Patch bump (#1408)

* xcode 16.0 with debug tests

* limit nproc for builds

* vmap bug

* assert bug

* run python tests in debug mode

* fix view, bool copies preserve bits'

* actual view fix

.circleci/config.yml

@@ -31,33 +31,35 @@ jobs:
           name: Install dependencies
           command: |
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install nanobind==2.1.0
             pip install numpy
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" CMAKE_BUILD_PARALLEL_LEVEL="" python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" CMAKE_BUILD_PARALLEL_LEVEL="" python3 setup.py develop
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              python3 setup.py build_ext --inplace
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              python3 setup.py develop
       - run:
           name: Generate package stubs
           command: |
             echo "stubs"
+            pip install typing_extensions
             python setup.py generate_stubs 
       - run:
           name: Run Python tests
           command: |
             python3 -m unittest discover python/tests -v
-      # TODO: Reenable when extension api becomes stable
-      # - run:
-      #     name: Build example extension
-      #     command: |
-      #       cd examples/extensions && python3 -m pip install . 
       - run:
           name: Build CPP only
           command: |
-            mkdir -p build && cd build && cmake .. -DMLX_BUILD_METAL=OFF && make -j
+            mkdir -p build && cd build 
+            cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
+            make -j `nproc`
       - run:
           name: Run CPP tests
           command: ./build/tests/tests
@@ -69,18 +71,19 @@ jobs:
         default: "15.2.0"
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
             brew install python@3.8
+            brew install openmpi
             python3.8 -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install nanobind==2.1.0
             pip install numpy
             pip install torch
             pip install tensorflow
@@ -89,11 +92,12 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL="" pip install -e . -v
+            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
+            pip install typing_extensions
             python setup.py generate_stubs 
       - run:
           name: Run Python tests
@@ -101,23 +105,47 @@ jobs:
             source env/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
-      # TODO: Reenable when extension api becomes stable
-      # - run:
-      #     name: Build example extension
-      #     command: |
-      #       cd examples/extensions && python3.11 -m pip install . 
+            mpirun -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
+      - run:
+          name: Build example extension
+          command: |
+            source env/bin/activate
+            cd examples/extensions
+            pip install -r requirements.txt
+            python setup.py build_ext -j8
       - store_test_results:
           path: test-results
       - run:
           name: Build CPP only
           command: |
             source env/bin/activate
-            mkdir -p build && cd build && cmake .. && make -j
+            mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
       - run:
           name: Run CPP tests
           command: |
             DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 ./build/tests/tests
-            DEVICE=cpu ./build/tests/tests
+      - run:
+          name: Build small binary
+          command: |
+            source env/bin/activate
+            cd build/
+            cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
+              -DBUILD_SHARED_LIBS=ON \
+              -DMLX_BUILD_CPU=OFF \
+              -DMLX_BUILD_SAFETENSORS=OFF \
+              -DMLX_BUILD_GGUF=OFF \
+              -DMLX_METAL_JIT=ON
+            make -j `sysctl -n hw.ncpu`
+      - run:
+          name: Run Python tests with JIT
+          command: |
+            source env/bin/activate
+            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
+              CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
+              pip install -e . -v
+            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
+              METAL_DEBUG_ERROR_MODE=0 \
+              python -m xmlrunner discover -v python/tests -o test-results/gpu_jit
 
   build_release:
     parameters:
@@ -132,18 +160,19 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
             brew install python@<< parameters.python_version >>
+            brew install openmpi
             python<< parameters.python_version >> -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install nanobind==2.1.0
             pip install --upgrade setuptools
             pip install numpy
             pip install twine
@@ -153,19 +182,20 @@ jobs:
           command: |
             source env/bin/activate
             DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
+            pip install typing_extensions
             python setup.py generate_stubs 
       - run:
           name: Build Python package
           command: |
             source env/bin/activate
             << parameters.build_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               python -m build -w
       - when:
           condition: << parameters.build_env >>
@@ -207,18 +237,19 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install nanobind==2.1.0
             pip install --upgrade setuptools
             pip install numpy
             pip install auditwheel
             pip install patchelf
             pip install build
             << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               pip install . -v
+            pip install typing_extensions
             python setup.py generate_stubs 
             << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL="" \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               python -m build --wheel
             auditwheel show dist/*
             auditwheel repair dist/* --plat manylinux_2_31_x86_64
@@ -239,7 +270,7 @@ workflows:
       - mac_build_and_test:
           matrix:
             parameters:
-              xcode_version: ["15.0.0", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test
 
   build_pypi_release:
@@ -274,7 +305,7 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              xcode_version: ["15.0.0", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
       - linux_build_and_test:
           requires: [ hold ]
   nightly_build:
@@ -298,7 +329,7 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["15.0.0", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0", "16.0.0"]
               build_env: ["DEV_RELEASE=1"]
   linux_test_release:
     when:

.github/workflows/pull_request.yml

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

.pre-commit-config.yaml

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

ACKNOWLEDGMENTS.md

@@ -7,15 +7,19 @@ with a short description of your contribution(s) below. For example:
 
 MLX was developed with contributions from the following individuals:
 
-- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops.
+- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops. Added `clip_grad_norm` along with `tree_reduce`.
 - Juarez Bochi: Fixed bug in cross attention.
 - Justin Deschenaux: Sine, Cosine, arange, randint, truncated normal, bernoulli, lion optimizer, Dropout2d, linear and logistic regression python example.
-- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream` and safetensor support.
+- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream`, safetensors support, `einsum`, and `einsum_path`.
 - Gabrijel Boduljak: Added `mlx.core.linalg`, implemented `norm` method and `InstanceNorm` layer. Implemented pooling layers and ``Upsample``.
 - Hinrik Snær Guðmundsson: Added `atleast_1d`, `atleast_2d`, `atleast_3d` ops.
 - Luca Arnaboldi: Added `Ceil` and `Floor` ops; implemented pickling, copy and deepcopy for mlx arrays.
 - Brian Keene & Atila Orhon, with Argmax Inc.: Added `fast.scaled_dot_product_attention`
-- AmirHossein Razlighi: Added chaining support for some of the ops in `nn.Module`.
+- AmirHossein Razlighi: Added chaining support for some of the ops in `nn.Module`. Comparison works for non array objects in `mlx.core.array`. Exception handling for invalid operations in `mlx.core.array`.
+- Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
+- Paul Paczuski: Improved stability of BCE loss calculation
+- Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
+
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>

CMakeLists.txt

@@ -15,12 +15,16 @@ 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 0.9.0)
+  set(MLX_VERSION 0.17.3)
 endif()
 
 # --------------------- Processor tests -------------------------
@@ -79,15 +83,17 @@ elseif (MLX_BUILD_METAL)
                   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 (${MACOS_VERSION} GREATER_EQUAL 14.0)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14_iOS17-beta.zip)
-  else()
+  if (${MACOS_VERSION} LESS 14.0)
     message(FATAL_ERROR "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON" )
   endif()
+  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
+
+  set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18-beta.zip)
+  # Get the metal version
+  execute_process(
+    COMMAND zsh "-c" "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal -E -x metal -P - | tail -1 | tr -d '\n'"
+    OUTPUT_VARIABLE MLX_METAL_VERSION
+    COMMAND_ERROR_IS_FATAL ANY)
 
   FetchContent_Declare(
     metal_cpp
@@ -101,55 +107,85 @@ elseif (MLX_BUILD_METAL)
     $<INSTALL_INTERFACE:include/metal_cpp>
   )
   target_link_libraries(
-    mlx
+    mlx PUBLIC
     ${METAL_LIB}
     ${FOUNDATION_LIB}
     ${QUARTZ_LIB})
+
+  add_compile_definitions("MLX_METAL_VERSION=${MLX_METAL_VERSION}")
 endif()
 
-find_library(ACCELERATE_LIBRARY Accelerate)
-if (MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
-  message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
-  set(MLX_BUILD_ACCELERATE ON)
-  target_link_libraries(mlx ${ACCELERATE_LIBRARY})
-  add_compile_definitions(ACCELERATE_NEW_LAPACK)
+if (MLX_BUILD_CPU)
+  find_library(ACCELERATE_LIBRARY Accelerate)
+  if (MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
+    message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
+    set(MLX_BUILD_ACCELERATE ON)
+    target_link_libraries(mlx PUBLIC ${ACCELERATE_LIBRARY})
+    add_compile_definitions(ACCELERATE_NEW_LAPACK)
+  else()
+    message(STATUS "Accelerate or arm neon not found, using default backend.")
+    set(MLX_BUILD_ACCELERATE OFF)
+    if(${CMAKE_HOST_APPLE})
+      # The blas shipped in macOS SDK is not supported, search homebrew for
+      # openblas instead.
+      set(BLA_VENDOR OpenBLAS)
+      set(LAPACK_ROOT "${LAPACK_ROOT};$ENV{LAPACK_ROOT};/usr/local/opt/openblas")
+    endif()
+    # Search and link with lapack.
+    find_package(LAPACK REQUIRED)
+    if (NOT LAPACK_FOUND)
+      message(FATAL_ERROR "Must have LAPACK installed")
+    endif()
+    find_path(LAPACK_INCLUDE_DIRS lapacke.h
+      /usr/include
+      /usr/local/include
+      /usr/local/opt/openblas/include)
+    message(STATUS "Lapack lib " ${LAPACK_LIBRARIES})
+    message(STATUS "Lapack include " ${LAPACK_INCLUDE_DIRS})
+    target_include_directories(mlx PRIVATE ${LAPACK_INCLUDE_DIRS})
+    target_link_libraries(mlx PUBLIC ${LAPACK_LIBRARIES})
+    # List blas after lapack otherwise we may accidentally incldue an old version
+    # of lapack.h from the include dirs of blas.
+    find_package(BLAS REQUIRED)
+    if (NOT BLAS_FOUND)
+      message(FATAL_ERROR "Must have BLAS installed")
+    endif()
+    # TODO find a cleaner way to do this
+    find_path(BLAS_INCLUDE_DIRS cblas.h
+      /usr/include
+      /usr/local/include
+      $ENV{BLAS_HOME}/include)
+    message(STATUS "Blas lib " ${BLAS_LIBRARIES})
+    message(STATUS "Blas include " ${BLAS_INCLUDE_DIRS})
+    target_include_directories(mlx PRIVATE ${BLAS_INCLUDE_DIRS})
+    target_link_libraries(mlx PUBLIC ${BLAS_LIBRARIES})
+  endif()
 else()
-  message(STATUS "Accelerate or arm neon not found, using default backend.")
   set(MLX_BUILD_ACCELERATE OFF)
-  if(${CMAKE_HOST_APPLE})
-    # The blas shipped in macOS SDK is not supported, search homebrew for
-    # openblas instead.
-    set(BLA_VENDOR OpenBLAS)
-    set(LAPACK_ROOT "${LAPACK_ROOT};$ENV{LAPACK_ROOT};/usr/local/opt/openblas")
-  endif()
-  # Search and link with lapack.
-  find_package(LAPACK REQUIRED)
-  if (NOT LAPACK_FOUND)
-    message(FATAL_ERROR "Must have LAPACK installed")
-  endif()
-  find_path(LAPACK_INCLUDE_DIRS lapacke.h
-    /usr/include
-    /usr/local/include
-    /usr/local/opt/openblas/include)
-  message(STATUS "Lapack lib " ${LAPACK_LIBRARIES})
-  message(STATUS "Lapack include " ${LAPACK_INCLUDE_DIRS})
-  target_include_directories(mlx PRIVATE ${LAPACK_INCLUDE_DIRS})
-  target_link_libraries(mlx ${LAPACK_LIBRARIES})
-  # List blas after lapack otherwise we may accidentally incldue an old version
-  # of lapack.h from the include dirs of blas.
-  find_package(BLAS REQUIRED)
-  if (NOT BLAS_FOUND)
-    message(FATAL_ERROR "Must have BLAS installed")
-  endif()
-  # TODO find a cleaner way to do this
-  find_path(BLAS_INCLUDE_DIRS cblas.h
-    /usr/include
-    /usr/local/include
-    $ENV{BLAS_HOME}/include)
-  message(STATUS "Blas lib " ${BLAS_LIBRARIES})
-  message(STATUS "Blas include " ${BLAS_INCLUDE_DIRS})
-  target_include_directories(mlx PRIVATE ${BLAS_INCLUDE_DIRS})
-  target_link_libraries(mlx ${BLAS_LIBRARIES})
+endif()
+
+find_package(MPI)
+if (MPI_FOUND)
+  execute_process(
+    COMMAND zsh "-c" "mpirun --version"
+    OUTPUT_VARIABLE MPI_VERSION
+    ERROR_QUIET
+  )
+  if (${MPI_VERSION} MATCHES ".*Open MPI.*")
+    target_include_directories(mlx PRIVATE ${MPI_INCLUDE_PATH})
+  elseif (MPI_VERSION STREQUAL "")
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING
+      "MPI found but mpirun is not available. Building without MPI."
+    )
+  else()
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING
+      "MPI which is not OpenMPI found. Building without MPI."
+    )
+  endif() 
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
@@ -161,6 +197,14 @@ target_include_directories(
   $<INSTALL_INTERFACE:include>
 )
 
+FetchContent_Declare(fmt
+  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+  GIT_TAG 10.2.1 
+  EXCLUDE_FROM_ALL
+)
+FetchContent_MakeAvailable(fmt)
+target_link_libraries(mlx PRIVATE fmt::fmt-header-only)
+
 if (MLX_BUILD_PYTHON_BINDINGS)
   message(STATUS "Building Python bindings.")
   find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)

README.md

@@ -88,13 +88,13 @@ for more information on building the C++ and Python APIs from source.
 
 ## Contributing 
 
-Check out the [contribution guidelines](CONTRIBUTING.md) for more information
+Check out the [contribution guidelines](https://github.com/ml-explore/mlx/tree/main/CONTRIBUTING.md) for more information
 on contributing to MLX. See the
 [docs](https://ml-explore.github.io/mlx/build/html/install.html) for more
 information on building from source, and running tests.
 
 We are grateful for all of [our
-contributors](ACKNOWLEDGMENTS.md#Individual-Contributors). If you contribute
+contributors](https://github.com/ml-explore/mlx/tree/main/ACKNOWLEDGMENTS.md#Individual-Contributors). If you contribute
 to MLX and wish to be acknowledged, please add your name to the list in your
 pull request.
 

benchmarks/cpp/time_utils.h

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

benchmarks/python/comparative/bench_torch.py

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

benchmarks/python/comparative/compare.py

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

benchmarks/python/compile_bench.py

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

benchmarks/python/conv1d_bench.py

@@ -0,0 +1,123 @@
+import argparse
+import math
+import os
+import subprocess
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+device_name = subprocess.check_output(["sysctl", "-n", "machdep.cpu.brand_string"])
+device_name = device_name.decode("utf-8").strip("\n")
+
+N_warmup = 10
+N_iter_bench = 100
+N_iter_func = 5
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+    torch.mps.synchronize()
+
+    s = time.perf_counter_ns()
+    for i in range(N_iter_bench):
+        f(a, b)
+    e = time.perf_counter_ns()
+    return (e - s) * 1e-9
+
+
+def make_mx_conv_1D(strides=1, padding=0, groups=1):
+    def mx_conv_1D(a, b):
+        ys = []
+        for _ in range(N_iter_func):
+            y = mx.conv1d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_1D
+
+
+def make_pt_conv_1D(strides=1, padding=0, groups=1):
+    @torch.no_grad()
+    def pt_conv_1D(a, b):
+        ys = []
+        for _ in range(N_iter_func):
+            y = torch.conv1d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        torch.mps.synchronize()
+        return ys
+
+    return pt_conv_1D
+
+
+def bench_shape(N, iH, C, wH, O, strides, padding, np_dtype, groups):
+    scale = 1.0 / math.sqrt(wH * C)
+    a_np = np.random.uniform(0, 0.5, (N, iH, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, wH, int(C / groups))).astype(np_dtype)
+
+    a_mx = mx.array(a_np)
+    b_mx = mx.array(b_np)
+
+    a_pt = torch.from_numpy(a_np.transpose((0, 2, 1))).to("mps")
+    b_pt = torch.from_numpy(b_np.transpose((0, 2, 1))).to("mps")
+
+    torch.mps.synchronize()
+
+    f_mx = make_mx_conv_1D(strides, padding, groups)
+    f_pt = make_pt_conv_1D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv1d(a_mx, b_mx, stride=strides, padding=padding, groups=groups)
+    out_pt = torch.conv1d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 1))
+    out_pt = out_pt.numpy(force=True)
+
+    atol = 2e-5 if np_dtype == np.float32 else 1e-4
+
+    if not np.allclose(out_pt, out_mx, atol=atol):
+        print(
+            f"Failed at {(N, iH, C)}, {(O, wH, C)} [strides = {strides}, padding = {padding}, groups = {groups}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
+        )
+
+    return time_mlx, time_torch
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run conv benchmarks")
+
+    dtypes = ("float32",)
+    shapes = (
+        (4, 32, 32, 5, 32, 1, 2, 1),
+        (4, 32, 32, 5, 32, 1, 2, 2),
+        (4, 32, 32, 5, 32, 1, 2, 4),
+        (4, 32, 32, 5, 32, 1, 2, 8),
+        (4, 32, 32, 5, 32, 1, 2, 8),
+        (4, 32, 32, 5, 32, 1, 2, 16),
+        (4, 32, 32, 5, 32, 1, 2, 32),
+        (4, 32, 256, 5, 512, 1, 2, 2),
+        (4, 32, 256, 5, 512, 1, 2, 128),
+        (4, 32, 256, 5, 512, 1, 2, 256),
+    )
+
+    for dtype in dtypes:
+        print("(N,  iH,  C),  (O,  wH,  C),   dtype,  stride, pads, groups, diff%")
+        for N, iH, C, wH, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, iH, C, wH, O, strides, padding, np_dtype, groups
+            )
+            diff = time_torch / time_mlx - 1.0
+
+            print(
+                f"({N}, {iH:3d}, {C:3d}), ({O:3d}, {wH:2d}, {C:3d}), {dtype}, {strides:5d}, {padding:4d}, {groups:6d}, {100. * diff:+5.2f}%"
+            )
+
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

benchmarks/python/conv_bench.py

@@ -28,11 +28,11 @@ def bench(f, a, b):
     return (e - s) * 1e-9
 
 
-def make_mx_conv_2D(strides=(1, 1), padding=(0, 0)):
+def make_mx_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
     def mx_conv_2D(a, b):
         ys = []
         for i in range(N_iter_func):
-            y = mx.conv2d(a, b, stride=strides, padding=padding)
+            y = mx.conv2d(a, b, stride=strides, padding=padding, groups=groups)
             ys.append(y)
         mx.eval(ys)
         return ys
@@ -40,12 +40,12 @@ def make_mx_conv_2D(strides=(1, 1), padding=(0, 0)):
     return mx_conv_2D
 
 
-def make_pt_conv_2D(strides=(1, 1), padding=(0, 0)):
+def make_pt_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
     @torch.no_grad()
     def pt_conv_2D(a, b):
         ys = []
         for i in range(N_iter_func):
-            y = torch.conv2d(a, b, stride=strides, padding=padding)
+            y = torch.conv2d(a, b, stride=strides, padding=padding, groups=groups)
             ys.append(y)
         torch.mps.synchronize()
         return ys
@@ -53,11 +53,12 @@ def make_pt_conv_2D(strides=(1, 1), padding=(0, 0)):
     return pt_conv_2D
 
 
-def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
-
+def bench_shape(N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype):
     scale = 1.0 / math.sqrt(kH * kH * C)
     a_np = np.random.uniform(0, 0.5, (N, H, W, C)).astype(np_dtype)
-    b_np = np.random.uniform(-scale, scale, (O, kH, kW, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kH, kW, int(C / groups))).astype(
+        np_dtype
+    )
 
     a_mx = mx.array(a_np)
     b_mx = mx.array(b_np)
@@ -67,15 +68,15 @@ def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
 
     torch.mps.synchronize()
 
-    f_mx = make_mx_conv_2D(strides, padding)
-    f_pt = make_pt_conv_2D(strides, padding)
+    f_mx = make_mx_conv_2D(strides, padding, groups)
+    f_pt = make_pt_conv_2D(strides, padding, groups)
 
     time_torch = bench(f_pt, a_pt, b_pt)
     time_mlx = bench(f_mx, a_mx, b_mx)
 
-    out_mx = mx.conv2d(a_mx, b_mx, stride=strides, padding=padding)
+    out_mx = mx.conv2d(a_mx, b_mx, stride=strides, padding=padding, groups=groups)
     out_pt = torch.conv2d(
-        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
     )
     out_pt = torch.permute(out_pt, (0, 2, 3, 1))
     out_pt = out_pt.numpy(force=True)
@@ -84,7 +85,7 @@ def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
 
     if not np.allclose(out_pt, out_mx, atol=atol):
         print(
-            f"Failed at {(N, H, W, C)}, {(O, kH, kW, C)} [strides = {strides}, padding = {padding}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
+            f"Failed at {(N, H, W, C)}, {(O, kH, kW, C)} [strides = {strides}, padding = {padding}, groups = {groups}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
         )
 
     return time_mlx, time_torch
@@ -95,35 +96,40 @@ if __name__ == "__main__":
 
     dtypes = ("float32",)
     shapes = (
-        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2)),
-        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2)),
-        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2)),
-        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2)),
-        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2)),
-        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2)),
-        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2)),
-        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2)),
-        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2)),
-        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2)),
-        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2)),
-        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2)),
-        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2)),
+        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2), 1),
+        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 2),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 16),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 64),
+        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2), 1),
     )
 
     for dtype in dtypes:
-        print("(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  diff%")
-        for N, H, W, C, kH, kW, O, strides, padding in shapes:
+        print(
+            "(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  groups, diff%"
+        )
+        for N, H, W, C, kH, kW, O, strides, padding, groups in shapes:
             np_dtype = getattr(np, dtype)
             time_mlx, time_torch = bench_shape(
-                N, H, W, C, kH, kW, O, strides, padding, np_dtype
+                N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype
             )
             diff = time_torch / time_mlx - 1.0
 
             print(
-                f"({N}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kH:2d}, {kW:2d}, {C:3d}), {dtype}, {strides}, {padding}, {100. * diff:+5.2f}%"
+                f"({N}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kH:2d}, {kW:2d}, {C:3d}), {dtype}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
             )
             if time_mlx >= 2.0 * time_torch:
                 print("ATTENTION ^^^^^^^")

benchmarks/python/conv_transpose_bench.py

@@ -0,0 +1,135 @@
+import argparse
+import math
+import os
+import subprocess
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_warmup = 10
+N_iter_bench = 100
+N_iter_func = 5
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+    torch.mps.synchronize()
+
+    s = time.perf_counter_ns()
+    for i in range(N_iter_bench):
+        f(a, b)
+    e = time.perf_counter_ns()
+    return (e - s) * 1e-9
+
+
+def make_mx_conv_transpose_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    def mx_conv_transpose_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv_transpose2d(
+                a, b, stride=strides, padding=padding, groups=groups
+            )
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_transpose_2D
+
+
+def make_pt_conv_transpose_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_transpose_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv_transpose2d(
+                a, b, stride=strides, padding=padding, groups=groups
+            )
+            ys.append(y)
+        torch.mps.synchronize()
+        return ys
+
+    return pt_conv_transpose_2D
+
+
+def bench_shape(N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype):
+    scale = 1.0 / math.sqrt(kH * kH * C)
+    a_np = np.random.uniform(0, 0.5, (N, H, W, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kH, kW, int(C / groups))).astype(
+        np_dtype
+    )
+
+    a_mx = mx.array(a_np)
+    b_mx = mx.array(b_np)
+
+    a_pt = torch.from_numpy(a_np.transpose((0, 3, 1, 2))).to("mps")
+    b_pt = torch.from_numpy(b_np.transpose((3, 0, 1, 2))).to("mps")
+
+    torch.mps.synchronize()
+
+    f_mx = make_mx_conv_transpose_2D(strides, padding, groups)
+    f_pt = make_pt_conv_transpose_2D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv_transpose2d(
+        a_mx, b_mx, stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.conv_transpose2d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 3, 1))
+    out_pt = out_pt.numpy(force=True)
+
+    atol = 2e-5 if np_dtype == np.float32 else 1e-4
+
+    if not np.allclose(out_pt, out_mx, atol=atol):
+        print(
+            f"Failed at {(N, H, W, C)}, {(O, kH, kW, C)} [strides = {strides}, padding = {padding}, groups = {groups}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
+        )
+
+    return time_mlx, time_torch
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run conv benchmarks")
+
+    dtypes = ("float32",)
+    shapes = (
+        (4, 32, 32, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2), 1),
+        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 1),
+        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2), 1),
+        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2), 1),
+        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2), 1),
+        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2), 1),
+        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2), 1),
+    )
+
+    for dtype in dtypes:
+        print(
+            "(N,   H,   W,   C), (  O, kH, kW,   C),   dtype, stride,   pads,  groups, diff%"
+        )
+        for N, H, W, C, kH, kW, O, strides, padding, groups in shapes:
+            np_dtype = getattr(np, dtype)
+            time_mlx, time_torch = bench_shape(
+                N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype
+            )
+            diff = time_torch / time_mlx - 1.0
+
+            print(
+                f"({N}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kH:2d}, {kW:2d}, {C:3d}), {dtype}, {strides}, {padding}, {groups:7d}, {100. * diff:+5.2f}%"
+            )
+            if time_mlx >= 2.0 * time_torch:
+                print("ATTENTION ^^^^^^^")

benchmarks/python/distributed_bench.py

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

benchmarks/python/einsum_bench.py

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

benchmarks/python/fft_bench.py

@@ -0,0 +1,118 @@
+# Copyright © 2024 Apple Inc.
+
+import matplotlib
+import mlx.core as mx
+import numpy as np
+import sympy
+import torch
+from time_utils import measure_runtime
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+
+def bandwidth_gb(runtime_ms, system_size):
+    bytes_per_fft = np.dtype(np.complex64).itemsize * 2
+    bytes_per_gb = 1e9
+    ms_per_s = 1e3
+    return system_size * bytes_per_fft / runtime_ms * ms_per_s / bytes_per_gb
+
+
+def run_bench(system_size, fft_sizes, backend="mlx", dim=1):
+    def fft_mlx(x):
+        if dim == 1:
+            out = mx.fft.fft(x)
+        elif dim == 2:
+            out = mx.fft.fft2(x)
+        mx.eval(out)
+        return out
+
+    def fft_mps(x):
+        if dim == 1:
+            out = torch.fft.fft(x)
+        elif dim == 2:
+            out = torch.fft.fft2(x)
+        torch.mps.synchronize()
+        return out
+
+    bandwidths = []
+    for n in fft_sizes:
+        batch_size = system_size // n**dim
+        shape = [batch_size] + [n for _ in range(dim)]
+        if backend == "mlx":
+            x_np = np.random.uniform(size=(system_size // n, n)).astype(np.complex64)
+            x = mx.array(x_np)
+            mx.eval(x)
+            fft = fft_mlx
+        elif backend == "mps":
+            x_np = np.random.uniform(size=(system_size // n, n)).astype(np.complex64)
+            x = torch.tensor(x_np, device="mps")
+            torch.mps.synchronize()
+            fft = fft_mps
+        else:
+            raise NotImplementedError()
+        runtime_ms = measure_runtime(fft, x=x)
+        bandwidth = bandwidth_gb(runtime_ms, np.prod(shape))
+        print(n, bandwidth)
+        bandwidths.append(bandwidth)
+
+    return np.array(bandwidths)
+
+
+def time_fft():
+    x = np.array(range(2, 512))
+    system_size = int(2**26)
+
+    print("MLX GPU")
+    with mx.stream(mx.gpu):
+        gpu_bandwidths = run_bench(system_size=system_size, fft_sizes=x)
+
+    print("MPS GPU")
+    mps_bandwidths = run_bench(system_size=system_size, fft_sizes=x, backend="mps")
+
+    print("CPU")
+    system_size = int(2**20)
+    with mx.stream(mx.cpu):
+        cpu_bandwidths = run_bench(system_size=system_size, fft_sizes=x)
+
+    x = np.array(x)
+
+    all_indices = x - x[0]
+    radix_2to13 = (
+        np.array([i for i in x if all(p <= 13 for p in sympy.primefactors(i))]) - x[0]
+    )
+    bluesteins = (
+        np.array([i for i in x if any(p > 13 for p in sympy.primefactors(i))]) - x[0]
+    )
+
+    for indices, name in [
+        (all_indices, "All"),
+        (radix_2to13, "Radix 2-13"),
+        (bluesteins, "Bluestein's"),
+    ]:
+        # plot bandwidths
+        print(name)
+        plt.scatter(x[indices], gpu_bandwidths[indices], color="green", label="GPU")
+        plt.scatter(x[indices], mps_bandwidths[indices], color="blue", label="MPS")
+        plt.scatter(x[indices], cpu_bandwidths[indices], color="red", label="CPU")
+        plt.title(f"MLX FFT Benchmark -- {name}")
+        plt.xlabel("N")
+        plt.ylabel("Bandwidth (GB/s)")
+        plt.legend()
+        plt.savefig(f"{name}.png")
+        plt.clf()
+
+    av_gpu_bandwidth = np.mean(gpu_bandwidths)
+    av_mps_bandwidth = np.mean(mps_bandwidths)
+    av_cpu_bandwidth = np.mean(cpu_bandwidths)
+    print("Average bandwidths:")
+    print("GPU:", av_gpu_bandwidth)
+    print("MPS:", av_mps_bandwidth)
+    print("CPU:", av_cpu_bandwidth)
+
+    portion_faster = len(np.where(gpu_bandwidths > mps_bandwidths)[0]) / len(x)
+    print("Percent MLX faster than MPS: ", portion_faster * 100)
+
+
+if __name__ == "__main__":
+    time_fft()

benchmarks/python/hadamard_bench.py

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

benchmarks/python/sdpa_bench.py

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

docs/Doxyfile

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

docs/README.md

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

docs/requirements.txt

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

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

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

docs/src/conf.py

@@ -22,6 +22,7 @@ extensions = [
     "sphinx.ext.autosummary",
     "sphinx.ext.intersphinx",
     "sphinx.ext.napoleon",
+    "breathe",
 ]
 
 python_use_unqualified_type_names = True
@@ -33,6 +34,9 @@ intersphinx_mapping = {
     "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"
@@ -79,3 +83,15 @@ def setup(app):
 # -- Options for LaTeX output ------------------------------------------------
 
 latex_documents = [(main_doc, "MLX.tex", "MLX Documentation", author, "manual")]
+latex_elements = {
+    "preamble": r"""
+    \usepackage{enumitem}
+    \setlistdepth{5}
+    \setlist[itemize,1]{label=$\bullet$}
+    \setlist[itemize,2]{label=$\bullet$}
+    \setlist[itemize,3]{label=$\bullet$}
+    \setlist[itemize,4]{label=$\bullet$}
+    \setlist[itemize,5]{label=$\bullet$}
+    \renewlist{itemize}{itemize}{5}
+""",
+}

docs/src/cpp/ops.rst

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

docs/src/dev/custom_metal_kernels.rst

@@ -0,0 +1,421 @@
+Custom Metal Kernels
+====================
+
+MLX supports writing custom Metal kernels through the Python and C++ APIs.
+
+Simple Example
+--------------
+
+Let's write a custom kernel that computes ``exp`` elementwise:
+
+.. code-block:: python
+
+  def exp_elementwise(a: mx.array):
+      source = """
+          uint elem = thread_position_in_grid.x;
+          T tmp = inp[elem];
+          out[elem] = metal::exp(tmp);
+      """
+
+      kernel = mx.fast.metal_kernel(
+          name="myexp",
+          input_names=["inp"],
+          output_names=["out"],
+          source=source,
+      )
+      outputs = kernel(
+          inputs=[a],
+          template=[("T", mx.float32)],
+          grid=(a.size, 1, 1),
+          threadgroup=(256, 1, 1),
+          output_shapes=[a.shape],
+          output_dtypes=[a.dtype],
+      )
+      return outputs[0]
+
+  a = mx.random.normal(shape=(4, 16)).astype(mx.float16)
+  b = exp_elementwise(a)
+  assert mx.allclose(b, mx.exp(a))
+
+.. note::
+    We are only required to pass the body of the Metal kernel in ``source``.
+
+The full function signature will be generated using:
+
+* The shapes/dtypes of ``inputs``
+    In the above, ``a`` is an ``mx.array`` of type ``mx.float16`` and we pass it with the key ``inp``
+    so we will add ``const device float16_t* inp`` to the signature.
+    ``inp_shape``, ``inp_strides`` and ``inp_ndim`` are also added for convenience if they are present
+    in ``source``.
+* The list of ``output_dtypes``
+    In the above, ``out`` is an ``mx.array`` of type ``mx.float16``
+    so we add ``device float16_t* out``.
+* Template parameters passed using ``template``
+    In the above, ``template=[("T", mx.float32)]`` adds a template of ``template <typename T>`` to the function
+    and instantiates the template with ``custom_kernel_myexp_float<float>``.
+    Template parameters can be ``mx.core.Dtype``, ``int`` or ``bool``.
+* Metal attributes used in ``source`` such as ``[[thread_position_in_grid]]``
+    These will be added as function arguments.
+    All the attributes defined in Table 5.8 of the `Metal Shading Language Specification <https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf>`_ are supported.
+
+Putting this all together, the generated function signature for ``myexp`` is as follows:
+
+.. code-block:: cpp
+
+  template <typename T>
+  [[kernel]] void custom_kernel_myexp_float(
+    const device float16_t* inp [[buffer(0)]],
+    device float16_t* out [[buffer(1)]],
+    uint3 thread_position_in_grid [[thread_position_in_grid]]) {
+
+          uint elem = thread_position_in_grid.x;
+          T tmp = inp[elem];
+          out[elem] = metal::exp(tmp);
+
+  }
+
+  template [[host_name("custom_kernel_myexp_float")]] [[kernel]] decltype(custom_kernel_myexp_float<float>) custom_kernel_myexp_float<float>;
+
+Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
+
+Using Shape/Strides
+-------------------
+
+``mx.fast.metal_kernel`` supports an argument ``ensure_row_contiguous`` which is ``True`` by default.
+This will copy the ``mx.array`` inputs if needed before the kernel is launched to ensure that the memory layout is row contiguous.
+Generally this makes writing the kernel easier, since we don't have to worry about gaps or the ordering of the dims
+when indexing.
+
+If we want to avoid this copy, ``metal_kernel`` automatically passes ``a_shape``, ``a_strides`` and ``a_ndim`` for each
+input array ``a`` if any are present in ``source``.
+We can then use MLX's built in indexing utils to fetch the right elements for each thread.
+
+Let's convert ``myexp`` above to support arbitrarily strided arrays without relying on a copy from ``ensure_row_contiguous``:
+
+.. code-block:: python
+
+  def exp_elementwise(a: mx.array):
+      source = """
+          uint elem = thread_position_in_grid.x;
+          // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
+          uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
+          T tmp = inp[loc];
+          // Output arrays are always row contiguous
+          out[elem] = metal::exp(tmp);
+      """
+
+      kernel = mx.fast.metal_kernel(
+          name="myexp_strided",
+          input_names=["inp"],
+          output_names=["out"],
+          source=source
+      )
+      outputs = kernel(
+          inputs=[a],
+          template=[("T", mx.float32)],
+          grid=(a.size, 1, 1),
+          threadgroup=(256, 1, 1),
+          output_shapes=[a.shape],
+          output_dtypes=[a.dtype],
+          ensure_row_contiguous=False,
+      )
+      return outputs[0]
+
+  a = mx.random.normal(shape=(4, 16)).astype(mx.float16)
+  # make non-contiguous
+  a = a[::2]
+  b = exp_elementwise(a)
+  assert mx.allclose(b, mx.exp(a))
+
+Complex Example
+-----------------------------
+
+Let's implement a more complex example: ``grid_sample`` in ``"bilinear"`` mode.
+
+We'll start with the following MLX implementation using standard ops:
+
+.. code-block:: python
+
+    def grid_sample_ref(x, grid):
+        N, H_in, W_in, _ = x.shape
+        ix = ((grid[..., 0] + 1) * W_in - 1) / 2
+        iy = ((grid[..., 1] + 1) * H_in - 1) / 2
+
+        ix_nw = mx.floor(ix).astype(mx.int32)
+        iy_nw = mx.floor(iy).astype(mx.int32)
+
+        ix_ne = ix_nw + 1
+        iy_ne = iy_nw
+
+        ix_sw = ix_nw
+        iy_sw = iy_nw + 1
+
+        ix_se = ix_nw + 1
+        iy_se = iy_nw + 1
+
+        nw = (ix_se - ix)    * (iy_se - iy)
+        ne = (ix    - ix_sw) * (iy_sw - iy)
+        sw = (ix_ne - ix)    * (iy    - iy_ne)
+        se = (ix    - ix_nw) * (iy    - iy_nw)
+
+        I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
+        I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
+        I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
+        I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
+
+        mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
+        mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
+        mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
+        mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
+
+        I_nw *= mask_nw[..., None]
+        I_ne *= mask_ne[..., None]
+        I_sw *= mask_sw[..., None]
+        I_se *= mask_se[..., None]
+
+        output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
+
+        return output
+
+Now let's use ``mx.custom_function`` together with ``mx.fast.metal_kernel``
+to write a fast GPU kernel for both the forward and backward passes.
+
+First we'll implement the forward pass as a fused kernel:
+
+.. code-block:: python
+
+    @mx.custom_function
+    def grid_sample(x, grid):
+
+        assert x.ndim == 4, "`x` must be 4D."
+        assert grid.ndim == 4, "`grid` must be 4D."
+
+        B, _, _, C = x.shape
+        _, gN, gM, D = grid.shape
+        out_shape = (B, gN, gM, C)
+
+        assert D == 2, "Last dim of `grid` must be size 2."
+
+        source = """
+            uint elem = thread_position_in_grid.x;
+            int H = x_shape[1];
+            int W = x_shape[2];
+            int C = x_shape[3];
+            int gH = grid_shape[1];
+            int gW = grid_shape[2];
+
+            int w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
+
+            uint grid_idx = elem / C * 2;
+            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+
+            int ix_nw = floor(ix);
+            int iy_nw = floor(iy);
+
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
+
+            int ix_sw = ix_nw;
+            int iy_sw = iy_nw + 1;
+
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
+
+            T nw = (ix_se - ix)    * (iy_se - iy);
+            T ne = (ix    - ix_sw) * (iy_sw - iy);
+            T sw = (ix_ne - ix)    * (iy    - iy_ne);
+            T se = (ix    - ix_nw) * (iy    - iy_nw);
+
+            int batch_idx = elem / C / gH / gW * b_stride;
+            int channel_idx = elem % C;
+            int base_idx = batch_idx + channel_idx;
+
+            T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
+            T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
+            T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
+            T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
+
+            I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
+            I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
+            I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
+            I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
+
+            out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
+        """
+        kernel = mx.fast.metal_kernel(
+            name="grid_sample",
+            input_names=["x", "grid"],
+            output_names=["out"],
+            source=source,
+        )
+        outputs = kernel(
+            inputs=[x, grid],
+            template=[("T", x.dtype)],
+            output_shapes=[out_shape],
+            output_dtypes=[x.dtype],
+            grid=(np.prod(out_shape), 1, 1),
+            threadgroup=(256, 1, 1),
+        )
+        return outputs[0]
+
+For a reasonably sized input such as:
+
+.. code-block:: python
+
+    x.shape = (8, 1024, 1024, 64)
+    grid.shape = (8, 256, 256, 2)
+
+On an M1 Max, we see a big performance improvement:
+
+``55.7ms -> 6.7ms => 8x speed up``
+
+Grid Sample VJP
+---------------
+
+Since we decorated ``grid_sample`` with ``mx.custom_function``, we can now define
+its custom vjp transform so MLX can differentiate it.
+
+The backwards pass requires atomically updating ``x_grad``/``grid_grad`` and so
+requires a few extra ``mx.fast.metal_kernel`` features:
+
+* ``init_value=0``
+    Initialize all of the kernel's outputs to this value before it runs. This allows us to update only part of the output arrays with the kernel.
+
+* ``atomic_outputs=True``
+    Designate all of the kernel outputs as ``atomic`` in the function signature. 
+    This means we can use Metal's ``atomic`` features to simultaneously update the ``x_grad`` and ``grid_grad`` arrays from multiple threadgroups. 
+    See section 6.15 of the `Metal Shading Language Specification <https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf>`_ for more details.
+
+We can then implement the backwards pass as follows:
+
+.. code-block:: python
+
+    @grid_sample.vjp
+    def grid_sample_vjp(primals, cotangent, _):
+        x, grid = primals
+        B, _, _, C = x.shape
+        _, gN, gM, D = grid.shape
+
+        assert D == 2, "Last dim of `grid` must be size 2."
+
+        source = """
+            uint elem = thread_position_in_grid.x;
+            int H = x_shape[1];
+            int W = x_shape[2];
+            int C = x_shape[3];
+            // Pad C to the nearest larger simdgroup size multiple
+            int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
+
+            int gH = grid_shape[1];
+            int gW = grid_shape[2];
+
+            int w_stride = C;
+            int h_stride = W * w_stride;
+            int b_stride = H * h_stride;
+
+            uint grid_idx = elem / C_padded * 2;
+            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+
+            int ix_nw = floor(ix);
+            int iy_nw = floor(iy);
+
+            int ix_ne = ix_nw + 1;
+            int iy_ne = iy_nw;
+
+            int ix_sw = ix_nw;
+            int iy_sw = iy_nw + 1;
+
+            int ix_se = ix_nw + 1;
+            int iy_se = iy_nw + 1;
+
+            T nw = (ix_se - ix)    * (iy_se - iy);
+            T ne = (ix    - ix_sw) * (iy_sw - iy);
+            T sw = (ix_ne - ix)    * (iy    - iy_ne);
+            T se = (ix    - ix_nw) * (iy    - iy_nw);
+
+            int batch_idx = elem / C_padded / gH / gW * b_stride;
+            int channel_idx = elem % C_padded;
+            int base_idx = batch_idx + channel_idx;
+
+            T gix = T(0);
+            T giy = T(0);
+            if (channel_idx < C) {
+                int cot_index = elem / C_padded * C + channel_idx;
+                T cot = cotangent[cot_index];
+                if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
+                    int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
+
+                    T I_nw = x[offset];
+                    gix -= I_nw * (iy_se - iy) * cot;
+                    giy -= I_nw * (ix_se - ix) * cot;
+                }
+                if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
+                    int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
+
+                    T I_ne = x[offset];
+                    gix += I_ne * (iy_sw - iy) * cot;
+                    giy -= I_ne * (ix - ix_sw) * cot;
+                }
+                if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
+                    int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
+
+                    T I_sw = x[offset];
+                    gix -= I_sw * (iy - iy_ne) * cot;
+                    giy += I_sw * (ix_ne - ix) * cot;
+                }
+                if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
+                    int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
+                    atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
+
+                    T I_se = x[offset];
+                    gix += I_se * (iy - iy_nw) * cot;
+                    giy += I_se * (ix - ix_nw) * cot;
+                }
+            }
+
+            T gix_mult = W / 2;
+            T giy_mult = H / 2;
+
+            // Reduce across each simdgroup first.
+            // This is much faster than relying purely on atomics.
+            gix = simd_sum(gix);
+            giy = simd_sum(giy);
+
+            if (thread_index_in_simdgroup == 0) {
+                atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
+                atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
+            }
+        """
+        kernel = mx.fast.metal_kernel(
+            name="grid_sample_grad",
+            input_names=["x", "grid", "cotangent"],
+            output_names=["x_grad", "grid_grad"],
+            source=source,
+            atomic_outputs=True,
+        )
+        # pad the output channels to simd group size
+        # so that our `simd_sum`s don't overlap.
+        simdgroup_size = 32
+        C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
+        grid_size = B * gN * gM * C_padded
+        outputs = kernel(
+            inputs=[x, grid, cotangent],
+            template=[("T", x.dtype)],
+            output_shapes=[x.shape, grid.shape],
+            output_dtypes=[x.dtype, x.dtype],
+            grid=(grid_size, 1, 1),
+            threadgroup=(256, 1, 1),
+            init_value=0,
+        )
+        return outputs[0], outputs[1]
+
+There's an even larger speed up for the vjp:
+
+``676.4ms -> 16.7ms => 40x speed up``

docs/src/dev/extensions.rst

@@ -1,24 +1,16 @@
-Developer Documentation
-=======================
+Custom Extensions in MLX
+========================
 
-MLX provides a open and flexible backend to which users may add operations 
-and specialized implementations without much hassle. While the library supplies
-efficient operations that can be used and composed for any number of 
-applications, there may arise cases where new functionalities or highly 
-optimized implementations are needed. For such cases, you may design and 
-implement your own operations that link to and build on top of :mod:`mlx.core`.
-We will introduce the inner-workings of MLX and go over a simple example to 
-learn the steps involved in adding new operations to MLX with your own CPU 
-and GPU implementations. 
+You can extend MLX with custom operations on the CPU or GPU. This guide
+explains how to do that with a simple example.
 
 Introducing the Example
 -----------------------
 
-Let's say that you would like an operation that takes in two arrays, 
-``x`` and ``y``, scales them both by some coefficients ``alpha`` and ``beta``
-respectively, and then adds them together to get the result 
-``z = alpha * x + beta * y``. Well, you can very easily do that by just 
-writing out a function as follows:
+Let's say you would like an operation that takes in two arrays, ``x`` and
+``y``, scales them both by coefficients ``alpha`` and ``beta`` respectively,
+and then adds them together to get the result ``z = alpha * x + beta * y``.
+You can do that in MLX directly:
 
 .. code-block:: python
 
@@ -27,44 +19,35 @@ writing out a function as follows:
     def simple_axpby(x: mx.array, y: mx.array, alpha: float, beta: float) -> mx.array:
         return alpha * x + beta * y
 
-This function performs that operation while leaving the implementations and 
-differentiation to MLX. 
+This function performs that operation while leaving the implementation and
+function transformations to MLX.
 
-However, you work with vector math libraries often and realize that the 
-``axpby`` routine defines the same operation ``Y = (alpha * X) + (beta * Y)``. 
-You would really like the part of your applications that does this operation 
-on the CPU to be very fast - so you decide that you want it to rely on the 
-``axpby`` routine provided by the Accelerate_ framework. Continuing to impose 
-our assumptions on to you, let's also assume that you want to learn how to add 
-your own implementation for the gradients of your new operation while going 
-over the ins-and-outs of the MLX framework. 
+However you may need to customize the underlying implementation, perhaps to
+make it faster or for custom differentiation. In this tutorial we will go
+through adding custom extensions. It will cover:
 
-Well, what a coincidence! You are in the right place. Over the course of this 
-example, we will learn:
-
-* The structure of the MLX library from the frontend API to the backend implementations.
-* How to implement your own CPU backend that redirects to Accelerate_ when appropriate (and a fallback if needed).
-* How to implement your own GPU implementation using metal.
-* How to add your own ``vjp`` and ``jvp``.
-* How to build your implementations, link them to MLX, and bind them to python.
+* The structure of the MLX library.
+* Implementing a CPU operation that redirects to Accelerate_ when appropriate.
+* Implementing a GPU operation using metal.
+* Adding the ``vjp`` and ``jvp`` function transformation.
+* Building a custom extension and binding it to python.
 
 Operations and Primitives
 -------------------------
 
-In one sentence, operations in MLX build the computation graph, and primitives 
-provide the rules for evaluation and transformations of said graph. Let's start 
-by discussing operations in more detail. 
+Operations in MLX build the computation graph. Primitives provide the rules for
+evaluating and transforming the graph. Let's start by discussing operations in
+more detail.
 
 Operations
 ^^^^^^^^^^^
 
-Operations are the frontend functions that operate on arrays. They are defined 
-in the C++ API (:ref:`cpp_ops`) and then we provide bindings to these 
-operations in the Python API (:ref:`ops`). 
+Operations are the front-end functions that operate on arrays. They are defined
+in the C++ API (:ref:`cpp_ops`), and the Python API (:ref:`ops`) binds them.
 
-We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and ``y``,
-and two scalars, ``alpha`` and ``beta``. This is how we would define it in the 
-C++ API:
+We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and
+``y``, and two scalars, ``alpha`` and ``beta``. This is how to define it in
+C++:
 
 .. code-block:: C++
 
@@ -83,10 +66,7 @@ C++ API:
         StreamOrDevice s = {} // Stream on which to schedule the operation
     );
 
-
-This operation itself can call other operations within it if needed. So, the 
-simplest way to go about implementing this operation would be do so in terms 
-of existing operations. 
+The simplest way to this operation is in terms of existing operations:
 
 .. code-block:: C++
 
@@ -100,25 +80,23 @@ of existing operations.
         // Scale x and y on the provided stream
         auto ax = multiply(array(alpha), x, s);
         auto by = multiply(array(beta), y, s);
-        
+
         // Add and return
         return add(ax, by, s);
     }
 
-However, as we discussed earlier, this is not our goal. The operations themselves 
-do not contain the implementations that act on the data, nor do they contain the
-rules of transformations. Rather, they are an easy to use interface that build 
-on top of the building blocks we call :class:`Primitive`. 
+The operations themselves do not contain the implementations that act on the
+data, nor do they contain the rules of transformations. Rather, they are an
+easy to use interface that use :class:`Primitive` building blocks.
 
 Primitives
 ^^^^^^^^^^^
 
-A :class:`Primitive` is part of the computation graph of an :class:`array`. It 
-defines how to create an output given a set of input :class:`array` . Further,
-a :class:`Primitive` is a class that contains rules on how it is evaluated 
-on the CPU or GPU, and how it acts under transformations such as ``vjp`` and 
-``jvp``. These words on their own can be a bit abstract, so lets take a step 
-back and go to our example to give ourselves a more concrete image. 
+A :class:`Primitive` is part of the computation graph of an :class:`array`. It
+defines how to create outputs arrays given a input arrays. Further, a
+:class:`Primitive` has methods to run on the CPU or GPU and for function
+transformations such as ``vjp`` and ``jvp``.  Lets go back to our example to be
+more concrete:
 
 .. code-block:: C++
 
@@ -134,11 +112,15 @@ back and go to our example to give ourselves a more concrete image.
         * To avoid unnecessary allocations, the evaluation function
         * is responsible for allocating space for the array.
         */
-        void eval_cpu(const std::vector<array>& inputs, array& out) override;
-        void eval_gpu(const std::vector<array>& inputs, array& out) override;
+        void eval_cpu(
+            const std::vector<array>& inputs,
+            std::vector<array>& outputs) override;
+        void eval_gpu(
+            const std::vector<array>& inputs,
+            std::vector<array>& outputs) override;
 
         /** The Jacobian-vector product. */
-        array jvp(
+        std::vector<array> jvp(
             const std::vector<array>& primals,
             const std::vector<array>& tangents,
             const std::vector<int>& argnums) override;
@@ -147,7 +129,8 @@ back and go to our example to give ourselves a more concrete image.
         std::vector<array> vjp(
             const std::vector<array>& primals,
             const array& cotan,
-            const std::vector<int>& argnums) override;
+            const std::vector<int>& argnums,
+            const std::vector<array>& outputs) override;
 
         /**
         * The primitive must know how to vectorize itself across
@@ -155,7 +138,7 @@ back and go to our example to give ourselves a more concrete image.
         * representing the vectorized computation and the axis which
         * corresponds to the output vectorized dimension.
         */
-        std::pair<array, int> vmap(
+        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) override;
 
@@ -175,22 +158,22 @@ back and go to our example to give ourselves a more concrete image.
         void eval(const std::vector<array>& inputs, array& out);
     };
 
-The :class:`Axpby` class derives from the base :class:`Primitive` class and 
-follows the above demonstrated interface. :class:`Axpby` treats ``alpha`` and 
-``beta`` as parameters. It then provides implementations of how the array ``out`` 
-is produced given ``inputs`` through :meth:`Axpby::eval_cpu` and 
-:meth:`Axpby::eval_gpu`. Further, it provides rules of transformations in 
-:meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and :meth:`Axpby::vmap`. 
+The :class:`Axpby` class derives from the base :class:`Primitive` class. The
+:class:`Axpby` treats ``alpha`` and ``beta`` as parameters. It then provides
+implementations of how the output array is produced given the inputs through
+:meth:`Axpby::eval_cpu` and :meth:`Axpby::eval_gpu`. It also provides rules
+of transformations in :meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and
+:meth:`Axpby::vmap`.
 
-Using the Primitives
-^^^^^^^^^^^^^^^^^^^^^
+Using the Primitive
+^^^^^^^^^^^^^^^^^^^
 
-Operations can use this :class:`Primitive` to add a new :class:`array` to 
-the computation graph. An :class:`array` can be constructed by providing its 
-data type, shape, the :class:`Primitive` that computes it, and the 
-:class:`array` inputs that are passed to the primitive.
+Operations can use this :class:`Primitive` to add a new :class:`array` to the
+computation graph. An :class:`array` can be constructed by providing its data
+type, shape, the :class:`Primitive` that computes it, and the :class:`array`
+inputs that are passed to the primitive.
 
-Let's re-implement our operation now in terms of our :class:`Axpby` primitive.
+Let's reimplement our operation now in terms of our :class:`Axpby` primitive.
 
 .. code-block:: C++
 
@@ -238,27 +221,26 @@ This operation now handles the following:
 Implementing the Primitive
 --------------------------
 
-No computation happens when we call the operation alone. In effect, the 
-operation only builds the computation graph. When we evaluate the output 
-array, MLX schedules the execution of the computation graph, and calls
-:meth:`Axpby::eval_cpu` or :meth:`Axpby::eval_gpu` depending on the 
-stream/device specified by the user. 
+No computation happens when we call the operation alone. The operation only
+builds the computation graph. When we evaluate the output array, MLX schedules
+the execution of the computation graph, and calls :meth:`Axpby::eval_cpu` or
+:meth:`Axpby::eval_gpu` depending on the stream/device specified by the user.
 
 .. warning::
     When :meth:`Primitive::eval_cpu` or :meth:`Primitive::eval_gpu` are called,
     no memory has been allocated for the output array. It falls on the implementation
-    of these functions to allocate memory as needed
+    of these functions to allocate memory as needed.
 
-Implementing the CPU Backend
+Implementing the CPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Let's start by trying to implement a naive and generic version of 
-:meth:`Axpby::eval_cpu`. We declared this as a private member function of 
-:class:`Axpby` earlier called :meth:`Axpby::eval`. 
+Let's start by implementing a naive and generic version of
+:meth:`Axpby::eval_cpu`. We declared this as a private member function of
+:class:`Axpby` earlier called :meth:`Axpby::eval`.
 
-Our naive method will go over each element of the output array, find the 
-corresponding input elements of ``x`` and ``y`` and perform the operation 
-pointwise. This is captured in the templated function :meth:`axpby_impl`. 
+Our naive method will go over each element of the output array, find the
+corresponding input elements of ``x`` and ``y`` and perform the operation
+point-wise. This is captured in the templated function :meth:`axpby_impl`.
 
 .. code-block:: C++
 
@@ -296,19 +278,19 @@ pointwise. This is captured in the templated function :meth:`axpby_impl`.
         }
     }
 
-Now, we would like our implementation to be able to do this pointwise operation 
-for all incoming floating point arrays. Accordingly, we add dispatches for 
-``float32``, ``float16``, ``bfloat16`` and ``complex64``. We throw an error 
-if we encounter an unexpected type.
+Our implementation should work for all incoming floating point arrays.
+Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
+``complex64``. We throw an error if we encounter an unexpected type.
 
 .. code-block:: C++
 
     /** Fall back implementation for evaluation on CPU */
-    void Axpby::eval(const std::vector<array>& inputs, array& out) {
-        // Check the inputs (registered in the op while constructing the out array)
-        assert(inputs.size() == 2);
+    void Axpby::eval(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Dispatch to the correct dtype
         if (out.dtype() == float32) {
@@ -321,28 +303,26 @@ if we encounter an unexpected type.
             return axpby_impl<complex64_t>(x, y, out, alpha_, beta_);
         } else {
             throw std::runtime_error(
-                "Axpby is only supported for floating point types.");
+                "[Axpby] Only supports floating point types.");
         }
     }
 
-We have a fallback implementation! Now, to do what we are really here to do. 
-Remember we wanted to use the ``axpby`` routine provided by the Accelerate_
-framework? Well, there are 3 complications to keep in mind:
+This is good as a fallback implementation. We can use the ``axpby`` routine
+provided by the Accelerate_ framework for a faster implementation in certain
+cases:
 
 #.  Accelerate does not provide implementations of ``axpby`` for half precision
-    floats. We can only direct to it for ``float32`` types 
-#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all elements
-    have fixed strides between them. Possibly due to broadcasts and transposes, 
-    we aren't guaranteed that the inputs fit this requirement. We can 
-    only direct to Accelerate if both ``x`` and ``y`` are row contiguous or 
-    column contiguous. 
-#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` inplace. 
-    MLX expects to write out the answer to a new array. We must copy the elements 
-    of ``y`` into the output array and use that as an input to ``axpby``
+    floats. We can only use it for ``float32`` types.
+#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all
+    elements have fixed strides between them. We only direct to Accelerate
+    if both ``x`` and ``y`` are row contiguous or column contiguous.
+#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` in-place.
+    MLX expects to write the output to a new array. We must copy the elements
+    of ``y`` into the output and use that as an input to ``axpby``.
 
-Let's write out an implementation that uses Accelerate in the right conditions. 
-It must simply allocate data for the output, copy elements of ``y`` into it, 
-and then call the :meth:`catlas_saxpby` from accelerate. 
+Let's write an implementation that uses Accelerate in the right conditions.
+It allocates data for the output, copies ``y`` into it, and then calls the
+:func:`catlas_saxpby` from accelerate.
 
 .. code-block:: C++
 
@@ -356,17 +336,7 @@ and then call the :meth:`catlas_saxpby` from accelerate.
         // Accelerate library provides catlas_saxpby which does
         // Y = (alpha * X) + (beta * Y) in place
         // To use it, we first copy the data in y over to the output array
-
-        // This specialization requires both x and y be contiguous in the same mode
-        // i.e: corresponding linear indices in both point to corresponding elements
-        // The data in the output array is allocated to match the strides in y
-        // such that x, y, and out are contiguous in the same mode and
-        // no transposition is needed
-        out.set_data(
-            allocator::malloc_or_wait(y.data_size() * out.itemsize()),
-            y.data_size(),
-            y.strides(),
-            y.flags());
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
         // We then copy over the elements using the contiguous vector specialization
         copy_inplace(y, out, CopyType::Vector);
@@ -389,18 +359,20 @@ and then call the :meth:`catlas_saxpby` from accelerate.
             /* INCY = */ 1);
     }
 
-Great! But what about the inputs that do not fit the criteria for accelerate?
-Luckily, we can always just direct back to :meth:`Axpby::eval`.
-
-With this in mind, lets finally implement our :meth:`Axpby::eval_cpu`.
+For inputs that do not fit the criteria for accelerate, we fall back to
+:meth:`Axpby::eval`. With this in mind, let's finish our
+:meth:`Axpby::eval_cpu`.
 
 .. code-block:: C++
 
     /** Evaluate primitive on CPU using accelerate specializations */
-    void Axpby::eval_cpu(const std::vector<array>& inputs, array& out) {
+    void Axpby::eval_cpu(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
         assert(inputs.size() == 2);
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Accelerate specialization for contiguous single precision float arrays
         if (out.dtype() == float32 &&
@@ -410,35 +382,33 @@ With this in mind, lets finally implement our :meth:`Axpby::eval_cpu`.
             return;
         }
 
-        // Fall back to common backend if specializations are not available
-        eval(inputs, out);
+        // Fall back to common back-end if specializations are not available
+        eval(inputs, outputs);
     }
 
-We have now hit a milestone! Just this much is enough to run the operation 
-:meth:`axpby` on a CPU stream! 
+Just this much is enough to run the operation :meth:`axpby` on a CPU stream! If
+you do not plan on running the operation on the GPU or using transforms on
+computation graphs that contain :class:`Axpby`, you can stop implementing the
+primitive here and enjoy the speed-ups you get from the Accelerate library.
 
-If you do not plan on running the operation on the GPU or using transforms on 
-computation graphs that contain :class:`Axpby`, you can stop implementing the 
-primitive here and enjoy the speed-ups you get from the Accelerate library. 
-
-Implementing the GPU Backend
+Implementing the GPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Apple silicon devices address their GPUs using the Metal_ shading language, and 
-all GPU kernels in MLX are written using metal. 
+Apple silicon devices address their GPUs using the Metal_ shading language, and
+GPU kernels in MLX are written using Metal.
 
 .. note::
 
-    Here are some helpful resources if you are new to metal!
+    Here are some helpful resources if you are new to Metal:
 
     * A walkthrough of the metal compute pipeline: `Metal Example`_
     * Documentation for metal shading language: `Metal Specification`_
     * Using metal from C++: `Metal-cpp`_
 
-Let's keep the GPU algorithm simple. We will launch exactly as many threads 
-as there are elements in the output. Each thread will pick the element it needs 
-from ``x`` and ``y``, do the pointwise operation, and then update its assigned 
-element in the output. 
+Let's keep the GPU kernel simple. We will launch exactly as many threads as
+there are elements in the output. Each thread will pick the element it needs
+from ``x`` and ``y``, do the point-wise operation, and update its assigned
+element in the output.
 
 .. code-block:: C++
 
@@ -457,15 +427,14 @@ element in the output.
         // Convert linear indices to offsets in array
         auto x_offset = elem_to_loc(index, shape, x_strides, ndim);
         auto y_offset = elem_to_loc(index, shape, y_strides, ndim);
-        
+
         // Do the operation and update the output
-        out[index] = 
+        out[index] =
             static_cast<T>(alpha) * x[x_offset] + static_cast<T>(beta) * y[y_offset];
     }
 
 We then need to instantiate this template for all floating point types and give
-each instantiation a unique host name so we can identify the right kernel for 
-each data type. 
+each instantiation a unique host name so we can identify it.
 
 .. code-block:: C++
 
@@ -488,29 +457,21 @@ each data type.
     instantiate_axpby(bfloat16, bfloat16_t);
     instantiate_axpby(complex64, complex64_t);
 
-This kernel will be compiled into a metal library ``mlx_ext.metallib`` as we 
-will see later in :ref:`Building with CMake`. In the following example, we 
-assume that the library ``mlx_ext.metallib`` will always be co-located with 
-the executable/ shared-library calling the :meth:`register_library` function. 
-The :meth:`register_library` function takes the library's name and potential 
-path (or in this case, a function that can produce the path of the metal 
-library) and tries to load that library if it hasn't already been registered 
-by the relevant static :class:`mlx::core::metal::Device` object. This is why, 
-it is important to package your C++ library with the metal library. We will 
-go over this process in more detail later. 
-
-The logic to determine the kernel, set the inputs, resolve the grid dimensions 
-and dispatch it to the GPU are contained in :meth:`Axpby::eval_gpu` as shown 
+The logic to determine the kernel, set the inputs, resolve the grid dimensions,
+and dispatch to the GPU are contained in :meth:`Axpby::eval_gpu` as shown
 below.
 
 .. code-block:: C++
 
     /** Evaluate primitive on GPU */
-    void Axpby::eval_gpu(const std::vector<array>& inputs, array& out) {
+    void Axpby::eval_gpu(
+      const std::vector<array>& inputs,
+      std::vector<array>& outputs) {
         // Prepare inputs
         assert(inputs.size() == 2);
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Each primitive carries the stream it should execute on
         // and each stream carries its device identifiers
@@ -518,22 +479,21 @@ below.
         // We get the needed metal device using the stream
         auto& d = metal::device(s.device);
 
-        // Allocate output memory 
+        // Allocate output memory
         out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
-        // Resolve name of kernel (corresponds to axpby.metal)
+        // Resolve name of kernel
         std::ostringstream kname;
         kname << "axpby_" << "general_" << type_to_name(out);
 
-        // Make sure the metal library is available and look for it
-        // in the same folder as this executable if needed
-        d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
+        // Make sure the metal library is available
+        d.register_library("mlx_ext");
 
         // Make a kernel from this metal library
         auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
         // Prepare to encode kernel
-        auto compute_encoder = d.get_command_encoder(s.index);
+        auto& compute_encoder = d.get_command_encoder(s.index);
         compute_encoder->setComputePipelineState(kernel);
 
         // Kernel parameters are registered with buffer indices corresponding to
@@ -542,17 +502,17 @@ below.
         size_t nelem = out.size();
 
         // Encode input arrays to kernel
-        set_array_buffer(compute_encoder, x, 0);
-        set_array_buffer(compute_encoder, y, 1);
+        compute_encoder.set_input_array(x, 0);
+        compute_encoder.set_input_array(y, 1);
 
         // Encode output arrays to kernel
-        set_array_buffer(compute_encoder, out, 2);
+        compute_encoder.set_output_array(out, 2);
 
         // Encode alpha and beta
         compute_encoder->setBytes(&alpha_, sizeof(float), 3);
         compute_encoder->setBytes(&beta_, sizeof(float), 4);
 
-        // Encode shape, strides and ndim 
+        // Encode shape, strides and ndim
         compute_encoder->setBytes(x.shape().data(), ndim * sizeof(int), 5);
         compute_encoder->setBytes(x.strides().data(), ndim * sizeof(size_t), 6);
         compute_encoder->setBytes(y.strides().data(), ndim * sizeof(size_t), 7);
@@ -570,33 +530,30 @@ below.
 
         // Launch the grid with the given number of threads divided among
         // the given threadgroups
-        compute_encoder->dispatchThreads(grid_dims, group_dims);
+        compute_encoder.dispatchThreads(grid_dims, group_dims);
     }
 
 We can now call the :meth:`axpby` operation on both the CPU and the GPU!
 
-A few things to note about MLX and metal before moving on. MLX keeps track 
-of the active ``compute_encoder``. We rely on :meth:`d.get_command_encoder` 
-to give us the active metal compute command encoder instead of building a 
-new one and calling :meth:`compute_encoder->end_encoding` at the end. 
-MLX keeps adding kernels (compute pipelines) to the active command encoder 
-until some specified limit is hit or the compute encoder needs to be flushed 
-for synchronization. MLX also handles enqueuing and committing the associated 
-command buffers as needed. We suggest taking a deeper dive into 
-:class:`metal::Device` if you would like to study this routine further.
+A few things to note about MLX and Metal before moving on. MLX keeps track of
+the active ``command_buffer`` and the ``MTLCommandBuffer`` to which it is
+associated. We rely on :meth:`d.get_command_encoder` to give us the active
+metal compute command encoder instead of building a new one and calling
+:meth:`compute_encoder->end_encoding` at the end. MLX adds kernels (compute
+pipelines) to the active command buffer until some specified limit is hit or
+the command buffer needs to be flushed for synchronization.
 
 Primitive Transforms
 ^^^^^^^^^^^^^^^^^^^^^
 
-Now that we have come this far, let's also learn how to add implementations to 
-transformations in a :class:`Primitive`. These transformations can be built on 
-top of our operations, including the one we just defined now. Which then gives 
-us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
+Next, let's add implementations for transformations in a :class:`Primitive`.
+These transformations can be built on top of other operations, including the
+one we just defined:
 
 .. code-block:: C++
 
     /** The Jacobian-vector product. */
-    array Axpby::jvp(
+    std::vector<array> Axpby::jvp(
             const std::vector<array>& primals,
             const std::vector<array>& tangents,
             const std::vector<int>& argnums) {
@@ -611,12 +568,12 @@ us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
         if (argnums.size() > 1) {
             auto scale = argnums[0] == 0 ? alpha_ : beta_;
             auto scale_arr = array(scale, tangents[0].dtype());
-            return multiply(scale_arr, tangents[0], stream());
+            return {multiply(scale_arr, tangents[0], stream())};
         }
         // If, argnums = {0, 1}, we take contributions from both
         // which gives us jvp = tangent_x * alpha + tangent_y * beta
         else {
-            return axpby(tangents[0], tangents[1], alpha_, beta_, stream());
+            return {axpby(tangents[0], tangents[1], alpha_, beta_, stream())};
         }
     }
 
@@ -625,34 +582,35 @@ us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
     /** The vector-Jacobian product. */
     std::vector<array> Axpby::vjp(
             const std::vector<array>& primals,
-            const array& cotan,
-            const std::vector<int>& argnums) {
+            const std::vector<array>& cotangents,
+            const std::vector<int>& argnums,
+            const std::vector<int>& /* unused */) {
         // Reverse mode diff
         std::vector<array> vjps;
         for (auto arg : argnums) {
             auto scale = arg == 0 ? alpha_ : beta_;
-            auto scale_arr = array(scale, cotan.dtype());
-            vjps.push_back(multiply(scale_arr, cotan, stream()));
+            auto scale_arr = array(scale, cotangents[0].dtype());
+            vjps.push_back(multiply(scale_arr, cotangents[0], stream()));
         }
         return vjps;
     }
 
-Finally, you need not have a transformation fully defined to start using your 
-own :class:`Primitive`.
+Note, a transformation does not need to be fully defined to start using
+the :class:`Primitive`.
 
 .. code-block:: C++
 
     /** Vectorize primitive along given axis */
-    std::pair<array, int> Axpby::vmap(
+    std::pair<std::vector<array>, std::vector<int>> Axpby::vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) {
-        throw std::runtime_error("Axpby has no vmap implementation.");
+        throw std::runtime_error("[Axpby] vmap not implemented.");
     }
 
 Building and Binding
 --------------------
 
-Let's look at the overall directory structure first. 
+Let's look at the overall directory structure first.
 
 | extensions
 | ├── axpby
@@ -666,40 +624,39 @@ Let's look at the overall directory structure first.
 | └── setup.py
 
 * ``extensions/axpby/`` defines the C++ extension library
-* ``extensions/mlx_sample_extensions`` sets out the structure for the 
-  associated python package
-* ``extensions/bindings.cpp`` provides python bindings for our operation
-* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and 
-  python bindings
+* ``extensions/mlx_sample_extensions`` sets out the structure for the
+  associated Python package
+* ``extensions/bindings.cpp`` provides Python bindings for our operation
+* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and
+  Python bindings
 * ``extensions/setup.py`` holds the ``setuptools`` rules to build and install
-  the python package
+  the Python package
 
 Binding to Python
 ^^^^^^^^^^^^^^^^^^
 
-We use PyBind11_ to build a Python API for the C++ library. Since bindings for
+We use nanobind_ to build a Python API for the C++ library. Since bindings for
 components such as :class:`mlx.core.array`, :class:`mlx.core.stream`, etc. are
-already provided, adding our :meth:`axpby` is simple!
+already provided, adding our :meth:`axpby` is simple.
 
 .. code-block:: C++
 
-    PYBIND11_MODULE(mlx_sample_extensions, m) {
-        m.doc() = "Sample C++ and metal extensions for MLX";
+   NB_MODULE(_ext, m) {
+        m.doc() = "Sample extension for MLX";
 
         m.def(
             "axpby",
             &axpby,
             "x"_a,
             "y"_a,
-            py::pos_only(),
             "alpha"_a,
             "beta"_a,
-            py::kw_only(),
-            "stream"_a = py::none(),
-            R"pbdoc(
+            nb::kw_only(),
+            "stream"_a = nb::none(),
+            R"(
                 Scale and sum two vectors element-wise
                 ``z = alpha * x + beta * y``
-                
+
                 Follows numpy style broadcasting between ``x`` and ``y``
                 Inputs are upcasted to floats if needed
 
@@ -711,17 +668,17 @@ already provided, adding our :meth:`axpby` is simple!
 
                 Returns:
                     array: ``alpha * x + beta * y``
-            )pbdoc");
+            )");
     }
 
-Most of the complexity in the above example comes from additional bells and 
+Most of the complexity in the above example comes from additional bells and
 whistles such as the literal names and doc-strings.
 
 .. warning::
 
-    :mod:`mlx.core` needs to be imported before importing 
-    :mod:`mlx_sample_extensions` as defined by the pybind11 module above to 
-    ensure that the casters for :mod:`mlx.core` components like 
+    :mod:`mlx.core` must be imported before importing
+    :mod:`mlx_sample_extensions` as defined by the nanobind module above to
+    ensure that the casters for :mod:`mlx.core` components like
     :class:`mlx.core.array` are available.
 
 .. _Building with CMake:
@@ -729,8 +686,8 @@ whistles such as the literal names and doc-strings.
 Building with CMake
 ^^^^^^^^^^^^^^^^^^^^
 
-Building the C++ extension library itself is simple, it only requires that you 
-``find_package(MLX CONFIG)`` and then link it to your library. 
+Building the C++ extension library only requires that you ``find_package(MLX
+CONFIG)`` and then link it to your library.
 
 .. code-block:: cmake
 
@@ -752,12 +709,12 @@ Building the C++ extension library itself is simple, it only requires that you
     # Link to mlx
     target_link_libraries(mlx_ext PUBLIC mlx)
 
-We also need to build the attached metal library. For convenience, we provide a 
-:meth:`mlx_build_metallib` function that builds a ``.metallib`` target given 
-sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and 
-automatically imported with MLX package). 
+We also need to build the attached Metal library. For convenience, we provide a
+:meth:`mlx_build_metallib` function that builds a ``.metallib`` target given
+sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and
+automatically imported with MLX package).
 
-Here is what that looks like in practice!
+Here is what that looks like in practice:
 
 .. code-block:: cmake
 
@@ -779,27 +736,29 @@ Here is what that looks like in practice!
 
     endif()
 
-Finally, we build the Pybind11_ bindings
+Finally, we build the nanobind_ bindings
 
 .. code-block:: cmake
 
-    pybind11_add_module(
-        mlx_sample_extensions
-        ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
+    nanobind_add_module(
+      _ext
+      NB_STATIC STABLE_ABI LTO NOMINSIZE
+      NB_DOMAIN mlx
+      ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
     )
-    target_link_libraries(mlx_sample_extensions PRIVATE mlx_ext)
+    target_link_libraries(_ext PRIVATE mlx_ext)
 
     if(BUILD_SHARED_LIBS)
-        target_link_options(mlx_sample_extensions PRIVATE -Wl,-rpath,@loader_path)
+      target_link_options(_ext PRIVATE -Wl,-rpath,@loader_path)
     endif()
 
 Building with ``setuptools``
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Once we have set out the CMake build rules as described above, we can use the
-build utilities defined in :mod:`mlx.extension` for a simple build process. 
+build utilities defined in :mod:`mlx.extension`:
 
-.. code-block:: python 
+.. code-block:: python
 
     from mlx import extension
     from setuptools import setup
@@ -809,48 +768,50 @@ build utilities defined in :mod:`mlx.extension` for a simple build process.
             name="mlx_sample_extensions",
             version="0.0.0",
             description="Sample C++ and Metal extensions for MLX primitives.",
-            ext_modules=[extension.CMakeExtension("mlx_sample_extensions")],
+            ext_modules=[extension.CMakeExtension("mlx_sample_extensions._ext")],
             cmdclass={"build_ext": extension.CMakeBuild},
-            packages = ["mlx_sample_extensions"],
-            package_dir = {"": "mlx_sample_extensions"},
-            package_data = {"mlx_sample_extensions" : ["*.so", "*.dylib", "*.metallib"]},
+            packages=["mlx_sample_extensions"],
+            package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
+            extras_require={"dev":[]},
             zip_safe=False,
-            python_requires=">=3.7",
+            python_requires=">=3.8",
         )
 
 .. note::
     We treat ``extensions/mlx_sample_extensions`` as the package directory
     even though it only contains a ``__init__.py`` to ensure the following:
-    
-    * :mod:`mlx.core` is always imported before importing  :mod:`mlx_sample_extensions`
-    * The C++ extension library and the metal library are co-located with the python 
-      bindings and copied together if the package is installed 
 
-You can build inplace for development using
+    * :mod:`mlx.core` must be imported before importing :mod:`_ext`
+    * The C++ extension library and the metal library are co-located with the python
+      bindings and copied together if the package is installed
+
+To build the package, first install the build dependencies with ``pip install
+-r requirements.txt``.  You can then build inplace for development using
 ``python setup.py build_ext -j8 --inplace`` (in ``extensions/``)
 
-This will result in a directory structure as follows:
+This results in the directory structure:
 
 | extensions
 | ├── mlx_sample_extensions
 | │   ├── __init__.py
 | │   ├── libmlx_ext.dylib # C++ extension library
 | │   ├── mlx_ext.metallib # Metal library
-| │   └── mlx_sample_extensions.cpython-3x-darwin.so # Python Binding
+| │   └── _ext.cpython-3x-darwin.so # Python Binding
 | ...
 
-When you try to install using the command ``python -m pip install .`` 
-(in ``extensions/``), the package will be installed with the same structure as 
-``extensions/mlx_sample_extensions`` and the C++ and metal library will be 
-copied along with the python binding since they are specified as ``package_data``.
+When you try to install using the command ``python -m pip install .`` (in
+``extensions/``), the package will be installed with the same structure as
+``extensions/mlx_sample_extensions`` and the C++ and Metal library will be
+copied along with the Python binding since they are specified as
+``package_data``.
 
 Usage
 -----
 
-After installing the extension as described above, you should be able to simply 
-import the python package and play with it as you would any other MLX operation!
+After installing the extension as described above, you should be able to simply
+import the Python package and play with it as you would any other MLX operation.
 
-Let's looks at a simple script and it's results!
+Let's look at a simple script and its results:
 
 .. code-block:: python
 
@@ -863,7 +824,7 @@ Let's looks at a simple script and it's results!
 
     print(f"c shape: {c.shape}")
     print(f"c dtype: {c.dtype}")
-    print(f"c correctness: {mx.all(c == 6.0).item()}")
+    print(f"c correct: {mx.all(c == 6.0).item()}")
 
 Output:
 
@@ -874,12 +835,12 @@ Output:
     c correctness: True
 
 Results
-^^^^^^^^^^^^^^^^
+^^^^^^^
 
-Let's run a quick benchmark and see how our new ``axpby`` operation compares 
-with the naive :meth:`simple_axpby` we defined at first on the CPU. 
+Let's run a quick benchmark and see how our new ``axpby`` operation compares
+with the naive :meth:`simple_axpby` we first defined on the CPU.
 
-.. code-block:: python 
+.. code-block:: python
 
     import mlx.core as mx
     from mlx_sample_extensions import axpby
@@ -898,7 +859,7 @@ with the naive :meth:`simple_axpby` we defined at first on the CPU.
     alpha = 4.0
     beta = 2.0
 
-    mx.eval((x, y))
+    mx.eval(x, y)
 
     def bench(f):
         # Warm up
@@ -919,30 +880,23 @@ with the naive :meth:`simple_axpby` we defined at first on the CPU.
 
     print(f"Simple axpby: {simple_time:.3f} s | Custom axpby: {custom_time:.3f} s")
 
-Results:
-
-.. code-block::
-
-    Simple axpby: 0.114 s | Custom axpby: 0.109 s
-
-We see some modest improvements right away! 
+The results are ``Simple axpby: 0.114 s | Custom axpby: 0.109 s``. We see
+modest improvements right away!
 
 This operation is now good to be used to build other operations, in
 :class:`mlx.nn.Module` calls, and also as a part of graph transformations like
-:meth:`grad`!
+:meth:`grad`.
 
 Scripts
 -------
 
 .. admonition:: Download the code
 
-   The full example code is available in `mlx <code>`_.
-
-.. code: `https://github.com/ml-explore/mlx/tree/main/examples/extensions/`_
+   The full example code is available in `mlx <https://github.com/ml-explore/mlx/tree/main/examples/extensions/>`_.
 
 .. _Accelerate: https://developer.apple.com/documentation/accelerate/blas?language=objc
 .. _Metal: https://developer.apple.com/documentation/metal?language=objc
 .. _Metal-cpp: https://developer.apple.com/metal/cpp/
 .. _`Metal Specification`: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
 .. _`Metal Example`: https://developer.apple.com/documentation/metal/performing_calculations_on_a_gpu?language=objc
-.. _PyBind11: https://pybind11.readthedocs.io/en/stable/
+.. _nanobind: https://nanobind.readthedocs.io/en/latest/

docs/src/dev/metal_debugger.rst

@@ -1,29 +1,45 @@
 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:
+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.
 
-The ``metal::start_capture`` function initiates a capture of all MLX GPU work.
+To build with debugging enabled in Python prepend
+``CMAKE_ARGS="-DMLX_METAL_DEBUG=ON"`` to the build call.
 
-.. code-block:: C++
+The :func:`metal.start_capture` function initiates a capture of all MLX GPU
+work.
 
-    int main() {
-        metal::start_capture("/Users/Jane/Developer/MLX.gputrace");
+.. note::
 
-        auto a = arange(10.f, 20.f, 1.f, float32);
-        auto b = arange(30.f, 40.f, 1.f, float32);
-        auto c = add(a, b);
+   To capture a GPU trace you must run the application with
+   ``MTL_CAPTURE_ENABLED=1``.
 
-        eval(c);
+.. code-block:: python
 
-        metal::stop_capture();
-    }
+    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
@@ -35,8 +51,8 @@ documentation`_ for more information.
 Xcode Workflow
 --------------
 
-You can skip saving to a path by running within Xcode. First, generate an Xcode
-project using CMake.
+You can skip saving to a path by running within Xcode. First, generate an
+Xcode project using CMake.
 
 .. code-block::
 

docs/src/examples/llama-inference.rst

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

docs/src/examples/mlp.rst

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

docs/src/index.rst

@@ -43,6 +43,7 @@ are the CPU and GPU.
    usage/function_transforms
    usage/compile
    usage/numpy
+   usage/distributed
    usage/using_streams
 
 .. toctree::
@@ -69,6 +70,7 @@ are the CPU and GPU.
    python/metal
    python/nn
    python/optimizers
+   python/distributed
    python/tree_utils
 
 .. toctree::
@@ -83,3 +85,4 @@ are the CPU and GPU.
 
    dev/extensions
    dev/metal_debugger
+   dev/custom_metal_kernels

docs/src/install.rst

@@ -70,36 +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
 
-Install `nanobind <https://nanobind.readthedocs.io/en/latest/>`_ with:
-
-.. code-block:: shell
-
-    pip install git+https://github.com/wjakob/nanobind.git
-
 Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-   env CMAKE_BUILD_PARALLEL_LEVEL="" pip install .
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
 
-For developing use an editable install:
+For developing, install the package with development dependencies, and use an
+editable install:
 
 .. code-block:: shell
 
-  env CMAKE_BUILD_PARALLEL_LEVEL="" pip install -e .
+  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
 
-To make sure the install is working run the tests with:
+Once the development dependencies are installed, you can build faster with:
+
+.. code-block:: shell
+
+ CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
+
+Run the tests with:
 
 .. code-block:: shell
 
-  pip install ".[testing]"
   python -m unittest discover python/tests
 
-Optional: Install stubs to enable auto completions and type checking from your IDE:
+Optional: Install stubs to enable auto completions and type checking from your
+IDE:
 
 .. code-block:: shell
 
-  pip install ".[dev]"
   python setup.py generate_stubs
 
 C++ API
@@ -120,7 +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:
 
@@ -139,7 +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
 
@@ -153,33 +153,67 @@ should point to the path to the built metal library.
      - OFF
    * - MLX_BUILD_METAL
      - ON
+   * - MLX_BUILD_CPU
+     - ON
    * - MLX_BUILD_PYTHON_BINDINGS
      - OFF
    * - MLX_METAL_DEBUG
      - OFF
-
+   * - MLX_BUILD_SAFETENSORS
+     - ON
+   * - MLX_BUILD_GGUF
+     - ON
+   * - MLX_METAL_JIT
+     - OFF
 
 .. note::
 
-    If you have multiple Xcode installations and wish to use 
-    a specific one while building, you can do so by adding the 
-    following environment variable before building 
+    If you have multiple Xcode installations and wish to use
+    a specific one while building, you can do so by adding the
+    following environment variable before building
 
     .. code-block:: shell
 
       export DEVELOPER_DIR="/path/to/Xcode.app/Contents/Developer/"
 
-    Further, you can use the following command to find out which 
+    Further, you can use the following command to find out which
     macOS SDK will be used
 
     .. code-block:: shell
 
       xcrun -sdk macosx --show-sdk-version
 
+Binary Size Minimization
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+To produce a smaller binary use the CMake flags ``CMAKE_BUILD_TYPE=MinSizeRel``
+and ``BUILD_SHARED_LIBS=ON``.
+
+The MLX CMake build has several additional options to make smaller binaries.
+For example, if you don't need the CPU backend or support for safetensors and
+GGUF, you can do:
+
+.. code-block:: shell
+
+  cmake .. \
+    -DCMAKE_BUILD_TYPE=MinSizeRel \
+    -DBUILD_SHARED_LIBS=ON \
+    -DMLX_BUILD_CPU=OFF \
+    -DMLX_BUILD_SAFETENSORS=OFF \
+    -DMLX_BUILD_GGUF=OFF \
+    -DMLX_METAL_JIT=ON
+
+THE ``MLX_METAL_JIT`` flag minimizes the size of the MLX Metal library which
+contains pre-built GPU kernels. This substantially reduces the size of the
+Metal library by run-time compiling kernels the first time they are used in MLX
+on a given machine. Note run-time compilation incurs a cold-start cost which can
+be anwywhere from a few hundred millisecond to a few seconds depending on the
+application. Once a kernel is compiled, it will be cached by the system. The
+Metal kernel cache persists accross reboots.
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 
-
 Metal not found
 ~~~~~~~~~~~~~~~
 
@@ -201,7 +235,7 @@ Then set the active developer directory:
 
   sudo xcode-select --switch /Applications/Xcode.app/Contents/Developer
 
-x86 Shell 
+x86 Shell
 ~~~~~~~~~
 
 .. _build shell:

docs/src/python/array.rst

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

docs/src/python/devices_and_streams.rst

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

docs/src/python/distributed.rst

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

docs/src/python/fast.rst

@@ -12,3 +12,5 @@ Fast
   layer_norm
   rope
   scaled_dot_product_attention
+  affine_quantize
+  metal_kernel

docs/src/python/linalg.rst

@@ -8,5 +8,10 @@ Linear Algebra
 .. autosummary:: 
    :toctree: _autosummary 
 
+    inv
+    tri_inv
     norm
+    cholesky
+    cholesky_inv
     qr
+    svd

docs/src/python/metal.rst

@@ -3,12 +3,17 @@ Metal
 
 .. currentmodule:: mlx.core.metal
 
-.. autosummary:: 
+.. autosummary::
   :toctree: _autosummary
 
   is_available
+  device_info
   get_active_memory
   get_peak_memory
+  reset_peak_memory
   get_cache_memory
   set_memory_limit
   set_cache_limit
+  clear_cache
+  start_capture
+  stop_capture

docs/src/python/nn.rst

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

docs/src/python/nn/functions.rst

@@ -17,6 +17,8 @@ simple functions.
    gelu_approx
    gelu_fast_approx
    glu
+   hard_shrink
+   hard_tanh
    hardswish
    leaky_relu
    log_sigmoid
@@ -29,6 +31,7 @@ simple functions.
    sigmoid
    silu
    softmax
+   softmin
    softplus
    softshrink
    step

docs/src/python/nn/layers.rst

@@ -15,15 +15,24 @@ Layers
    BatchNorm
    Conv1d
    Conv2d
+   Conv3d
+   ConvTranspose1d
+   ConvTranspose2d
+   ConvTranspose3d
    Dropout
    Dropout2d
    Dropout3d
    Embedding
    GELU
+   GLU
    GroupNorm
    GRU
+   HardShrink
+   HardTanh
+   Hardswish
    InstanceNorm
    LayerNorm
+   LeakyReLU
    Linear
    LSTM
    MaxPool1d
@@ -31,16 +40,23 @@ Layers
    Mish
    MultiHeadAttention
    PReLU
+   QuantizedEmbedding
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU6
    RNN
    RoPE
    SELU
    Sequential
    SiLU
    SinusoidalPositionalEncoding
+   Softmin
    Softshrink
+   Softsign
+   Softmax
+   Softplus
    Step
+   Tanh
    Transformer
-   Upsample
+   Upsample

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
@@ -19,22 +20,34 @@ 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
@@ -42,30 +55,40 @@ Operations
    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
    inner
+   isfinite
    isclose
    isinf
    isnan
    isneginf
    isposinf
+   issubdtype
+   left_shift
    less
    less_equal
    linspace
@@ -83,22 +106,29 @@ Operations
    max
    maximum
    mean
+   meshgrid
    min
    minimum
    moveaxis
    multiply
+   nan_to_num
    negative
+   not_equal
    ones
    ones_like
    outer
    partition
    pad
+   power
    prod
    quantize
    quantized_matmul
+   radians
    reciprocal
+   remainder
    repeat
    reshape
+   right_shift
    round
    rsqrt
    save
@@ -117,6 +147,7 @@ Operations
    square
    squeeze
    stack
+   std
    stop_gradient
    subtract
    sum
@@ -128,11 +159,13 @@ Operations
    tensordot
    tile
    topk
+   trace
    transpose
    tri
    tril
    triu
    var
+   view
    where
    zeros
    zeros_like

docs/src/python/optimizers.rst

@@ -1,5 +1,7 @@
 .. _optimizers:
 
+.. currentmodule:: mlx.optimizers
+
 Optimizers
 ==========
 
@@ -29,8 +31,48 @@ model's parameters and the **optimizer state**.
             # Compute the new parameters but also the optimizer state.
             mx.eval(model.parameters(), optimizer.state)
 
+Saving and Loading
+------------------
+
+To serialize an optimizer, save its state. To load an optimizer, load and set
+the saved state. Here's a simple example:
+
+.. code-block:: python
+
+   import mlx.core as mx
+   from mlx.utils import tree_flatten, tree_unflatten
+   import mlx.optimizers as optim
+
+   optimizer = optim.Adam(learning_rate=1e-2)
+
+   # Perform some updates with the optimizer
+   model = {"w" : mx.zeros((5, 5))}
+   grads = {"w" : mx.ones((5, 5))}
+   optimizer.update(model, grads)
+
+   # Save the state
+   state = tree_flatten(optimizer.state)
+   mx.save_safetensors("optimizer.safetensors", dict(state))
+
+   # Later on, for example when loading from a checkpoint,
+   # recreate the optimizer and load the state
+   optimizer = optim.Adam(learning_rate=1e-2)
+
+   state = tree_unflatten(list(mx.load("optimizer.safetensors").items()))
+   optimizer.state = state
+
+Note, not every optimizer configuation parameter is saved in the state. For
+example, for Adam the learning rate is saved but the ``betas`` and ``eps``
+parameters are not. A good rule of thumb is if the parameter can be scheduled
+then it will be included in the optimizer state.
+
 .. toctree::
 
    optimizers/optimizer
    optimizers/common_optimizers
    optimizers/schedulers
+
+.. autosummary::
+   :toctree: _autosummary
+
+   clip_grad_norm

docs/src/python/random.rst

@@ -38,8 +38,10 @@ 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

docs/src/python/transforms.rst

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

docs/src/python/tree_utils.rst

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

docs/src/usage/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/lazy_evaluation.rst

@@ -18,7 +18,7 @@ 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.
 

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

docs/src/usage/saving_and_loading.rst

@@ -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:

examples/cpp/CMakeLists.txt

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

examples/cpp/distributed.cpp

@@ -0,0 +1,22 @@
+// Copyright © 2024 Apple Inc.
+
+#include <iostream>
+
+#include "mlx/mlx.h"
+
+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

@@ -8,9 +8,10 @@
 using namespace mlx::core;
 
 int main() {
-  // Enable the MLX_METAL_DEBUG CMake option to enhance the capture with groups,
-  // labels, etc.
-  assert(metal::start_capture());
+  // 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

examples/cpp/tutorial.cpp

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

examples/extensions/CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.27)
 
-project(mlx_sample_extensions LANGUAGES CXX)
+project(_ext LANGUAGES CXX)
 
 # ----------------------------- Setup -----------------------------
 set(CMAKE_CXX_STANDARD 17)
@@ -11,8 +11,12 @@ option(BUILD_SHARED_LIBS "Build extensions as a shared library" ON)
 
 # ----------------------------- Dependencies -----------------------------
 find_package(MLX CONFIG REQUIRED)
-find_package(Python COMPONENTS Interpreter Development)
-find_package(pybind11 CONFIG REQUIRED)
+find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)
+execute_process(
+  COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
+  OUTPUT_STRIP_TRAILING_WHITESPACE OUTPUT_VARIABLE NB_DIR)
+list(APPEND CMAKE_PREFIX_PATH "${NB_DIR}")
+find_package(nanobind CONFIG REQUIRED)
 
 # ----------------------------- Extensions -----------------------------
 
@@ -38,7 +42,6 @@ target_link_libraries(mlx_ext PUBLIC mlx)
 
 # Build metallib
 if(MLX_BUILD_METAL)
-
   mlx_build_metallib(
     TARGET mlx_ext_metallib
     TITLE mlx_ext
@@ -54,13 +57,15 @@ if(MLX_BUILD_METAL)
 
 endif()
 
-# ----------------------------- Pybind -----------------------------
-pybind11_add_module(
-  mlx_sample_extensions
+# ----------------------------- Python Bindings -----------------------------
+nanobind_add_module(
+  _ext
+  NB_STATIC STABLE_ABI LTO NOMINSIZE
+  NB_DOMAIN mlx 
   ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
 )
-target_link_libraries(mlx_sample_extensions PRIVATE mlx_ext)
+target_link_libraries(_ext PRIVATE mlx_ext)
 
 if(BUILD_SHARED_LIBS)
-  target_link_options(mlx_sample_extensions PRIVATE -Wl,-rpath,@loader_path)
+  target_link_options(_ext PRIVATE -Wl,-rpath,@loader_path)
 endif()

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);
 
@@ -106,12 +106,12 @@ void axpby_impl(
 /** Fall back implementation for evaluation on CPU */
 void Axpby::eval(
     const std::vector<array>& inputs,
-    std::vector<array>& out_arr) {
-  auto out = out_arr[0];
+    std::vector<array>& outputs) {
   // Check the inputs (registered in the op while constructing the out array)
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Dispatch to the correct dtype
   if (out.dtype() == float32) {
@@ -150,11 +150,7 @@ void axpby_impl_accelerate(
   // The data in the output array is allocated to match the strides in y
   // such that x, y, and out are contiguous in the same mode and
   // no transposition is needed
-  out.set_data(
-      allocator::malloc_or_wait(y.data_size() * out.itemsize()),
-      y.data_size(),
-      y.strides(),
-      y.flags());
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
   // We then copy over the elements using the contiguous vector specialization
   copy_inplace(y, out, CopyType::Vector);
@@ -180,11 +176,11 @@ void axpby_impl_accelerate(
 /** Evaluate primitive on CPU using accelerate specializations */
 void Axpby::eval_cpu(
     const std::vector<array>& inputs,
-    std::vector<array>& outarr) {
-  auto out = outarr[0];
+    std::vector<array>& outputs) {
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Accelerate specialization for contiguous single precision float arrays
   if (out.dtype() == float32 &&
@@ -195,7 +191,7 @@ void Axpby::eval_cpu(
   }
 
   // Fall back to common backend if specializations are not available
-  eval(inputs, outarr);
+  eval(inputs, outputs);
 }
 
 #else // Accelerate not available
@@ -203,8 +199,8 @@ void Axpby::eval_cpu(
 /** Evaluate primitive on CPU falling back to common backend */
 void Axpby::eval_cpu(
     const std::vector<array>& inputs,
-    std::vector<array>& out) {
-  eval(inputs, out);
+    const std::vector<array>& outputs) {
+  eval(inputs, outputs);
 }
 
 #endif
@@ -218,12 +214,12 @@ void Axpby::eval_cpu(
 /** Evaluate primitive on GPU */
 void Axpby::eval_gpu(
     const std::vector<array>& inputs,
-    std::vector<array>& outarr) {
+    std::vector<array>& outputs) {
   // Prepare inputs
-  auto out = outarr[0];
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Each primitive carries the stream it should execute on
   // and each stream carries its device identifiers
@@ -253,15 +249,14 @@ void Axpby::eval_gpu(
   kname << (contiguous_kernel ? "contiguous_" : "general_");
   kname << type_to_name(out);
 
-  // Make sure the metal library is available and look for it
-  // in the same folder as this executable if needed
-  d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
+  // Make sure the metal library is available
+  d.register_library("mlx_ext");
 
   // Make a kernel from this metal library
   auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
   // Prepare to encode kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
 
   // Kernel parameters are registered with buffer indices corresponding to
@@ -270,11 +265,11 @@ void Axpby::eval_gpu(
   size_t nelem = out.size();
 
   // Encode input arrays to kernel
-  set_array_buffer(compute_encoder, x, 0);
-  set_array_buffer(compute_encoder, y, 1);
+  compute_encoder.set_input_array(x, 0);
+  compute_encoder.set_input_array(y, 1);
 
   // Encode output arrays to kernel
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode alpha and beta
   compute_encoder->setBytes(&alpha_, sizeof(float), 3);
@@ -300,7 +295,7 @@ void Axpby::eval_gpu(
 
   // Launch the grid with the given number of threads divided among
   // the given threadgroups
-  compute_encoder->dispatchThreads(grid_dims, group_dims);
+  compute_encoder.dispatchThreads(grid_dims, group_dims);
 }
 
 #else // Metal is not available
@@ -372,4 +367,4 @@ bool Axpby::is_equivalent(const Primitive& other) const {
   return alpha_ == r_other.alpha_ && beta_ == r_other.beta_;
 }
 
-} // namespace mlx::core
+} // namespace mlx::core

examples/extensions/axpby/axpby.h

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

examples/extensions/axpby/axpby.metal

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

examples/extensions/bindings.cpp

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

examples/extensions/mlx_sample_extensions/__init__.py

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

examples/extensions/pyproject.toml

@@ -1,3 +1,8 @@
 [build-system]
-requires = ["setuptools>=42", "pybind11>=2.10", "cmake>=3.24", "mlx @ git+https://github.com/mlx-explore/mlx@main"]
-build-backend = "setuptools.build_meta"
+requires = [
+  "setuptools>=42",
+  "cmake>=3.24",
+  "mlx>=0.17.0",
+  "nanobind==2.1.0",
+]
+build-backend = "setuptools.build_meta"

examples/extensions/requirements.txt

@@ -0,0 +1,4 @@
+setuptools>=42
+cmake>=3.24
+mlx>=0.17.0
+nanobind==2.1.0

examples/extensions/setup.py

@@ -1,4 +1,4 @@
-# Copyright © 2023 Apple Inc.
+# Copyright © 2023-2024 Apple Inc.
 
 from setuptools import setup
 
@@ -9,10 +9,9 @@ if __name__ == "__main__":
         name="mlx_sample_extensions",
         version="0.0.0",
         description="Sample C++ and Metal extensions for MLX primitives.",
-        ext_modules=[extension.CMakeExtension("mlx_sample_extensions")],
+        ext_modules=[extension.CMakeExtension("mlx_sample_extensions._ext")],
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
-        package_dir={"": "."},
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
         zip_safe=False,
         python_requires=">=3.8",

examples/extensions/test.py

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

mlx/CMakeLists.txt

@@ -6,6 +6,7 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/compile.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/dtype.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/einsum.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/fast.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/ops.cpp
@@ -19,11 +20,17 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/metal.h
 )
 
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
+if (MLX_BUILD_CPU)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
+else()
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_cpu)
+endif()
+
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/distributed)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
 if (MLX_BUILD_ACCELERATE)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/accelerate)
-else()
+elseif(MLX_BUILD_CPU)
   target_sources(
     mlx
     PRIVATE

mlx/allocator.cpp

@@ -23,11 +23,22 @@ void free(Buffer buffer) {
 }
 
 Buffer CommonAllocator::malloc(size_t size, bool) {
-  return Buffer{std::malloc(size)};
+  void* ptr = std::malloc(size + sizeof(size_t));
+  if (ptr != nullptr) {
+    *static_cast<size_t*>(ptr) = size;
+  }
+  return Buffer{ptr};
 }
 
 void CommonAllocator::free(Buffer buffer) {
-  std::free(buffer.raw_ptr());
+  std::free(buffer.ptr());
+}
+
+size_t CommonAllocator::size(Buffer buffer) const {
+  if (buffer.ptr() == nullptr) {
+    return 0;
+  }
+  return *static_cast<size_t*>(buffer.ptr());
 }
 
 Buffer malloc_or_wait(size_t size) {

mlx/allocator.h

@@ -14,7 +14,7 @@ class Buffer {
   void* ptr_;
 
  public:
-  Buffer(void* ptr) : ptr_(ptr){};
+  Buffer(void* ptr) : ptr_(ptr) {};
 
   // Get the raw data pointer from the buffer
   void* raw_ptr();
@@ -41,6 +41,7 @@ class Allocator {
  public:
   virtual Buffer malloc(size_t size, bool allow_swap = false) = 0;
   virtual void free(Buffer buffer) = 0;
+  virtual size_t size(Buffer buffer) const = 0;
 
   Allocator() = default;
   Allocator(const Allocator& other) = delete;
@@ -57,6 +58,7 @@ class CommonAllocator : public Allocator {
  public:
   virtual Buffer malloc(size_t size, bool allow_swap = false) override;
   virtual void free(Buffer buffer) override;
+  virtual size_t size(Buffer buffer) const override;
 
  private:
   CommonAllocator() = default;

mlx/array.cpp

@@ -1,5 +1,4 @@
 // Copyright © 2023-2024 Apple Inc.
-
 #include <functional>
 
 #include "mlx/array.h"
@@ -18,6 +17,10 @@ bool in_tracing() {
   return detail::InTracing::in_tracing();
 }
 
+bool retain_graph() {
+  return detail::RetainGraph::retain_graph();
+}
+
 } // namespace
 
 array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
@@ -93,11 +96,17 @@ void array::detach() {
 }
 
 void array::eval() {
-  mlx::core::eval({*this});
+  // Ensure the array is ready to be read
+  if (status() == Status::scheduled) {
+    event().wait();
+    set_status(Status::available);
+  } else if (status() == Status::unscheduled) {
+    mlx::core::eval({*this});
+  }
 }
 
 bool array::is_tracer() const {
-  return array_desc_->is_tracer && in_tracing();
+  return array_desc_->is_tracer && in_tracing() || retain_graph();
 }
 
 void array::set_data(allocator::Buffer buffer, deleter_t d) {
@@ -161,6 +170,40 @@ void array::move_shared_buffer(array other) {
   move_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
+array::~array() {
+  if (array_desc_ == nullptr) {
+    return;
+  }
+
+  // Ignore arrays that might be detached during eval
+  if (status() == array::Status::scheduled) {
+    return;
+  }
+
+  // Break circular reference for non-detached arrays with siblings
+  if (auto n = siblings().size(); n > 0) {
+    bool do_detach = true;
+    // If all siblings have siblings.size() references except
+    // the one we are currently destroying (which has siblings.size() + 1)
+    // then there are no more external references
+    do_detach &= (array_desc_.use_count() == (n + 1));
+    for (auto& s : siblings()) {
+      do_detach &= (s.array_desc_.use_count() == n);
+      if (!do_detach) {
+        break;
+      }
+    }
+    if (do_detach) {
+      for (auto& s : siblings()) {
+        for (auto& ss : s.siblings()) {
+          ss.array_desc_ = nullptr;
+        }
+        s.array_desc_->siblings.clear();
+      }
+    }
+  }
+}
+
 void array::ArrayDesc::init() {
   strides.resize(shape.size());
   size = 1;
@@ -168,13 +211,13 @@ void array::ArrayDesc::init() {
     strides[i] = size;
     size *= shape[i];
   }
-  for (auto& in : inputs) {
+  for (const auto& in : inputs) {
     is_tracer |= in.is_tracer();
   }
 }
 
 array::ArrayDesc::ArrayDesc(std::vector<int> shape, Dtype dtype)
-    : shape(std::move(shape)), dtype(dtype) {
+    : shape(std::move(shape)), dtype(dtype), status(Status::available) {
   init();
 }
 
@@ -185,11 +228,42 @@ array::ArrayDesc::ArrayDesc(
     std::vector<array> inputs)
     : shape(std::move(shape)),
       dtype(dtype),
+      status(Status::unscheduled),
       primitive(std::move(primitive)),
       inputs(std::move(inputs)) {
   init();
 }
 
+array::ArrayDesc::~ArrayDesc() {
+  // When an array description is destroyed it will delete a bunch of arrays
+  // that may also destroy their corresponding descriptions and so on and so
+  // forth.
+  //
+  // This calls recursively the destructor and can result in stack overflow, we
+  // instead put them in a vector and destroy them one at a time resulting in a
+  // max stack depth of 2.
+  std::vector<std::shared_ptr<ArrayDesc>> for_deletion;
+
+  for (array& a : inputs) {
+    if (a.array_desc_.use_count() == 1) {
+      for_deletion.push_back(std::move(a.array_desc_));
+    }
+  }
+
+  while (!for_deletion.empty()) {
+    // top is going to be deleted at the end of the block *after* the arrays
+    // with inputs have been moved into the vector
+    auto top = std::move(for_deletion.back());
+    for_deletion.pop_back();
+
+    for (array& a : top->inputs) {
+      if (a.array_desc_.use_count() == 1) {
+        for_deletion.push_back(std::move(a.array_desc_));
+      }
+    }
+  }
+}
+
 array::ArrayIterator::ArrayIterator(const array& arr, int idx)
     : arr(arr), idx(idx) {
   if (arr.ndim() == 0) {

mlx/array.h

@@ -9,6 +9,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/dtype.h"
+#include "mlx/event.h"
 
 namespace mlx::core {
 
@@ -72,32 +73,32 @@ class array {
       this->array_desc_ = other.array_desc_;
     }
     return *this;
-  };
+  }
 
   /** The size of the array's datatype in bytes. */
   size_t itemsize() const {
     return size_of(dtype());
-  };
+  }
 
   /** The number of elements in the array. */
   size_t size() const {
     return array_desc_->size;
-  };
+  }
 
   /** The number of bytes in the array. */
   size_t nbytes() const {
     return size() * itemsize();
-  };
+  }
 
   /** The number of dimensions of the array. */
   size_t ndim() const {
     return array_desc_->shape.size();
-  };
+  }
 
   /** The shape of the array as a vector of integers. */
   const std::vector<int>& shape() const {
     return array_desc_->shape;
-  };
+  }
 
   /**
    *  Get the size of the corresponding dimension.
@@ -106,17 +107,26 @@ class array {
    *  bounds checking. */
   int shape(int dim) const {
     return shape().at(dim < 0 ? dim + ndim() : dim);
-  };
+  }
 
   /** The strides of the array. */
   const std::vector<size_t>& strides() const {
     return array_desc_->strides;
-  };
+  }
+
+  /**
+   *  Get the stride of the corresponding dimension.
+   *
+   *  This function supports negative indexing and provides
+   *  bounds checking. */
+  size_t strides(int dim) const {
+    return strides().at(dim < 0 ? dim + ndim() : dim);
+  }
 
   /** Get the arrays data type. */
   Dtype dtype() const {
     return array_desc_->dtype;
-  };
+  }
 
   /** Evaluate the array. */
   void eval();
@@ -150,10 +160,10 @@ class array {
 
     friend bool operator==(const ArrayIterator& a, const ArrayIterator& b) {
       return a.arr.id() == b.arr.id() && a.idx == b.idx;
-    };
+    }
     friend bool operator!=(const ArrayIterator& a, const ArrayIterator& b) {
       return !(a == b);
-    };
+    }
 
    private:
     const array& arr;
@@ -199,7 +209,7 @@ class array {
     allocator::Buffer buffer;
     deleter_t d;
     Data(allocator::Buffer buffer, deleter_t d = allocator::free)
-        : buffer(buffer), d(d){};
+        : buffer(buffer), d(d) {}
     // Not copyable
     Data(const Data& d) = delete;
     Data& operator=(const Data& d) = delete;
@@ -209,33 +219,45 @@ class array {
   };
 
   struct Flags {
-    // True if there are no gaps in the underlying data. Each item
+    // True iff there are no gaps in the underlying data. Each item
     // in the underlying data buffer belongs to at least one index.
+    //
+    // True iff:
+    // prod(shape[i] for i in range(ndim) if strides[i] > 0) == data_size()
     bool contiguous : 1;
 
+    // True iff:
+    // strides[-1] == 1 and
+    // all(strides[i] == (shape[i+1]*strides[i+1]) or shape[i] == 1 for i in
+    // range(ndim - 1))
     bool row_contiguous : 1;
+
+    // True iff:
+    // strides[0] == 1 and
+    // all(strides[i] == (shape[i-1]*strides[i-1]) or shape[i] == 1 for i in
+    // range(1, ndim))
     bool col_contiguous : 1;
   };
 
   /** The array's primitive. */
   Primitive& primitive() const {
     return *(array_desc_->primitive);
-  };
+  }
 
   /** A shared pointer to the array's primitive. */
   std::shared_ptr<Primitive>& primitive_ptr() const {
     return array_desc_->primitive;
-  };
+  }
 
   /** Check if the array has an attached primitive or is a leaf node. */
   bool has_primitive() const {
     return array_desc_->primitive != nullptr;
-  };
+  }
 
   /** The array's inputs. */
   const std::vector<array>& inputs() const {
     return array_desc_->inputs;
-  };
+  }
 
   std::vector<array>& inputs() {
     return array_desc_->inputs;
@@ -249,24 +271,18 @@ class array {
   /** The array's siblings. */
   const std::vector<array>& siblings() const {
     return array_desc_->siblings;
-  };
+  }
+
+  /** The array's siblings. */
+  std::vector<array>& siblings() {
+    return array_desc_->siblings;
+  }
 
   void set_siblings(std::vector<array> siblings, uint16_t position) {
     array_desc_->siblings = std::move(siblings);
     array_desc_->position = position;
   }
 
-  /** The i-th output of the array's primitive. */
-  const array& output(int i) const {
-    if (i == array_desc_->position) {
-      return *this;
-    } else if (i < array_desc_->position) {
-      return siblings()[i];
-    } else {
-      return siblings()[i + 1];
-    }
-  };
-
   /** The outputs of the array's primitive (i.e. this array and
    * its siblings) in the order the primitive expects. */
   std::vector<array> outputs() const {
@@ -277,7 +293,7 @@ class array {
     outputs.push_back(*this);
     outputs.insert(outputs.end(), siblings().begin() + idx, siblings().end());
     return outputs;
-  };
+  }
 
   /** Detach the array from the graph. */
   void detach();
@@ -285,19 +301,32 @@ class array {
   /** Get the Flags bit-field. */
   const Flags& flags() const {
     return array_desc_->flags;
-  };
+  }
 
-  /** The size (in elements) of the underlying buffer the array points to. */
+  /** The size (in elements) of the underlying buffer the array points to.
+   *
+   * This can be different than the actual size of the array if the array has
+   * been broadcast or irregularly strided.  If ``first`` is the offset into
+   * the data buffer of the first element of the array (i.e. the offset
+   * corresponding to ``arr[0, 0, ...]``) and last is the offset into the
+   * data buffer of the last element of the array (i.e. the offset
+   * corresponding to ``arr[-1, -1, ...]``) then ``data_size = last - first``.
+   * Note, ``data_size`` is in units of ``item_size`` (not bytes).
+   **/
   size_t data_size() const {
     return array_desc_->data_size;
-  };
+  }
 
   allocator::Buffer& buffer() {
     return array_desc_->data->buffer;
-  };
+  }
   const allocator::Buffer& buffer() const {
     return array_desc_->data->buffer;
-  };
+  }
+
+  size_t buffer_size() const {
+    return allocator::allocator().size(buffer());
+  }
 
   // Return a copy of the shared pointer
   // to the array::Data struct
@@ -308,16 +337,35 @@ class array {
   template <typename T>
   T* data() {
     return static_cast<T*>(array_desc_->data_ptr);
-  };
+  }
 
   template <typename T>
   const T* data() const {
     return static_cast<T*>(array_desc_->data_ptr);
-  };
+  }
 
-  // Check if the array has been evaluated
-  bool is_evaled() const {
-    return array_desc_->data != nullptr;
+  enum Status { unscheduled, scheduled, available };
+
+  bool is_available() const {
+    return status() == Status::available;
+  }
+
+  Status status() const {
+    return array_desc_->status;
+  }
+
+  void set_status(Status s) const {
+    array_desc_->status = s;
+  }
+
+  // Get the array's shared event
+  Event& event() const {
+    return array_desc_->event;
+  }
+
+  // Attach an event to a not yet evaluated array
+  void attach_event(Event e) const {
+    array_desc_->event = std::move(e);
   }
 
   // Mark the array as a tracer array (true) or not.
@@ -358,6 +406,8 @@ class array {
     array_desc_ = other.array_desc_;
   }
 
+  ~array();
+
  private:
   // Initialize the arrays data
   template <typename It>
@@ -370,6 +420,11 @@ class array {
     Dtype dtype;
     std::shared_ptr<Primitive> primitive;
 
+    Status status;
+
+    // An event on the array used for synchronization
+    Event event;
+
     // Indicates an array is being used in a graph transform
     // and should not be detached from the graph
     bool is_tracer{false};
@@ -382,8 +437,6 @@ class array {
     void* data_ptr{nullptr};
 
     // The size in elements of the data buffer the array accesses
-    // This can be different than the actual size of the array if it
-    // has been broadcast or irregularly strided.
     size_t data_size;
 
     // Contains useful meta data about the array
@@ -404,6 +457,8 @@ class array {
         std::shared_ptr<Primitive> primitive,
         std::vector<array> inputs);
 
+    ~ArrayDesc();
+
    private:
     // Initialize size, strides, and other metadata
     void init();
@@ -468,10 +523,11 @@ T array::item() const {
   if (size() != 1) {
     throw std::invalid_argument("item can only be called on arrays of size 1.");
   }
-  if (!is_evaled()) {
+  if (status() == Status::unscheduled) {
     throw std::invalid_argument(
         "item() const can only be called on evaled arrays");
   }
+  const_cast<array*>(this)->eval();
   return *data<T>();
 }
 

mlx/backend/accelerate/conv.cpp

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

mlx/backend/accelerate/matmul.cpp

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

mlx/backend/accelerate/primitives.cpp

@@ -3,8 +3,7 @@
 #include <cassert>
 #include <cmath>
 
-#include <vecLib/vDSP.h>
-#include <vecLib/vForce.h>
+#include <Accelerate/Accelerate.h>
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/binary.h"
@@ -31,11 +30,13 @@ DEFAULT(ArgPartition)
 DEFAULT(ArgReduce)
 DEFAULT(ArgSort)
 DEFAULT(AsStrided)
+DEFAULT(BlockMaskedMM)
 DEFAULT(Broadcast)
 DEFAULT(Ceil)
 DEFAULT(Concatenate)
+DEFAULT(Conjugate)
 DEFAULT(Copy)
-DEFAULT_MULTI(CustomVJP)
+DEFAULT_MULTI(CustomTransforms)
 DEFAULT_MULTI(Depends)
 DEFAULT_MULTI(DivMod)
 DEFAULT(NumberOfElements)
@@ -45,8 +46,11 @@ DEFAULT(ErfInv)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Gather)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
 DEFAULT(Less)
 DEFAULT(LessEqual)
 DEFAULT(Load)
@@ -76,6 +80,7 @@ DEFAULT(StopGradient)
 DEFAULT_MULTI(SVD)
 DEFAULT(Transpose)
 DEFAULT(Inverse)
+DEFAULT(Cholesky)
 
 void Abs::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
@@ -97,7 +102,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -112,7 +117,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vadd((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
         });
   } else if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -127,7 +132,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vaddi((const int*)a, 1, (const int*)b, 1, (int*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x + y; });
+    eval(inputs, out);
   }
 }
 
@@ -191,6 +196,26 @@ void ArcTan::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  if (out.dtype() == float32 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    if (a.is_donatable()) {
+      out.copy_shared_buffer(a);
+    } else if (b.is_donatable()) {
+      out.copy_shared_buffer(b);
+    } else {
+      out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    }
+    int size = a.data_size();
+    vvatan2f(out.data<float>(), a.data<float>(), b.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
 void ArcTanh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -262,7 +287,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -275,7 +300,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdivi((const int*)b, 1, (const int*)a, 1, (int*)o, 1, n);
         });
   } else if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -290,7 +315,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdiv((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x / y; });
+    eval(inputs, out);
   }
 }
 
@@ -301,12 +326,21 @@ void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
     set_unary_output_data(in, out);
     auto size = in.data_size();
     vvexpf(out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, [](auto x) { return std::exp(x); });
   } else {
-    throw std::invalid_argument(
-        "[exp] Cannot exponentiate elements in array"
-        " with non floating point type.");
+    eval(inputs, out);
+  }
+}
+
+void Expm1::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vvexpm1f(
+        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+  } else {
+    eval(inputs, out);
   }
 }
 
@@ -355,12 +389,8 @@ void Log1p::eval_cpu(const std::vector<array>& inputs, array& out) {
     auto size = in.data_size();
     vvlog1pf(
         out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, [](auto x) { return std::log1p(x); });
   } else {
-    throw std::invalid_argument(
-        "[log1p] Cannot compute log of elements in array with"
-        " non floating point type.");
+    eval(inputs, out);
   }
 }
 
@@ -370,7 +400,7 @@ void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -385,7 +415,7 @@ void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vmul((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x * y; });
+    eval(inputs, out);
   }
 }
 
@@ -396,7 +426,7 @@ void Negative::eval_cpu(const std::vector<array>& inputs, array& out) {
     set_unary_output_data(in, out);
     vDSP_vneg(in.data<float>(), 1, out.data<float>(), 1, in.data_size());
   } else {
-    unary(in, out, [](auto x) { return -x; });
+    eval(inputs, out);
   }
 }
 
@@ -483,7 +513,7 @@ void Square::eval_cpu(const std::vector<array>& inputs, array& out) {
     auto size = in.data_size();
     vDSP_vsq(in.data<float>(), 1, out.data<float>(), 1, size);
   } else {
-    unary(in, out, [](auto x) { return x * x; });
+    eval(inputs, out);
   }
 }
 
@@ -509,7 +539,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -527,7 +557,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vsub((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
         });
   } else if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -539,7 +569,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
         },
         UseDefaultBinaryOp());
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x - y; });
+    eval(inputs, out);
   }
 }
 

mlx/backend/accelerate/reduce.cpp

@@ -2,8 +2,8 @@
 
 #include <cassert>
 
+#include <Accelerate/Accelerate.h>
 #include <simd/vector.h>
-#include <vecLib/vDSP.h>
 
 #include "mlx/backend/common/reduce.h"
 #include "mlx/primitives.h"

mlx/backend/accelerate/softmax.cpp

@@ -1,9 +1,12 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <limits>
 
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 #include <arm_neon.h>
+#endif
+
 #include <simd/math.h>
 #include <simd/vector.h>
 
@@ -53,25 +56,25 @@ inline simd_float16 simd_fast_exp(simd_float16 x) {
   return (*(simd_float16*)&epart) * x;
 }
 
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 /**
  * The ARM neon equivalent of the fast exp above.
  */
 inline float16x8_t neon_fast_exp(float16x8_t x) {
-  x = vmulq_f16(x, vdupq_n_f16(1.442695)); // multiply with log_2(e)
-  x = vmaxq_f16(x, vdupq_n_f16(-14)); // clamp under with -14
-  x = vminq_f16(x, vdupq_n_f16(14)); // clamp over with 14
+  x = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
+  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f))); // clamp under with -14
+  x = vminq_f16(x, vdupq_n_f16(float16_t(14.f))); // clamp over with 14
 
-  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(0.5)));
+  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(float16_t(0.5f))));
   float16x8_t fpart = vsubq_f16(x, ipart);
 
-  x = vdupq_n_f16(1.535336188319500e-4f);
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(9.618437357674640e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(5.550332471162809e-2f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(2.402264791363012e-1f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(6.931472028550421e-1f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.000000000000000f), x, fpart);
+  x = vdupq_n_f16(float16_t(1.535336188319500e-4f));
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.000000000000000f)), x, fpart);
 
   // generate 2**ipart in the floating point representation using integer
   // bitshifting
@@ -107,53 +110,6 @@ inline float16_t neon_reduce_add(float16x8_t x) {
   return vget_lane_f16(y, 0);
 }
 
-template <typename T, typename VT>
-struct AccelerateSimdOps {
-  VT init(T a) {
-    return a;
-  }
-
-  VT load(const T* a) {
-    return *(VT*)a;
-  }
-
-  void store(T* dst, VT x) {
-    *(VT*)dst = x;
-  }
-
-  VT max(VT a, VT b) {
-    return simd_max(a, b);
-  };
-
-  VT exp(VT x) {
-    return simd_fast_exp(x);
-  }
-
-  VT add(VT a, VT b) {
-    return a + b;
-  }
-
-  VT sub(VT a, T b) {
-    return a - b;
-  }
-
-  VT mul(VT a, VT b) {
-    return a * b;
-  }
-
-  VT mul(VT a, T b) {
-    return a * b;
-  }
-
-  T reduce_max(VT x) {
-    return simd_reduce_max(x);
-  }
-
-  T reduce_add(VT x) {
-    return simd_reduce_add(x);
-  }
-};
-
 template <typename T, typename VT>
 struct NeonFp16SimdOps {
   VT init(T a) {
@@ -170,7 +126,7 @@ struct NeonFp16SimdOps {
 
   VT max(VT a, VT b) {
     return vmaxq_f16(a, b);
-  };
+  }
 
   VT exp(VT x) {
     return neon_fast_exp(x);
@@ -201,7 +157,56 @@ struct NeonFp16SimdOps {
   }
 };
 
-template <typename T, typename VT, typename Ops, int N>
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+template <typename T, typename VT>
+struct AccelerateSimdOps {
+  VT init(T a) {
+    return a;
+  }
+
+  VT load(const T* a) {
+    return *(VT*)a;
+  }
+
+  void store(T* dst, VT x) {
+    *(VT*)dst = x;
+  }
+
+  VT max(VT a, VT b) {
+    return simd_max(a, b);
+  }
+
+  VT exp(VT x) {
+    return simd_fast_exp(x);
+  }
+
+  VT add(VT a, VT b) {
+    return a + b;
+  }
+
+  VT sub(VT a, T b) {
+    return a - b;
+  }
+
+  VT mul(VT a, VT b) {
+    return a * b;
+  }
+
+  VT mul(VT a, T b) {
+    return a * b;
+  }
+
+  T reduce_max(VT x) {
+    return simd_reduce_max(x);
+  }
+
+  T reduce_add(VT x) {
+    return simd_reduce_add(x);
+  }
+};
+
+template <typename T, typename AccT, typename VT, typename Ops, int N>
 void softmax(const array& in, array& out) {
   Ops ops;
 
@@ -218,13 +223,21 @@ void softmax(const array& in, array& out) {
     VT vmaximum = ops.init(-std::numeric_limits<float>::infinity());
     size_t s = M;
     while (s >= N) {
-      vmaximum = ops.max(ops.load(current_in_ptr), vmaximum);
+      VT vals;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vals = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vals[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vmaximum = ops.max(vals, vmaximum);
       current_in_ptr += N;
       s -= N;
     }
-    T maximum = ops.reduce_max(vmaximum);
+    AccT maximum = ops.reduce_max(vmaximum);
     while (s-- > 0) {
-      maximum = std::max(maximum, *current_in_ptr);
+      maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
       current_in_ptr++;
     }
 
@@ -234,18 +247,29 @@ void softmax(const array& in, array& out) {
     current_in_ptr = in_ptr;
     s = M;
     while (s >= N) {
-      VT vexp = ops.exp(ops.sub(*(VT*)current_in_ptr, maximum));
-      ops.store(current_out_ptr, vexp);
-      *(VT*)current_out_ptr = vexp;
+      VT vexp;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vexp = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vexp = ops.exp(ops.sub(vexp, maximum));
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, vexp);
+      }
       vnormalizer = ops.add(vnormalizer, vexp);
       current_in_ptr += N;
       current_out_ptr += N;
       s -= N;
     }
-    T normalizer = ops.reduce_add(vnormalizer);
+    AccT normalizer = ops.reduce_add(vnormalizer);
     while (s-- > 0) {
-      T _exp = std::exp(*current_in_ptr - maximum);
-      *current_out_ptr = _exp;
+      AccT _exp = std::exp(*current_in_ptr - maximum);
+      if (std::is_same<T, AccT>::value) {
+        *current_out_ptr = _exp;
+      }
       normalizer += _exp;
       current_in_ptr++;
       current_out_ptr++;
@@ -254,14 +278,33 @@ void softmax(const array& in, array& out) {
 
     // Normalize
     current_out_ptr = out_ptr;
+    current_in_ptr = in_ptr;
     s = M;
     while (s >= N) {
-      ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
+      } else {
+        VT vexp;
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+        vexp = ops.mul(ops.exp(ops.sub(vexp, maximum)), normalizer);
+        for (int i = 0; i < N; ++i) {
+          current_out_ptr[i] = vexp[i];
+        }
+        current_in_ptr += N;
+      }
       current_out_ptr += N;
       s -= N;
     }
     while (s-- > 0) {
-      *current_out_ptr *= normalizer;
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr *= normalizer;
+      } else {
+        AccT _exp = std::exp(*current_in_ptr - maximum);
+        *current_out_ptr = static_cast<T>(_exp * normalizer);
+        current_in_ptr++;
+      }
       current_out_ptr++;
     }
   }
@@ -308,15 +351,33 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
           "Softmax is defined only for floating point types");
       break;
     case float32:
-      softmax<float, simd_float16, AccelerateSimdOps<float, simd_float16>, 16>(
-          in, out);
+      softmax<
+          float,
+          float,
+          simd_float16,
+          AccelerateSimdOps<float, simd_float16>,
+          16>(in, out);
       break;
     case float16:
-      softmax<
-          float16_t,
-          float16x8_t,
-          NeonFp16SimdOps<float16_t, float16x8_t>,
-          8>(in, out);
+      if (precise_) {
+        softmax<
+            float16_t,
+            float,
+            simd_float16,
+            AccelerateSimdOps<float, simd_float16>,
+            16>(in, out);
+      } else {
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        softmax<
+            float16_t,
+            float16_t,
+            float16x8_t,
+            NeonFp16SimdOps<float16_t, float16x8_t>,
+            8>(in, out);
+#else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        eval(inputs, out); // Redirect to common backend for consistency
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+      }
       break;
     case bfloat16:
       eval(inputs, out);

mlx/backend/accelerate/utils.h

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

mlx/backend/common/CMakeLists.txt

@@ -37,15 +37,20 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/common.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/erf.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/masked_mm.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/reduce_utils.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
@@ -54,6 +59,7 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/svd.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/cholesky.cpp
   ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp
 )
 

mlx/backend/common/binary.cpp

@@ -196,6 +196,20 @@ void LogAddExp::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void LogicalAnd::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalAnd());
+}
+
+void LogicalOr::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalOr requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalOr());
+}
+
 void Maximum::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
@@ -236,4 +250,82 @@ void Subtract::eval(const std::vector<array>& inputs, array& out) {
   binary(a, b, out, detail::Subtract());
 }
 
+void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto dispatch_type = [&a, &b, &out](auto op) {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool>(a, b, out, op);
+      case uint8:
+        binary_op<uint8_t>(a, b, out, op);
+        break;
+      case uint16:
+        binary_op<uint16_t>(a, b, out, op);
+        break;
+      case uint32:
+        binary_op<uint32_t>(a, b, out, op);
+        break;
+      case uint64:
+        binary_op<uint64_t>(a, b, out, op);
+        break;
+      case int8:
+        binary_op<int8_t>(a, b, out, op);
+        break;
+      case int16:
+        binary_op<int16_t>(a, b, out, op);
+        break;
+      case int32:
+        binary_op<int32_t>(a, b, out, op);
+        break;
+      case int64:
+        binary_op<int64_t>(a, b, out, op);
+        break;
+      default:
+        throw std::runtime_error(
+            "[BitwiseBinary::eval_cpu] Type not supported");
+        break;
+    }
+  };
+  switch (op_) {
+    case BitwiseBinary::And:
+      dispatch_type(detail::BitwiseAnd());
+      break;
+    case BitwiseBinary::Or:
+      dispatch_type(detail::BitwiseOr());
+      break;
+    case BitwiseBinary::Xor:
+      dispatch_type(detail::BitwiseXor());
+      break;
+    case BitwiseBinary::LeftShift:
+      dispatch_type(detail::LeftShift());
+      break;
+    case BitwiseBinary::RightShift:
+      dispatch_type(detail::RightShift());
+      break;
+  }
+}
+
+void ArcTan2::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  if (out.dtype() == float32) {
+    binary_op<float>(a, b, out, detail::ArcTan2());
+  } else if (out.dtype() == float16) {
+    binary_op<float16_t>(a, b, out, detail::ArcTan2());
+  } else if (out.dtype() == bfloat16) {
+    binary_op<bfloat16_t>(a, b, out, detail::ArcTan2());
+  } else if (issubdtype(out.dtype(), inexact)) {
+    std::ostringstream err;
+    err << "[arctan2] Does not support " << out.dtype();
+    throw std::invalid_argument(err.str());
+  } else {
+    throw std::invalid_argument(
+        "[arctan2] Cannot compute inverse tangent for arrays"
+        " with non floating point type.");
+  }
+}
+
 } // namespace mlx::core

mlx/backend/common/binary.h

@@ -1,6 +1,8 @@
 // Copyright © 2023 Apple Inc.
 
 #pragma once
+#include <cassert>
+
 #include "mlx/allocator.h"
 #include "mlx/array.h"
 #include "mlx/backend/common/utils.h"
@@ -41,13 +43,15 @@ void set_binary_op_output_data(
     array& out,
     BinaryOpType bopt,
     bool donate_with_move = false) {
+  bool b_donatable = is_donatable(b, out);
+  bool a_donatable = is_donatable(a, out);
   switch (bopt) {
     case BinaryOpType::ScalarScalar:
       out.set_data(
           allocator::malloc_or_wait(out.itemsize()), 1, a.strides(), a.flags());
       break;
     case BinaryOpType::ScalarVector:
-      if (b.is_donatable() && b.itemsize() == out.itemsize()) {
+      if (b_donatable) {
         if (donate_with_move) {
           out.move_shared_buffer(b);
         } else {
@@ -62,7 +66,7 @@ void set_binary_op_output_data(
       }
       break;
     case BinaryOpType::VectorScalar:
-      if (a.is_donatable() && a.itemsize() == out.itemsize()) {
+      if (a_donatable) {
         if (donate_with_move) {
           out.move_shared_buffer(a);
         } else {
@@ -77,13 +81,13 @@ void set_binary_op_output_data(
       }
       break;
     case BinaryOpType::VectorVector:
-      if (a.is_donatable() && a.itemsize() == out.itemsize()) {
+      if (a_donatable) {
         if (donate_with_move) {
           out.move_shared_buffer(a);
         } else {
           out.copy_shared_buffer(a);
         }
-      } else if (b.is_donatable() && b.itemsize() == out.itemsize()) {
+      } else if (b_donatable) {
         if (donate_with_move) {
           out.move_shared_buffer(b);
         } else {
@@ -98,16 +102,14 @@ void set_binary_op_output_data(
       }
       break;
     case BinaryOpType::General:
-      if (a.is_donatable() && a.flags().row_contiguous &&
-          a.itemsize() == out.itemsize() && a.size() == out.size()) {
+      if (a_donatable && a.flags().row_contiguous && a.size() == out.size()) {
         if (donate_with_move) {
           out.move_shared_buffer(a);
         } else {
           out.copy_shared_buffer(a);
         }
       } else if (
-          b.is_donatable() && b.flags().row_contiguous &&
-          b.itemsize() == out.itemsize() && b.size() == out.size()) {
+          b_donatable && b.flags().row_contiguous && b.size() == out.size()) {
         if (donate_with_move) {
           out.move_shared_buffer(b);
         } else {

mlx/backend/common/cholesky.cpp

@@ -0,0 +1,101 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/copy.h"
+#include "mlx/linalg.h"
+#include "mlx/primitives.h"
+
+#ifdef ACCELERATE_NEW_LAPACK
+#include <Accelerate/Accelerate.h>
+#else
+#include <lapack.h>
+#endif
+
+namespace mlx::core {
+
+namespace {
+
+// Delegate to the Cholesky factorization taking into account differences in
+// LAPACK implementations (basically how to pass the 'uplo' string to fortran).
+int spotrf_wrapper(char uplo, float* matrix, int N) {
+  int info;
+
+#ifdef LAPACK_FORTRAN_STRLEN_END
+  spotrf_(
+      /* uplo = */ &uplo,
+      /* n = */ &N,
+      /* a = */ matrix,
+      /* lda = */ &N,
+      /* info = */ &info,
+      /* uplo_len = */ static_cast<size_t>(1));
+#else
+  spotrf_(
+      /* uplo = */ &uplo,
+      /* n = */ &N,
+      /* a = */ matrix,
+      /* lda = */ &N,
+      /* info = */ &info);
+#endif
+
+  return info;
+}
+
+} // namespace
+
+void cholesky_impl(const array& a, array& factor, bool upper) {
+  // Lapack uses the column-major convention. We take advantage of the fact that
+  // the matrix should be symmetric:
+  //   (A)ᵀ = A
+  // and that a column-major lower triangular matrix is a row-major upper
+  // triangular matrix, so uplo is the opposite of what we would expect from
+  // upper
+
+  char uplo = (upper) ? 'L' : 'U';
+
+  // The decomposition is computed in place, so just copy the input to the
+  // output.
+  copy(
+      a,
+      factor,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+
+  const int N = a.shape(-1);
+  const size_t num_matrices = a.size() / (N * N);
+
+  float* matrix = factor.data<float>();
+
+  for (int i = 0; i < num_matrices; i++) {
+    // Compute Cholesky factorization.
+    int info = spotrf_wrapper(uplo, matrix, N);
+
+    // TODO: We do nothing when the matrix is not positive semi-definite
+    // because throwing an error would result in a crash. If we figure out how
+    // to catch errors from the implementation we should throw.
+    if (info < 0) {
+      std::stringstream msg;
+      msg << "[cholesky] Cholesky decomposition failed with error code "
+          << info;
+      throw std::runtime_error(msg.str());
+    }
+
+    // Zero out the upper/lower triangle while advancing the pointer to the
+    // next matrix at the same time.
+    for (int row = 0; row < N; row++) {
+      if (upper) {
+        std::fill(matrix, matrix + row, 0);
+      } else {
+        std::fill(matrix + row + 1, matrix + N, 0);
+      }
+      matrix += N;
+    }
+  }
+}
+
+void Cholesky::eval(const std::vector<array>& inputs, array& output) {
+  if (inputs[0].dtype() != float32) {
+    throw std::runtime_error("[Cholesky::eval] only supports float32.");
+  }
+  cholesky_impl(inputs[0], output, upper_);
+}
+
+} // namespace mlx::core

mlx/backend/common/common.cpp

@@ -0,0 +1,304 @@
+// Copyright © 2024 Apple Inc.
+#include <cassert>
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+void AsStrided::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+
+  auto& in = inputs[0];
+
+  if (!in.flags().row_contiguous) {
+    // Just ensuring that inputs[0] came from the ops which would ensure the
+    // input is row contiguous.
+    throw std::runtime_error(
+        "AsStrided must be used with row contiguous arrays only.");
+  }
+
+  // Compute the flags given the shape and strides
+  bool row_contiguous = true, col_contiguous = true;
+  size_t r = 1, c = 1;
+  for (int i = strides_.size() - 1, j = 0; i >= 0; i--, j++) {
+    row_contiguous &= (r == strides_[i]) || (shape_[i] == 1);
+    col_contiguous &= (c == strides_[j]) || (shape_[j] == 1);
+    r *= shape_[i];
+    c *= shape_[j];
+  }
+  auto flags = in.flags();
+  // TODO: Compute the contiguous flag in a better way cause now we are
+  //       unnecessarily strict.
+  flags.contiguous = row_contiguous || col_contiguous;
+  flags.row_contiguous = row_contiguous;
+  flags.col_contiguous = col_contiguous;
+
+  // There is no easy way to compute the actual data size so we use out.size().
+  // The contiguous flag will almost certainly not be set so no code should
+  // rely on data_size anyway.
+  size_t data_size = out.size();
+
+  return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
+}
+
+void Broadcast::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+  std::vector<size_t> strides(out.ndim(), 0);
+  int diff = out.ndim() - in.ndim();
+  for (int i = in.ndim() - 1; i >= 0; --i) {
+    strides[i + diff] = (in.shape()[i] == 1) ? 0 : in.strides()[i];
+  }
+  auto flags = in.flags();
+  if (out.size() > in.size()) {
+    flags.row_contiguous = flags.col_contiguous = false;
+  }
+  out.copy_shared_buffer(in, strides, flags, in.data_size());
+}
+
+void Copy::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  out.copy_shared_buffer(inputs[0]);
+}
+
+void CustomTransforms::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() > outputs.size());
+  for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
+       i++, j++) {
+    outputs[i].copy_shared_buffer(inputs[j]);
+  }
+}
+
+void Depends::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() > outputs.size());
+  for (int i = 0; i < outputs.size(); i++) {
+    outputs[i].copy_shared_buffer(inputs[i]);
+  }
+}
+
+void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  double numel = 1;
+  for (auto ax : axes_) {
+    numel *= inputs[0].shape(ax);
+  }
+
+  if (inverted_) {
+    numel = 1.0 / numel;
+  }
+
+  switch (out.dtype()) {
+    case bool_:
+      *out.data<bool>() = static_cast<bool>(numel);
+      break;
+    case uint8:
+      *out.data<uint8_t>() = static_cast<uint8_t>(numel);
+      break;
+    case uint16:
+      *out.data<uint16_t>() = static_cast<uint16_t>(numel);
+      break;
+    case uint32:
+      *out.data<uint32_t>() = static_cast<uint32_t>(numel);
+      break;
+    case uint64:
+      *out.data<uint64_t>() = static_cast<uint64_t>(numel);
+      break;
+    case int8:
+      *out.data<int8_t>() = static_cast<int8_t>(numel);
+      break;
+    case int16:
+      *out.data<int16_t>() = static_cast<int16_t>(numel);
+      break;
+    case int32:
+      *out.data<int32_t>() = static_cast<int32_t>(numel);
+      break;
+    case int64:
+      *out.data<int64_t>() = static_cast<int64_t>(numel);
+      break;
+    case float16:
+      *out.data<float16_t>() = static_cast<float16_t>(numel);
+      break;
+    case float32:
+      *out.data<float>() = static_cast<float>(numel);
+      break;
+    case bfloat16:
+      *out.data<bfloat16_t>() = static_cast<bfloat16_t>(numel);
+      break;
+    case complex64:
+      *out.data<complex64_t>() = static_cast<complex64_t>(numel);
+      break;
+  }
+}
+
+std::pair<bool, std::vector<size_t>> Reshape::prepare_reshape(
+    const array& in,
+    const array& out) {
+  // Special case for empty arrays or row contiguous arrays
+  if (in.size() == 0 || in.flags().row_contiguous) {
+    return {false, out.strides()};
+  }
+
+  // Special case for scalars
+  if (in.ndim() == 0) {
+    std::vector<size_t> out_strides(out.ndim(), 0);
+    return {false, out_strides};
+  }
+
+  // Firstly let's collapse all the contiguous dimensions of the input
+  auto [shape, _strides] = collapse_contiguous_dims(in);
+  auto& strides = _strides[0];
+
+  // If shapes fit exactly in the contiguous dims then no copy is necessary so
+  // let's check.
+  std::vector<size_t> out_strides;
+  bool copy_necessary = false;
+  int j = 0;
+  for (int i = 0; i < out.ndim(); i++) {
+    int N = out.shape(i);
+    if (j < shape.size() && shape[j] % N == 0) {
+      shape[j] /= N;
+      out_strides.push_back(shape[j] * strides[j]);
+      j += (shape[j] == 1);
+    } else if (N == 1) {
+      // i > 0 because otherwise j < shape.size() && shape[j] % 1 == 0
+      out_strides.push_back(out_strides.back());
+    } else {
+      copy_necessary = true;
+      break;
+    }
+  }
+
+  return {copy_necessary, out_strides};
+}
+
+void Reshape::shared_buffer_reshape(
+    const array& in,
+    const std::vector<size_t>& out_strides,
+    array& out) {
+  auto flags = in.flags();
+  if (flags.row_contiguous) {
+    // For row contiguous reshapes:
+    // - Shallow copy the buffer
+    // - If reshaping into a vector (all singleton dimensions except one) it
+    //    becomes col contiguous again.
+    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
+    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
+  }
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
+}
+
+void Split::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+
+  auto& in = inputs[0];
+
+  auto compute_new_flags = [](const auto& shape,
+                              const auto& strides,
+                              size_t in_data_size,
+                              auto flags) {
+    size_t data_size = 1;
+    size_t f_stride = 1;
+    size_t b_stride = 1;
+    flags.row_contiguous = true;
+    flags.col_contiguous = true;
+    for (int i = 0, ri = shape.size() - 1; ri >= 0; i++, ri--) {
+      flags.col_contiguous &= strides[i] == f_stride || shape[i] == 1;
+      flags.row_contiguous &= strides[ri] == b_stride || shape[ri] == 1;
+      f_stride *= shape[i];
+      b_stride *= shape[ri];
+      if (strides[i] > 0) {
+        data_size *= shape[i];
+      }
+    }
+
+    if (data_size == 1) {
+      // Broadcasted scalar array is contiguous.
+      flags.contiguous = true;
+    } else if (data_size == in_data_size) {
+      // Means we sliced a broadcasted dimension so leave the "no holes" flag
+      // alone.
+    } else {
+      // We sliced something. So either we are row or col contiguous or we
+      // punched a hole.
+      flags.contiguous &= flags.row_contiguous || flags.col_contiguous;
+    }
+
+    return std::pair<decltype(flags), size_t>{flags, data_size};
+  };
+
+  std::vector<int> indices(1, 0);
+  indices.insert(indices.end(), indices_.begin(), indices_.end());
+  for (int i = 0; i < indices.size(); i++) {
+    size_t offset = indices[i] * in.strides()[axis_];
+    auto [new_flags, data_size] = compute_new_flags(
+        outputs[i].shape(), in.strides(), in.data_size(), in.flags());
+    outputs[i].copy_shared_buffer(
+        in, in.strides(), new_flags, data_size, offset);
+  }
+}
+
+std::tuple<int64_t, std::vector<int64_t>> SliceUpdate::prepare_slice(
+    const array& in) {
+  int64_t data_offset = 0;
+  std::vector<int64_t> inp_strides(in.ndim(), 0);
+  for (int i = 0; i < in.ndim(); ++i) {
+    data_offset += start_indices_[i] * in.strides()[i];
+    inp_strides[i] = in.strides()[i] * strides_[i];
+  }
+
+  return std::make_tuple(data_offset, inp_strides);
+}
+
+void StopGradient::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  out.copy_shared_buffer(inputs[0]);
+}
+
+void Transpose::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  std::vector<size_t> out_strides(out.ndim());
+  auto& in = inputs[0];
+  for (int ax = 0; ax < axes_.size(); ++ax) {
+    out_strides[ax] = in.strides()[axes_[ax]];
+  }
+
+  // Conditions for {row/col}_contiguous
+  // - array must be contiguous (no gaps)
+  // - underlying buffer size should have the same size as the array
+  // - cumulative product of shapes is equal to the strides (we can ignore axes
+  //   with size == 1)
+  //   - in the forward direction (column contiguous)
+  //   - in the reverse direction (row contiguous)
+  // - vectors are both row and col contiguous (hence if both row/col are
+  //   true, they stay true)
+  auto flags = in.flags();
+  if (flags.contiguous && in.data_size() == in.size()) {
+    size_t f_stride = 1;
+    size_t b_stride = 1;
+    flags.col_contiguous = true;
+    flags.row_contiguous = true;
+    for (int i = 0, ri = out.ndim() - 1; i < out.ndim(); ++i, --ri) {
+      flags.col_contiguous &= (out_strides[i] == f_stride || out.shape(i) == 1);
+      f_stride *= out.shape(i);
+      flags.row_contiguous &=
+          (out_strides[ri] == b_stride || out.shape(ri) == 1);
+      b_stride *= out.shape(ri);
+    }
+  }
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
+}
+
+} // namespace mlx::core

mlx/backend/common/compiled.cpp

@@ -18,7 +18,8 @@ void print_constant(std::ostream& os, const array& x) {
     case complex64:
       return print_complex_constant<complex64_t>(os, x);
     case int8:
-      return print_int_constant<int8_t>(os, x);
+      os << static_cast<int32_t>(x.item<int8_t>());
+      return;
     case int16:
       return print_int_constant<int16_t>(os, x);
     case int32:
@@ -26,7 +27,8 @@ void print_constant(std::ostream& os, const array& x) {
     case int64:
       return print_int_constant<int64_t>(os, x);
     case uint8:
-      return print_int_constant<uint8_t>(os, x);
+      os << static_cast<uint32_t>(x.item<uint8_t>());
+      return;
     case uint16:
       return print_int_constant<uint16_t>(os, x);
     case uint32:
@@ -126,4 +128,102 @@ std::string build_lib_name(
   return os.str();
 }
 
+bool compiled_check_contiguity(
+    const std::vector<array>& inputs,
+    const std::vector<int>& shape) {
+  bool contiguous = true;
+  bool all_contig = true;
+  bool all_row_contig = true;
+  bool all_col_contig = true;
+  int non_scalar_inputs = 0;
+  for (const auto& x : inputs) {
+    if (is_scalar(x)) {
+      continue;
+    }
+    non_scalar_inputs++;
+    bool shape_eq = x.shape() == shape;
+    all_contig &= (x.flags().contiguous && shape_eq);
+    all_row_contig &= (x.flags().row_contiguous && shape_eq);
+    all_col_contig &= (x.flags().col_contiguous && shape_eq);
+  }
+  if (non_scalar_inputs > 1 && !all_row_contig && !all_col_contig) {
+    contiguous = false;
+  } else if (non_scalar_inputs == 1 && !all_contig) {
+    contiguous = false;
+  } else if (non_scalar_inputs == 0 && !shape.empty()) {
+    contiguous = false;
+  }
+  return contiguous;
+}
+
+void compiled_allocate_outputs(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::vector<array>& inputs_,
+    const std::unordered_set<uintptr_t>& constant_ids_,
+    bool contiguous,
+    bool move_buffers /* = false */) {
+  if (contiguous) {
+    int o = 0;
+    std::vector<size_t> strides;
+    size_t data_size;
+    array::Flags flags;
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+      auto& in = inputs[i];
+      // Conditions for donation
+      // - Correct size
+      // - Not a scalar
+      // - Donatable
+      // - Not a constant
+      if (in.itemsize() == outputs[o].itemsize() && !is_scalar(in) &&
+          in.is_donatable() &&
+          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+        if (move_buffers) {
+          outputs[o++].move_shared_buffer(in);
+        } else {
+          outputs[o++].copy_shared_buffer(in);
+        }
+      }
+      // Get representative input flags to properly set non-donated outputs
+      if (strides.empty() && in.size() == outputs[0].size()) {
+        strides = in.strides();
+        flags = in.flags();
+        data_size = in.data_size();
+      }
+    }
+    for (; o < outputs.size(); ++o) {
+      outputs[o].set_data(
+          allocator::malloc_or_wait(data_size * outputs[o].itemsize()),
+          data_size,
+          strides,
+          flags);
+    }
+  } else {
+    int o = 0;
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+      auto& in = inputs[i];
+      // Conditions for donation
+      // - Row contiguous
+      // - Donatable
+      // - Correct size
+      // - Not a constant
+      if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
+          in.itemsize() == outputs[o].itemsize() && in.is_donatable() &&
+          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+        if (move_buffers) {
+          outputs[o].move_shared_buffer(
+              in, outputs[o].strides(), in.flags(), in.data_size());
+        } else {
+          outputs[o].copy_shared_buffer(
+              in, outputs[o].strides(), in.flags(), in.data_size());
+        }
+        o++;
+      }
+    }
+    for (; o < outputs.size(); ++o) {
+      outputs[o].set_data(allocator::malloc_or_wait(outputs[o].nbytes()));
+    }
+  }
+}
+
 } // namespace mlx::core

mlx/backend/common/compiled.h

@@ -53,4 +53,18 @@ inline bool is_scalar(const array& x) {
   return x.ndim() == 0;
 }
 
+// Check if we can use a contiguous operation given inputs and the output shape
+bool compiled_check_contiguity(
+    const std::vector<array>& inputs,
+    const std::vector<int>& shape);
+
+// Allocate space for the outputs possibly with input donation
+void compiled_allocate_outputs(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::vector<array>& inputs_,
+    const std::unordered_set<uintptr_t>& constant_ids_,
+    bool contiguous,
+    bool move_buffers = false);
+
 } // namespace mlx::core

mlx/backend/common/compiled_cpu.cpp

@@ -2,6 +2,7 @@
 
 #include <dlfcn.h>
 #include <filesystem>
+#include <fstream>
 #include <list>
 
 #include "mlx/backend/common/compiled.h"
@@ -52,8 +53,25 @@ void* compile(
     return nullptr;
   }
 
+  std::string kernel_file_name;
+
+  // Deal with long kernel names. Maximum length for files on macOS is 255
+  // characters. Clip file name with a little extra room and append a 16
+  // character hash.
+  constexpr int max_file_name_length = 245;
+  if (kernel_name.size() > max_file_name_length) {
+    std::ostringstream file_name;
+    file_name
+        << std::string_view(kernel_name).substr(0, max_file_name_length - 16);
+    auto file_id = std::hash<std::string>{}(kernel_name);
+    file_name << "_" << std::hex << std::setw(16) << file_id << std::dec;
+    kernel_file_name = file_name.str();
+  } else {
+    kernel_file_name = kernel_name;
+  }
+
   std::ostringstream shared_lib_name;
-  shared_lib_name << "lib" << kernel_name << ".so";
+  shared_lib_name << "lib" << kernel_file_name << ".so";
   auto shared_lib_path = get_temp_file(shared_lib_name.str());
   bool lib_exists = false;
   {
@@ -64,7 +82,7 @@ void* compile(
   if (!lib_exists) {
     // Open source file and write source code to it
     std::ostringstream source_file_name;
-    source_file_name << kernel_name << ".cpp";
+    source_file_name << kernel_file_name << ".cpp";
     auto source_file_path = get_temp_file(source_file_name.str());
 
     std::ofstream source_file(source_file_path);
@@ -248,28 +266,7 @@ void Compiled::eval_cpu(
 
   // Figure out which kernel we are using
   auto& shape = outputs[0].shape();
-  bool contiguous = true;
-  {
-    bool all_contig = true;
-    bool all_row_contig = true;
-    bool all_col_contig = true;
-    int non_scalar_inputs = 0;
-    for (auto& x : inputs) {
-      if (is_scalar(x)) {
-        continue;
-      }
-      non_scalar_inputs++;
-      bool shape_eq = x.shape() == shape;
-      all_contig &= (x.flags().contiguous && shape_eq);
-      all_row_contig &= (x.flags().row_contiguous && shape_eq);
-      all_col_contig &= (x.flags().col_contiguous && shape_eq);
-    }
-    if (non_scalar_inputs > 1 && !all_row_contig && !all_col_contig) {
-      contiguous = false;
-    } else if (non_scalar_inputs == 1 && !all_contig) {
-      contiguous = false;
-    }
-  }
+  bool contiguous = compiled_check_contiguity(inputs, shape);
 
   // Handle all broadcasting and collect function input arguments
   std::vector<void*> args;
@@ -342,58 +339,8 @@ void Compiled::eval_cpu(
     fn_ptr = compile(kernel_name, kernel.str());
   }
 
-  // Allocate space for the outputs possibly with input donation
-  if (contiguous) {
-    int o = 0;
-    std::vector<size_t> strides;
-    size_t data_size;
-    array::Flags flags;
-    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
-      auto& in = inputs[i];
-      // Conditions for donation
-      // - Contiguous
-      // - Donatable
-      // - Correct size
-      // - Not a constant
-      if (in.flags().contiguous && !is_scalar(in) && in.is_donatable() &&
-          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        outputs[o++].copy_shared_buffer(in);
-      }
-      // Get representative input flags to properly set non-donated outputs
-      if (strides.empty() && in.size() == outputs[0].size()) {
-        strides = in.strides();
-        flags = in.flags();
-        data_size = in.data_size();
-      }
-    }
-    for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(
-          allocator::malloc_or_wait(data_size * outputs[o].itemsize()),
-          data_size,
-          strides,
-          flags);
-    }
-  } else {
-    int o = 0;
-    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
-      auto& in = inputs[i];
-      // Conditions for donation
-      // - Row contiguous
-      // - Donatable
-      // - Correct size
-      // - Not a constant
-      if (in.flags().row_contiguous && in.nbytes() == outputs[o].nbytes() &&
-          in.is_donatable() &&
-          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        outputs[o].copy_shared_buffer(
-            in, outputs[o].strides(), in.flags(), in.data_size());
-        o++;
-      }
-    }
-    for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(allocator::malloc_or_wait(outputs[o].nbytes()));
-    }
-  }
+  compiled_allocate_outputs(
+      inputs, outputs, inputs_, constant_ids_, contiguous, false);
 
   for (auto& x : outputs) {
     args.push_back(x.data<void>());

mlx/backend/common/conv.cpp

@@ -38,11 +38,15 @@ void slow_conv_1D(
 
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
   const int iH = 1 + in_dilation[0] * (in.shape(1) - 1); // Input spatial dim
+  const int C = in.shape(2); // Input channels
   const int oH = out.shape(1); // Output spatial dim
   const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(2); // In channels
   const int wH = wt.shape(1); // Weight spatial dim
 
+  const int groups = C / wt.shape(2);
+  const int C_per_group = wt.shape(2);
+  const int O_per_group = O / groups;
+
   const size_t in_stride_N = in.strides()[0];
   const size_t in_stride_H = in.strides()[1];
   const size_t in_stride_C = in.strides()[2];
@@ -57,35 +61,36 @@ void slow_conv_1D(
 
   for (int n = 0; n < N; ++n) {
     for (int oh = 0; oh < oH; ++oh) {
-      for (int o = 0; o < O; ++o) {
-        const T* filter_wt_ptr = start_wt_ptr + o * wt_stride_O;
-        float r = 0.;
+      for (int g = 0; g < groups; ++g) {
+        for (int o = g * O_per_group; o < (g + 1) * O_per_group; ++o) {
+          const T* filter_wt_ptr = start_wt_ptr + o * wt_stride_O;
+          float r = 0.;
 
-        for (int wh = 0; wh < wH; ++wh) {
-          const T* wt_ptr = filter_wt_ptr + wh * wt_stride_H;
+          for (int wh = 0; wh < wH; ++wh) {
+            const T* wt_ptr = filter_wt_ptr + wh * wt_stride_H;
 
-          int wh_flip = flip ? (wH - wh - 1) : wh;
-          int ih = oh * wt_strides[0] - padding[0] + wh_flip * wt_dilation[0];
+            int wh_flip = flip ? (wH - wh - 1) : wh;
+            int ih = oh * wt_strides[0] - padding[0] + wh_flip * wt_dilation[0];
 
-          auto ih_div = std::div(ih, in_dilation[0]);
+            auto ih_div = std::div(ih, in_dilation[0]);
 
-          if (ih >= 0 && ih < iH && ih_div.rem == 0) {
-            for (int c = 0; c < C; ++c) {
-              r += static_cast<float>(
-                       in_ptr[ih_div.quot * in_stride_H + c * in_stride_C]) *
-                  static_cast<float>(wt_ptr[c * wt_stride_C]);
-            } // c
+            if (ih >= 0 && ih < iH && ih_div.rem == 0) {
+              for (int c = g * C_per_group; c < (g + 1) * C_per_group; ++c) {
+                r += static_cast<float>(
+                         in_ptr[ih_div.quot * in_stride_H + c * in_stride_C]) *
+                    static_cast<float>(wt_ptr[(c % C_per_group) * wt_stride_C]);
+              } // c
 
-          } // ih check
-        } // wh
+            } // ih check
+          } // wh
 
-        out_ptr[oh * out_stride_H + o * out_stride_O] = static_cast<T>(r);
-      } // o
+          out_ptr[oh * out_stride_H + o * out_stride_O] = static_cast<T>(r);
+        } // o
+      } // g
     } // oh
 
     in_ptr += in_stride_N;
     out_ptr += out_stride_N;
-
   } // n
 }
 
@@ -106,13 +111,17 @@ void slow_conv_2D(
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
   const int iH = 1 + in_dilation[0] * (in.shape(1) - 1); // Input spatial dim
   const int iW = 1 + in_dilation[1] * (in.shape(2) - 1); // Input spatial dim
+  const int C = in.shape(3); // In channels
   const int oH = out.shape(1); // Output spatial dim
   const int oW = out.shape(2); // Output spatial dim
   const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(3); // In channels
   const int wH = wt.shape(1); // Weight spatial dim
   const int wW = wt.shape(2); // Weight spatial dim
 
+  const int groups = C / wt.shape(3);
+  const int C_per_group = wt.shape(3);
+  const int O_per_group = O / groups;
+
   const size_t in_stride_N = in.strides()[0];
   const size_t in_stride_H = in.strides()[1];
   const size_t in_stride_W = in.strides()[2];
@@ -136,33 +145,35 @@ void slow_conv_2D(
         int ih_base = oh * wt_strides[0] - padding[0];
         int iw_base = ow * wt_strides[1] - padding[1];
 
-        for (int o = 0; o < O; ++o) {
-          float r = 0.;
+        for (int g = 0; g < groups; ++g) {
+          for (int o = g * O_per_group; o < (g + 1) * O_per_group; ++o) {
+            float r = 0.;
 
-          for (int wh = 0; wh < wH; ++wh) {
-            for (int ww = 0; ww < wW; ++ww) {
-              int wh_flip = flip ? wH - wh - 1 : wh;
-              int ww_flip = flip ? wW - ww - 1 : ww;
-              int ih = ih_base + wh_flip * wt_dilation[0];
-              int iw = iw_base + ww_flip * wt_dilation[1];
+            for (int wh = 0; wh < wH; ++wh) {
+              for (int ww = 0; ww < wW; ++ww) {
+                int wh_flip = flip ? wH - wh - 1 : wh;
+                int ww_flip = flip ? wW - ww - 1 : ww;
+                int ih = ih_base + wh_flip * wt_dilation[0];
+                int iw = iw_base + ww_flip * wt_dilation[1];
 
-              const T* wt_ptr_pt = wt_ptr + wh * wt_stride_H + ww * wt_stride_W;
-              const T* in_ptr_pt = in_ptr + ih * in_stride_H + iw * in_stride_W;
+                const T* wt_ptr_pt =
+                    wt_ptr + wh * wt_stride_H + ww * wt_stride_W;
+                const T* in_ptr_pt =
+                    in_ptr + ih * in_stride_H + iw * in_stride_W;
 
-              for (int c = 0; c < C; ++c) {
-                r += static_cast<float>(in_ptr_pt[0]) *
-                    static_cast<float>(wt_ptr_pt[0]);
-                in_ptr_pt += in_stride_C;
-                wt_ptr_pt += wt_stride_C;
-              } // c
+                for (int c = g * C_per_group; c < (g + 1) * C_per_group; ++c) {
+                  r += static_cast<float>(in_ptr_pt[c * in_stride_C]) *
+                      static_cast<float>(
+                           wt_ptr_pt[(c % C_per_group) * wt_stride_C]);
+                } // c
+              } // ww
+            } // wh
 
-            } // ww
-          } // wh
-
-          out_ptr[0] = static_cast<T>(r);
-          out_ptr += out_stride_O;
-          wt_ptr += wt_stride_O;
-        } // o
+            out_ptr[0] = static_cast<T>(r);
+            out_ptr += out_stride_O;
+            wt_ptr += wt_stride_O;
+          } // o
+        } // g
       };
 
   int jump_h = flip ? -wt_dilation[0] : wt_dilation[0];
@@ -214,41 +225,43 @@ void slow_conv_2D(
         int wh_base = base_h[oh % f_out_jump_h];
         int ww_base = base_w[ow % f_out_jump_w];
 
-        for (int o = 0; o < O; ++o) {
-          float r = 0.;
+        for (int g = 0; g < groups; ++g) {
+          for (int o = g * O_per_group; o < (g + 1) * O_per_group; ++o) {
+            float r = 0.;
 
-          for (int wh = wh_base; wh < wH; wh += f_wgt_jump_h) {
-            for (int ww = ww_base; ww < wW; ww += f_wgt_jump_w) {
-              int wh_flip = flip ? wH - wh - 1 : wh;
-              int ww_flip = flip ? wW - ww - 1 : ww;
-              int ih = ih_base + wh_flip * wt_dilation[0];
-              int iw = iw_base + ww_flip * wt_dilation[1];
+            for (int wh = wh_base; wh < wH; wh += f_wgt_jump_h) {
+              for (int ww = ww_base; ww < wW; ww += f_wgt_jump_w) {
+                int wh_flip = flip ? wH - wh - 1 : wh;
+                int ww_flip = flip ? wW - ww - 1 : ww;
+                int ih = ih_base + wh_flip * wt_dilation[0];
+                int iw = iw_base + ww_flip * wt_dilation[1];
 
-              if (ih >= 0 && ih < iH && iw >= 0 && iw < iW) {
-                const T* wt_ptr_pt =
-                    wt_ptr + wh * wt_stride_H + ww * wt_stride_W;
+                if (ih >= 0 && ih < iH && iw >= 0 && iw < iW) {
+                  const T* wt_ptr_pt =
+                      wt_ptr + wh * wt_stride_H + ww * wt_stride_W;
 
-                int ih_dil = !is_idil_one ? (ih / in_dilation[0]) : ih;
-                int iw_dil = !is_idil_one ? (iw / in_dilation[1]) : iw;
+                  int ih_dil = !is_idil_one ? (ih / in_dilation[0]) : ih;
+                  int iw_dil = !is_idil_one ? (iw / in_dilation[1]) : iw;
 
-                const T* in_ptr_pt =
-                    in_ptr + ih_dil * in_stride_H + iw_dil * in_stride_W;
+                  const T* in_ptr_pt =
+                      in_ptr + ih_dil * in_stride_H + iw_dil * in_stride_W;
 
-                for (int c = 0; c < C; ++c) {
-                  r += static_cast<float>(in_ptr_pt[0]) *
-                      static_cast<float>(wt_ptr_pt[0]);
-                  in_ptr_pt += in_stride_C;
-                  wt_ptr_pt += wt_stride_C;
-                } // c
+                  for (int c = g * C_per_group; c < (g + 1) * C_per_group;
+                       ++c) {
+                    r += static_cast<float>(in_ptr_pt[c * in_stride_C]) *
+                        static_cast<float>(
+                             wt_ptr_pt[(c % C_per_group) * wt_stride_C]);
+                  } // c
 
-              } // ih, iw check
-            } // ww
-          } // wh
+                } // ih, iw check
+              } // ww
+            } // wh
 
-          out_ptr[0] = static_cast<T>(r);
-          out_ptr += out_stride_O;
-          wt_ptr += wt_stride_O;
-        } // o
+            out_ptr[0] = static_cast<T>(r);
+            out_ptr += out_stride_O;
+            wt_ptr += wt_stride_O;
+          } // o
+        } // g
       };
 
   int oH_border_0 = 0;
@@ -305,6 +318,296 @@ void slow_conv_2D(
   } // n
 }
 
+template <typename T>
+void slow_conv_3D(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    const std::vector<int>& in_dilation,
+    bool flip) {
+  const T* st_wt_ptr = wt.data<T>();
+  const T* st_in_ptr = in.data<T>();
+  T* st_out_ptr = out.data<T>();
+
+  const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
+  const int iD = 1 + in_dilation[0] * (in.shape(1) - 1); // Input spatial dim
+  const int iH = 1 + in_dilation[1] * (in.shape(2) - 1); // Input spatial dim
+  const int iW = 1 + in_dilation[2] * (in.shape(3) - 1); // Input spatial dim
+  const int oD = out.shape(1); // Output spatial dim
+  const int oH = out.shape(2); // Output spatial dim
+  const int oW = out.shape(3); // Output spatial dim
+  const int O = wt.shape(0); // Out channels
+  const int C = wt.shape(4); // In channels
+  const int wD = wt.shape(1); // Weight spatial dim
+  const int wH = wt.shape(2); // Weight spatial dim
+  const int wW = wt.shape(3); // Weight spatial dim
+
+  const size_t in_stride_N = in.strides()[0];
+  const size_t in_stride_D = in.strides()[1];
+  const size_t in_stride_H = in.strides()[2];
+  const size_t in_stride_W = in.strides()[3];
+  const size_t in_stride_C = in.strides()[4];
+
+  const size_t wt_stride_O = wt.strides()[0];
+  const size_t wt_stride_D = wt.strides()[1];
+  const size_t wt_stride_H = wt.strides()[2];
+  const size_t wt_stride_W = wt.strides()[3];
+  const size_t wt_stride_C = wt.strides()[4];
+
+  const size_t out_stride_N = out.strides()[0];
+  const size_t out_stride_D = out.strides()[1];
+  const size_t out_stride_H = out.strides()[2];
+  const size_t out_stride_W = out.strides()[3];
+  const size_t out_stride_O = out.strides()[4];
+
+  bool is_idil_one =
+      in_dilation[0] == 1 && in_dilation[1] == 1 && in_dilation[2] == 1;
+
+  auto pt_conv_no_checks = [&](const T* in_ptr,
+                               const T* wt_ptr,
+                               T* out_ptr,
+                               int od,
+                               int oh,
+                               int ow) {
+    out_ptr += od * out_stride_D + oh * out_stride_H + ow * out_stride_W;
+    int id_base = od * wt_strides[0] - padding[0];
+    int ih_base = oh * wt_strides[1] - padding[1];
+    int iw_base = ow * wt_strides[2] - padding[2];
+
+    for (int o = 0; o < O; ++o) {
+      float r = 0.;
+
+      for (int wd = 0; wd < wD; ++wd) {
+        for (int wh = 0; wh < wH; ++wh) {
+          for (int ww = 0; ww < wW; ++ww) {
+            int wd_flip = flip ? wD - wd - 1 : wd;
+            int wh_flip = flip ? wH - wh - 1 : wh;
+            int ww_flip = flip ? wW - ww - 1 : ww;
+            int id = id_base + wd_flip * wt_dilation[0];
+            int ih = ih_base + wh_flip * wt_dilation[1];
+            int iw = iw_base + ww_flip * wt_dilation[2];
+
+            const T* wt_ptr_pt =
+                wt_ptr + wd * wt_stride_D + wh * wt_stride_H + ww * wt_stride_W;
+            const T* in_ptr_pt =
+                in_ptr + id * in_stride_D + ih * in_stride_H + iw * in_stride_W;
+
+            for (int c = 0; c < C; ++c) {
+              r += static_cast<float>(in_ptr_pt[0]) *
+                  static_cast<float>(wt_ptr_pt[0]);
+              in_ptr_pt += in_stride_C;
+              wt_ptr_pt += wt_stride_C;
+            } // c
+
+          } // ww
+        } // wh
+      } // wd
+
+      out_ptr[0] = static_cast<T>(r);
+      out_ptr += out_stride_O;
+      wt_ptr += wt_stride_O;
+    } // o
+  };
+
+  int jump_d = flip ? -wt_dilation[0] : wt_dilation[0];
+  int jump_h = flip ? -wt_dilation[1] : wt_dilation[1];
+  int jump_w = flip ? -wt_dilation[2] : wt_dilation[2];
+
+  int init_d = (flip ? (wD - 1) * wt_dilation[0] : 0);
+  int init_h = (flip ? (wH - 1) * wt_dilation[1] : 0);
+  int init_w = (flip ? (wW - 1) * wt_dilation[2] : 0);
+
+  int f_wgt_jump_d = std::lcm(in_dilation[0], wt_dilation[0]) / wt_dilation[0];
+  int f_wgt_jump_h = std::lcm(in_dilation[1], wt_dilation[1]) / wt_dilation[1];
+  int f_wgt_jump_w = std::lcm(in_dilation[2], wt_dilation[2]) / wt_dilation[2];
+
+  int f_out_jump_d = std::lcm(in_dilation[0], wt_strides[0]) / wt_strides[0];
+  int f_out_jump_h = std::lcm(in_dilation[1], wt_strides[1]) / wt_strides[1];
+  int f_out_jump_w = std::lcm(in_dilation[2], wt_strides[2]) / wt_strides[2];
+
+  std::vector<int> base_d(f_out_jump_d);
+  std::vector<int> base_h(f_out_jump_h);
+  std::vector<int> base_w(f_out_jump_w);
+
+  for (int i = 0; i < f_out_jump_d; ++i) {
+    int id_loop = i * wt_strides[0] - padding[0] + init_d;
+
+    int wd_base = 0;
+    while (wd_base < wD && id_loop % in_dilation[0] != 0) {
+      wd_base++;
+      id_loop += jump_d;
+    }
+
+    base_d[i] = wd_base;
+  }
+
+  for (int i = 0; i < f_out_jump_h; ++i) {
+    int ih_loop = i * wt_strides[1] - padding[1] + init_h;
+
+    int wh_base = 0;
+    while (wh_base < wH && ih_loop % in_dilation[1] != 0) {
+      wh_base++;
+      ih_loop += jump_h;
+    }
+
+    base_h[i] = wh_base;
+  }
+
+  for (int j = 0; j < f_out_jump_w; ++j) {
+    int iw_loop = j * wt_strides[2] - padding[2] + init_w;
+
+    int ww_base = 0;
+    while (ww_base < wW && iw_loop % in_dilation[2] != 0) {
+      ww_base++;
+      iw_loop += jump_w;
+    }
+
+    base_w[j] = ww_base;
+  }
+
+  auto pt_conv_all_checks = [&](const T* in_ptr,
+                                const T* wt_ptr,
+                                T* out_ptr,
+                                int od,
+                                int oh,
+                                int ow) {
+    out_ptr += od * out_stride_D + oh * out_stride_H + ow * out_stride_W;
+
+    int id_base = od * wt_strides[0] - padding[0];
+    int ih_base = oh * wt_strides[1] - padding[1];
+    int iw_base = ow * wt_strides[2] - padding[2];
+
+    int wd_base = base_d[od % f_out_jump_d];
+    int wh_base = base_h[oh % f_out_jump_h];
+    int ww_base = base_w[ow % f_out_jump_w];
+
+    for (int o = 0; o < O; ++o) {
+      float r = 0.;
+
+      for (int wd = wd_base; wd < wD; wd += f_wgt_jump_d) {
+        for (int wh = wh_base; wh < wH; wh += f_wgt_jump_h) {
+          for (int ww = ww_base; ww < wW; ww += f_wgt_jump_w) {
+            int wd_flip = flip ? wD - wd - 1 : wd;
+            int wh_flip = flip ? wH - wh - 1 : wh;
+            int ww_flip = flip ? wW - ww - 1 : ww;
+            int id = id_base + wd_flip * wt_dilation[0];
+            int ih = ih_base + wh_flip * wt_dilation[1];
+            int iw = iw_base + ww_flip * wt_dilation[2];
+
+            if (id >= 0 && id < iD && ih >= 0 && ih < iH && iw >= 0 &&
+                iw < iW) {
+              const T* wt_ptr_pt = wt_ptr + wd * wt_stride_D +
+                  wh * wt_stride_H + ww * wt_stride_W;
+
+              int id_dil = !is_idil_one ? (id / in_dilation[0]) : id;
+              int ih_dil = !is_idil_one ? (ih / in_dilation[1]) : ih;
+              int iw_dil = !is_idil_one ? (iw / in_dilation[2]) : iw;
+
+              const T* in_ptr_pt = in_ptr + id_dil * in_stride_D +
+                  ih_dil * in_stride_H + iw_dil * in_stride_W;
+
+              for (int c = 0; c < C; ++c) {
+                r += static_cast<float>(in_ptr_pt[0]) *
+                    static_cast<float>(wt_ptr_pt[0]);
+                in_ptr_pt += in_stride_C;
+                wt_ptr_pt += wt_stride_C;
+              } // c
+
+            } // iD, ih, iw check
+          } // ww
+        } // wh
+      } // wd
+
+      out_ptr[0] = static_cast<T>(r);
+      out_ptr += out_stride_O;
+      wt_ptr += wt_stride_O;
+    } // o
+  };
+
+  int oD_border_0 = 0;
+  int oD_border_1 =
+      is_idil_one ? ((padding[0] + wt_strides[0] - 1) / wt_strides[0]) : oD;
+  int oD_border_2 = std::max(
+      oD_border_1, (iD + padding[0] - wD * wt_dilation[0]) / wt_strides[0]);
+  int oD_border_3 = oD;
+
+  int oH_border_0 = 0;
+  int oH_border_1 =
+      is_idil_one ? ((padding[1] + wt_strides[1] - 1) / wt_strides[1]) : oH;
+  int oH_border_2 = std::max(
+      oH_border_1, (iH + padding[1] - wH * wt_dilation[1]) / wt_strides[1]);
+  int oH_border_3 = oH;
+
+  int oW_border_0 = 0;
+  int oW_border_1 =
+      is_idil_one ? ((padding[2] + wt_strides[2] - 1) / wt_strides[2]) : oW;
+  int oW_border_2 = std::max(
+      oW_border_1, (iW + padding[2] - wW * wt_dilation[2]) / wt_strides[2]);
+  int oW_border_3 = oW;
+
+  for (int n = 0; n < N; ++n) {
+    // Case 1: od might put us out of bounds
+    for (int od = oD_border_0; od < oD_border_1; ++od) {
+      for (int oh = 0; oh < oH; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+    } // od
+
+    // Case 2: od in bounds
+    for (int od = oD_border_1; od < oD_border_2; ++od) {
+      // Case 2.1: oh might put us out of bounds
+      for (int oh = oH_border_0; oh < oH_border_1; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+
+      // Case 2.2: oh in bounds
+      for (int oh = oH_border_1; oh < oH_border_2; ++oh) {
+        // Case 2.2.1: ow might put us out of bounds
+        for (int ow = oW_border_0; ow < oW_border_1; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+
+        // Case 2.2.2: ow in bounds
+        for (int ow = oW_border_1; ow < oW_border_2; ++ow) {
+          pt_conv_no_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+
+        // Case 2.2.3: ow might put us out of bounds
+        for (int ow = oW_border_2; ow < oW_border_3; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+
+      // Case 2.3: oh might put us out of bounds
+      for (int oh = oH_border_2; oh < oH_border_3; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+    } // od
+
+    // Case 3: od might put us out of bounds
+    for (int od = oD_border_2; od < oD_border_3; ++od) {
+      for (int oh = 0; oh < oH; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+    } // od
+
+    st_in_ptr += in_stride_N;
+    st_out_ptr += out_stride_N;
+
+  } // n
+}
+
 void dispatch_slow_conv_1D(
     const array& in,
     const array& wt,
@@ -353,6 +656,30 @@ void dispatch_slow_conv_2D(
   }
 }
 
+void dispatch_slow_conv_3D(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    const std::vector<int>& in_dilation,
+    bool flip) {
+  if (in.dtype() == float32) {
+    return slow_conv_3D<float>(
+        in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+  } else if (in.dtype() == float16) {
+    return slow_conv_3D<float16_t>(
+        in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+  } else if (in.dtype() == bfloat16) {
+    return slow_conv_3D<bfloat16_t>(
+        in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+  } else {
+    throw std::invalid_argument(
+        "[Convolution::eval] got unsupported data type.");
+  }
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Explicit gemm conv
 ///////////////////////////////////////////////////////////////////////////////
@@ -366,11 +693,15 @@ void explicit_gemm_conv_1D_cpu(
     const std::vector<int>& wt_dilation) {
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
   const int iH = in.shape(1); // Input spatial dim
+  const int C = in.shape(2); // Input channels
   const int oH = out.shape(1); // Output spatial dim
   const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(2); // In channels
   const int wH = wt.shape(1); // Weight spatial dim
 
+  const int groups = C / wt.shape(2);
+  const int C_per_group = wt.shape(2);
+  const int O_per_group = O / groups;
+
   auto conv_dtype = float32;
 
   // Pad input
@@ -402,6 +733,11 @@ void explicit_gemm_conv_1D_cpu(
       in_padded.strides()[1],
       in_padded.strides()[2]};
   auto flags = in_padded.flags();
+  if (groups > 1) {
+    // Transpose the last two dimensions for grouped convolutions
+    std::swap(strided_shape[2], strided_shape[3]);
+    std::swap(strided_strides[2], strided_strides[3]);
+  }
 
   array in_strided_view(strided_shape, in_padded.dtype(), nullptr, {});
   in_strided_view.copy_shared_buffer(
@@ -416,7 +752,19 @@ void explicit_gemm_conv_1D_cpu(
   auto gemm_wt = wt;
   auto gemm_out = out;
 
-  if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
+  if (groups > 1) {
+    // Transpose the last two dimensions for grouped convolutions
+    array wt_transpose(
+        {wt.shape(0), wt.shape(2), wt.shape(1)}, wt.dtype(), nullptr, {});
+    wt_transpose.copy_shared_buffer(
+        wt,
+        {wt.strides(0), wt.strides(2), wt.strides(1)},
+        wt.flags(),
+        wt.size(),
+        0);
+    gemm_wt = array(wt_transpose.shape(), float32, nullptr, {});
+    copy(wt_transpose, gemm_wt, CopyType::General);
+  } else if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
@@ -428,27 +776,29 @@ void explicit_gemm_conv_1D_cpu(
     gemm_out.set_data(allocator::malloc_or_wait(gemm_out.nbytes()));
   }
 
-  // Perform gemm
-  cblas_sgemm(
-      CblasRowMajor,
-      CblasNoTrans, // no trans A
-      CblasTrans, // transB
-      strided_reshape[0], // M
-      O, // N
-      strided_reshape[1], // K
-      1.0f, // alpha
-      in_strided.data<float>(),
-      strided_reshape[1], // lda
-      gemm_wt.data<float>(),
-      strided_reshape[1], // ldb
-      0.0f, // beta
-      gemm_out.data<float>(),
-      O // ldc
-  );
+  for (int g = 0; g < groups; ++g) {
+    // Perform gemm
+    cblas_sgemm(
+        CblasRowMajor,
+        CblasNoTrans, // no trans A
+        CblasTrans, // transB
+        strided_reshape[0], // M
+        O_per_group, // N
+        C_per_group * wH, // K
+        1.0f, // alpha
+        in_strided.data<float>() + g * C_per_group * wH, // A
+        wH * C, // lda
+        gemm_wt.data<float>() + g * O_per_group * C_per_group * wH, // B
+        wH * C_per_group, // ldb
+        0.0f, // beta
+        gemm_out.data<float>() + g * O_per_group, // C
+        O // ldc
+    );
 
-  // Copy results if needed
-  if (out.dtype() != float32) {
-    copy(gemm_out, out, CopyType::Vector);
+    // Copy results if needed
+    if (out.dtype() != float32) {
+      copy(gemm_out, out, CopyType::Vector);
+    }
   }
 }
 
@@ -554,6 +904,131 @@ void explicit_gemm_conv_2D_cpu(
   }
 }
 
+void explicit_gemm_conv_ND_cpu(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation) {
+  const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
+  const auto iDim = std::vector<int>(
+      in.shape().begin() + 1, in.shape().end() - 1); // Input spatial dim
+  const auto oDim = std::vector<int>(
+      out.shape().begin() + 1, out.shape().end() - 1); // Output spatial dim
+  const int O = wt.shape(0); // Out channels
+  const int C = wt.shape(-1); // In channels
+  const auto wDim = std::vector<int>(
+      wt.shape().begin() + 1, wt.shape().end() - 1); // Weight spatial dim
+
+  auto conv_dtype = float32;
+
+  // Pad input
+  std::vector<int> padded_shape(in.shape().size());
+  padded_shape.front() = N;
+  for (size_t i = 0; i < iDim.size(); i++) {
+    padded_shape[i + 1] = iDim[i] + 2 * padding[i];
+  }
+  padded_shape.back() = C;
+  array in_padded(padded_shape, conv_dtype, nullptr, {});
+
+  // Fill with zeros
+  copy(array(0, conv_dtype), in_padded, CopyType::Scalar);
+
+  // Pick input slice from padded
+  size_t data_offset = 0;
+  for (size_t i = 0; i < padding.size(); i++) {
+    data_offset += padding[i] * in_padded.strides()[i + 1];
+  }
+  array in_padded_slice(in.shape(), in_padded.dtype(), nullptr, {});
+  in_padded_slice.copy_shared_buffer(
+      in_padded,
+      in_padded.strides(),
+      in_padded.flags(),
+      in_padded_slice.size(),
+      data_offset);
+
+  // Copy input values into the slice
+  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral);
+
+  // Make strided view
+  std::vector<int> strided_shape(oDim.size() + wDim.size() + 2);
+  strided_shape.front() = N;
+  for (size_t i = 0; i < oDim.size(); i++) {
+    strided_shape[i + 1] = oDim[i];
+  }
+  for (size_t i = 0; i < wDim.size(); i++) {
+    strided_shape[i + 1 + oDim.size()] = wDim[i];
+  }
+  strided_shape.back() = C;
+
+  std::vector<size_t> strided_strides(in.shape().size() * 2 - 2);
+  strided_strides[0] = in_padded.strides()[0];
+  for (size_t i = 0; i < wt_strides.size(); i++) {
+    strided_strides[i + 1] = in_padded.strides()[i + 1] * wt_strides[i];
+  }
+  for (size_t i = 1; i < in_padded.strides().size(); i++) {
+    strided_strides[i + wt_strides.size()] = in_padded.strides()[i];
+  }
+
+  auto flags = in_padded.flags();
+
+  array in_strided_view(strided_shape, in_padded.dtype(), nullptr, {});
+  in_strided_view.copy_shared_buffer(
+      in_padded, strided_strides, flags, in_strided_view.size(), 0);
+
+  // Materialize strided view
+  std::vector<int> strided_reshape = {N, C};
+  for (const auto& o : oDim) {
+    strided_reshape[0] *= o;
+  }
+  for (const auto& w : wDim) {
+    strided_reshape[1] *= w;
+  }
+
+  array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
+  copy(in_strided_view, in_strided, CopyType::General);
+
+  // Check wt dtype and prepare
+  auto gemm_wt = wt;
+  auto gemm_out = out;
+
+  if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
+    auto ctype =
+        wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
+    gemm_wt = array(wt.shape(), float32, nullptr, {});
+    copy(wt, gemm_wt, ctype);
+  }
+
+  if (out.dtype() != float32) {
+    gemm_out = array(out.shape(), float32, nullptr, {});
+    gemm_out.set_data(allocator::malloc_or_wait(gemm_out.nbytes()));
+  }
+
+  // Perform gemm
+  cblas_sgemm(
+      CblasRowMajor,
+      CblasNoTrans, // no trans A
+      CblasTrans, // transB
+      strided_reshape[0], // M
+      O, // N
+      strided_reshape[1], // K
+      1.0f, // alpha
+      in_strided.data<float>(),
+      strided_reshape[1], // lda
+      gemm_wt.data<float>(),
+      strided_reshape[1], // ldb
+      0.0f, // beta
+      gemm_out.data<float>(),
+      O // ldc
+  );
+
+  // Copy results if needed
+  if (out.dtype() != float32) {
+    copy(gemm_out, out, CopyType::Vector);
+  }
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Conv routing
 ///////////////////////////////////////////////////////////////////////////////
@@ -589,6 +1064,19 @@ void conv_2D_cpu(
       in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
 }
 
+void conv_3D_cpu(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    const std::vector<int>& in_dilation,
+    bool flip) {
+  return dispatch_slow_conv_3D(
+      in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+}
+
 } // namespace
 
 void Convolution::eval(const std::vector<array>& inputs, array& out) {
@@ -597,8 +1085,20 @@ void Convolution::eval(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
   auto& wt = inputs[1];
 
+  // 3D convolution
+  if (in.ndim() == (3 + 2)) {
+    return conv_3D_cpu(
+        in,
+        wt,
+        out,
+        padding_,
+        kernel_strides_,
+        kernel_dilation_,
+        input_dilation_,
+        flip_);
+  }
   // 2D convolution
-  if (in.ndim() == (2 + 2)) {
+  else if (in.ndim() == (2 + 2)) {
     return conv_2D_cpu(
         in,
         wt,
@@ -625,7 +1125,7 @@ void Convolution::eval(const std::vector<array>& inputs, array& out) {
   else {
     std::ostringstream msg;
     msg << "[Convolution::eval] Convolution currently only supports"
-        << " 1D and 2D convolutions. Got inputs with " << in.ndim() - 2
+        << " 1D, 2D and 3D convolutions. Got inputs with " << in.ndim() - 2
         << " spatial dimensions";
     throw std::invalid_argument(msg.str());
   }

mlx/backend/common/copy.cpp

@@ -4,6 +4,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/copy.h"
+#include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
@@ -142,29 +143,31 @@ void copy_general(
     const std::vector<int>& data_shape,
     const std::vector<stride_t>& i_strides,
     int64_t i_offset) {
-  switch (src.ndim()) {
+  auto [new_shape, new_strides] = collapse_contiguous_dims(
+      data_shape, std::vector<std::vector<stride_t>>{i_strides});
+  switch (new_shape.size()) {
     case 1:
       copy_general_dim1<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
     case 2:
       copy_general_dim2<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
     case 3:
       copy_general_dim3<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
     case 4:
       copy_general_dim4<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
   }
 
   auto src_ptr = src.data<SrcT>() + i_offset;
   auto dst_ptr = dst.data<DstT>();
   for (size_t i = 0; i < dst.size(); ++i) {
-    stride_t src_elem = elem_to_loc(i, data_shape, i_strides);
+    stride_t src_elem = elem_to_loc(i, new_shape, new_strides[0]);
     dst_ptr[i] = static_cast<DstT>(src_ptr[src_elem]);
   }
 }
@@ -195,10 +198,10 @@ inline void copy_general_general_dims(
     const std::vector<int>& data_shape,
     const std::vector<stride_t>& i_strides,
     const std::vector<stride_t>& o_strides,
-    stride_t i_offset,
-    stride_t o_offset) {
+    int64_t i_offset,
+    int64_t o_offset) {
   if constexpr (D > 1) {
-    int axis = src.ndim() - D;
+    int axis = data_shape.size() - D;
     auto stride_src = i_strides[axis];
     auto stride_dst = o_strides[axis];
     auto N = data_shape[axis];
@@ -209,7 +212,7 @@ inline void copy_general_general_dims(
       o_offset += stride_dst;
     }
   } else {
-    int axis = src.ndim() - 1;
+    int axis = data_shape.size() - 1;
     auto stride_src = i_strides[axis];
     auto stride_dst = o_strides[axis];
     auto N = data_shape[axis];
@@ -230,38 +233,76 @@ void copy_general_general(
     const std::vector<int>& data_shape,
     const std::vector<stride_t>& i_strides,
     const std::vector<stride_t>& o_strides,
-    stride_t i_offset,
-    stride_t o_offset) {
-  switch (src.ndim()) {
+    int64_t i_offset,
+    int64_t o_offset) {
+  auto [new_shape, new_strides] = collapse_contiguous_dims(
+      data_shape, std::vector<std::vector<stride_t>>{i_strides, o_strides});
+  switch (new_shape.size()) {
     case 1:
       copy_general_general_dims<SrcT, DstT, stride_t, 1>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 2:
       copy_general_general_dims<SrcT, DstT, stride_t, 2>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 3:
       copy_general_general_dims<SrcT, DstT, stride_t, 3>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 4:
       copy_general_general_dims<SrcT, DstT, stride_t, 4>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 5:
       copy_general_general_dims<SrcT, DstT, stride_t, 5>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
   }
 
   int size = std::accumulate(
-      data_shape.begin() - 5, data_shape.end(), 1, std::multiplies<int>());
+      new_shape.end() - 5, new_shape.end(), 1, std::multiplies<int>());
   for (int i = 0; i < src.size(); i += size) {
-    stride_t src_offset = i_offset + elem_to_loc(i, data_shape, i_strides);
-    stride_t dst_offset = o_offset + elem_to_loc(i, dst.shape(), o_strides);
+    stride_t src_offset = i_offset + elem_to_loc(i, new_shape, new_strides[0]);
+    stride_t dst_offset = o_offset + elem_to_loc(i, new_shape, new_strides[1]);
     copy_general_general_dims<SrcT, DstT, stride_t, 5>(
-        src, dst, data_shape, i_strides, o_strides, src_offset, dst_offset);
+        src,
+        dst,
+        new_shape,
+        new_strides[0],
+        new_strides[1],
+        src_offset,
+        dst_offset);
   }
 }
 
@@ -272,7 +313,7 @@ inline void copy_general_general(const array& src, array& dst) {
 }
 
 template <typename SrcT, typename DstT, typename... Args>
-void copy(const array& src, array& dst, CopyType ctype, Args... args) {
+void copy(const array& src, array& dst, CopyType ctype, Args&&... args) {
   switch (ctype) {
     case CopyType::Scalar:
       copy_single<SrcT, DstT>(src, dst);
@@ -281,54 +322,54 @@ void copy(const array& src, array& dst, CopyType ctype, Args... args) {
       copy_vector<SrcT, DstT>(src, dst);
       return;
     case CopyType::General:
-      copy_general<SrcT, DstT>(src, dst, args...);
+      copy_general<SrcT, DstT>(src, dst, std::forward<Args>(args)...);
       return;
     case CopyType::GeneralGeneral:
-      copy_general_general<SrcT, DstT>(src, dst, args...);
+      copy_general_general<SrcT, DstT>(src, dst, std::forward<Args>(args)...);
   }
 }
 
 template <typename SrcT, typename... Args>
-void copy(const array& src, array& dst, CopyType ctype, Args... args) {
+void copy(const array& src, array& dst, CopyType ctype, Args&&... args) {
   switch (dst.dtype()) {
     case bool_:
-      copy<SrcT, bool>(src, dst, ctype, args...);
+      copy<SrcT, bool>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint8:
-      copy<SrcT, uint8_t>(src, dst, ctype, args...);
+      copy<SrcT, uint8_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint16:
-      copy<SrcT, uint16_t>(src, dst, ctype, args...);
+      copy<SrcT, uint16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint32:
-      copy<SrcT, uint32_t>(src, dst, ctype, args...);
+      copy<SrcT, uint32_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint64:
-      copy<SrcT, uint64_t>(src, dst, ctype, args...);
+      copy<SrcT, uint64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int8:
-      copy<SrcT, int8_t>(src, dst, ctype, args...);
+      copy<SrcT, int8_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int16:
-      copy<SrcT, int16_t>(src, dst, ctype, args...);
+      copy<SrcT, int16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int32:
-      copy<SrcT, int32_t>(src, dst, ctype, args...);
+      copy<SrcT, int32_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int64:
-      copy<SrcT, int64_t>(src, dst, ctype, args...);
+      copy<SrcT, int64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case float16:
-      copy<SrcT, float16_t>(src, dst, ctype, args...);
+      copy<SrcT, float16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case float32:
-      copy<SrcT, float>(src, dst, ctype, args...);
+      copy<SrcT, float>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case bfloat16:
-      copy<SrcT, bfloat16_t>(src, dst, ctype, args...);
+      copy<SrcT, bfloat16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case complex64:
-      copy<SrcT, complex64_t>(src, dst, ctype, args...);
+      copy<SrcT, complex64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
   }
 }
@@ -338,46 +379,46 @@ inline void copy_inplace_dispatch(
     const array& src,
     array& dst,
     CopyType ctype,
-    Args... args) {
+    Args&&... args) {
   switch (src.dtype()) {
     case bool_:
-      copy<bool>(src, dst, ctype, args...);
+      copy<bool>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint8:
-      copy<uint8_t>(src, dst, ctype, args...);
+      copy<uint8_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint16:
-      copy<uint16_t>(src, dst, ctype, args...);
+      copy<uint16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint32:
-      copy<uint32_t>(src, dst, ctype, args...);
+      copy<uint32_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint64:
-      copy<uint64_t>(src, dst, ctype, args...);
+      copy<uint64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int8:
-      copy<int8_t>(src, dst, ctype, args...);
+      copy<int8_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int16:
-      copy<int16_t>(src, dst, ctype, args...);
+      copy<int16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int32:
-      copy<int32_t>(src, dst, ctype, args...);
+      copy<int32_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int64:
-      copy<int64_t>(src, dst, ctype, args...);
+      copy<int64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case float16:
-      copy<float16_t>(src, dst, ctype, args...);
+      copy<float16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case float32:
-      copy<float>(src, dst, ctype, args...);
+      copy<float>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case bfloat16:
-      copy<bfloat16_t>(src, dst, ctype, args...);
+      copy<bfloat16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case complex64:
-      copy<complex64_t>(src, dst, ctype, args...);
+      copy<complex64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
   }
 }
@@ -444,8 +485,17 @@ void copy_inplace(
   }
 }
 
-template <>
-void copy_inplace<int64_t>(
+template void copy_inplace<size_t>(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<size_t>& i_strides,
+    const std::vector<size_t>& o_strides,
+    int64_t i_offset,
+    int64_t o_offset,
+    CopyType ctype);
+
+template void copy_inplace<int64_t>(
     const array& src,
     array& dst,
     const std::vector<int>& data_shape,
@@ -453,24 +503,6 @@ void copy_inplace<int64_t>(
     const std::vector<int64_t>& o_strides,
     int64_t i_offset,
     int64_t o_offset,
-    CopyType ctype) {
-  switch (ctype) {
-    case CopyType::General:
-    case CopyType::GeneralGeneral:
-      return copy_inplace_dispatch(
-          src,
-          dst,
-          ctype,
-          data_shape,
-          i_strides,
-          o_strides,
-          i_offset,
-          o_offset);
-
-    case CopyType::Scalar:
-    case CopyType::Vector:
-      return copy_inplace_dispatch(src, dst, ctype);
-  }
-}
+    CopyType ctype);
 
 } // namespace mlx::core

mlx/backend/common/default_primitives.cpp

@@ -5,7 +5,6 @@
 #else
 #include <cblas.h>
 #endif
-
 #include <cstring>
 
 #include "mlx/array.h"
@@ -34,6 +33,7 @@ DEFAULT(ArcCosh)
 DEFAULT(ArcSin)
 DEFAULT(ArcSinh)
 DEFAULT(ArcTan)
+DEFAULT(ArcTan2)
 DEFAULT(ArcTanh)
 DEFAULT(ArgPartition)
 DEFAULT(ArgReduce)
@@ -41,14 +41,18 @@ DEFAULT(ArgSort)
 DEFAULT(AsType)
 DEFAULT(AsStrided)
 DEFAULT(Broadcast)
+DEFAULT(BlockMaskedMM)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT_MULTI(DivMod)
 DEFAULT(Ceil)
 DEFAULT(Concatenate)
+DEFAULT(Conjugate)
 DEFAULT(Convolution)
 DEFAULT(Copy)
 DEFAULT(Cos)
 DEFAULT(Cosh)
-DEFAULT_MULTI(CustomVJP)
+DEFAULT_MULTI(CustomTransforms)
 DEFAULT_MULTI(Depends)
 DEFAULT(Divide)
 DEFAULT(NumberOfElements)
@@ -57,12 +61,14 @@ DEFAULT(Equal)
 DEFAULT(Erf)
 DEFAULT(ErfInv)
 DEFAULT(Exp)
+DEFAULT(Expm1)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Full)
 DEFAULT(Gather)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
 DEFAULT(Less)
 DEFAULT(LessEqual)
 DEFAULT(Load)
@@ -107,6 +113,7 @@ DEFAULT(Tan)
 DEFAULT(Tanh)
 DEFAULT(Transpose)
 DEFAULT(Inverse)
+DEFAULT(Cholesky)
 
 namespace {
 

mlx/backend/common/hadamard.cpp

@@ -0,0 +1,107 @@
+// Copyright © 2024 Apple Inc.
+
+#include <cassert>
+
+#include "mlx/backend/common/copy.h"
+#include "mlx/backend/common/hadamard.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+// n = 2^k component
+template <typename T>
+void hadamard_n(array& out, int n, int m, float scale) {
+  for (int b = 0; b < out.size() / n; b++) {
+    size_t loc = b * n;
+    T* data_ptr = out.data<T>() + loc;
+    int h = 1;
+    int n_over_2 = n / 2;
+    while (h < n) {
+      for (int i = 0; i < n / 2; i++) {
+        int k = i & (h - 1);
+        int j = ((i - k) << 1) + k;
+        float x = *(data_ptr + j);
+        float y = *(data_ptr + j + h);
+        *(data_ptr + j) = x + y;
+        *(data_ptr + j + h) = x - y;
+        if (h == n_over_2) {
+          *(data_ptr + j) *= scale;
+          *(data_ptr + j + h) *= scale;
+        }
+      }
+      h <<= 1;
+    }
+  }
+}
+
+// m component
+template <typename T>
+void hadamard_m(array& out, int n, int m, float scale) {
+  auto h_matrices = hadamard_matrices();
+  auto& matrix = h_matrices[m];
+  auto start = 1;
+  auto end = matrix.find('\n', start);
+  std::vector<bool> hmat_vec;
+  while (end != std::string_view::npos) {
+    auto row = matrix.substr(start, end - start);
+    for (int i = 0; i < row.length(); i++) {
+      hmat_vec.push_back(row[i] == '+');
+    }
+    start = end + 1;
+    end = matrix.find('\n', start);
+  }
+
+  for (int b = 0; b < out.size() / m / n; b++) {
+    size_t loc = b * n * m;
+    T* data_ptr = out.data<T>() + loc;
+    for (int i = 0; i < n; i++) {
+      std::vector<float> out(m);
+      for (int j = 0; j < m; j++) {
+        for (int k = 0; k < m; k++) {
+          float x = *(data_ptr + i + k * n);
+          if (hmat_vec[k + j * m]) {
+            out[j] += x;
+          } else {
+            out[j] -= x;
+          }
+        }
+      }
+      for (int j = 0; j < m; j++) {
+        *(data_ptr + i + j * n) = out[j] * scale;
+      }
+    }
+  }
+}
+
+template <typename T>
+void hadamard(array& out, int n, int m, float scale) {
+  float n_scale = m > 1 ? 1.0 : scale;
+  hadamard_n<T>(out, n, m, n_scale);
+  if (m > 1) {
+    hadamard_m<T>(out, n, m, scale);
+  }
+}
+
+void Hadamard::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+
+  // Copy input to output
+  copy(in, out, CopyType::General);
+
+  int axis = out.ndim() - 1;
+  auto [n, m] = decompose_hadamard(out.shape(axis));
+
+  switch (in.dtype()) {
+    case float32:
+      return hadamard<float>(out, n, m, scale_);
+    case float16:
+      return hadamard<float16_t>(out, n, m, scale_);
+    case bfloat16:
+      return hadamard<bfloat16_t>(out, n, m, scale_);
+    default:
+      throw std::invalid_argument("[hadamard] Unsupported type.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/hadamard.h

@@ -0,0 +1,105 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#include <map>
+
+#include "mlx/utils.h"
+
+namespace mlx::core {
+
+// From http://neilsloane.com/hadamard/
+constexpr std::string_view h12 = R"(
++-++++++++++
+--+-+-+-+-+-
++++-++----++
++---+--+-++-
++++++-++----
++-+---+--+-+
+++--+++-++--
++--++---+--+
+++----+++-++
++--+-++---+-
+++++----+++-
++-+--+-++---
+)";
+
+constexpr std::string_view h20 = R"(
++----+----++--++-++-
+-+----+---+++---+-++
+--+----+---+++-+-+-+
+---+----+---+++++-+-
+----+----++--++-++-+
+-+++++-----+--+++--+
++-+++-+---+-+--+++--
+++-++--+---+-+--+++-
++++-+---+---+-+--+++
+++++-----++--+-+--++
+--++-+-++-+-----++++
+---++-+-++-+---+-+++
++---++-+-+--+--++-++
+++---++-+----+-+++-+
+-++---++-+----+++++-
+-+--+--++-+----+----
++-+-----++-+----+---
+-+-+-+---+--+----+--
+--+-+++------+----+-
++--+--++------+----+
+)";
+
+constexpr std::string_view h28 = R"(
++------++----++-+--+-+--++--
+-+-----+++-----+-+--+-+--++-
+--+-----+++---+-+-+----+--++
+---+-----+++---+-+-+-+--+--+
+----+-----+++---+-+-+++--+--
+-----+-----++++--+-+--++--+-
+------++----++-+--+-+--++--+
+--++++-+-------++--+++-+--+-
+---++++-+-----+-++--+-+-+--+
++---+++--+----++-++--+-+-+--
+++---++---+----++-++--+-+-+-
++++---+----+----++-++--+-+-+
+++++--------+-+--++-++--+-+-
+-++++--------+++--++--+--+-+
+-+-++-++--++--+--------++++-
++-+-++--+--++--+--------++++
+-+-+-++--+--++--+----+---+++
++-+-+-++--+--+---+---++---++
+++-+-+-++--+------+--+++---+
+-++-+-+-++--+------+-++++---
++-++-+---++--+------+-++++--
+-++--++-+-++-+++----++------
++-++--++-+-++-+++-----+-----
+++-++---+-+-++-+++-----+----
+-++-++-+-+-+-+--+++-----+---
+--++-++++-+-+----+++-----+--
++--++-+-++-+-+----+++-----+-
+++--++-+-++-+-+----++------+
+)";
+
+inline const std::map<int, std::string_view> hadamard_matrices() {
+  return {{12, h12}, {20, h20}, {28, h28}};
+}
+
+inline std::pair<int, int> decompose_hadamard(int n) {
+  // n = m*2^k
+  int m = 1;
+  if (!is_power_of_2(n)) {
+    auto h_matrices = hadamard_matrices();
+    for (auto [factor, _] : h_matrices) {
+      if (n % factor == 0) {
+        m = factor;
+        n /= factor;
+        break;
+      }
+    }
+    if (m == 1) {
+      throw std::invalid_argument(
+          "[hadamard] Only supports n = m*2^k where m in (1, 12, 20, 28).");
+    }
+  }
+  return {n, m};
+}
+
+} // namespace mlx::core

mlx/backend/common/inverse.cpp

@@ -2,7 +2,6 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/copy.h"
-#include "mlx/linalg.h"
 #include "mlx/primitives.h"
 
 #ifdef ACCELERATE_NEW_LAPACK
@@ -11,9 +10,106 @@
 #include <lapack.h>
 #endif
 
+// Wrapper to account for differences in
+// LAPACK implementations (basically how to pass the 'uplo' string to fortran).
+int strtri_wrapper(char uplo, char diag, float* matrix, int N) {
+  int info;
+
+#ifdef LAPACK_FORTRAN_STRLEN_END
+  strtri_(
+      /* uplo = */ &uplo,
+      /* diag = */ &diag,
+      /* N = */ &N,
+      /* a = */ matrix,
+      /* lda = */ &N,
+      /* info = */ &info,
+      /* uplo_len = */ static_cast<size_t>(1),
+      /* diag_len = */ static_cast<size_t>(1));
+#else
+  strtri_(
+      /* uplo = */ &uplo,
+      /* diag = */ &diag,
+      /* N = */ &N,
+      /* a = */ matrix,
+      /* lda = */ &N,
+      /* info = */ &info);
+#endif
+
+  return info;
+}
+
 namespace mlx::core {
 
-void inverse_impl(const array& a, array& inv) {
+void general_inv(array& inv, int N, int i) {
+  int info;
+  auto ipiv = array::Data{allocator::malloc_or_wait(sizeof(int) * N)};
+  // Compute LU factorization.
+  sgetrf_(
+      /* m = */ &N,
+      /* n = */ &N,
+      /* a = */ inv.data<float>() + N * N * i,
+      /* lda = */ &N,
+      /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
+      /* info = */ &info);
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "inverse_impl: LU factorization failed with error code " << info;
+    throw std::runtime_error(ss.str());
+  }
+
+  static const int lwork_query = -1;
+  float workspace_size = 0;
+
+  // Compute workspace size.
+  sgetri_(
+      /* m = */ &N,
+      /* a = */ nullptr,
+      /* lda = */ &N,
+      /* ipiv = */ nullptr,
+      /* work = */ &workspace_size,
+      /* lwork = */ &lwork_query,
+      /* info = */ &info);
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "inverse_impl: LU workspace calculation failed with error code "
+       << info;
+    throw std::runtime_error(ss.str());
+  }
+
+  const int lwork = workspace_size;
+  auto scratch = array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
+
+  // Compute inverse.
+  sgetri_(
+      /* m = */ &N,
+      /* a = */ inv.data<float>() + N * N * i,
+      /* lda = */ &N,
+      /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
+      /* work = */ static_cast<float*>(scratch.buffer.raw_ptr()),
+      /* lwork = */ &lwork,
+      /* info = */ &info);
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "inverse_impl: inversion failed with error code " << info;
+    throw std::runtime_error(ss.str());
+  }
+}
+
+void tri_inv(array& inv, int N, int i, bool upper) {
+  const char uplo = upper ? 'L' : 'U';
+  const char diag = 'N';
+  int info = strtri_wrapper(uplo, diag, inv.data<float>() + N * N * i, N);
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "inverse_impl: triangular inversion failed with error code " << info;
+    throw std::runtime_error(ss.str());
+  }
+}
+
+void inverse_impl(const array& a, array& inv, bool tri, bool upper) {
   // Lapack uses the column-major convention. We take advantage of the following
   // identity to avoid transposing (see
   // https://math.stackexchange.com/a/340234):
@@ -25,63 +121,11 @@ void inverse_impl(const array& a, array& inv) {
   const int N = a.shape(-1);
   const size_t num_matrices = a.size() / (N * N);
 
-  int info;
-  auto ipiv = array::Data{allocator::malloc_or_wait(sizeof(int) * N)};
-
   for (int i = 0; i < num_matrices; i++) {
-    // Compute LU factorization.
-    sgetrf_(
-        /* m = */ &N,
-        /* n = */ &N,
-        /* a = */ inv.data<float>() + N * N * i,
-        /* lda = */ &N,
-        /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "inverse_impl: LU factorization failed with error code " << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    static const int lwork_query = -1;
-    float workspace_size = 0;
-
-    // Compute workspace size.
-    sgetri_(
-        /* m = */ &N,
-        /* a = */ nullptr,
-        /* lda = */ &N,
-        /* ipiv = */ nullptr,
-        /* work = */ &workspace_size,
-        /* lwork = */ &lwork_query,
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "inverse_impl: LU workspace calculation failed with error code "
-         << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    const int lwork = workspace_size;
-    auto scratch =
-        array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
-
-    // Compute inverse.
-    sgetri_(
-        /* m = */ &N,
-        /* a = */ inv.data<float>() + N * N * i,
-        /* lda = */ &N,
-        /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
-        /* work = */ static_cast<float*>(scratch.buffer.raw_ptr()),
-        /* lwork = */ &lwork,
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "inverse_impl: inversion failed with error code " << info;
-      throw std::runtime_error(ss.str());
+    if (tri) {
+      tri_inv(inv, N, i, upper);
+    } else {
+      general_inv(inv, N, i);
     }
   }
 }
@@ -90,15 +134,7 @@ void Inverse::eval(const std::vector<array>& inputs, array& output) {
   if (inputs[0].dtype() != float32) {
     throw std::runtime_error("[Inverse::eval] only supports float32.");
   }
-  inverse_impl(inputs[0], output);
-}
-
-std::pair<std::vector<array>, std::vector<int>> Inverse::vmap(
-    const std::vector<array>& inputs,
-    const std::vector<int>& axes) {
-  auto ax = axes[0] >= 0 ? 0 : -1;
-  auto a = axes[0] > 0 ? moveaxis(inputs[0], axes[0], 0, stream()) : inputs[0];
-  return {{linalg::inv(a, stream())}, {ax}};
+  inverse_impl(inputs[0], output, tri_, upper_);
 }
 
 } // namespace mlx::core

mlx/backend/common/load.cpp

@@ -5,11 +5,9 @@
 #include <utility>
 
 #include "mlx/allocator.h"
-#include "mlx/io/load.h"
+#include "mlx/backend/common/load.h"
 #include "mlx/primitives.h"
 
-namespace mlx::core {
-
 namespace {
 
 template <const uint8_t scalar_size>
@@ -29,12 +27,14 @@ void swap_endianness(uint8_t* data_bytes, size_t N) {
 
 } // namespace
 
-void Load::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 0);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+namespace mlx::core {
 
-  reader_->seek(offset_, std::ios_base::beg);
-  reader_->read(out.data<char>(), out.nbytes());
+void load(
+    array& out,
+    size_t offset,
+    const std::shared_ptr<io::Reader>& reader,
+    bool swap_endianness_) {
+  reader->read(out.data<char>(), out.nbytes(), offset);
 
   if (swap_endianness_) {
     switch (out.itemsize()) {
@@ -51,4 +51,11 @@ void Load::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void Load::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 0);
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  load(out, offset_, reader_, swap_endianness_);
+}
+
 } // namespace mlx::core

mlx/backend/common/load.h

@@ -0,0 +1,14 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/array.h"
+#include "mlx/io/load.h"
+
+namespace mlx::core {
+
+void load(
+    array& out,
+    size_t offset,
+    const std::shared_ptr<io::Reader>& reader,
+    bool swap_endianess);
+
+} // namespace mlx::core

mlx/backend/common/make_compiled_preamble.sh

@@ -11,7 +11,7 @@ GCC=$2
 SRCDIR=$3
 CLANG=$4
 
-if [ $CLANG = "TRUE" ]; then
+if [ "$CLANG" = "TRUE" ]; then
   read -r -d '' INCLUDES <<- EOM
   #include <cmath>
   #include <complex>
@@ -21,13 +21,14 @@ EOM
 
 fi
 
-CONTENT=$($GCC -I $SRCDIR -E $SRCDIR/mlx/backend/common/compiled_preamble.h 2>/dev/null)
+CONTENT=$($GCC -I "$SRCDIR" -E "$SRCDIR/mlx/backend/common/compiled_preamble.h" 2>/dev/null)
 
 cat << EOF > "$OUTPUT_FILE"
 const char* get_kernel_preamble() {
 return R"preamble(
 $INCLUDES
 $CONTENT
+using namespace mlx::core;
 using namespace mlx::core::detail;
 )preamble";
 }

mlx/backend/common/masked_mm.cpp

@@ -0,0 +1,305 @@
+// Copyright © 2024 Apple Inc.
+
+#ifdef ACCELERATE_NEW_LAPACK
+#include <Accelerate/Accelerate.h>
+#else
+#include <cblas.h>
+#endif
+
+#include <cstring>
+
+#include "mlx/array.h"
+#include "mlx/backend/common/copy.h"
+#include "mlx/backend/common/utils.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+namespace {
+
+template <typename T, typename mask_t>
+inline void mask_matrix(
+    T* data,
+    const mask_t* mask,
+    int block_size,
+    const int X,
+    const int Y,
+    const size_t X_data_str,
+    const size_t Y_data_str,
+    const size_t X_mask_str,
+    const size_t Y_mask_str,
+    const size_t mask_offset) {
+  int tX = (X + block_size - 1) / block_size;
+  int tY = (Y + block_size - 1) / block_size;
+
+  for (int i = 0; i < tX; i++) {
+    for (int j = 0; j < tY; j++) {
+      mask_t do_mask = mask[mask_offset + i * X_mask_str + j * Y_mask_str];
+      if (do_mask != 1) {
+        int loc_x = i * block_size;
+        int loc_y = j * block_size;
+        T* data_block = data + loc_x * X_data_str + loc_y * Y_data_str;
+
+        int size_x = std::min(block_size, X - loc_x);
+        int size_y = std::min(block_size, Y - loc_y);
+        for (int ii = 0; ii < size_x; ii++) {
+          for (int jj = 0; jj < size_y; jj++) {
+            if constexpr (std::is_same_v<mask_t, bool>) {
+              data_block[ii * X_data_str + jj * Y_data_str] = T(0.);
+            } else {
+              data_block[ii * X_data_str + jj * Y_data_str] *= do_mask;
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+} // namespace
+
+void BlockMaskedMM::eval(const std::vector<array>& inputs, array& out) {
+  if (out.dtype() != float32) {
+    throw std::runtime_error(
+        "[BlockMaskedMM::eval] Currently only supports float32.");
+  }
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  auto& a_pre = inputs[0];
+  auto& b_pre = inputs[1];
+
+  auto check_transpose =
+      [](const array& arr, bool do_copy, bool expand_all = false) {
+        auto stx = arr.strides()[arr.ndim() - 2];
+        auto sty = arr.strides()[arr.ndim() - 1];
+        if (!expand_all && stx == arr.shape(-1) && sty == 1) {
+          if (do_copy) {
+            array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+            copy(arr, arr_copy, CopyType::Vector);
+            return std::make_tuple(false, stx, arr_copy);
+          }
+          return std::make_tuple(false, stx, arr);
+        } else if (!expand_all && stx == 1 && sty == arr.shape(-2)) {
+          if (do_copy) {
+            array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+            copy(arr, arr_copy, CopyType::Vector);
+            return std::make_tuple(true, sty, arr_copy);
+          }
+          return std::make_tuple(true, sty, arr);
+        } else {
+          array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+          copy(arr, arr_copy, CopyType::General);
+          size_t stx = arr.shape(-1);
+          return std::make_tuple(false, stx, arr_copy);
+        }
+      };
+
+  bool has_op_mask = inputs.size() > 3;
+  bool has_out_mask = inputs.size() == 3 || inputs.size() == 5;
+  auto [a_transposed, lda, a] =
+      check_transpose(a_pre, has_op_mask, inputs.back().dtype() != bool_);
+  auto [b_transposed, ldb, b] =
+      check_transpose(b_pre, has_op_mask, inputs.back().dtype() != bool_);
+
+  size_t M = a.shape(-2);
+  size_t N = b.shape(-1);
+  size_t K = a.shape(-1);
+
+  if (M == 0 || N == 0) {
+    return;
+  }
+
+  if (K == 0) {
+    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
+    return;
+  }
+
+  auto mask_array = [](const array& mask,
+                       float* data,
+                       int block_size,
+                       int batch_idx,
+                       int X,
+                       int Y,
+                       size_t X_data_str,
+                       size_t Y_data_str) {
+    size_t mask_offset = elem_to_loc(
+        mask.shape(-1) * mask.shape(-2) * batch_idx,
+        mask.shape(),
+        mask.strides());
+
+    size_t X_mask_str = mask.strides()[mask.ndim() - 2];
+    size_t Y_mask_str = mask.strides()[mask.ndim() - 1];
+
+    if (mask.dtype() == bool_) {
+      return mask_matrix(
+          data,
+          mask.data<bool>(),
+          block_size,
+          X,
+          Y,
+          X_data_str,
+          Y_data_str,
+          X_mask_str,
+          Y_mask_str,
+          mask_offset);
+    } else {
+      return mask_matrix(
+          data,
+          mask.data<float>(),
+          block_size,
+          X,
+          Y,
+          X_data_str,
+          Y_data_str,
+          X_mask_str,
+          Y_mask_str,
+          mask_offset);
+    }
+  };
+
+  for (int i = 0; i < (out.size() / (M * size_t(N))); ++i) {
+    // Adjust pointer
+    float* ai =
+        a.data<float>() + elem_to_loc(M * K * i, a.shape(), a.strides());
+    float* bi =
+        b.data<float>() + elem_to_loc(K * N * i, b.shape(), b.strides());
+    float* ci = out.data<float>() + M * N * i;
+
+    // Zero out blocks in a and b if needed
+    if (has_op_mask) {
+      auto& a_mask = inputs[inputs.size() - 2];
+      mask_array(
+          a_mask,
+          ai,
+          block_size_,
+          i,
+          M,
+          K,
+          a_transposed ? 1 : lda,
+          a_transposed ? lda : 1);
+
+      auto& b_mask = inputs[inputs.size() - 1];
+      mask_array(
+          b_mask,
+          bi,
+          block_size_,
+          i,
+          K,
+          N,
+          b_transposed ? 1 : ldb,
+          b_transposed ? ldb : 1);
+    }
+
+    // Do matmul
+    cblas_sgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        1.0, // alpha
+        ai,
+        lda,
+        bi,
+        ldb,
+        0.0, // beta
+        ci,
+        out.shape(-1) // ldc
+    );
+
+    // Zero out blocks in out
+    if (has_out_mask) {
+      mask_array(inputs[2], ci, block_size_, i, M, N, N, 1);
+    }
+  }
+}
+
+void GatherMM::eval(const std::vector<array>& inputs, array& out) {
+  if (out.dtype() != float32) {
+    throw std::runtime_error(
+        "[GatherMM::eval] Currently only supports float32.");
+  }
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  auto& a_pre = inputs[0];
+  auto& b_pre = inputs[1];
+
+  auto check_transpose = [](const array& arr) {
+    auto stx = arr.strides()[arr.ndim() - 2];
+    auto sty = arr.strides()[arr.ndim() - 1];
+    if (stx == arr.shape(-1) && sty == 1) {
+      return std::make_tuple(false, stx, arr);
+    } else if (stx == 1 && sty == arr.shape(-2)) {
+      return std::make_tuple(true, sty, arr);
+    } else {
+      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+      copy(arr, arr_copy, CopyType::General);
+      size_t stx = arr.shape(-1);
+      return std::make_tuple(false, stx, arr_copy);
+    }
+  };
+
+  auto [a_transposed, lda, a] = check_transpose(a_pre);
+  auto [b_transposed, ldb, b] = check_transpose(b_pre);
+
+  size_t M = a.shape(-2);
+  size_t N = b.shape(-1);
+  size_t K = a.shape(-1);
+
+  if (M == 0 || N == 0) {
+    return;
+  }
+
+  if (K == 0) {
+    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
+    return;
+  }
+
+  // Get batch dims
+  auto batch_size_out = out.size() / (M * N);
+  size_t matrix_stride_out = M * N;
+
+  auto get_batch_dims = [](const auto& v) {
+    return decltype(v){v.begin(), v.end() - 2};
+  };
+
+  auto& lhs_indices = inputs[2];
+  auto& rhs_indices = inputs[3];
+
+  std::vector<int> batch_shape = get_batch_dims(out.shape());
+  int batch_ndim = batch_shape.size();
+
+  std::vector<int> batch_shape_A = get_batch_dims(a.shape());
+  std::vector<size_t> batch_strides_A = get_batch_dims(a.strides());
+  std::vector<int> batch_shape_B = get_batch_dims(b.shape());
+  std::vector<size_t> batch_strides_B = get_batch_dims(b.strides());
+
+  const uint32_t* lhs_indices_ptr = lhs_indices.data<uint32_t>();
+  const uint32_t* rhs_indices_ptr = rhs_indices.data<uint32_t>();
+
+  for (int i = 0; i < batch_size_out; i++) {
+    // Get index
+    uint32_t indx_A = lhs_indices_ptr[elem_to_loc(i, lhs_indices)];
+    uint32_t indx_B = rhs_indices_ptr[elem_to_loc(i, rhs_indices)];
+
+    cblas_sgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        1.0f, // alpha
+        a.data<float>() + elem_to_loc(indx_A, batch_shape_A, batch_strides_A),
+        lda,
+        b.data<float>() + elem_to_loc(indx_B, batch_shape_B, batch_strides_B),
+        ldb,
+        0.0f, // beta
+        out.data<float>() + matrix_stride_out * i,
+        out.shape(-1) // ldc
+    );
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/ops.h

@@ -108,213 +108,233 @@ struct Abs {
   template <typename T>
   T operator()(T x) {
     return std::abs(x);
-  };
+  }
   uint8_t operator()(uint8_t x) {
     return x;
-  };
+  }
   uint16_t operator()(uint16_t x) {
     return x;
-  };
+  }
   uint32_t operator()(uint32_t x) {
     return x;
-  };
+  }
   uint64_t operator()(uint64_t x) {
     return x;
-  };
+  }
   bool operator()(bool x) {
     return x;
-  };
+  }
 };
 
 struct ArcCos {
   template <typename T>
   T operator()(T x) {
     return std::acos(x);
-  };
+  }
 };
 
 struct ArcCosh {
   template <typename T>
   T operator()(T x) {
     return std::acosh(x);
-  };
+  }
 };
 
 struct ArcSin {
   template <typename T>
   T operator()(T x) {
     return std::asin(x);
-  };
+  }
 };
 
 struct ArcSinh {
   template <typename T>
   T operator()(T x) {
     return std::asinh(x);
-  };
+  }
 };
 
 struct ArcTan {
   template <typename T>
   T operator()(T x) {
     return std::atan(x);
-  };
+  }
+};
+
+struct ArcTan2 {
+  template <typename T>
+  T operator()(T y, T x) {
+    return std::atan2(y, x);
+  }
 };
 
 struct ArcTanh {
   template <typename T>
   T operator()(T x) {
     return std::atanh(x);
-  };
+  }
 };
 
 struct Ceil {
   template <typename T>
   T operator()(T x) {
     return std::ceil(x);
-  };
+  }
   int8_t operator()(int8_t x) {
     return x;
-  };
+  }
   int16_t operator()(int16_t x) {
     return x;
-  };
+  }
   int32_t operator()(int32_t x) {
     return x;
-  };
+  }
   int64_t operator()(int64_t x) {
     return x;
-  };
+  }
   uint8_t operator()(uint8_t x) {
     return x;
-  };
+  }
   uint16_t operator()(uint16_t x) {
     return x;
-  };
+  }
   uint32_t operator()(uint32_t x) {
     return x;
-  };
+  }
   uint64_t operator()(uint64_t x) {
     return x;
-  };
+  }
   bool operator()(bool x) {
     return x;
-  };
+  }
+};
+
+struct Conjugate {
+  complex64_t operator()(complex64_t x) {
+    return std::conj(x);
+  }
 };
 
 struct Cos {
   template <typename T>
   T operator()(T x) {
     return std::cos(x);
-  };
+  }
 };
 
 struct Cosh {
   template <typename T>
   T operator()(T x) {
     return std::cosh(x);
-  };
+  }
 };
 
 struct Erf {
   template <typename T>
   T operator()(T x) {
     return static_cast<T>(fast_erf(static_cast<float>(x)));
-  };
+  }
 };
 
 struct ErfInv {
   template <typename T>
   T operator()(T x) {
     return static_cast<T>(fast_erfinv(static_cast<float>(x)));
-  };
+  }
 };
 
 struct Exp {
   template <typename T>
   T operator()(T x) {
     return fast_exp(x);
-  };
+  }
 
   complex64_t operator()(complex64_t x) {
     return std::exp(x);
   }
 };
 
+struct Expm1 {
+  template <typename T>
+  T operator()(T x) {
+    return expm1(x);
+  }
+};
+
 struct Floor {
   template <typename T>
   T operator()(T x) {
     return std::floor(x);
-  };
+  }
   int8_t operator()(int8_t x) {
     return x;
-  };
+  }
   int16_t operator()(int16_t x) {
     return x;
-  };
+  }
   int32_t operator()(int32_t x) {
     return x;
-  };
+  }
   int64_t operator()(int64_t x) {
     return x;
-  };
+  }
   uint8_t operator()(uint8_t x) {
     return x;
-  };
+  }
   uint16_t operator()(uint16_t x) {
     return x;
-  };
+  }
   uint32_t operator()(uint32_t x) {
     return x;
-  };
+  }
   uint64_t operator()(uint64_t x) {
     return x;
-  };
+  }
   bool operator()(bool x) {
     return x;
-  };
+  }
 };
 
 struct Log {
   template <typename T>
   T operator()(T x) {
     return std::log(x);
-  };
+  }
 };
 
 struct Log2 {
   template <typename T>
   T operator()(T x) {
     return std::log2(x);
-  };
+  }
 };
 
 struct Log10 {
   template <typename T>
   T operator()(T x) {
     return std::log10(x);
-  };
+  }
 };
 
 struct Log1p {
   template <typename T>
   T operator()(T x) {
     return log1p(x);
-  };
+  }
 };
 
 struct LogicalNot {
   template <typename T>
   T operator()(T x) {
     return !x;
-  };
+  }
 };
 
 struct Negative {
   template <typename T>
   T operator()(T x) {
     return -x;
-  };
+  }
 };
 
 struct Round {
@@ -353,55 +373,59 @@ struct Sign {
   uint64_t operator()(uint64_t x) {
     return x != 0;
   }
+
+  complex64_t operator()(complex64_t x) {
+    return x == complex64_t(0) ? x : x / std::abs(x);
+  }
 };
 
 struct Sin {
   template <typename T>
   T operator()(T x) {
     return std::sin(x);
-  };
+  }
 };
 
 struct Sinh {
   template <typename T>
   T operator()(T x) {
     return std::sinh(x);
-  };
+  }
 };
 
 struct Square {
   template <typename T>
   T operator()(T x) {
     return x * x;
-  };
+  }
 };
 
 struct Sqrt {
   template <typename T>
   T operator()(T x) {
     return std::sqrt(x);
-  };
+  }
 };
 
 struct Rsqrt {
   template <typename T>
   T operator()(T x) {
     return static_cast<decltype(x)>(1.0) / std::sqrt(x);
-  };
+  }
 };
 
 struct Tan {
   template <typename T>
   T operator()(T x) {
     return std::tan(x);
-  };
+  }
 };
 
 struct Tanh {
   template <typename T>
   T operator()(T x) {
     return std::tanh(x);
-  };
+  }
 };
 
 struct Add {
@@ -534,7 +558,7 @@ struct LogAddExp {
         ? maxval
         : static_cast<decltype(x)>(
               maxval + std::log1p(fast_exp(minval - maxval)));
-  };
+  }
 };
 
 struct Multiply {
@@ -582,14 +606,14 @@ struct LogicalAnd {
   template <typename T>
   T operator()(T x, T y) {
     return x && y;
-  };
+  }
 };
 
 struct LogicalOr {
   template <typename T>
   T operator()(T x, T y) {
     return x || y;
-  };
+  }
 };
 
 struct Select {
@@ -599,4 +623,39 @@ struct Select {
   }
 };
 
+struct BitwiseAnd {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x & y;
+  }
+};
+
+struct BitwiseOr {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x | y;
+  }
+};
+
+struct BitwiseXor {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x ^ y;
+  }
+};
+
+struct LeftShift {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x << y;
+  }
+};
+
+struct RightShift {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x >> y;
+  }
+};
+
 } // namespace mlx::core::detail

mlx/backend/common/primitives.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <algorithm>
 #include <cassert>
@@ -8,9 +8,9 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/arange.h"
-#include "mlx/backend/common/binary.h"
 #include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/ops.h"
+#include "mlx/backend/common/slicing.h"
 #include "mlx/backend/common/threefry.h"
 #include "mlx/backend/common/unary.h"
 #include "mlx/backend/common/utils.h"
@@ -113,61 +113,6 @@ void AsType::eval(const std::vector<array>& inputs, array& out) {
   copy(in, out, ctype);
 }
 
-void AsStrided::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-
-  auto& in = inputs[0];
-
-  if (!in.flags().row_contiguous) {
-    // Just ensuring that inputs[0] came from the ops which would ensure the
-    // input is row contiguous.
-    throw std::runtime_error(
-        "AsStrided must be used with row contiguous arrays only.");
-  }
-
-  // Compute the flags given the shape and strides
-  bool row_contiguous = true, col_contiguous = true;
-  size_t r = 1, c = 1;
-  for (int i = strides_.size() - 1, j = 0; i >= 0; i--, j++) {
-    row_contiguous &= (r == strides_[i]) || (shape_[i] == 1);
-    col_contiguous &= (c == strides_[j]) || (shape_[j] == 1);
-    r *= shape_[i];
-    c *= shape_[j];
-  }
-  auto flags = in.flags();
-  // TODO: Compute the contiguous flag in a better way cause now we are
-  //       unnecessarily strict.
-  flags.contiguous = row_contiguous || col_contiguous;
-  flags.row_contiguous = row_contiguous;
-  flags.col_contiguous = col_contiguous;
-
-  // There is no easy way to compute the actual data size so we use out.size().
-  // The contiguous flag will almost certainly not be set so no code should
-  // rely on data_size anyway.
-  size_t data_size = out.size();
-
-  return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
-}
-
-void Broadcast::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-  std::vector<size_t> strides(out.ndim(), 0);
-  int diff = out.ndim() - in.ndim();
-  for (int i = in.ndim() - 1; i >= 0; --i) {
-    strides[i + diff] = (in.shape()[i] == 1) ? 0 : in.strides()[i];
-  }
-  auto flags = in.flags();
-  if (out.size() > in.size()) {
-    flags.row_contiguous = flags.col_contiguous = false;
-  }
-  out.copy_shared_buffer(in, strides, flags, in.data_size());
-}
-
 void Ceil::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
@@ -203,9 +148,15 @@ void Concatenate::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void Copy::eval(const std::vector<array>& inputs, array& out) {
+void Conjugate::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  out.copy_shared_buffer(inputs[0]);
+  const auto& in = inputs[0];
+  if (out.dtype() == complex64) {
+    unary_fp(in, out, detail::Conjugate());
+  } else {
+    throw std::invalid_argument(
+        "[conjugate] conjugate must be called on complex input.");
+  }
 }
 
 void Cos::eval(const std::vector<array>& inputs, array& out) {
@@ -232,81 +183,6 @@ void Cosh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void CustomVJP::eval(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() > outputs.size());
-  for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
-       i++, j++) {
-    outputs[i].copy_shared_buffer(inputs[j]);
-  }
-}
-
-void Depends::eval(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() > outputs.size());
-  for (int i = 0; i < outputs.size(); i++) {
-    outputs[i].copy_shared_buffer(inputs[i]);
-  }
-}
-
-void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  double numel = 1;
-  for (auto ax : axes_) {
-    numel *= inputs[0].shape(ax);
-  }
-
-  if (inverted_) {
-    numel = 1.0 / numel;
-  }
-
-  switch (out.dtype()) {
-    case bool_:
-      *out.data<bool>() = static_cast<bool>(numel);
-      break;
-    case uint8:
-      *out.data<uint8_t>() = static_cast<uint8_t>(numel);
-      break;
-    case uint16:
-      *out.data<uint16_t>() = static_cast<uint16_t>(numel);
-      break;
-    case uint32:
-      *out.data<uint32_t>() = static_cast<uint32_t>(numel);
-      break;
-    case uint64:
-      *out.data<uint64_t>() = static_cast<uint64_t>(numel);
-      break;
-    case int8:
-      *out.data<int8_t>() = static_cast<int8_t>(numel);
-      break;
-    case int16:
-      *out.data<int16_t>() = static_cast<int16_t>(numel);
-      break;
-    case int32:
-      *out.data<int32_t>() = static_cast<int32_t>(numel);
-      break;
-    case int64:
-      *out.data<int64_t>() = static_cast<int64_t>(numel);
-      break;
-    case float16:
-      *out.data<float16_t>() = static_cast<float16_t>(numel);
-      break;
-    case float32:
-      *out.data<float>() = static_cast<float>(numel);
-      break;
-    case bfloat16:
-      *out.data<bfloat16_t>() = static_cast<bfloat16_t>(numel);
-      break;
-    case complex64:
-      *out.data<complex64_t>() = static_cast<complex64_t>(numel);
-      break;
-  }
-}
-
 void Erf::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -359,6 +235,18 @@ void Exp::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void Expm1::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (issubdtype(out.dtype(), inexact)) {
+    unary_fp(in, out, detail::Expm1());
+  } else {
+    throw std::invalid_argument(
+        "[expm1] Cannot exponentiate elements in array"
+        " with non floating point type.");
+  }
+}
+
 void Floor::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
@@ -425,20 +313,6 @@ void LogicalNot::eval(const std::vector<array>& inputs, array& out) {
   unary(in, out, detail::LogicalNot());
 }
 
-void LogicalAnd::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
-  auto& in1 = inputs[0];
-  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalAnd());
-}
-
-void LogicalOr::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2); // LogicalOr requires two input arrays
-  auto& in1 = inputs[0];
-  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalOr());
-}
-
 void Negative::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
@@ -524,63 +398,6 @@ void RandomBits::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-std::pair<bool, std::vector<size_t>> Reshape::prepare_reshape(
-    const array& in,
-    const array& out) {
-  // Special case for empty arrays or row contiguous arrays
-  if (in.size() == 0 || in.flags().row_contiguous) {
-    return {false, out.strides()};
-  }
-
-  // Special case for scalars
-  if (in.ndim() == 0) {
-    std::vector<size_t> out_strides(out.ndim(), 0);
-    return {false, out_strides};
-  }
-
-  // Firstly let's collapse all the contiguous dimensions of the input
-  auto [shape, _strides] = collapse_contiguous_dims(in);
-  auto& strides = _strides[0];
-
-  // If shapes fit exactly in the contiguous dims then no copy is necessary so
-  // let's check.
-  std::vector<size_t> out_strides;
-  bool copy_necessary = false;
-  int j = 0;
-  for (int i = 0; i < out.ndim(); i++) {
-    int N = out.shape(i);
-    if (j < shape.size() && shape[j] % N == 0) {
-      shape[j] /= N;
-      out_strides.push_back(shape[j] * strides[j]);
-      j += (shape[j] == 1);
-    } else if (N == 1) {
-      // i > 0 because otherwise j < shape.size() && shape[j] % 1 == 0
-      out_strides.push_back(out_strides.back());
-    } else {
-      copy_necessary = true;
-      break;
-    }
-  }
-
-  return {copy_necessary, out_strides};
-}
-
-void Reshape::shared_buffer_reshape(
-    const array& in,
-    const std::vector<size_t>& out_strides,
-    array& out) {
-  auto flags = in.flags();
-  if (flags.row_contiguous) {
-    // For row contiguous reshapes:
-    // - Shallow copy the buffer
-    // - If reshaping into a vector (all singleton dimensions except one) it
-    //    becomes col contiguous again.
-    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
-    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
-  }
-  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
-}
-
 void Reshape::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -588,7 +405,17 @@ void Reshape::eval(const std::vector<array>& inputs, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
 
   if (copy_necessary) {
-    copy(in, out, in.data_size() == 1 ? CopyType::Scalar : CopyType::General);
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    auto out_strides = make_contiguous_strides<size_t>(in.shape());
+    copy_inplace<size_t>(
+        in,
+        out,
+        in.shape(),
+        in.strides(),
+        out_strides,
+        0,
+        0,
+        CopyType::General);
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -651,49 +478,6 @@ void Sinh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-std::tuple<bool, int64_t, std::vector<int64_t>> Slice::prepare_slice(
-    const array& in) {
-  int64_t data_offset = 0;
-  bool copy_needed = false;
-  std::vector<int64_t> inp_strides(in.ndim(), 0);
-  for (int i = 0; i < in.ndim(); ++i) {
-    data_offset += start_indices_[i] * in.strides()[i];
-    inp_strides[i] = in.strides()[i] * strides_[i];
-
-    copy_needed |= strides_[i] < 0;
-  }
-
-  return std::make_tuple(copy_needed, data_offset, inp_strides);
-}
-
-void Slice::shared_buffer_slice(
-    const array& in,
-    const std::vector<size_t>& out_strides,
-    size_t data_offset,
-    array& out) {
-  // Compute row/col contiguity
-  auto [data_size, is_row_contiguous, is_col_contiguous] =
-      check_contiguity(out.shape(), out_strides);
-
-  auto flags = in.flags();
-  flags.row_contiguous = is_row_contiguous;
-  flags.col_contiguous = is_col_contiguous;
-
-  if (data_size == 1) {
-    // Broadcasted scalar array is contiguous.
-    flags.contiguous = true;
-  } else if (data_size == in.data_size()) {
-    // Means we sliced a broadcasted dimension so leave the "no holes" flag
-    // alone.
-  } else {
-    // We sliced something. So either we are row or col contiguous or we
-    // punched a hole.
-    flags.contiguous &= flags.row_contiguous || flags.col_contiguous;
-  }
-
-  out.copy_shared_buffer(in, out_strides, flags, data_size, data_offset);
-}
-
 void Slice::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   if (out.size() == 0) {
@@ -704,7 +488,8 @@ void Slice::eval(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Calculate out strides, initial offset and if copy needs to be made
-  auto [copy_needed, data_offset, inp_strides] = prepare_slice(in);
+  auto [copy_needed, data_offset, inp_strides] =
+      prepare_slice(in, start_indices_, strides_);
 
   // Do copy if needed
   if (copy_needed) {
@@ -720,23 +505,19 @@ void Slice::eval(const std::vector<array>& inputs, array& out) {
         /* int64_t o_offset = */ 0,
         /* CopyType ctype = */ CopyType::General);
   } else {
+    size_t data_end = 1;
+    for (int i = 0; i < end_indices_.size(); ++i) {
+      if (in.shape()[i] > 1) {
+        auto end_idx = start_indices_[i] + out.shape()[i] * strides_[i] - 1;
+        data_end += end_idx * in.strides()[i];
+      }
+    }
+    size_t data_size = data_end - data_offset;
     std::vector<size_t> ostrides{inp_strides.begin(), inp_strides.end()};
-    shared_buffer_slice(in, ostrides, data_offset, out);
+    shared_buffer_slice(in, ostrides, data_offset, data_size, out);
   }
 }
 
-std::tuple<int64_t, std::vector<int64_t>> SliceUpdate::prepare_slice(
-    const array& in) {
-  int64_t data_offset = 0;
-  std::vector<int64_t> inp_strides(in.ndim(), 0);
-  for (int i = 0; i < in.ndim(); ++i) {
-    data_offset += start_indices_[i] * in.strides()[i];
-    inp_strides[i] = in.strides()[i] * strides_[i];
-  }
-
-  return std::make_tuple(data_offset, inp_strides);
-}
-
 void SliceUpdate::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   if (out.size() == 0) {
@@ -774,58 +555,6 @@ void SliceUpdate::eval(const std::vector<array>& inputs, array& out) {
       /* CopyType ctype = */ CopyType::GeneralGeneral);
 }
 
-void Split::eval(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 1);
-
-  auto& in = inputs[0];
-
-  auto compute_new_flags = [](const auto& shape,
-                              const auto& strides,
-                              size_t in_data_size,
-                              auto flags) {
-    size_t data_size = 1;
-    size_t f_stride = 1;
-    size_t b_stride = 1;
-    flags.row_contiguous = true;
-    flags.col_contiguous = true;
-    for (int i = 0, ri = shape.size() - 1; ri >= 0; i++, ri--) {
-      flags.col_contiguous &= strides[i] == f_stride || shape[i] == 1;
-      flags.row_contiguous &= strides[ri] == b_stride || shape[ri] == 1;
-      f_stride *= shape[i];
-      b_stride *= shape[ri];
-      if (strides[i] > 0) {
-        data_size *= shape[i];
-      }
-    }
-
-    if (data_size == 1) {
-      // Broadcasted scalar array is contiguous.
-      flags.contiguous = true;
-    } else if (data_size == in_data_size) {
-      // Means we sliced a broadcasted dimension so leave the "no holes" flag
-      // alone.
-    } else {
-      // We sliced something. So either we are row or col contiguous or we
-      // punched a hole.
-      flags.contiguous &= flags.row_contiguous || flags.col_contiguous;
-    }
-
-    return std::pair<decltype(flags), size_t>{flags, data_size};
-  };
-
-  std::vector<int> indices(1, 0);
-  indices.insert(indices.end(), indices_.begin(), indices_.end());
-  for (int i = 0; i < indices.size(); i++) {
-    size_t offset = indices[i] * in.strides()[axis_];
-    auto [new_flags, data_size] = compute_new_flags(
-        outputs[i].shape(), in.strides(), in.data_size(), in.flags());
-    outputs[i].copy_shared_buffer(
-        in, in.strides(), new_flags, data_size, offset);
-  }
-}
-
 void Square::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
@@ -842,11 +571,6 @@ void Sqrt::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void StopGradient::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  out.copy_shared_buffer(inputs[0]);
-}
-
 void Tan::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -871,38 +595,43 @@ void Tanh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void Transpose::eval(const std::vector<array>& inputs, array& out) {
+void View::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  std::vector<size_t> out_strides(out.ndim());
   auto& in = inputs[0];
-  for (int ax = 0; ax < axes_.size(); ++ax) {
-    out_strides[ax] = in.strides()[axes_[ax]];
-  }
-
-  // Conditions for {row/col}_contiguous
-  // - array must be contiguous (no gaps)
-  // - underlying buffer size should have the same size as the array
-  // - cumulative product of shapes is equal to the strides (we can ignore axes
-  //   with size == 1)
-  //   - in the forward direction (column contiguous)
-  //   - in the reverse direction (row contiguous)
-  // - vectors are both row and col contiguous (hence if both row/col are
-  //   true, they stay true)
-  auto flags = in.flags();
-  if (flags.contiguous && in.data_size() == in.size()) {
-    size_t f_stride = 1;
-    size_t b_stride = 1;
-    flags.col_contiguous = true;
-    flags.row_contiguous = true;
-    for (int i = 0, ri = out.ndim() - 1; i < out.ndim(); ++i, --ri) {
-      flags.col_contiguous &= (out_strides[i] == f_stride || out.shape(i) == 1);
-      f_stride *= out.shape(i);
-      flags.row_contiguous &=
-          (out_strides[ri] == b_stride || out.shape(ri) == 1);
-      b_stride *= out.shape(ri);
+  auto ibytes = size_of(in.dtype());
+  auto obytes = size_of(out.dtype());
+  // Conditions for buffer copying (disjunction):
+  // - type size is the same
+  // - type size is smaller and the last axis is contiguous
+  // - the entire array is row contiguous
+  if (ibytes == obytes || obytes < ibytes && in.strides().back() == 1 ||
+      in.flags().row_contiguous) {
+    auto strides = in.strides();
+    for (int i = 0; i < strides.size() - 1; ++i) {
+      strides[i] *= ibytes;
+      strides[i] /= obytes;
     }
+    out.copy_shared_buffer(
+        in, strides, in.flags(), in.data_size() * ibytes / obytes);
+  } else {
+    auto tmp = array(
+        in.shape(), in.dtype() == bool_ ? uint8 : in.dtype(), nullptr, {});
+    tmp.set_data(allocator::malloc_or_wait(tmp.nbytes()));
+    if (in.dtype() == bool_) {
+      auto in_tmp = array(in.shape(), uint8, nullptr, {});
+      in_tmp.copy_shared_buffer(in);
+      copy_inplace(in_tmp, tmp, CopyType::General);
+    } else {
+      copy_inplace(in, tmp, CopyType::General);
+    }
+
+    auto flags = out.flags();
+    flags.contiguous = true;
+    flags.row_contiguous = true;
+    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
+    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
+    out.move_shared_buffer(tmp, out.strides(), flags, out.size());
   }
-  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
 }
 
 } // namespace mlx::core

mlx/backend/common/quantized.cpp

@@ -192,7 +192,7 @@ void _qmm_dispatch_typed(
 }
 
 void _qmm_dispatch(
-    array out,
+    array& out,
     const array& x,
     const array& w,
     const array& scales,
@@ -253,6 +253,81 @@ void _qmm_dispatch(
   }
 }
 
+void _bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& biases,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    int bits,
+    int group_size,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.shape(-2);
+  int N = out.shape(-1);
+
+  int w_els = w.shape(-1) * w.shape(-2);
+  int g_els = scales.shape(-1) * scales.shape(-2);
+
+  const uint32_t* lhs_indices_data = lhs_indices.data<uint32_t>();
+  const uint32_t* rhs_indices_data = rhs_indices.data<uint32_t>();
+
+  for (int i = 0; i < lhs_indices.size(); i++) {
+    int x_idx = lhs_indices_data[elem_to_loc(i, lhs_indices)];
+    int w_idx = rhs_indices_data[elem_to_loc(i, rhs_indices)];
+
+    switch (x.dtype()) {
+      case float32:
+        _qmm_dispatch_typed<float>(
+            out.data<float>() + i * M * N,
+            x.data<float>() + elem_to_loc(x_idx * M * K, x),
+            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
+            scales.data<float>() + elem_to_loc(w_idx * g_els, scales),
+            biases.data<float>() + elem_to_loc(w_idx * g_els, biases),
+            M,
+            N,
+            K,
+            bits,
+            group_size,
+            transposed_w);
+        break;
+      case float16:
+        _qmm_dispatch_typed<float16_t>(
+            out.data<float16_t>() + i * M * N,
+            x.data<float16_t>() + elem_to_loc(x_idx * M * K, x),
+            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
+            scales.data<float16_t>() + elem_to_loc(w_idx * g_els, scales),
+            biases.data<float16_t>() + elem_to_loc(w_idx * g_els, biases),
+            M,
+            N,
+            K,
+            bits,
+            group_size,
+            transposed_w);
+        break;
+      case bfloat16:
+        _qmm_dispatch_typed<bfloat16_t>(
+            out.data<bfloat16_t>() + i * M * N,
+            x.data<bfloat16_t>() + elem_to_loc(x_idx * M * K, x),
+            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
+            scales.data<bfloat16_t>() + elem_to_loc(w_idx * g_els, scales),
+            biases.data<bfloat16_t>() + elem_to_loc(w_idx * g_els, biases),
+            M,
+            N,
+            K,
+            bits,
+            group_size,
+            transposed_w);
+        break;
+      default:
+        throw std::invalid_argument(
+            "[quantized_matmul] only floating types are supported");
+    }
+  }
+}
+
 } // namespace
 
 void QuantizedMatmul::eval(const std::vector<array>& inputs, array& out) {
@@ -282,4 +357,45 @@ void QuantizedMatmul::eval(const std::vector<array>& inputs, array& out) {
   _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
 }
 
+void GatherQMM::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 6);
+
+  auto& x_pre = inputs[0];
+  auto& w_pre = inputs[1];
+  auto& scales_pre = inputs[2];
+  auto& biases_pre = inputs[3];
+  auto& lhs_indices = inputs[4];
+  auto& rhs_indices = inputs[5];
+
+  auto ensure_row_contiguous_last_dims = [](const array& arr) {
+    auto stride_0 = arr.strides()[arr.ndim() - 2];
+    auto stride_1 = arr.strides()[arr.ndim() - 1];
+    if (stride_0 == arr.shape(-1) && stride_1 == 1) {
+      return arr;
+    } else {
+      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+      copy(arr, arr_copy, CopyType::General);
+      return arr_copy;
+    }
+  };
+
+  auto x = ensure_row_contiguous_last_dims(x_pre);
+  auto w = ensure_row_contiguous_last_dims(w_pre);
+  auto scales = ensure_row_contiguous_last_dims(scales_pre);
+  auto biases = ensure_row_contiguous_last_dims(biases_pre);
+
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  _bs_qmm_dispatch(
+      out,
+      x,
+      w,
+      scales,
+      biases,
+      lhs_indices,
+      rhs_indices,
+      group_size_,
+      bits_,
+      transpose_);
+}
+
 } // namespace mlx::core

mlx/backend/common/reduce.cpp

@@ -87,6 +87,38 @@ struct OrReduce {
   }
 };
 
+struct MaxReduce {
+  template <typename T>
+  std::enable_if_t<std::is_integral_v<T>> operator()(T* y, T x) {
+    (*y) = (*y > x) ? *y : x;
+  };
+
+  template <typename T>
+  std::enable_if_t<!std::is_integral_v<T>> operator()(T* y, T x) {
+    if (std::isnan(x)) {
+      *y = x;
+    } else {
+      (*y) = (*y > x) ? *y : x;
+    }
+  };
+};
+
+struct MinReduce {
+  template <typename T>
+  std::enable_if_t<std::is_integral_v<T>> operator()(T* y, T x) {
+    (*y) = (*y < x) ? *y : x;
+  };
+
+  template <typename T>
+  std::enable_if_t<!std::is_integral_v<T>> operator()(T* y, T x) {
+    if (std::isnan(x)) {
+      *y = x;
+    } else {
+      (*y) = (*y < x) ? *y : x;
+    }
+  };
+};
+
 template <typename InT>
 void reduce_dispatch_out(
     const array& in,
@@ -104,63 +136,27 @@ void reduce_dispatch_out(
     }
     case Reduce::Sum: {
       auto op = [](auto y, auto x) { (*y) = (*y) + x; };
-      switch (out.dtype()) {
-        case bool_:
-          reduction_op<InT, bool>(in, out, axes, false, op);
-          break;
-        case uint8:
-          reduction_op<InT, uint8_t>(in, out, axes, 0, op);
-          break;
-        case uint16:
-          reduction_op<InT, uint16_t>(in, out, axes, 0, op);
-          break;
-        case uint32:
-          reduction_op<InT, uint32_t>(in, out, axes, 0, op);
-          break;
-        case uint64:
-          reduction_op<InT, uint64_t>(in, out, axes, 0, op);
-          break;
-        case int8:
-          reduction_op<InT, int8_t>(in, out, axes, 0, op);
-          break;
-        case int16:
-          reduction_op<InT, int16_t>(in, out, axes, 0, op);
-          break;
-        case int32:
-          reduction_op<InT, int32_t>(in, out, axes, 0, op);
-          break;
-        case int64:
-          reduction_op<InT, int64_t>(in, out, axes, 0, op);
-          break;
-        case float16:
-          reduction_op<InT, float16_t>(in, out, axes, 0.0f, op);
-          break;
-        case float32:
-          reduction_op<InT, float>(in, out, axes, 0.0f, op);
-          break;
-        case bfloat16:
-          reduction_op<InT, bfloat16_t>(in, out, axes, 0.0f, op);
-          break;
-        case complex64:
-          reduction_op<InT, complex64_t>(in, out, axes, complex64_t{0.0f}, op);
-          break;
+      if (out.dtype() == int32) {
+        // special case since the input type can be bool
+        reduction_op<InT, int32_t>(in, out, axes, 0, op);
+      } else {
+        reduction_op<InT, InT>(in, out, axes, 0, op);
       }
-    } break;
+      break;
+    }
     case Reduce::Prod: {
       auto op = [](auto y, auto x) { (*y) *= x; };
       reduction_op<InT, InT>(in, out, axes, 1, op);
       break;
     }
     case Reduce::Max: {
-      auto op = [](auto y, auto x) { (*y) = (*y > x) ? *y : x; };
       auto init = Limits<InT>::min;
-      reduction_op<InT, InT>(in, out, axes, init, op);
+      reduction_op<InT, InT>(in, out, axes, init, MaxReduce());
       break;
     }
     case Reduce::Min: {
-      auto op = [](auto y, auto x) { (*y) = (*y < x) ? *y : x; };
       auto init = Limits<InT>::max;
-      reduction_op<InT, InT>(in, out, axes, init, op);
+      reduction_op<InT, InT>(in, out, axes, init, MinReduce());
       break;
     }
   }
@@ -168,6 +164,29 @@ void reduce_dispatch_out(
 
 } // namespace
 
+void nd_loop(
+    std::function<void(int)> callback,
+    const std::vector<int>& shape,
+    const std::vector<size_t>& strides) {
+  std::function<void(int, int)> loop_inner;
+  loop_inner = [&](int dim, int offset) {
+    if (dim < shape.size() - 1) {
+      int size = shape[dim];
+      size_t stride = strides[dim];
+      for (int i = 0; i < size; i++) {
+        loop_inner(dim + 1, offset + i * stride);
+      }
+    } else {
+      int size = shape[dim];
+      size_t stride = strides[dim];
+      for (int i = 0; i < size; i++) {
+        callback(offset + i * stride);
+      }
+    }
+  };
+  loop_inner(0, 0);
+}
+
 void Reduce::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];

mlx/backend/common/reduce.h

@@ -49,47 +49,18 @@ struct ReductionPlan {
   ReductionPlan(ReductionOpType type_) : type(type_) {}
 };
 
-namespace {
+ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes);
 
 // Helper for the ndimensional strided loop
 // Should this be in utils?
-inline void nd_loop(
+void nd_loop(
     std::function<void(int)> callback,
     const std::vector<int>& shape,
-    const std::vector<size_t>& strides) {
-  std::function<void(int, int)> loop_inner;
-  loop_inner = [&](int dim, int offset) {
-    if (dim < shape.size() - 1) {
-      int size = shape[dim];
-      size_t stride = strides[dim];
-      for (int i = 0; i < size; i++) {
-        loop_inner(dim + 1, offset + i * stride);
-      }
-    } else {
-      int size = shape[dim];
-      size_t stride = strides[dim];
-      for (int i = 0; i < size; i++) {
-        callback(offset + i * stride);
-      }
-    }
-  };
-  loop_inner(0, 0);
-}
+    const std::vector<size_t>& strides);
 
 std::pair<std::vector<int>, std::vector<size_t>> shapes_without_reduction_axes(
     const array& x,
-    const std::vector<int>& axes) {
-  std::vector<int> shape = x.shape();
-  std::vector<size_t> strides = x.strides();
-
-  for (int i = axes.size() - 1; i >= 0; i--) {
-    int a = axes[i];
-    shape.erase(shape.begin() + a);
-    strides.erase(strides.begin() + a);
-  }
-
-  return std::make_pair(shape, strides);
-}
+    const std::vector<int>& axes);
 
 template <typename T, typename U, typename Op>
 struct DefaultStridedReduce {
@@ -123,102 +94,6 @@ struct DefaultContiguousReduce {
   }
 };
 
-ReductionPlan get_reduction_plan(const array& x, const std::vector<int> axes) {
-  // The data is all there and we are reducing over everything
-  if (x.size() == x.data_size() && axes.size() == x.ndim() &&
-      x.flags().contiguous) {
-    return ContiguousAllReduce;
-  }
-
-  // Row contiguous input so the output is row contiguous
-  if (x.flags().row_contiguous) {
-    // Merge consecutive axes
-    std::vector<int> shape = {x.shape(axes[0])};
-    std::vector<size_t> strides = {x.strides()[axes[0]]};
-    for (int i = 1; i < axes.size(); i++) {
-      if (axes[i] - 1 == axes[i - 1]) {
-        shape.back() *= x.shape(axes[i]);
-        strides.back() = x.strides()[axes[i]];
-      } else {
-        shape.push_back(x.shape(axes[i]));
-        strides.push_back(x.strides()[axes[i]]);
-      }
-    }
-
-    if (strides.back() == 1) {
-      return ReductionPlan(ContiguousReduce, shape, strides);
-    } else if (strides.back() > 1) {
-      return ReductionPlan(ContiguousStridedReduce, shape, strides);
-    }
-  }
-
-  // Let's check if we can optimize our access patterns
-  //
-  // 1. We have a reduction axis with stride 1. Simply call
-  //    GeneralContiguousReduce and be done with it.
-  // 2. We have transpositions and we are not reducing over the axis with
-  //    stride 1. However, we are reducing over an axis where everything is
-  //    contiguous in memory to the right of that axis. We can call strided
-  //    reduce and be done with it.
-  // 2. We have weird transpositions and expands. Copy the strides to the
-  //    output, then call strided reduce.
-
-  // Sort reduction axes by stride in order to merge them and figure out if we
-  // have a contiguous reduction.
-  std::vector<std::pair<int, size_t>> reductions;
-  for (auto a : axes) {
-    reductions.push_back(std::make_pair(x.shape(a), x.strides()[a]));
-  }
-  std::sort(reductions.begin(), reductions.end(), [](auto a, auto b) {
-    return a.second > b.second;
-  });
-  // Extract the two smallest and try to merge them in case the contiguous
-  // reduction can be bigger than just the last axis.
-  for (int i = reductions.size() - 1; i >= 1; i--) {
-    auto a = reductions[i];
-    auto b = reductions[i - 1];
-
-    // b.stride = a.shape * a.stride then a and b are contiguous
-    if (b.second == a.first * a.second) {
-      reductions.erase(reductions.begin() + i);
-      reductions[i - 1] = std::make_pair(a.first * b.first, a.second);
-    }
-  }
-
-  std::vector<int> shape;
-  std::vector<size_t> strides;
-  for (auto r : reductions) {
-    shape.push_back(r.first);
-    strides.push_back(r.second);
-  }
-
-  // We can call the contiguous reduction op for every weird way the input is
-  // structured in the rest of the axes.
-  if (strides.back() == 1) {
-    return ReductionPlan(GeneralContiguousReduce, shape, strides);
-  }
-
-  // Delegate to the general strided reduction op if the axes after
-  // strides.back() are contiguous.
-  if (strides.back() > 1) {
-    int size = 1;
-    for (int i = x.ndim() - 1; i >= 0; i--) {
-      if (axes.back() == i) {
-        continue;
-      }
-      if (x.strides()[i] != size) {
-        break;
-      }
-      size *= x.shape(i);
-    }
-    if (size >= strides.back()) {
-      return ReductionPlan(GeneralStridedReduce, shape, strides);
-    }
-  }
-
-  return ReductionPlan(GeneralReduce, shape, strides);
-}
-
 template <typename T, typename U, typename OpS, typename OpC, typename Op>
 void reduction_op(
     const array& x,
@@ -361,6 +236,4 @@ void reduction_op(
   reduction_op<T, U>(x, out, axes, init, ops, opc, op);
 }
 
-} // namespace
-
 } // namespace mlx::core

mlx/backend/common/reduce_utils.cpp

@@ -0,0 +1,147 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/common/reduce.h"
+
+namespace mlx::core {
+
+std::pair<std::vector<int>, std::vector<size_t>> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes) {
+  std::vector<int> shape = x.shape();
+  std::vector<size_t> strides = x.strides();
+
+  for (int i = axes.size() - 1; i >= 0; i--) {
+    int a = axes[i];
+    shape.erase(shape.begin() + a);
+    strides.erase(strides.begin() + a);
+  }
+
+  return std::make_pair(shape, strides);
+}
+
+ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
+  // The data is all there and we are reducing over everything
+  if (x.size() == x.data_size() && axes.size() == x.ndim() &&
+      x.flags().contiguous) {
+    return ContiguousAllReduce;
+  }
+
+  // Row contiguous input so the output is row contiguous
+  if (x.flags().row_contiguous) {
+    // Merge consecutive axes
+    std::vector<int> shape = {x.shape(axes[0])};
+    std::vector<size_t> strides = {x.strides()[axes[0]]};
+    for (int i = 1; i < axes.size(); i++) {
+      if (axes[i] - 1 == axes[i - 1] && x.shape(axes[i]) > 1) {
+        shape.back() *= x.shape(axes[i]);
+        strides.back() = x.strides()[axes[i]];
+      } else {
+        shape.push_back(x.shape(axes[i]));
+        strides.push_back(x.strides()[axes[i]]);
+      }
+    }
+
+    // Remove singleton axes from the plan
+    for (int i = shape.size() - 1; i >= 0; i--) {
+      if (shape[i] == 1) {
+        shape.erase(shape.begin() + i);
+        strides.erase(strides.begin() + i);
+      }
+    }
+
+    if (strides.back() == 1) {
+      return ReductionPlan(ContiguousReduce, shape, strides);
+    } else if (strides.back() > 1) {
+      return ReductionPlan(ContiguousStridedReduce, shape, strides);
+    }
+  }
+
+  // Let's check if we can optimize our access patterns
+  //
+  // 1. We have a reduction axis with stride 1. Simply call
+  //    GeneralContiguousReduce and be done with it.
+  // 2. We have transpositions and we are not reducing over the axis with
+  //    stride 1. However, we are reducing over an axis where everything is
+  //    contiguous in memory to the right of that axis. We can call strided
+  //    reduce and be done with it.
+  // 2. We have weird transpositions and expands. Copy the strides to the
+  //    output, then call strided reduce.
+
+  // Sort reduction axes by stride in order to merge them and figure out if we
+  // have a contiguous reduction.
+  std::vector<std::pair<int, size_t>> reductions;
+  for (auto a : axes) {
+    if (x.shape(a) > 1) {
+      reductions.push_back(std::make_pair(x.shape(a), x.strides()[a]));
+    }
+  }
+  std::sort(reductions.begin(), reductions.end(), [](auto a, auto b) {
+    bool a_is_zero = a.second == 0;
+    bool b_is_zero = b.second == 0;
+    return (a_is_zero != b_is_zero) ? a.second < b.second : a.second > b.second;
+  });
+  // Extract the two smallest and try to merge them in case the contiguous
+  // reduction can be bigger than just the last axis.
+  for (int i = reductions.size() - 1; i >= 1; i--) {
+    auto a = reductions[i];
+    auto b = reductions[i - 1];
+
+    // b.stride = a.shape * a.stride then a and b are contiguous
+    if (b.second == a.first * a.second) {
+      reductions.erase(reductions.begin() + i);
+      reductions[i - 1] = std::make_pair(a.first * b.first, a.second);
+    }
+  }
+
+  std::vector<int> shape;
+  std::vector<size_t> strides;
+  for (auto r : reductions) {
+    shape.push_back(r.first);
+    strides.push_back(r.second);
+  }
+
+  // We can call the contiguous reduction op for every weird way the input is
+  // structured in the rest of the axes.
+  if (strides.back() == 1) {
+    return ReductionPlan(GeneralContiguousReduce, shape, strides);
+  }
+
+  // Delegate to the general strided reduction op if the axes after
+  // strides.back() are contiguous.
+  if (strides.back() > 1) {
+    int size = 1;
+    bool have_expand = false;
+    for (int i = x.ndim() - 1; i >= 0; i--) {
+      if (axes.back() == i) {
+        continue;
+      }
+
+      size_t stride_i = x.strides()[i];
+      int shape_i = x.shape(i);
+      if (stride_i == 0) {
+        if (shape_i == 1) {
+          continue;
+        }
+
+        have_expand = true;
+        break;
+      }
+
+      if (stride_i != size && shape_i != 1) {
+        break;
+      }
+      size *= shape_i;
+    }
+    // In the case of an expanded dimension we are being conservative and
+    // require the smallest reduction stride to be smaller than the maximum row
+    // contiguous size. The reason is that we can't easily know if the reduced
+    // axis is before or after an expanded dimension.
+    if (size > strides.back() || (size == strides.back() && !have_expand)) {
+      return ReductionPlan(GeneralStridedReduce, shape, strides);
+    }
+  }
+
+  return ReductionPlan(GeneralReduce, shape, strides);
+}
+
+} // namespace mlx::core

mlx/backend/common/scan.cpp

@@ -234,7 +234,7 @@ void scan_dispatch(
       auto op = [](U* o, const U* y, const T* x) { *o = (*x < *y) ? *y : *x; };
       auto init = (issubdtype(input.dtype(), floating))
           ? static_cast<U>(-std::numeric_limits<float>::infinity())
-          : std::numeric_limits<U>::max();
+          : std::numeric_limits<U>::min();
       auto opcs = DefaultContiguousScan<T, U, decltype(op)>(op, init);
       auto opss = DefaultStridedScan<T, U, decltype(op)>(op, init);
       scan_op<T, U>(opcs, opss, input, output, axis, reverse, inclusive);