+          source=source,
+      )
+      outputs = kernel(
+          inputs=[a],
+          template=[("T", mx.float32)],
+          grid=(a.size, 1, 1),
+          threadgroup=(256, 1, 1),
+          output_shapes=[a.shape],
+          output_dtypes=[a.dtype],
+      )
+      return outputs[0]
+
+  a = mx.random.normal(shape=(4, 16)).astype(mx.float16)
+  b = exp_elementwise(a)
+  assert mx.allclose(b, mx.exp(a))
+
+.. note::
+    We are only required to pass the body of the Metal kernel in ``source``.
+
+The full function signature will be generated using:
+
+* The shapes/dtypes of ``inputs``
+    In the above, ``a`` is an ``mx.array`` of type ``mx.float16`` and we pass it with the key ``inp``
+    so we will add ``const device float16_t* inp`` to the signature.
+    ``inp_shape``, ``inp_strides`` and ``inp_ndim`` are also added for convenience if they are present
+    in ``source``.
+* The list of ``output_dtypes``
+    In the above, ``out`` is an ``mx.array`` of type ``mx.float16``
+    so we add ``device float16_t* out``.
+* Template parameters passed using ``template``
+    In the above, ``template=[("T", mx.float32)]`` adds a template of ``template <typename T>`` to the function
+    and instantiates the template with ``custom_kernel_myexp_float<float>``.
+    Template parameters can be ``mx.core.Dtype``, ``int`` or ``bool``.
+* Metal attributes used in ``source`` such as ``[[thread_position_in_grid]]``
+    These will be added as function arguments.
+    All the attributes defined in Table 5.8 of the `Metal Shading Language Specification <https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf>`_ are supported.
+
+Putting this all together, the generated function signature for ``myexp`` is as follows:
+
+.. code-block:: cpp
+
+  template <typename T>
+  [[kernel]] void custom_kernel_myexp_float(
+    const device float16_t* inp [[buffer(0)]],
+    device float16_t* out [[buffer(1)]],
+    uint3 thread_position_in_grid [[thread_position_in_grid]]) {
+
+          uint elem = thread_position_in_grid.x;
+          T tmp = inp[elem];
+          out[elem] = metal::exp(tmp);
+
+  }
+
+  template [[host_name("custom_kernel_myexp_float")]] [[kernel]] decltype(custom_kernel_myexp_float<float>) custom_kernel_myexp_float<float>;
+
+Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads <https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_ function.
+This means we will launch ``mx.prod(grid)`` threads, subdivided into ``threadgroup`` size threadgroups.
+For optimal performance, each thread group dimension should be less than or equal to the corresponding grid dimension.
+
+Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
+
+Using Shape/Strides
+-------------------
+
+``mx.fast.metal_kernel`` supports an argument ``ensure_row_contiguous`` which is ``True`` by default.
+This will copy the ``mx.array`` inputs if needed before the kernel is launched to ensure that the memory layout is row contiguous.
+Generally this makes writing the kernel easier, since we don't have to worry about gaps or the ordering of the dims
+when indexing.
+
+If we want to avoid this copy, ``metal_kernel`` automatically passes ``a_shape``, ``a_strides`` and ``a_ndim`` for each
+input array ``a`` if any are present in ``source``.
+We can then use MLX's built in indexing utils to fetch the right elements for each thread.
+
+Let's convert ``myexp`` above to support arbitrarily strided arrays without relying on a copy from ``ensure_row_contiguous``:
+
+.. code-block:: python
+
+  def exp_elementwise(a: mx.array):
+      source = """
+          uint elem = thread_position_in_grid.x;
+          // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
+          uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
+          T tmp = inp[loc];
+          // Output arrays are always row contiguous
+          out[elem] = metal::exp(tmp);
+      """
+
+      kernel = mx.fast.metal_kernel(
+          name="myexp_strided",
+          input_names=["inp"],
+          output_names=["out"],
+          source=source
+      )
+      outputs = kernel(
+          inputs=[a],
+          template=[("T", mx.float32)],
+          grid=(a.size, 1, 1),
+          threadgroup=(256, 1, 1),
+          output_shapes=[a.shape],
+          output_dtypes=[a.dtype],
+          ensure_row_contiguous=False,
+      )
+      return outputs[0]
+
+  a = mx.random.normal(shape=(4, 16)).astype(mx.float16)
+  # make non-contiguous
+  a = a[::2]
+  b = exp_elementwise(a)
+  assert mx.allclose(b, mx.exp(a))
+
+Complex Example
+-----------------------------
+
+Let's implement a more complex example: ``grid_sample`` in ``"bilinear"`` mode.
+
+We'll start with the following MLX implementation using standard ops:
+
+.. code-block:: python
+
+    def grid_sample_ref(x, grid):
+        N, H_in, W_in, _ = x.shape
+        ix = ((grid[..., 0] + 1) * W_in - 1) / 2
+        iy = ((grid[..., 1] + 1) * H_in - 1) / 2
+
+        ix_nw = mx.floor(ix).astype(mx.int32)
+        iy_nw = mx.floor(iy).astype(mx.int32)
+
+        ix_ne = ix_nw + 1
+        iy_ne = iy_nw
+
+        ix_sw = ix_nw
+        iy_sw = iy_nw + 1
+
+        ix_se = ix_nw + 1
+        iy_se = iy_nw + 1
+
+        nw = (ix_se - ix)    * (iy_se - iy)
+        ne = (ix    - ix_sw) * (iy_sw - iy)
+        sw = (ix_ne - ix)    * (iy    - iy_ne)
+        se = (ix    - ix_nw) * (iy    - iy_nw)
+
+        I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
+        I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
+        I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
+        I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
+
+        mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
+        mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
+        mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
+        mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
+
+        I_nw *= mask_nw[..., None]
+        I_ne *= mask_ne[..., None]
+        I_sw *= mask_sw[..., None]
+        I_se *= mask_se[..., None]
+
+        output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
+
+        return output
+
+Now let's use ``mx.custom_function`` together with ``mx.fast.metal_kernel``
+to write a fast GPU kernel for both the forward and backward passes.
+
+First we'll implement the forward pass as a fused kernel:
+
+.. code-block:: python
+
+    @mx.custom_function
+    def grid_sample(x, grid):
+
+        assert x.ndim == 4, "`x` must be 4D."
+        assert grid.ndim == 4, "`grid` must be 4D."
+
+        B, _, _, C = x.shape
+        _, gN, gM, D = grid.shape
+        out_shape = (B, gN, gM, C)
+
+        assert D == 2, "Last dim of `grid` must be size 2."
+
+        source = """
+            uint elem = thread_position_in_grid.x;
+            int H = x_shape[1];
+            int W = x_shape[2];
+            int C = x_shape[3];
+            int gH = grid_shape[1];
+            int gW = grid_shape[2];
+
+            int w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
+
+            uint grid_idx = elem / C * 2;
+            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+
+            int ix_nw = floor(ix);
+            int iy_nw = floor(iy);
+
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
+
+            int ix_sw = ix_nw;
+            int iy_sw = iy_nw + 1;
+
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
+
+            T nw = (ix_se - ix)    * (iy_se - iy);
+            T ne = (ix    - ix_sw) * (iy_sw - iy);
+            T sw = (ix_ne - ix)    * (iy    - iy_ne);
+            T se = (ix    - ix_nw) * (iy    - iy_nw);
+
+            int batch_idx = elem / C / gH / gW * b_stride;
+            int channel_idx = elem % C;
+            int base_idx = batch_idx + channel_idx;
+
+            T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
+            T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
+            T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
+            T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
+
+            I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
+            I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
+            I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
+            I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
+
+            out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
+        """
+        kernel = mx.fast.metal_kernel(
+            name="grid_sample",
+            input_names=["x", "grid"],
+            output_names=["out"],
+            source=source,
+        )
+        outputs = kernel(
+            inputs=[x, grid],
+            template=[("T", x.dtype)],
+            output_shapes=[out_shape],
+            output_dtypes=[x.dtype],
+            grid=(np.prod(out_shape), 1, 1),
+            threadgroup=(256, 1, 1),
+        )
+        return outputs[0]
+
+For a reasonably sized input such as:
+
+.. code-block:: python
+
+    x.shape = (8, 1024, 1024, 64)
+    grid.shape = (8, 256, 256, 2)
+
+On an M1 Max, we see a big performance improvement:
+
+``55.7ms -> 6.7ms => 8x speed up``
+
+Grid Sample VJP
+---------------
+
+Since we decorated ``grid_sample`` with ``mx.custom_function``, we can now define
+its custom vjp transform so MLX can differentiate it.
+
+The backwards pass requires atomically updating ``x_grad``/``grid_grad`` and so
+requires a few extra ``mx.fast.metal_kernel`` features:
+
+* ``init_value=0``
+    Initialize all of the kernel's outputs to this value before it runs. This allows us to update only part of the output arrays with the kernel.
+
+* ``atomic_outputs=True``
+    Designate all of the kernel outputs as ``atomic`` in the function signature. 
+    This means we can use Metal's ``atomic`` features to simultaneously update the ``x_grad`` and ``grid_grad`` arrays from multiple threadgroups. 
+    See section 6.15 of the `Metal Shading Language Specification <https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf>`_ for more details.
+
+We can then implement the backwards pass as follows:
+
+.. code-block:: python
+
+    @grid_sample.vjp
+    def grid_sample_vjp(primals, cotangent, _):
+        x, grid = primals
+        B, _, _, C = x.shape
+        _, gN, gM, D = grid.shape
+
+        assert D == 2, "Last dim of `grid` must be size 2."
+
+        source = """
+            uint elem = thread_position_in_grid.x;
+            int H = x_shape[1];
+            int W = x_shape[2];
+            int C = x_shape[3];
+            // Pad C to the nearest larger simdgroup size multiple
+            int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
+
+            int gH = grid_shape[1];
+            int gW = grid_shape[2];
+
+            int w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
+
+            uint grid_idx = elem / C_padded * 2;
+            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+
+            int ix_nw = floor(ix);
+            int iy_nw = floor(iy);
+
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
+
+            int ix_sw = ix_nw;
+            int iy_sw = iy_nw + 1;
+
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
+
+            T nw = (ix_se - ix)    * (iy_se - iy);
+            T ne = (ix    - ix_sw) * (iy_sw - iy);
+            T sw = (ix_ne - ix)    * (iy    - iy_ne);
+            T se = (ix    - ix_nw) * (iy    - iy_nw);
+
+            int batch_idx = elem / C_padded / gH / gW * b_stride;
+            int channel_idx = elem % C_padded;
+            int base_idx = batch_idx + channel_idx;
+
+            T gix = T(0);
+            T giy = T(0);
+            if (channel_idx < C) {
+                int cot_index = elem / C_padded * C + channel_idx;
+                T cot = cotangent[cot_index];
+                if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
+                    int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
+
+                    T I_nw = x[offset];
+                    gix -= I_nw * (iy_se - iy) * cot;
+                    giy -= I_nw * (ix_se - ix) * cot;
+                }
+                if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
+                    int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
+
+                    T I_ne = x[offset];
+                    gix += I_ne * (iy_sw - iy) * cot;
+                    giy -= I_ne * (ix - ix_sw) * cot;
+                }
+                if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
+                    int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
+
+                    T I_sw = x[offset];
+                    gix -= I_sw * (iy - iy_ne) * cot;
+                    giy += I_sw * (ix_ne - ix) * cot;
+                }
+                if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
+                    int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
+
+                    T I_se = x[offset];
+                    gix += I_se * (iy - iy_nw) * cot;
+                    giy += I_se * (ix - ix_nw) * cot;
+                }
+            }
+
+            T gix_mult = W / 2;
+            T giy_mult = H / 2;
+
+            // Reduce across each simdgroup first.
+            // This is much faster than relying purely on atomics.
+            gix = simd_sum(gix);
+            giy = simd_sum(giy);
+
+            if (thread_index_in_simdgroup == 0) {
+                atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
+                atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
+            }
+        """
+        kernel = mx.fast.metal_kernel(
+            name="grid_sample_grad",
+            input_names=["x", "grid", "cotangent"],
+            output_names=["x_grad", "grid_grad"],
+            source=source,
+            atomic_outputs=True,
+        )
+        # pad the output channels to simd group size
+        # so that our `simd_sum`s don't overlap.
+        simdgroup_size = 32
+        C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
+        grid_size = B * gN * gM * C_padded
+        outputs = kernel(
+            inputs=[x, grid, cotangent],
+            template=[("T", x.dtype)],
+            output_shapes=[x.shape, grid.shape],
+            output_dtypes=[x.dtype, x.dtype],
+            grid=(grid_size, 1, 1),
+            threadgroup=(256, 1, 1),
+            init_value=0,
+        )
+        return outputs[0], outputs[1]
+
+There's an even larger speed up for the vjp:
+
+``676.4ms -> 16.7ms => 40x speed up``

docs/src/dev/extensions.rst

@@ -1,24 +1,16 @@
-Developer Documentation
+Custom Extensions in MLX
-=======================
+========================
 
-MLX provides a open and flexible backend to which users may add operations 
+You can extend MLX with custom operations on the CPU or GPU. This guide
-and specialized implementations without much hassle. While the library supplies
+explains how to do that with a simple example.
-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. 
 
 Introducing the Example
 -----------------------
 
-Let's say that you would like an operation that takes in two arrays, 
+Let's say you would like an operation that takes in two arrays, ``x`` and
-``x`` and ``y``, scales them both by some coefficients ``alpha`` and ``beta``
+``y``, scales them both by coefficients ``alpha`` and ``beta`` respectively,
-respectively, and then adds them together to get the result 
+and then adds them together to get the result ``z = alpha * x + beta * y``.
-``z = alpha * x + beta * y``. Well, you can very easily do that by just 
+You can do that in MLX directly:
-writing out a function as follows:
 
 .. 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 
+This function performs that operation while leaving the implementation and
-differentiation to MLX. 
+function transformations to MLX.
 
-However, you work with vector math libraries often and realize that the 
+However you may need to customize the underlying implementation, perhaps to
-``axpby`` routine defines the same operation ``Y = (alpha * X) + (beta * Y)``. 
+make it faster or for custom differentiation. In this tutorial we will go
-You would really like the part of your applications that does this operation 
+through adding custom extensions. It will cover:
-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 add 
-your own implementation for the gradients of your new operation while going 
-over the ins-and-outs of the MLX framework. 
 
-Well, what a coincidence! You are in the right place. Over the course of this 
+* The structure of the MLX library.
-example, we will learn:
+* Implementing a CPU operation that redirects to Accelerate_ when appropriate.
-
+* Implementing a GPU operation using metal.
-* The structure of the MLX library from the frontend API to the backend implementations.
+* Adding the ``vjp`` and ``jvp`` function transformation.
-* How to implement your own CPU backend that redirects to Accelerate_ when appropriate (and a fallback if needed).
+* Building a custom extension and binding it to python.
-* 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.
 
 Operations and Primitives
 -------------------------
 
-In one sentence, operations in MLX build the computation graph, and primitives 
+Operations in MLX build the computation graph. Primitives provide the rules for
-provide the rules for evaluation and transformations of said graph. Let's start 
+evaluating and transforming the graph. Let's start by discussing operations in
-by discussing operations in more detail. 
+more detail.
 
 Operations
 ^^^^^^^^^^^
 
-Operations are the frontend functions that operate on arrays. They are defined 
+Operations are the front-end functions that operate on arrays. They are defined
-in the C++ API (:ref:`cpp_ops`) and then we provide bindings to these 
+in the C++ API (:ref:`cpp_ops`), and the Python API (:ref:`ops`) binds them.
-operations in the Python API (:ref:`ops`). 
 
-We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and ``y``,
+We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and
-and two scalars, ``alpha`` and ``beta``. This is how we would define it in the 
+``y``, and two scalars, ``alpha`` and ``beta``. This is how to define it in
-C++ API:
+C++:
 
 .. code-block:: C++
 
@@ -83,10 +66,7 @@ C++ API:
         StreamOrDevice s = {} // Stream on which to schedule the operation
     );
 
-
+The simplest way to this operation is in terms of existing operations:
-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. 
 
 .. 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 
+The operations themselves do not contain the implementations that act on the
-do not contain the implementations that act on the data, nor do they contain the
+data, nor do they contain the rules of transformations. Rather, they are an
-rules of transformations. Rather, they are an easy to use interface that build 
+easy to use interface that use :class:`Primitive` building blocks.
-on top of the building blocks we call :class:`Primitive`. 
 
 Primitives
 ^^^^^^^^^^^
 
-A :class:`Primitive` is part of the computation graph of an :class:`array`. It 
+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,
+defines how to create outputs arrays given a input arrays. Further, a
-a :class:`Primitive` is a class that contains rules on how it is evaluated 
+:class:`Primitive` has methods to run on the CPU or GPU and for function
-on the CPU or GPU, and how it acts under transformations such as ``vjp`` and 
+transformations such as ``vjp`` and ``jvp``.  Lets go back to our example to be
-``jvp``. These words on their own can be a bit abstract, so lets take a step 
+more concrete:
-back and go to our example to give ourselves a more concrete image. 
 
 .. 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_cpu(
-        void eval_gpu(const std::vector<array>& inputs, array& out) override;
+            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 
+The :class:`Axpby` class derives from the base :class:`Primitive` class. The
-follows the above demonstrated interface. :class:`Axpby` treats ``alpha`` and 
+:class:`Axpby` treats ``alpha`` and ``beta`` as parameters. It then provides
-``beta`` as parameters. It then provides implementations of how the array ``out`` 
+implementations of how the output array is produced given the inputs through
-is produced given ``inputs`` through :meth:`Axpby::eval_cpu` and 
+:meth:`Axpby::eval_cpu` and :meth:`Axpby::eval_gpu`. It also provides rules
-:meth:`Axpby::eval_gpu`. Further, it provides rules of transformations in 
+of transformations in :meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and
-:meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and :meth:`Axpby::vmap`. 
+:meth:`Axpby::vmap`.
 
-Using the Primitives
+Using the Primitive
-^^^^^^^^^^^^^^^^^^^^^
+^^^^^^^^^^^^^^^^^^^
 
-Operations can use this :class:`Primitive` to add a new :class:`array` to 
+Operations can use this :class:`Primitive` to add a new :class:`array` to the
-the computation graph. An :class:`array` can be constructed by providing its 
+computation graph. An :class:`array` can be constructed by providing its data
-data type, shape, the :class:`Primitive` that computes it, and the 
+type, shape, the :class:`Primitive` that computes it, and the :class:`array`
-:class:`array` inputs that are passed to the primitive.
+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 
+No computation happens when we call the operation alone. The operation only
-operation only builds the computation graph. When we evaluate the output 
+builds the computation graph. When we evaluate the output array, MLX schedules
-array, MLX schedules the execution of the computation graph, and calls
+the execution of the computation graph, and calls :meth:`Axpby::eval_cpu` or
-:meth:`Axpby::eval_cpu` or :meth:`Axpby::eval_gpu` depending on the 
+:meth:`Axpby::eval_gpu` depending on the stream/device specified by the user.
-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 
+Let's start by implementing a naive and generic version of
-:meth:`Axpby::eval_cpu`. We declared this as a private member function of 
+:meth:`Axpby::eval_cpu`. We declared this as a private member function of
-:class:`Axpby` earlier called :meth:`Axpby::eval`. 
+:class:`Axpby` earlier called :meth:`Axpby::eval`.
 
-Our naive method will go over each element of the output array, find the 
+Our naive method will go over each element of the output array, find the
-corresponding input elements of ``x`` and ``y`` and perform the operation 
+corresponding input elements of ``x`` and ``y`` and perform the operation
-pointwise. This is captured in the templated function :meth:`axpby_impl`. 
+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 
+Our implementation should work for all incoming floating point arrays.
-for all incoming floating point arrays. Accordingly, we add dispatches for 
+Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
-``float32``, ``float16``, ``bfloat16`` and ``complex64``. We throw an error 
+``complex64``. We throw an error if we encounter an unexpected type.
-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) {
+    void Axpby::eval(
-        // Check the inputs (registered in the op while constructing the out array)
+      const std::vector<array>& inputs,
-        assert(inputs.size() == 2);
+      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. 
+This is good as a fallback implementation. We can use the ``axpby`` routine
-Remember we wanted to use the ``axpby`` routine provided by the Accelerate_
+provided by the Accelerate_ framework for a faster implementation in certain
-framework? Well, there are 3 complications to keep in mind:
+cases:
 
 #.  Accelerate does not provide implementations of ``axpby`` for half precision
-    floats. We can only direct to it for ``float32`` types 
+    floats. We can only use it for ``float32`` types.
-#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all elements
+#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all
-    have fixed strides between them. Possibly due to broadcasts and transposes, 
+    elements have fixed strides between them. We only direct to Accelerate
-    we aren't guaranteed that the inputs fit this requirement. We can 
+    if both ``x`` and ``y`` are row contiguous or column contiguous.
-    only direct to Accelerate if both ``x`` and ``y`` are row contiguous or 
+#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` in-place.
-    column contiguous. 
+    MLX expects to write the output to a new array. We must copy the elements
-#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` inplace. 
+    of ``y`` into the output and use that as an input to ``axpby``.
-    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``
 
-Let's write out an implementation that uses Accelerate in the right conditions. 
+Let's write an implementation that uses Accelerate in the right conditions.
-It must simply allocate data for the output, copy elements of ``y`` into it, 
+It allocates data for the output, copies ``y`` into it, and then calls the
-and then call the :meth:`catlas_saxpby` from accelerate. 
+: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
-
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
-        // 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());
 
         // 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?
+For inputs that do not fit the criteria for accelerate, we fall back to
-Luckily, we can always just direct back to :meth:`Axpby::eval`.
+:meth:`Axpby::eval`. With this in mind, let's finish our
-
+:meth:`Axpby::eval_cpu`.
-With this in mind, lets finally implement 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
+        // Fall back to common back-end if specializations are not available
-        eval(inputs, out);
+        eval(inputs, outputs);
     }
 
-We have now hit a milestone! Just this much is enough to run the operation 
+Just this much is enough to run the operation :meth:`axpby` on a CPU stream! If
-:meth:`axpby` on a CPU stream! 
+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 
+Implementing the GPU Back-end
-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
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Apple silicon devices address their GPUs using the Metal_ shading language, and 
+Apple silicon devices address their GPUs using the Metal_ shading language, and
-all GPU kernels in MLX are written using metal. 
+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 
+Let's keep the GPU kernel simple. We will launch exactly as many threads as
-as there are elements in the output. Each thread will pick the element it needs 
+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 
+from ``x`` and ``y``, do the point-wise operation, and update its assigned
-element in the output. 
+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 instantiation a unique host name so we can identify it.
-each data type. 
 
 .. 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 
+The logic to determine the kernel, set the inputs, resolve the grid dimensions,
-will see later in :ref:`Building with CMake`. In the following example, we 
+and dispatch to the GPU are contained in :meth:`Axpby::eval_gpu` as shown
-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 
 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
@@ -518,23 +479,22 @@ 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
+        // Make sure the metal library is available
-        // in the same folder as this executable if needed
+        d.register_library("mlx_ext");
-        d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
 
         // 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);
+        compute_encoder.set_compute_pipeline_state(kernel);
 
         // Kernel parameters are registered with buffer indices corresponding to
         // those in the kernel declaration at axpby.metal
@@ -542,21 +502,21 @@ below.
         size_t nelem = out.size();
 
         // Encode input arrays to kernel
-        set_array_buffer(compute_encoder, x, 0);
+        compute_encoder.set_input_array(x, 0);
-        set_array_buffer(compute_encoder, y, 1);
+        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.set_bytes(alpha_, 3);
-        compute_encoder->setBytes(&beta_, sizeof(float), 4);
+        compute_encoder.set_bytes(beta_, 4);
 
-        // Encode shape, strides and ndim 
+        // Encode shape, strides and ndim
-        compute_encoder->setBytes(x.shape().data(), ndim * sizeof(int), 5);
+        compute_encoder.set_vector_bytes(x.shape(), 5);
-        compute_encoder->setBytes(x.strides().data(), ndim * sizeof(size_t), 6);
+        compute_encoder.set_vector_bytes(x.strides(), 6);
-        compute_encoder->setBytes(y.strides().data(), ndim * sizeof(size_t), 7);
+        compute_encoder.set_bytes(y.strides(), 7);
-        compute_encoder->setBytes(&ndim, sizeof(int), 8);
+        compute_encoder.set_bytes(ndim, 8);
 
         // We launch 1 thread for each input and make sure that the number of
         // threads in any given threadgroup is not higher than the max allowed
@@ -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.dispatch_threads(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 
+A few things to note about MLX and Metal before moving on. MLX keeps track of
-of the active ``compute_encoder``. We rely on :meth:`d.get_command_encoder` 
+the active ``command_buffer`` and the ``MTLCommandBuffer`` to which it is
-to give us the active metal compute command encoder instead of building a 
+associated. We rely on :meth:`d.get_command_encoder` to give us the active
-new one and calling :meth:`compute_encoder->end_encoding` at the end. 
+metal compute command encoder instead of building a new one and calling
-MLX keeps adding kernels (compute pipelines) to the active command encoder 
+:meth:`compute_encoder->end_encoding` at the end. MLX adds kernels (compute
-until some specified limit is hit or the compute encoder needs to be flushed 
+pipelines) to the active command buffer until some specified limit is hit or
-for synchronization. MLX also handles enqueuing and committing the associated 
+the command buffer needs to be flushed for synchronization.
-command buffers as needed. We suggest taking a deeper dive into 
-:class:`metal::Device` if you would like to study this routine further.
 
 Primitive Transforms
 ^^^^^^^^^^^^^^^^^^^^^
 
-Now that we have come this far, let's also learn how to add implementations to 
+Next, let's add implementations for transformations in a :class:`Primitive`.
-transformations in a :class:`Primitive`. These transformations can be built on 
+These transformations can be built on top of other operations, including the
-top of our operations, including the one we just defined now. Which then gives 
+one we just defined:
-us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
 
 .. 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<array>& cotangents,
-            const std::vector<int>& argnums) {
+            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());
+            auto scale_arr = array(scale, cotangents[0].dtype());
-            vjps.push_back(multiply(scale_arr, cotan, stream()));
+            vjps.push_back(multiply(scale_arr, cotangents[0], stream()));
         }
         return vjps;
     }
 
-Finally, you need not have a transformation fully defined to start using your 
+Note, a transformation does not need to be fully defined to start using
-own :class:`Primitive`.
+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 
+* ``extensions/mlx_sample_extensions`` sets out the structure for the
-  associated python package
+  associated Python package
-* ``extensions/bindings.cpp`` provides python bindings for our operation
+* ``extensions/bindings.cpp`` provides Python bindings for our operation
-* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and 
+* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and
-  python bindings
+  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 
+We use nanobind_ to build a Python API for the C++ library. Since bindings for
-for all needed components such as `mlx.core.array`, `mlx.core.stream`, etc. 
+components such as :class:`mlx.core.array`, :class:`mlx.core.stream`, etc. are
-are already provided, adding our :meth:`axpby` becomes very simple!
+already provided, adding our :meth:`axpby` is simple.
 
 .. code-block:: C++
 
-    PYBIND11_MODULE(mlx_sample_extensions, m) {
+   NB_MODULE(_ext, m) {
-        m.doc() = "Sample C++ and metal extensions for MLX";
+        m.doc() = "Sample extension for MLX";
 
         m.def(
             "axpby",
             &axpby,
             "x"_a,
             "y"_a,
-            py::pos_only(),
             "alpha"_a,
             "beta"_a,
-            py::kw_only(),
+            nb::kw_only(),
-            "stream"_a = py::none(),
+            "stream"_a = nb::none(),
-            R"pbdoc(
+            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 @@ are already provided, adding our :meth:`axpby` becomes very 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.core` must be imported before importing
-    :mod:`mlx_sample_extensions` as defined by the pybind11 module above to 
+    :mod:`mlx_sample_extensions` as defined by the nanobind module above to
-    ensure that the casters for :mod:`mlx.core` components like 
+    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 
+Building the C++ extension library only requires that you ``find_package(MLX
-``find_package(MLX CONFIG)`` and then link it to your library. 
+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 
+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 
+:meth:`mlx_build_metallib` function that builds a ``.metallib`` target given
-sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and 
+sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and
-automatically imported with MLX package). 
+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(
+    nanobind_add_module(
-        mlx_sample_extensions
+      _ext
-        ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
+      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"],
+            packages=["mlx_sample_extensions"],
-            package_dir = {"": "mlx_sample_extensions"},
+            package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-            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 .`` 
+When you try to install using the command ``python -m pip install .`` (in
-(in ``extensions/``), the package will be installed with the same structure as 
+``extensions/``), the package will be installed with the same structure as
-``extensions/mlx_sample_extensions`` and the C++ and metal library will be 
+``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``.
+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 
+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!
+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 
+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. 
+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:
+The results are ``Simple axpby: 0.114 s | Custom axpby: 0.109 s``. We see
+modest improvements right away!
 
-.. code-block::
+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
-    Simple axpby: 0.114 s | Custom axpby: 0.109 s
+:meth:`grad`.
-
-We see some 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 such as :meth:`grad` and :meth:`simplify`!
 
 Scripts
 -------
 
 .. admonition:: Download the code
 
-   The full example code is available in `mlx-examples <code>`_.
+   The full example code is available in `mlx <https://github.com/ml-explore/mlx/tree/main/examples/extensions/>`_.
-
-.. code: `TODO_LINK/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

@@ -41,7 +41,9 @@ are the CPU and GPU.
    usage/indexing
    usage/saving_and_loading
    usage/function_transforms
+   usage/compile
    usage/numpy
+   usage/distributed
    usage/using_streams
 
 .. toctree::
@@ -57,14 +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::
@@ -78,3 +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
+- Using a native Python >= 3.9
-- macOS >= 13.3
+- 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>`_
+Then simply build and install MLX using pip:
-installed. You can install ``pybind11`` with ``pip``, ``brew`` or ``conda`` as follows:
 
 .. code-block:: shell
 
-    pip install "pybind11[global]"
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
-    conda install pybind11
-    brew install pybind11
 
-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 
+    If you have multiple Xcode installations and wish to use
-    a specific one while building, you can do so by adding the 
+    a specific one while building, you can do so by adding the
-    following environment variable before building 
+    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 across 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

@@ -9,9 +9,11 @@ Devices and Streams
   :toctree: _autosummary
 
    Device
+   Stream
    default_device
    set_default_device
-   Stream
    default_stream
    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 +1,15 @@
+.. _fast:
+
+Fast
+====
+
+.. currentmodule:: mlx.core.fast
+
+.. autosummary:: 
+  :toctree: _autosummary
+
+  rms_norm
+  layer_norm
+  rope
+  scaled_dot_product_attention
+  metal_kernel

docs/src/python/linalg.rst

@@ -8,4 +8,13 @@ Linear Algebra
 .. autosummary:: 
    :toctree: _autosummary 
 
+    inv
+    tri_inv
     norm
+    cholesky
+    cholesky_inv
+    cross
+    qr
+    svd
+    eigvalsh
+    eigh

docs/src/python/metal.rst

@@ -0,0 +1,20 @@
+Metal
+=====
+
+.. currentmodule:: mlx.core.metal
+
+.. 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::
 
@@ -180,3 +181,4 @@ In detail:
    nn/layers
    nn/functions
    nn/losses
+   nn/init

docs/src/python/nn/functions.rst

@@ -12,12 +12,28 @@ simple functions.
    :toctree: _autosummary_functions
    :template: nn-module-template.rst
 
+   elu
+   celu
    gelu
    gelu_approx
    gelu_fast_approx
+   glu
+   hard_shrink
+   hard_tanh
+   hardswish
+   leaky_relu
+   log_sigmoid
+   log_softmax
    mish
    prelu
    relu
+   relu6
    selu
+   sigmoid
    silu
+   softmax
+   softmin
+   softplus
+   softshrink
    step
+   tanh

docs/src/python/nn/init.rst

@@ -0,0 +1,45 @@
+.. _init:
+
+.. currentmodule:: mlx.nn.init
+
+Initializers
+------------
+
+The ``mlx.nn.init`` package contains commonly used initializers for neural
+network parameters. Initializers return a function which can be applied to any
+input :obj:`mlx.core.array` to produce an initialized output.
+
+For example:
+
+.. code:: python
+
+   import mlx.core as mx
+   import mlx.nn as nn
+
+   init_fn = nn.init.uniform()
+
+   # Produces a [2, 2] uniform matrix
+   param = init_fn(mx.zeros((2, 2)))
+
+To re-initialize all the parameter in an :obj:`mlx.nn.Module` from say a uniform 
+distribution, you can do:
+
+.. code:: python
+  
+   import mlx.nn as nn
+   model = nn.Sequential(nn.Linear(5, 10), nn.ReLU(), nn.Linear(10, 5))
+   init_fn = nn.init.uniform(low=-0.1, high=0.1)
+   model.apply(init_fn)
+   
+
+.. autosummary::
+   :toctree: _autosummary
+
+   constant
+   normal
+   uniform
+   identity
+   glorot_normal
+   glorot_uniform
+   he_normal
+   he_uniform

docs/src/python/nn/layers.rst

@@ -10,28 +10,60 @@ Layers
    :template: nn-module-template.rst
 
    ALiBi
+   AvgPool1d
+   AvgPool2d
+   AvgPool3d
    BatchNorm
+   CELU
    Conv1d
    Conv2d
+   Conv3d
+   ConvTranspose1d
+   ConvTranspose2d
+   ConvTranspose3d
    Dropout
    Dropout2d
    Dropout3d
    Embedding
+   ELU
    GELU
+   GLU
    GroupNorm
+   GRU
+   HardShrink
+   HardTanh
+   Hardswish
    InstanceNorm
    LayerNorm
+   LeakyReLU
    Linear
+   LogSigmoid
+   LogSoftmax
+   LSTM
+   MaxPool1d
+   MaxPool2d
+   MaxPool3d
    Mish
    MultiHeadAttention
    PReLU
+   QuantizedEmbedding
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU6
+   RNN
    RoPE
    SELU
    Sequential
+   Sigmoid
    SiLU
    SinusoidalPositionalEncoding
+   Softmin
+   Softshrink
+   Softsign
+   Softmax
+   Softplus
    Step
+   Tanh
    Transformer
+   Upsample

docs/src/python/nn/losses.rst

@@ -18,6 +18,7 @@ Loss Functions
    kl_div_loss
    l1_loss
    log_cosh_loss
+   margin_ranking_loss
    mse_loss
    nll_loss
    smooth_l1_loss

docs/src/python/nn/module.rst

@@ -11,6 +11,7 @@ Module
       :toctree: _autosummary
    
       Module.training
+      Module.state
    
    .. rubric:: Methods
 
@@ -29,6 +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
@@ -19,42 +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
-   isnan
+   isfinite
-   isposinf
+   isclose
-   isneginf
    isinf
+   isnan
+   isneginf
+   isposinf
+   issubdtype
+   left_shift
    less
    less_equal
    linspace
@@ -72,22 +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
@@ -106,6 +151,7 @@ Operations
    square
    squeeze
    stack
+   std
    stop_gradient
    subtract
    sum
@@ -115,11 +161,15 @@ Operations
    tan
    tanh
    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
 ==========
 
@@ -29,19 +31,48 @@ model's parameters and the **optimizer state**.
             # Compute the new parameters but also the optimizer state.
             mx.eval(model.parameters(), optimizer.state)
 
-.. currentmodule:: mlx.optimizers
+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
-   :template: optimizers-template.rst
 
-   OptimizerState
+   clip_grad_norm
-   Optimizer
-   SGD
-   RMSprop
-   Adagrad
-   AdaDelta
-   Adam
-   AdamW
-   Adamax
-   Lion

docs/src/python/optimizers/common_optimizers.rst

@@ -0,0 +1,20 @@
+.. _common_optimizers:
+
+Common Optimizers
+=================
+
+.. currentmodule:: mlx.optimizers
+
+.. autosummary::
+   :toctree: _autosummary
+   :template: optimizers-template.rst
+
+   SGD
+   RMSprop
+   Adagrad
+   Adafactor
+   AdaDelta
+   Adam
+   AdamW
+   Adamax
+   Lion

docs/src/python/optimizers/optimizer.rst

@@ -0,0 +1,23 @@
+Optimizer
+=========
+
+.. currentmodule:: mlx.optimizers
+
+.. autoclass:: Optimizer 
+
+
+   .. rubric:: Attributes
+
+   .. autosummary::
+      :toctree: _autosummary
+
+      Optimizer.state
+   
+   .. rubric:: Methods
+
+   .. autosummary::
+      :toctree: _autosummary
+   
+      Optimizer.apply_gradients
+      Optimizer.init
+      Optimizer.update

docs/src/python/optimizers/schedulers.rst

@@ -0,0 +1,15 @@
+.. _schedulers:
+
+Schedulers
+==========
+
+.. currentmodule:: mlx.optimizers
+
+.. autosummary::
+   :toctree: _autosummary
+
+   cosine_decay    
+   exponential_decay    
+   join_schedules
+   linear_schedule
+   step_decay    

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

@@ -9,9 +9,12 @@ Transforms
   :toctree: _autosummary
 
    eval
+   compile
+   custom_function
+   disable_compile
+   enable_compile
    grad
    value_and_grad
    jvp
    vjp
    vmap
-   simplify

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

@@ -0,0 +1,423 @@
+.. _compile:
+
+Compilation
+===========
+
+.. currentmodule:: mlx.core
+
+MLX has a :func:`compile` function transformation which compiles computation
+graphs. Function compilation results in smaller graphs by merging common work
+and fusing certain operations. In many cases this can lead to big improvements
+in run-time and memory use.
+
+Getting started with :func:`compile` is simple, but there are some edge cases
+that are good to be aware of for more complex graphs and advanced usage.
+
+Basics of Compile
+-----------------
+
+Let's start with a simple example:
+
+.. code-block:: python
+
+  def fun(x, y):
+      return mx.exp(-x) + y
+
+  x = mx.array(1.0)
+  y = mx.array(2.0)
+
+  # Regular call, no compilation
+  # Prints: array(2.36788, dtype=float32)
+  print(fun(x, y))
+
+  # Compile the function
+  compiled_fun = mx.compile(fun)
+
+  # 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
+function multiple times will not initiate a new compilation. This means you
+should typically compile functions that you plan to use more than once.
+
+.. code-block:: python
+
+  def fun(x, y):
+      return mx.exp(-x) + y
+
+  x = mx.array(1.0)
+  y = mx.array(2.0)
+
+  compiled_fun = mx.compile(fun)
+
+  # Compiled here
+  compiled_fun(x, y)
+
+  # Not compiled again
+  compiled_fun(x, y)
+
+  # Not compiled again
+  mx.compile(fun)(x, y)
+
+There are some important cases to be aware of that can cause a function to
+be recompiled:
+
+* Changing the shape or number of dimensions
+* Changing the type of any of the inputs
+* Changing the number of inputs to the function
+
+In certain cases only some of the compilation stack will be rerun (for
+example when changing the shapes) and in other cases the full compilation
+stack will be rerun (for example when changing the types). In general you
+should avoid compiling functions too frequently.
+
+Another idiom to watch out for is compiling functions which get created and
+destroyed frequently. This can happen, for example, when compiling an anonymous
+function in a loop:
+
+.. code-block:: python
+
+  a = mx.array(1.0)
+  # Don't do this, compiles lambda at each iteration
+  for _ in range(5):
+      mx.compile(lambda x: mx.exp(mx.abs(x)))(a)
+
+Example Speedup
+---------------
+
+The :func:`mlx.nn.gelu` is a nonlinear activation function commonly used with
+Transformer-based models. The implementation involves several unary and binary
+element-wise operations:
+
+.. code-block:: python
+
+  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
+use it with large arrays it will be memory bandwidth bound.  However, all of
+the operations in the ``gelu`` are fusible into a single kernel with
+:func:`compile`. This can speedup both cases considerably.
+
+Let's compare the runtime of the regular function versus the compiled
+function. We'll use the following timing helper which does a warm up and
+handles synchronization:
+
+.. code-block:: python
+
+  import time
+
+  def timeit(fun, x):
+      # warm up
+      for _ in range(10):
+          mx.eval(fun(x))
+
+      tic = time.perf_counter()
+      for _ in range(100):
+          mx.eval(fun(x))
+      toc = time.perf_counter()
+      tpi = 1e3 * (toc - tic) / 100
+      print(f"Time per iteration {tpi:.3f} (ms)")
+
+
+Now make an array, and benchmark both functions:
+
+.. code-block:: python
+
+  x = mx.random.uniform(shape=(32, 1000, 4096))
+  timeit(nn.gelu, x)
+  timeit(mx.compile(nn.gelu), x)
+
+On an M1 Max the times are 15.5 and 3.1 milliseconds. The compiled ``gelu`` is
+five times faster.
+
+Debugging
+---------
+
+When a compiled function is first called, it is traced with placeholder
+inputs. This means you can't evaluate arrays (for example to print their
+contents) inside compiled functions.
+
+.. code-block:: python
+
+  @mx.compile
+  def fun(x):
+      z = -x
+      print(z)  # Crash
+      return mx.exp(z)
+
+  fun(mx.array(5.0))
+
+For debugging, inspecting arrays can be helpful. One way to do that is to
+globally disable compilation using the :func:`disable_compile` function or
+``MLX_DISABLE_COMPILE`` flag. For example the following is okay even though
+``fun`` is compiled:
+
+.. code-block:: python
+
+  @mx.compile
+  def fun(x):
+      z = -x
+      print(z) # Okay
+      return mx.exp(z)
+
+  mx.disable_compile()
+  fun(mx.array(5.0))
+
+
+Pure Functions
+--------------
+
+Compiled functions are intended to be *pure*; that is they should not have side
+effects. For example:
+
+.. code-block:: python
+
+  state = []
+
+  @mx.compile
+  def fun(x, y):
+      z = x + y
+      state.append(z)
+      return mx.exp(z)
+
+  fun(mx.array(1.0), mx.array(2.0))
+  # Crash!
+  print(state)
+
+After the first call of ``fun``, the ``state`` list will hold a placeholder
+array. The placeholder does not have any data; it is only used to build the
+computation graph. Printing such an array results in a crash.
+
+You have two options to deal with this. The first option is to simply return
+``state`` as an output:
+
+.. code-block:: python
+
+   state = []
+
+   @mx.compile
+   def fun(x, y):
+      z = x + y
+      state.append(z)
+      return mx.exp(z), state
+
+    _, state = fun(mx.array(1.0), mx.array(2.0))
+    # Prints [array(3, dtype=float32)]
+    print(state)
+
+In some cases returning updated state can be pretty inconvenient. Hence,
+:func:`compile` has a parameter to capture implicit outputs:
+
+.. code-block:: python
+
+  from functools import partial
+
+  state = []
+
+  # Tell compile to capture state as an output
+  @partial(mx.compile, outputs=state)
+  def fun(x, y):
+      z = x + y
+      state.append(z)
+      return mx.exp(z), state
+
+  fun(mx.array(1.0), mx.array(2.0))
+  # Prints [array(3, dtype=float32)]
+  print(state)
+
+This is particularly useful for compiling a function which includes an update
+to a container of arrays, as is commonly done when training the parameters of a
+:class:`mlx.nn.Module`.
+
+Compiled functions will also treat any inputs not in the parameter list as
+constants. For example:
+
+.. code-block:: python
+
+  state = [mx.array(1.0)]
+
+  @mx.compile
+  def fun(x):
+      return x + state[0]
+
+  # Prints array(2, dtype=float32)
+  print(fun(mx.array(1.0)))
+
+  # Update state
+  state[0] = mx.array(5.0)
+
+  # Still prints array(2, dtype=float32)
+  print(fun(mx.array(1.0)))
+
+In order to have the change of state reflected in the outputs of ``fun`` you
+again have two options. The first option is to simply pass ``state`` as input
+to the function. In some cases this can be pretty inconvenient. Hence,
+:func:`compile` also has a parameter to capture implicit inputs:
+
+.. code-block:: python
+
+  from functools import partial
+  state = [mx.array(1.0)]
+
+  # Tell compile to capture state as an input
+  @partial(mx.compile, inputs=state)
+  def fun(x):
+      return x + state[0]
+
+  # Prints array(2, dtype=float32)
+  print(fun(mx.array(1.0)))
+
+  # Update state
+  state[0] = mx.array(5.0)
+
+  # Prints array(6, dtype=float32)
+  print(fun(mx.array(1.0)))
+
+
+Compiling Training Graphs
+-------------------------
+
+This section will step through how to use :func:`compile` with a simple example
+of a common setup: training a model with :obj:`mlx.nn.Module` using an
+:obj:`mlx.optimizers.Optimizer` with state. We will show how to compile the
+full forward, backward, and update with :func:`compile`.
+
+To start, here is the simple example without any compilation:
+
+.. code-block:: python
+
+  import mlx.core as mx
+  import mlx.nn as nn
+  import mlx.optimizers as optim
+
+  # 4 examples with 10 features each
+  x = mx.random.uniform(shape=(4, 10))
+
+  # 0, 1 targets
+  y = mx.array([0, 1, 0, 1])
+
+  # Simple linear model
+  model = nn.Linear(10, 1)
+
+  # SGD with momentum
+  optimizer = optim.SGD(learning_rate=0.1, momentum=0.8)
+
+  def loss_fn(model, x, y):
+      logits = model(x).squeeze()
+      return nn.losses.binary_cross_entropy(logits, y)
+
+  loss_and_grad_fn = nn.value_and_grad(model, loss_fn)
+
+  # Perform 10 steps of gradient descent
+  for it in range(10):
+      loss, grads = loss_and_grad_fn(model, x, y)
+      optimizer.update(model, grads)
+      mx.eval(model.parameters(), optimizer.state)
+
+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
+
+  import mlx.core as mx
+  import mlx.nn as nn
+  import mlx.optimizers as optim
+  from functools import partial
+
+  # 4 examples with 10 features each
+  x = mx.random.uniform(shape=(4, 10))
+
+  # 0, 1 targets
+  y = mx.array([0, 1, 0, 1])
+
+  # Simple linear model
+  model = nn.Linear(10, 1)
+
+  # SGD with momentum
+  optimizer = optim.SGD(learning_rate=0.1, momentum=0.8)
+
+  def loss_fn(model, x, y):
+      logits = model(x).squeeze()
+      return nn.losses.binary_cross_entropy(logits, y)
+
+  # 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)
+      loss, grads = loss_and_grad_fn(model, x, y)
+      optimizer.update(model, grads)
+      return loss
+
+  # Perform 10 steps of gradient descent
+  for it in range(10):
+      loss = step(x, y)
+      # Evaluate the model and optimizer state
+      mx.eval(state)
+      print(loss)
+
+
+.. note::
+
+  If you are using a module which performs random sampling such as
+  :func:`mlx.nn.Dropout`, make sure you also include ``mx.random.state`` in the
+  ``state`` captured by :func:`compile`, i.e. ``state = [model.state,
+  optimizer.state, mx.random.state]``.
+
+
+.. note::
+
+   For more examples of compiling full training graphs checkout the  `MLX
+   Examples <https://github.com/ml-explore/mlx-examples>`_ GitHub repo.
+
+Transformations with Compile
+----------------------------
+
+In MLX function transformations are composable. You can apply any function
+transformation to the output of any other function transformation. For more on
+this, see the documentation on :ref:`function transforms
+<function_transforms>`.
+
+Compiling transformed functions works just as expected:
+
+.. code-block:: python
+
+  grad_fn = mx.grad(mx.exp)
+
+  compiled_grad_fn = mx.compile(grad_fn)
+
+  # Prints: array(2.71828, dtype=float32)
+  print(grad_fn(mx.array(1.0)))
+
+  # Also prints: array(2.71828, dtype=float32)
+  print(compiled_grad_fn(mx.array(1.0)))
+
+.. note::
+
+   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`.
+
+You can also compile functions which themselves call compiled functions. A
+good practice is to compile the outer most function to give :func:`compile`
+the most opportunity to optimize the computation graph:
+
+.. code-block:: python
+
+  @mx.compile
+  def inner(x):
+      return mx.exp(-mx.abs(x))
+
+  def outer(x):
+      inner(inner(x))
+
+  # Compiling the outer function is good to do as it will likely
+  # be faster even though the inner functions are compiled
+  fun = mx.compile(outer)

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

@@ -5,9 +5,12 @@ Function Transforms
 
 .. currentmodule:: mlx.core
 
-MLX uses composable function transformations for automatic differentiation and
+MLX uses composable function transformations for automatic differentiation,
-vectorization. The key idea behind composable function transformations is that
+vectorization, and compute graph optimizations. To see the complete list of
-every transformation returns a function which can be further transformed. 
+function transformations check-out the :ref:`API documentation <transforms>`.
+
+The key idea behind composable function transformations is that every
+transformation returns a function which can be further transformed.
 
 Here is a simple example:
 
@@ -22,7 +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
 
@@ -36,10 +39,10 @@ Using :func:`grad` on the output of :func:`grad` is always ok. You keep
 getting higher order derivatives.
 
 Any of the MLX function transformations can be composed in any order to any
-depth. To see the complete list of function transformations check-out the
+depth. See the following sections for more information on :ref:`automatic
-:ref:`API documentation <transforms>`. See the following sections for more
+differentiation <auto diff>` and :ref:`automatic vectorization <vmap>`.
-information on :ref:`automatic differentiaion <auto diff>` and
+For more information on :func:`compile` see the :ref:`compile documentation <compile>`.
-:ref:`automatic vectorization <vmap>`.
+
 
 Automatic Differentiation
 -------------------------
@@ -47,7 +50,7 @@ Automatic Differentiation
 .. _auto diff:
 
 Automatic differentiation in MLX works on functions rather than on implicit
-graphs. 
+graphs.
 
 .. note::
 
@@ -111,7 +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)}
@@ -129,7 +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.
 
 
@@ -158,19 +161,19 @@ A naive way to add the elements from two sets of vectors is with a loop:
   ys = mx.random.uniform(shape=(100, 4096))
 
   def naive_add(xs, ys):
-      return [xs[i] + ys[:, i] for i in range(xs.shape[1])]
+      return [xs[i] + ys[:, i] for i in range(xs.shape[0])]
 
 Instead you can use :func:`vmap` to automatically vectorize the addition:
 
 .. code-block:: python
-   
+
    # Vectorize over the second dimension of x and the
    # first dimension of y
-   vmap_add = mx.vmap(lambda x, y: x + y, in_axes=(1, 0))
+   vmap_add = mx.vmap(lambda x, y: x + y, in_axes=(0, 1))
 
 The ``in_axes`` parameter can be used to specify which dimensions of the
 corresponding input to vectorize over. Similarly, use ``out_axes`` to specify
-where the vectorized axes should be in the outputs. 
+where the vectorized axes should be in the outputs.
 
 Let's time these two different versions:
 
@@ -181,8 +184,8 @@ Let's time these two different versions:
   print(timeit.timeit(lambda: mx.eval(naive_add(xs, ys)), number=100))
   print(timeit.timeit(lambda: mx.eval(vmap_add(xs, ys)), number=100))
 
-On an M1 Max the naive version takes in total ``0.390`` seconds whereas the
+On an M1 Max the naive version takes in total ``5.639`` seconds whereas the
-vectorized version takes only ``0.025`` seconds, more than ten times faster.
+vectorized version takes only ``0.024`` seconds, more than 200 times faster.
 
 Of course, this operation is quite contrived. A better approach is to simply do
 ``xs + ys.T``, but for more complex functions :func:`vmap` can be quite handy.

docs/src/usage/indexing.rst

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

docs/src/usage/lazy_evaluation.rst

@@ -13,14 +13,14 @@ 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 like :func:`simplify`.
+:func:`vmap` and graph optimizations.
 
 Currently, MLX does not compile and rerun compute graphs. They are all
 generated dynamically. However, lazy evaluation makes it much easier to
@@ -109,14 +109,14 @@ Here is a concrete example:
 
 An important behavior to be aware of is when the graph will be implicitly
 evaluated. Anytime you ``print`` an array, convert it to an
-:obj:`numpy.ndarray`, or otherwise access it's memory via :obj:`memoryview`,
+:obj:`numpy.ndarray`, or otherwise access its memory via :obj:`memoryview`,
 the graph will be evaluated. Saving arrays via :func:`save` (or any other MLX
 saving functions) will also evaluate the array.
 
 
 Calling :func:`array.item` on a scalar array will also evaluate it. In the
 example above, printing the loss (``print(loss)``) or adding the loss scalar to
-a list (``losses.append(loss.item())``) would cause a graph evaluation. If 
+a list (``losses.append(loss.item())``) would cause a graph evaluation. If
 these lines are before ``mx.eval(loss, model.parameters())`` then this
 will be a partial evaluation, computing only the forward pass.
 

docs/src/usage/numpy.rst

@@ -3,7 +3,11 @@
 Conversion to NumPy and Other Frameworks
 ========================================
 
-MLX array implements the `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_.
+MLX array supports conversion between other frameworks with either:
+
+* The `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_.
+* `DLPack <https://dmlc.github.io/dlpack/latest/>`_.
+
 Let's convert an array to NumPy and back.
 
 .. code-block:: python
@@ -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 
+   * - Format
-     - Extension 
+     - Extension
      - Function
-     - Notes 
+     - Notes
-   * - NumPy 
+   * - NumPy
-     - ``.npy`` 
+     - ``.npy``
      - :func:`save`
      - Single arrays only
-   * - NumPy archive 
+   * - NumPy archive
-     - ``.npz`` 
+     - ``.npz``
      - :func:`savez` and :func:`savez_compressed`
-     - Multiple arrays 
+     - Multiple arrays
    * - Safetensors
-     - ``.safetensors`` 
+     - ``.safetensors``
      - :func:`save_safetensors`
-     - Multiple arrays 
+     - Multiple arrays
-   * - GGUF 
+   * - GGUF
-     - ``.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);
+  auto d2fdx2 = grad(grad(fn))(x);
-  // df2dx2 is 2
+  // d2fdx2 is 2
 }
 
 int main() {

examples/extensions/CMakeLists.txt

@@ -1,6 +1,6 @@
-cmake_minimum_required(VERSION 3.24)
+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(
-find_package(pybind11 CONFIG REQUIRED)
+  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(
+target_sources(mlx_ext PUBLIC ${CMAKE_CURRENT_LIST_DIR}/axpby/axpby.cpp)
-  mlx_ext
-  PUBLIC
-  ${CMAKE_CURRENT_LIST_DIR}/axpby/axpby.cpp
-)
 
 # Add include headers
-target_include_directories(
+target_include_directories(mlx_ext PUBLIC ${CMAKE_CURRENT_LIST_DIR})
-  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
+    TARGET
-    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
-  )
+    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 -----------------------------
+# ----------------------------- Python Bindings -----------------------------
-pybind11_add_module(
+nanobind_add_module(
-  mlx_sample_extensions
+  _ext
-  ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
+  NB_STATIC
-)
+  STABLE_ABI
-target_link_libraries(mlx_sample_extensions PRIVATE mlx_ext)
+  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()
+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) {
+    std::vector<array>& outputs) {
-  auto out = out_arr[0];
   // 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(
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-      allocator::malloc_or_wait(y.data_size() * out.itemsize()),
-      y.data_size(),
-      y.strides(),
-      y.flags());
 
   // 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) {
+    std::vector<array>& outputs) {
-  auto out = outarr[0];
   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) {
+    const std::vector<array>& outputs) {
-  eval(inputs, out);
+  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
@@ -253,16 +249,15 @@ void Axpby::eval_gpu(
   kname << (contiguous_kernel ? "contiguous_" : "general_");
   kname << type_to_name(out);
 
-  // Make sure the metal library is available and look for it
+  // Make sure the metal library is available
-  // in the same folder as this executable if needed
+  d.register_library("mlx_ext");
-  d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
 
   // 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);
+  compute_encoder.set_compute_pipeline_state(kernel);
 
   // Kernel parameters are registered with buffer indices corresponding to
   // those in the kernel declaration at axpby.metal
@@ -270,22 +265,22 @@ void Axpby::eval_gpu(
   size_t nelem = out.size();
 
   // Encode input arrays to kernel
-  set_array_buffer(compute_encoder, x, 0);
+  compute_encoder.set_input_array(x, 0);
-  set_array_buffer(compute_encoder, y, 1);
+  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.set_bytes(alpha_, 3);
-  compute_encoder->setBytes(&beta_, sizeof(float), 4);
+  compute_encoder.set_bytes(beta_, 4);
 
   // Encode shape, strides and ndim if needed
   if (!contiguous_kernel) {
-    compute_encoder->setBytes(x.shape().data(), ndim * sizeof(int), 5);
+    compute_encoder.set_vector_bytes(x.shape(), 5);
-    compute_encoder->setBytes(x.strides().data(), ndim * sizeof(size_t), 6);
+    compute_encoder.set_vector_bytes(x.strides(), 6);
-    compute_encoder->setBytes(y.strides().data(), ndim * sizeof(size_t), 7);
+    compute_encoder.set_bytes(y.strides(), 7);
-    compute_encoder->setBytes(&ndim, sizeof(int), 8);
+    compute_encoder.set_bytes(ndim, 8);
   }
 
   // We launch 1 thread for each input and make sure that the number of
@@ -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.dispatch_threads(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

@@ -2,7 +2,6 @@
 
 #include <metal_stdlib>
 
-#include "mlx/backend/metal/kernels/bf16.h"
 #include "mlx/backend/metal/kernels/utils.h"
 
 template <typename T>
@@ -19,7 +18,7 @@ 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];
 }
 
@@ -31,33 +30,33 @@ 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)            \
+#define instantiate_axpby(type_name, type)                               \
-  template [[host_name("axpby_general_" #type_name)]] \
+  template [[host_name("axpby_general_" #type_name)]] [[kernel]] void    \
-  [[kernel]] void axpby_general<type>(                \
+  axpby_general<type>(                                                   \
-      device const type* x [[buffer(0)]],             \
+      device const type* x [[buffer(0)]],                                \
-      device const type* y [[buffer(1)]],             \
+      device const type* y [[buffer(1)]],                                \
-      device type* out [[buffer(2)]],                 \
+      device type* out [[buffer(2)]],                                    \
-      constant const float& alpha [[buffer(3)]],      \
+      constant const float& alpha [[buffer(3)]],                         \
-      constant const float& beta [[buffer(4)]],       \
+      constant const float& beta [[buffer(4)]],                          \
-      constant const int* shape [[buffer(5)]],        \
+      constant const int* shape [[buffer(5)]],                           \
-      constant const size_t* x_strides [[buffer(6)]], \
+      constant const size_t* x_strides [[buffer(6)]],                    \
-      constant const size_t* y_strides [[buffer(7)]], \
+      constant const size_t* y_strides [[buffer(7)]],                    \
-      constant const int& ndim [[buffer(8)]],         \
+      constant const int& ndim [[buffer(8)]],                            \
-      uint index [[thread_position_in_grid]]);        \
+      uint index [[thread_position_in_grid]]);                           \
-  template [[host_name("axpby_contiguous_" #type_name)]] \
+  template [[host_name("axpby_contiguous_" #type_name)]] [[kernel]] void \
-  [[kernel]] void axpby_contiguous<type>(                \
+  axpby_contiguous<type>(                                                \
-      device const type* x [[buffer(0)]],                \
+      device const type* x [[buffer(0)]],                                \
-      device const type* y [[buffer(1)]],                \
+      device const type* y [[buffer(1)]],                                \
-      device type* out [[buffer(2)]],                    \
+      device type* out [[buffer(2)]],                                    \
-      constant const float& alpha [[buffer(3)]],         \
+      constant const float& alpha [[buffer(3)]],                         \
-      constant const float& beta [[buffer(4)]],          \
+      constant const float& beta [[buffer(4)]],                          \
       uint index [[thread_position_in_grid]]);
 
 instantiate_axpby(float32, float);
 instantiate_axpby(float16, half);
 instantiate_axpby(bfloat16, bfloat16_t);
-instantiate_axpby(complex64, complex64_t);
+instantiate_axpby(complex64, complex64_t);

examples/extensions/bindings.cpp

@@ -1,31 +1,31 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
-#include <pybind11/pybind11.h>
+#include <nanobind/nanobind.h>
-#include <pybind11/stl.h>
+#include <nanobind/stl/variant.h>
 
 #include "axpby/axpby.h"
 
-namespace py = pybind11;
+namespace nb = nanobind;
-using namespace py::literals;
+using namespace nb::literals;
+
 using namespace mlx::core;
 
-PYBIND11_MODULE(mlx_sample_extensions, m) {
+NB_MODULE(_ext, m) {
-  m.doc() = "Sample C++ and metal extensions for MLX";
+  m.doc() = "Sample extension for MLX";
 
   m.def(
       "axpby",
       &axpby,
       "x"_a,
       "y"_a,
-      py::pos_only(),
       "alpha"_a,
       "beta"_a,
-      py::kw_only(),
+      nb::kw_only(),
-      "stream"_a = py::none(),
+      "stream"_a = nb::none(),
-      R"pbdoc(
+      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"]
+requires = [
-build-backend = "setuptools.build_meta"
+  "setuptools>=42",
+  "cmake>=3.24",
+  "mlx>=0.18.0",
+  "nanobind==2.2.0",
+]
+build-backend = "setuptools.build_meta"