version bump (#1154)

* Float mask update

* Update CPU impl

.circleci/config.yml

@@ -31,7 +31,7 @@ jobs:
           name: Install dependencies
           command: |
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install numpy
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
@@ -49,11 +49,6 @@ jobs:
           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: |
@@ -69,18 +64,19 @@ 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 git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install numpy
             pip install torch
             pip install tensorflow
@@ -101,11 +97,14 @@ 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:
@@ -117,7 +116,13 @@ jobs:
           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
 
   build_release:
     parameters:
@@ -132,7 +137,7 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
@@ -143,7 +148,7 @@ 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 git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install --upgrade setuptools
             pip install numpy
             pip install twine
@@ -207,7 +212,7 @@ 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 git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install --upgrade setuptools
             pip install numpy
             pip install auditwheel

.pre-commit-config.yaml

@@ -1,11 +1,11 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v17.0.6
+    rev: v18.1.4
     hooks:
     -   id: clang-format
 # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 24.2.0
+    rev: 24.4.2
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort

ACKNOWLEDGMENTS.md

@@ -7,7 +7,7 @@ with a short description of your contribution(s) below. For example:
 
 MLX was developed with contributions from the following individuals:
 
-- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops.
+- 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.
@@ -15,7 +15,9 @@ MLX was developed with contributions from the following individuals:
 - 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.
+
 <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 0.14.0)
 endif()
 
 # --------------------- Processor tests -------------------------
@@ -82,9 +86,13 @@ elseif (MLX_BUILD_METAL)
   message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
 
   if (${MACOS_VERSION} GREATER_EQUAL 14.2)
+    set(METAL_CPP_PATCH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/metal.14.2.diff)
     set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14.2_iOS17.2.zip)
+    set(MLX_METAL_VERSION METAL_3_1)
   elseif (${MACOS_VERSION} GREATER_EQUAL 14.0)
+    set(METAL_CPP_PATCH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/metal.14.0.diff)
     set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14_iOS17-beta.zip)
+    set(MLX_METAL_VERSION METAL_3_0)
   else()
     message(FATAL_ERROR "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON" )
   endif()
@@ -92,6 +100,7 @@ elseif (MLX_BUILD_METAL)
   FetchContent_Declare(
     metal_cpp
     URL ${METAL_CPP_URL}
+    PATCH_COMMAND /usr/bin/patch -N -i ${METAL_CPP_PATCH} || true
   )
 
   FetchContent_MakeAvailable(metal_cpp)
@@ -101,55 +110,66 @@ 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})
 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)
+    target_include_directories(mlx PRIVATE ${MPI_INCLUDE_PATH})
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
@@ -161,6 +181,14 @@ 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/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 +53,13 @@ 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 +69,15 @@ 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 +86,7 @@ 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 +97,40 @@ 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/fft_bench.py

@@ -0,0 +1,57 @@
+# Copyright © 2024 Apple Inc.
+
+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 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):
+    def fft(x):
+        out = mx.fft.fft(x)
+        mx.eval(out)
+        return out
+
+    bandwidths = []
+    for k in range(4, 12):
+        n = 2**k
+        x = mx.random.uniform(shape=(system_size // n, n)).astype(mx.float32)
+        x = x.astype(mx.complex64)
+        mx.eval(x)
+        runtime_ms = measure_runtime(fft, x=x)
+        bandwidths.append(bandwidth_gb(runtime_ms, system_size))
+
+    return bandwidths
+
+
+def time_fft():
+
+    with mx.stream(mx.cpu):
+        cpu_bandwidths = run_bench(system_size=int(2**22))
+
+    with mx.stream(mx.gpu):
+        gpu_bandwidths = run_bench(system_size=int(2**29))
+
+    # plot bandwidths
+    x = [2**k for k in range(4, 12)]
+    plt.scatter(x, gpu_bandwidths, color="green", label="GPU")
+    plt.scatter(x, cpu_bandwidths, color="red", label="CPU")
+    plt.title("MLX FFT Benchmark")
+    plt.xlabel("N")
+    plt.ylabel("Bandwidth (GB/s)")
+    plt.legend()
+    plt.savefig("fft_plot.png")
+
+
+if __name__ == "__main__":
+    time_fft()

cmake/metal.14.0.diff

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

cmake/metal.14.2.diff

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

docs/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 +1,3 @@
+sphinx
+breathe
+sphinx-book-theme

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"

docs/src/cpp/ops.rst

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

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
@@ -518,10 +479,10 @@ 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);
 
@@ -533,7 +494,7 @@ below.
         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 +503,17 @@ 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 +531,30 @@ 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 +569,12 @@ 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 +583,35 @@ 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 +625,39 @@ 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 +669,17 @@ 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 +687,8 @@ 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 +710,12 @@ 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 +737,29 @@ 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 +769,50 @@ 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 +825,7 @@ 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 +836,12 @@ 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 +860,7 @@ 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 +881,23 @@ 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/install.rst

@@ -74,7 +74,7 @@ Install `nanobind <https://nanobind.readthedocs.io/en/latest/>`_ with:
 
 .. code-block:: shell
 
-    pip install git+https://github.com/wjakob/nanobind.git
+    pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
 
 Then simply build and install MLX using pip:
 
@@ -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/devices_and_streams.rst

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

docs/src/python/linalg.rst

@@ -8,5 +8,8 @@ Linear Algebra
 .. autosummary:: 
    :toctree: _autosummary 
 
+    inv
     norm
+    cholesky
     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/layers.rst

@@ -15,6 +15,7 @@ Layers
    BatchNorm
    Conv1d
    Conv2d
+   Conv3d
    Dropout
    Dropout2d
    Dropout3d
@@ -31,6 +32,7 @@ Layers
    Mish
    MultiHeadAttention
    PReLU
+   QuantizedEmbedding
    QuantizedLinear
    RMSNorm
    ReLU
@@ -43,4 +45,4 @@ Layers
    Softshrink
    Step
    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
@@ -19,19 +20,27 @@ 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
@@ -42,6 +51,7 @@ Operations
    cummin
    cumprod
    cumsum
+   degrees
    dequantize
    diag
    diagonal
@@ -51,12 +61,15 @@ Operations
    erf
    erfinv
    exp
+   expm1
    expand_dims
    eye
    flatten
    floor
    floor_divide
    full
+   gather_mm
+   gather_qmm
    greater
    greater_equal
    identity
@@ -66,6 +79,8 @@ Operations
    isnan
    isneginf
    isposinf
+   issubdtype
+   left_shift
    less
    less_equal
    linspace
@@ -83,22 +98,28 @@ Operations
    max
    maximum
    mean
+   meshgrid
    min
    minimum
    moveaxis
    multiply
    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 +138,7 @@ Operations
    square
    squeeze
    stack
+   std
    stop_gradient
    subtract
    sum
@@ -128,6 +150,7 @@ Operations
    tensordot
    tile
    topk
+   trace
    transpose
    tri
    tril

docs/src/python/optimizers.rst

@@ -1,5 +1,7 @@
 .. _optimizers:
 
+.. currentmodule:: mlx.optimizers
+
 Optimizers
 ==========
 
@@ -34,3 +36,8 @@ model's parameters and the **optimizer state**.
    optimizers/optimizer
    optimizers/common_optimizers
    optimizers/schedulers
+
+.. autosummary::
+   :toctree: _autosummary
+
+   clip_grad_norm

docs/src/python/random.rst

@@ -38,6 +38,7 @@ we use a splittable version of Threefry, which is a counter-based PRNG.
    gumbel
    key
    normal
+   multivariate_normal
    randint
    seed
    split

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/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/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_reduce_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
@@ -261,7 +257,7 @@ void Axpby::eval_gpu(
   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 +266,11 @@ 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 +296,7 @@ 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 +368,4 @@ 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>=0.9.0",
+  "nanobind@git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4",
+]
+build-backend = "setuptools.build_meta"

examples/extensions/requirements.txt

@@ -0,0 +1,4 @@
+setuptools>=42
+cmake>=3.24
+mlx>=0.9.0
+nanobind@git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4

examples/extensions/setup.py

@@ -1,4 +1,4 @@
-# Copyright © 2023 Apple Inc.
+# Copyright © 2023-2024 Apple Inc.
 
 from setuptools import setup
 
@@ -9,11 +9,11 @@ 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"]},
+        extras_require={"dev": []},
         zip_safe=False,
         python_requires=">=3.8",
     )

examples/extensions/test.py

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

mlx/CMakeLists.txt

@@ -19,11 +19,17 @@ 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.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();

mlx/array.cpp

@@ -1,5 +1,4 @@
 // Copyright © 2023-2024 Apple Inc.
-
 #include <functional>
 
 #include "mlx/array.h"
@@ -93,7 +92,13 @@ 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 {
@@ -161,6 +166,39 @@ 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 will be detached
+  if (status() != array::Status::unscheduled) {
+    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;
@@ -174,7 +212,7 @@ void array::ArrayDesc::init() {
 }
 
 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 +223,42 @@ 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 destory 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 {
 
@@ -113,6 +114,15 @@ class array {
     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;
@@ -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;
@@ -251,22 +261,16 @@ class array {
     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 {
@@ -315,9 +319,27 @@ class array {
     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;
+  }
+  const 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 +380,8 @@ class array {
     array_desc_ = other.array_desc_;
   }
 
+  ~array();
+
  private:
   // Initialize the arrays data
   template <typename It>
@@ -370,6 +394,11 @@ 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};
@@ -404,6 +433,8 @@ class array {
         std::shared_ptr<Primitive> primitive,
         std::vector<array> inputs);
 
+    ~ArrayDesc();
+
    private:
     // Initialize size, strides, and other metadata
     void init();
@@ -468,10 +499,11 @@ 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/matmul.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 
@@ -196,6 +196,40 @@ 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

@@ -31,9 +31,11 @@ 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(Depends)
@@ -45,6 +47,8 @@ DEFAULT(ErfInv)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Gather)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
 DEFAULT(Less)
@@ -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);
@@ -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];
@@ -310,6 +335,19 @@ void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void Expm1::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vvexpm1f(
+        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+  } else {
+    eval(inputs, out);
+  }
+}
+
 void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];

mlx/backend/accelerate/softmax.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <limits>
@@ -201,7 +201,7 @@ struct NeonFp16SimdOps {
   }
 };
 
-template <typename T, typename VT, typename Ops, int N>
+template <typename T, typename AccT, typename VT, typename Ops, int N>
 void softmax(const array& in, array& out) {
   Ops ops;
 
@@ -218,13 +218,21 @@ 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 +242,29 @@ 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 +273,33 @@ 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 +346,29 @@ 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 {
+        softmax<
+            float16_t,
+            float16_t,
+            float16x8_t,
+            NeonFp16SimdOps<float16_t, float16x8_t>,
+            8>(in, out);
+      }
       break;
     case bfloat16:
       eval(inputs, out);

mlx/backend/common/CMakeLists.txt

@@ -37,10 +37,12 @@ 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}/masked_mm.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
@@ -54,6 +56,7 @@ 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

@@ -236,4 +236,82 @@ 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"

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 +1,347 @@
+// 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 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;
+  }
+}
+
+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<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);
+}
+
+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

@@ -126,4 +126,102 @@ 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.nbytes() == outputs[o].nbytes() &&
+          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

@@ -52,8 +52,25 @@ 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 +81,7 @@ 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 +265,7 @@ 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 +338,8 @@ 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,

mlx/backend/common/copy.cpp

@@ -256,7 +256,7 @@ void copy_general_general(
   }
 
   int size = std::accumulate(
-      data_shape.begin() - 5, data_shape.end(), 1, std::multiplies<int>());
+      data_shape.end() - 5, data_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);
@@ -272,7 +272,7 @@ 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 +281,54 @@ 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 +338,46 @@ 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;
   }
 }

mlx/backend/common/default_primitives.cpp

@@ -34,6 +34,7 @@ DEFAULT(ArcCosh)
 DEFAULT(ArcSin)
 DEFAULT(ArcSinh)
 DEFAULT(ArcTan)
+DEFAULT(ArcTan2)
 DEFAULT(ArcTanh)
 DEFAULT(ArgPartition)
 DEFAULT(ArgReduce)
@@ -41,9 +42,13 @@ 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)
@@ -57,6 +62,7 @@ DEFAULT(Equal)
 DEFAULT(Erf)
 DEFAULT(ErfInv)
 DEFAULT(Exp)
+DEFAULT(Expm1)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Full)
@@ -107,6 +113,7 @@ DEFAULT(Tan)
 DEFAULT(Tanh)
 DEFAULT(Transpose)
 DEFAULT(Inverse)
+DEFAULT(Cholesky)
 
 namespace {
 

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
@@ -93,12 +92,4 @@ void Inverse::eval(const std::vector<array>& inputs, array& output) {
   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}};
-}
-
 } // 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>

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

@@ -161,6 +161,13 @@ struct ArcTan {
   };
 };
 
+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) {
@@ -202,6 +209,12 @@ struct Ceil {
   };
 };
 
+struct Conjugate {
+  complex64_t operator()(complex64_t x) {
+    return std::conj(x);
+  }
+};
+
 struct Cos {
   template <typename T>
   T operator()(T x) {
@@ -241,6 +254,13 @@ struct Exp {
   }
 };
 
+struct Expm1 {
+  template <typename T>
+  T operator()(T x) {
+    return expm1(x);
+  };
+};
+
 struct Floor {
   template <typename T>
   T operator()(T x) {
@@ -599,4 +619,39 @@ 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>
@@ -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];
@@ -524,63 +412,6 @@ 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];
@@ -651,49 +482,6 @@ 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) {
@@ -725,18 +513,6 @@ void Slice::eval(const std::vector<array>& inputs, array& 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 +550,6 @@ 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 +566,6 @@ 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 +590,4 @@ void Tanh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-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/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/softmax.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <cmath>
@@ -10,7 +10,7 @@ namespace mlx::core {
 
 namespace {
 
-template <typename T>
+template <typename T, typename AccT>
 void softmax(const array& in, array& out) {
   const T* in_ptr = in.data<T>();
   T* out_ptr = out.data<T>();
@@ -22,26 +22,36 @@ void softmax(const array& in, array& out) {
   for (int i = 0; i < M; i++, in_ptr += N, out_ptr += N) {
     // Find the maximum
     current_in_ptr = in_ptr;
-    T maximum = *current_in_ptr;
+    AccT maximum = *current_in_ptr;
     for (int j = 0; j < N; j++, current_in_ptr++) {
-      maximum = (maximum < *current_in_ptr) ? *current_in_ptr : maximum;
+      maximum = (maximum < *current_in_ptr) ? static_cast<AccT>(*current_in_ptr)
+                                            : maximum;
     }
 
     // Compute the normalizer and the exponentials
-    T normalizer = 0;
+    AccT normalizer = 0;
     current_out_ptr = out_ptr;
     current_in_ptr = in_ptr;
     for (int j = 0; j < N; j++, current_out_ptr++, current_in_ptr++) {
-      T expv = std::exp(*current_in_ptr - maximum);
+      AccT expv = std::exp(*current_in_ptr - maximum);
       normalizer += expv;
-      *current_out_ptr = expv;
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr = expv;
+      }
     }
     normalizer = 1 / normalizer;
 
     // Normalize
+    current_in_ptr = in_ptr;
     current_out_ptr = out_ptr;
     for (int j = 0; j < N; j++, current_out_ptr++) {
-      *current_out_ptr *= normalizer;
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr *= normalizer;
+      } else {
+        auto v = std::exp(*current_in_ptr - maximum);
+        *current_out_ptr = static_cast<T>(v * normalizer);
+        current_in_ptr++;
+      }
     }
   }
 }
@@ -91,13 +101,21 @@ void Softmax::eval(const std::vector<array>& inputs, array& out) {
           "Softmax is defined only for floating point types");
       break;
     case float32:
-      softmax<float>(in, out);
+      softmax<float, float>(in, out);
       break;
     case float16:
-      softmax<float16_t>(in, out);
+      if (precise_) {
+        softmax<float16_t, float>(in, out);
+      } else {
+        softmax<float16_t, float16_t>(in, out);
+      }
       break;
     case bfloat16:
-      softmax<bfloat16_t>(in, out);
+      if (precise_) {
+        softmax<bfloat16_t, float>(in, out);
+      } else {
+        softmax<bfloat16_t, bfloat16_t>(in, out);
+      }
       break;
     case complex64:
       throw std::invalid_argument(

mlx/backend/common/svd.cpp

@@ -3,7 +3,6 @@
 #include "mlx/allocator.h"
 #include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/lapack_helper.h"
-#include "mlx/linalg.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
@@ -145,12 +144,4 @@ void SVD::eval(const std::vector<array>& inputs, std::vector<array>& outputs) {
   svd_impl(inputs[0], outputs[0], outputs[1], outputs[2]);
 }
 
-std::pair<std::vector<array>, std::vector<int>> SVD::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::svd(a, stream())}, {ax, ax, ax}};
-}
-
 } // namespace mlx::core

mlx/backend/metal/CMakeLists.txt

@@ -1,31 +1,130 @@
-add_custom_command(
-    OUTPUT  compiled_preamble.cpp
+function(make_jit_source SRC_FILE)
+  # This function takes a metal header file,
+  # runs the C preprocessesor on it, and makes
+  # the processed contents available as a string in a C++ function
+  # mlx::core::metal::${SRC_NAME}()
+  #
+  # To use the function, declare it in jit/includes.h and
+  # include jit/includes.h.
+  #
+  # Additional arguments to this function are treated as dependencies
+  # in the Cmake build system.
+  get_filename_component(SRC_NAME ${SRC_FILE} NAME)
+  add_custom_command(
+    OUTPUT  jit/${SRC_NAME}.cpp
     COMMAND /bin/bash
               ${CMAKE_CURRENT_SOURCE_DIR}/make_compiled_preamble.sh
-              ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp
+              ${CMAKE_CURRENT_BINARY_DIR}/jit
               ${CMAKE_C_COMPILER}
               ${PROJECT_SOURCE_DIR}
+              ${SRC_FILE}
+              "-D${MLX_METAL_VERSION}"
     DEPENDS make_compiled_preamble.sh
-            kernels/compiled_preamble.h
-            kernels/unary.h
-            kernels/binary.h
-)
+            kernels/${SRC_FILE}.h
+            ${ARGN}
+  )
+  add_custom_target(${SRC_NAME} DEPENDS jit/${SRC_NAME}.cpp)
+  add_dependencies(mlx ${SRC_NAME})
+  target_sources(
+    mlx
+    PRIVATE
+    ${CMAKE_CURRENT_BINARY_DIR}/jit/${SRC_NAME}.cpp
+  )
+endfunction(make_jit_source)
 
-add_custom_target(
-  compiled_preamble
-  DEPENDS compiled_preamble.cpp
+make_jit_source(
+  utils
+  kernels/bf16.h
+  kernels/complex.h
+  kernels/defines.h
 )
+make_jit_source(
+  unary_ops
+  kernels/erf.h
+  kernels/expm1f.h
+)
+make_jit_source(binary_ops)
+make_jit_source(ternary_ops)
+make_jit_source(
+  reduce_utils
+  kernels/atomic.h
+  kernels/reduction/ops.h
+)
+make_jit_source(scatter)
+make_jit_source(gather)
 
-add_dependencies(mlx compiled_preamble)
+if (MLX_METAL_JIT) 
+  target_sources(
+    mlx
+    PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/jit_kernels.cpp
+  )
+  make_jit_source(arange)
+  make_jit_source(copy)
+  make_jit_source(unary)
+  make_jit_source(binary)
+  make_jit_source(binary_two)
+  make_jit_source(ternary)
+  make_jit_source(softmax)
+  make_jit_source(scan)
+  make_jit_source(sort)
+  make_jit_source(
+    reduce
+    kernels/reduction/reduce_all.h
+    kernels/reduction/reduce_col.h
+    kernels/reduction/reduce_row.h
+  )
+  make_jit_source(
+    steel/gemm/gemm
+    kernels/steel/utils.h
+    kernels/steel/gemm/loader.h
+    kernels/steel/gemm/mma.h
+    kernels/steel/gemm/params.h
+    kernels/steel/gemm/transforms.h
+  )
+  make_jit_source(steel/gemm/kernels/steel_gemm_fused)
+  make_jit_source(
+    steel/gemm/kernels/steel_gemm_masked
+    kernels/steel/defines.h
+  )
+  make_jit_source(steel/gemm/kernels/steel_gemm_splitk)
+  make_jit_source(
+    steel/conv/conv
+    kernels/steel/utils.h
+    kernels/steel/defines.h
+    kernels/steel/gemm/mma.h
+    kernels/steel/gemm/transforms.h
+    kernels/steel/conv/params.h
+    kernels/steel/conv/loader.h
+    kernels/steel/conv/loaders/loader_channel_l.h
+    kernels/steel/conv/loaders/loader_channel_n.h
+  )
+  make_jit_source(
+    steel/conv/kernels/steel_conv
+  )
+  make_jit_source(
+    steel/conv/kernels/steel_conv_general
+    kernels/steel/defines.h
+    kernels/steel/conv/loaders/loader_general.h
+  )
+else()
+  target_sources(
+    mlx
+    PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/nojit_kernels.cpp
+  )
+endif()
 
 target_sources(
   mlx
   PRIVATE
   ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/event.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
@@ -39,7 +138,8 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
-  ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/unary.cpp
 )
 
 if (NOT MLX_METAL_PATH)

mlx/backend/metal/allocator.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2023-2024 Apple Inc.
 #include "mlx/backend/metal/allocator.h"
 #include "mlx/backend/metal/metal.h"
+#include "mlx/backend/metal/metal_impl.h"
 
 #include <mach/vm_page_size.h>
 #include <unistd.h>
@@ -139,10 +140,15 @@ void BufferCache::remove_from_list(BufferCache::BufferHolder* to_remove) {
 
 MetalAllocator::MetalAllocator()
     : device_(device(mlx::core::Device::gpu).mtl_device()),
-      buffer_cache_(device_),
-      block_limit_(1.5 * device_->recommendedMaxWorkingSetSize()),
-      gc_limit_(0.95 * device_->recommendedMaxWorkingSetSize()),
-      max_pool_size_(block_limit_) {}
+      buffer_cache_(device_) {
+  auto memsize = std::get<size_t>(device_info()["memory_size"]);
+  block_limit_ =
+      std::min(1.5 * device_->recommendedMaxWorkingSetSize(), 0.95 * memsize);
+  gc_limit_ = std::min(
+      static_cast<size_t>(0.95 * device_->recommendedMaxWorkingSetSize()),
+      block_limit_);
+  max_pool_size_ = block_limit_;
+}
 
 size_t MetalAllocator::set_cache_limit(size_t limit) {
   std::swap(limit, max_pool_size_);
@@ -164,6 +170,15 @@ Buffer MetalAllocator::malloc(size_t size, bool allow_swap /* = false */) {
     return Buffer{nullptr};
   }
 
+  // More helpful message if maximum buffer length is exceeded
+  if (size > device_->maxBufferLength()) {
+    std::ostringstream msg;
+    msg << "Attempting to allocate " << size << " bytes which is greater than"
+        << " the maximum allowed buffer size of " << device_->maxBufferLength()
+        << " bytes.";
+    throw std::runtime_error(msg.str());
+  }
+
   // Align up memory
   if (size > vm_page_size) {
     size = vm_page_size * ((size + vm_page_size - 1) / vm_page_size);
@@ -208,6 +223,11 @@ Buffer MetalAllocator::malloc(size_t size, bool allow_swap /* = false */) {
   return Buffer{static_cast<void*>(buf)};
 }
 
+void MetalAllocator::clear_cache() {
+  std::unique_lock lk(mutex_);
+  buffer_cache_.clear();
+}
+
 void MetalAllocator::free(Buffer buffer) {
   auto buf = static_cast<MTL::Buffer*>(buffer.ptr());
   std::unique_lock lk(mutex_);
@@ -238,9 +258,15 @@ size_t get_active_memory() {
 size_t get_peak_memory() {
   return allocator().get_peak_memory();
 }
+void reset_peak_memory() {
+  allocator().reset_peak_memory();
+}
 size_t get_cache_memory() {
   return allocator().get_cache_memory();
 }
+void clear_cache() {
+  return allocator().clear_cache();
+}
 
 } // namespace metal
 

mlx/backend/metal/allocator.h

@@ -26,6 +26,7 @@ class BufferCache {
   size_t cache_size() {
     return pool_size_;
   }
+  void clear();
 
  private:
   struct BufferHolder {
@@ -37,7 +38,6 @@ class BufferCache {
     MTL::Buffer* buf;
   };
 
-  void clear();
   void add_at_head(BufferHolder* to_add);
   void remove_from_list(BufferHolder* to_remove);
 
@@ -62,11 +62,16 @@ class MetalAllocator : public allocator::Allocator {
   size_t get_peak_memory() {
     return peak_memory_;
   };
+  void reset_peak_memory() {
+    std::unique_lock lk(mutex_);
+    peak_memory_ = 0;
+  };
   size_t get_cache_memory() {
     return buffer_cache_.cache_size();
   };
   size_t set_cache_limit(size_t limit);
   size_t set_memory_limit(size_t limit, bool relaxed);
+  void clear_cache();
 
  private:
   MTL::Device* device_;

mlx/backend/metal/binary.cpp

@@ -0,0 +1,322 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/kernels.h"
+#include "mlx/backend/metal/utils.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+constexpr int MAX_BINARY_SPECIALIZED_DIMS = 5;
+
+void binary_op(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::string op) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, outputs[0], bopt, true);
+  set_binary_op_output_data(a, b, outputs[1], bopt, true);
+
+  auto& out = outputs[0];
+  if (out.size() == 0) {
+    return;
+  }
+
+  // Try to collapse contiguous dims
+  auto [shape, strides] = collapse_contiguous_dims(a, b, out);
+  auto& strides_a = strides[0];
+  auto& strides_b = strides[1];
+  auto& strides_out = strides[2];
+
+  std::string kernel_name;
+  {
+    std::ostringstream kname;
+    switch (bopt) {
+      case BinaryOpType::ScalarScalar:
+        kname << "ss";
+        break;
+      case BinaryOpType::ScalarVector:
+        kname << "sv";
+        break;
+      case BinaryOpType::VectorScalar:
+        kname << "vs";
+        break;
+      case BinaryOpType::VectorVector:
+        kname << "vv";
+        break;
+      case BinaryOpType::General:
+        kname << "g";
+        if (shape.size() <= MAX_BINARY_SPECIALIZED_DIMS) {
+          kname << shape.size();
+        } else {
+          kname << "n";
+        }
+        break;
+    }
+    kname << op << type_to_name(a);
+    kernel_name = kname.str();
+  }
+
+  auto& s = out.primitive().stream();
+  auto& d = metal::device(s.device);
+
+  auto kernel = get_binary_two_kernel(d, kernel_name, a, outputs[0]);
+
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder->setComputePipelineState(kernel);
+
+  // - If a is donated it goes to the first output
+  // - If b is donated it goes to the first output if a was not donated
+  //   otherwise it goes to the second output
+  bool donate_a = a.data_shared_ptr() == nullptr;
+  bool donate_b = b.data_shared_ptr() == nullptr;
+  compute_encoder.set_input_array(donate_a ? outputs[0] : a, 0);
+  compute_encoder.set_input_array(
+      donate_b ? (donate_a ? outputs[1] : outputs[0]) : b, 1);
+  compute_encoder.set_output_array(outputs[0], 2);
+  compute_encoder.set_output_array(outputs[1], 3);
+
+  if (bopt == BinaryOpType::General) {
+    auto ndim = shape.size();
+    if (ndim > 3) {
+      compute_encoder->setBytes(shape.data(), ndim * sizeof(int), 4);
+      compute_encoder->setBytes(strides_a.data(), ndim * sizeof(size_t), 5);
+      compute_encoder->setBytes(strides_b.data(), ndim * sizeof(size_t), 6);
+    } else {
+      // The shape is implicit in the grid for <= 3D
+      compute_encoder->setBytes(strides_a.data(), ndim * sizeof(size_t), 4);
+      compute_encoder->setBytes(strides_b.data(), ndim * sizeof(size_t), 5);
+    }
+
+    if (ndim > MAX_BINARY_SPECIALIZED_DIMS) {
+      compute_encoder->setBytes(&ndim, sizeof(int), 7);
+    }
+
+    // Launch up to 3D grid of threads
+    size_t dim0 = ndim > 0 ? shape[ndim - 1] : 1;
+    size_t dim1 = ndim > 1 ? shape[ndim - 2] : 1;
+    size_t rest = out.size() / (dim0 * dim1);
+    NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
+    if (thread_group_size != 1024) {
+      throw std::runtime_error("[Metal::binary] Must use 1024 sized block");
+    }
+    auto group_dims = get_block_dims(dim0, dim1, rest);
+    MTL::Size grid_dims = MTL::Size(dim0, dim1, rest);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
+  } else {
+    // Launch a 1D grid of threads
+    size_t nthreads = out.data_size();
+    MTL::Size grid_dims = MTL::Size(nthreads, 1, 1);
+    NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
+    if (thread_group_size > nthreads) {
+      thread_group_size = nthreads;
+    }
+    MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
+  }
+}
+
+void binary_op(
+    const std::vector<array>& inputs,
+    array& out,
+    const std::string op) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt, true);
+  if (out.size() == 0) {
+    return;
+  }
+
+  // Try to collapse contiguous dims
+  auto [shape, strides] = collapse_contiguous_dims(a, b, out);
+  auto& strides_a = strides[0];
+  auto& strides_b = strides[1];
+  auto& strides_out = strides[2];
+
+  std::string kernel_name;
+  {
+    std::ostringstream kname;
+    switch (bopt) {
+      case BinaryOpType::ScalarScalar:
+        kname << "ss";
+        break;
+      case BinaryOpType::ScalarVector:
+        kname << "sv";
+        break;
+      case BinaryOpType::VectorScalar:
+        kname << "vs";
+        break;
+      case BinaryOpType::VectorVector:
+        kname << "vv";
+        break;
+      case BinaryOpType::General:
+        kname << "g";
+        if (shape.size() <= MAX_BINARY_SPECIALIZED_DIMS) {
+          kname << shape.size();
+        } else {
+          kname << "n";
+        }
+        break;
+    }
+    kname << op << type_to_name(a);
+    kernel_name = kname.str();
+  }
+
+  auto& s = out.primitive().stream();
+  auto& d = metal::device(s.device);
+
+  auto kernel = get_binary_kernel(d, kernel_name, a, out);
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder->setComputePipelineState(kernel);
+  bool donate_a = a.data_shared_ptr() == nullptr;
+  bool donate_b = b.data_shared_ptr() == nullptr;
+  compute_encoder.set_input_array(donate_a ? out : a, 0);
+  compute_encoder.set_input_array(donate_b ? out : b, 1);
+  compute_encoder.set_output_array(out, 2);
+
+  if (bopt == BinaryOpType::General) {
+    auto ndim = shape.size();
+    if (ndim > 3) {
+      compute_encoder->setBytes(shape.data(), ndim * sizeof(int), 3);
+      compute_encoder->setBytes(strides_a.data(), ndim * sizeof(size_t), 4);
+      compute_encoder->setBytes(strides_b.data(), ndim * sizeof(size_t), 5);
+    } else {
+      // The shape is implicit in the grid for <= 3D
+      compute_encoder->setBytes(strides_a.data(), ndim * sizeof(size_t), 3);
+      compute_encoder->setBytes(strides_b.data(), ndim * sizeof(size_t), 4);
+    }
+
+    if (ndim > MAX_BINARY_SPECIALIZED_DIMS) {
+      compute_encoder->setBytes(&ndim, sizeof(int), 6);
+    }
+
+    // Launch up to 3D grid of threads
+    size_t dim0 = ndim > 0 ? shape[ndim - 1] : 1;
+    size_t dim1 = ndim > 1 ? shape[ndim - 2] : 1;
+    size_t rest = out.size() / (dim0 * dim1);
+    NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
+    if (thread_group_size != 1024) {
+      throw std::runtime_error("[Metal::binary] Must use 1024 sized block");
+    }
+    auto group_dims = get_block_dims(dim0, dim1, rest);
+    MTL::Size grid_dims = MTL::Size(dim0, dim1, rest);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
+  } else {
+    // Launch a 1D grid of threads
+    size_t nthreads =
+        bopt == BinaryOpType::General ? out.size() : out.data_size();
+    MTL::Size grid_dims = MTL::Size(nthreads, 1, 1);
+    NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
+    if (thread_group_size > nthreads) {
+      thread_group_size = nthreads;
+    }
+    MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
+  }
+}
+
+void Add::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "add");
+}
+
+void ArcTan2::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "arctan2");
+}
+
+void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_op(inputs, out, "bitwise_and");
+      break;
+    case BitwiseBinary::Or:
+      binary_op(inputs, out, "bitwise_or");
+      break;
+    case BitwiseBinary::Xor:
+      binary_op(inputs, out, "bitwise_xor");
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_op(inputs, out, "left_shift");
+      break;
+    case BitwiseBinary::RightShift:
+      binary_op(inputs, out, "right_shift");
+      break;
+  }
+}
+
+void Divide::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "div");
+}
+
+void DivMod::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  binary_op(inputs, outputs, "divmod");
+}
+
+void Remainder::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "rem");
+}
+
+void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, equal_nan_ ? "naneq" : "eq");
+}
+
+void Greater::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "ge");
+}
+
+void GreaterEqual::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "geq");
+}
+
+void Less::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "le");
+}
+
+void LessEqual::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "leq");
+}
+
+void LogicalAnd::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "land");
+}
+
+void LogicalOr::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "lor");
+}
+
+void LogAddExp::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "lae");
+}
+
+void Maximum::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "max");
+}
+
+void Minimum::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "min");
+}
+
+void Multiply::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "mul");
+}
+
+void NotEqual::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "neq");
+}
+
+void Power::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "pow");
+}
+
+void Subtract::eval_gpu(const std::vector<array>& inputs, array& out) {
+  binary_op(inputs, out, "sub");
+}
+
+} // namespace mlx::core

mlx/backend/metal/compiled.cpp

@@ -4,8 +4,8 @@
 
 #include "mlx/backend/common/compiled.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/metal/compiled_preamble.h"
 #include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/jit/includes.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/graph_utils.h"
 #include "mlx/primitives.h"
@@ -190,7 +190,8 @@ void Compiled::eval_gpu(
   // If not we have to build it ourselves
   if (lib == nullptr) {
     std::ostringstream kernel;
-    kernel << metal::get_kernel_preamble() << std::endl;
+    kernel << metal::utils() << metal::unary_ops() << metal::binary_ops()
+           << metal::ternary_ops();
     build_kernel(
         kernel,
         kernel_lib_ + "_contiguous",
@@ -229,14 +230,7 @@ void Compiled::eval_gpu(
 
   // Figure out which kernel we are using
   auto& output_shape = outputs[0].shape();
-  bool contiguous = true;
-  for (auto& x : inputs) {
-    if ((!x.flags().row_contiguous || x.shape() != output_shape) &&
-        !is_scalar(x)) {
-      contiguous = false;
-      break;
-    }
-  }
+  bool contiguous = compiled_check_contiguity(inputs, output_shape);
 
   // Collapse contiguous dims to route to a faster kernel if possible. Also
   // handle all broadcasting.
@@ -296,7 +290,7 @@ void Compiled::eval_gpu(
     }
   }
   auto kernel = d.get_kernel(kernel_name, lib);
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
 
   // Put the inputs in
@@ -307,7 +301,7 @@ void Compiled::eval_gpu(
       continue;
     }
     auto& x = inputs[i];
-    set_array_buffer(compute_encoder, x, cnt++);
+    compute_encoder.set_input_array(x, cnt++);
     if (!contiguous && !is_scalar(x)) {
       compute_encoder->setBytes(
           strides[stride_idx].data(),
@@ -317,32 +311,12 @@ void Compiled::eval_gpu(
     }
   }
 
-  // Allocate space for the outputs possibly with input donation
-  {
-    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].move_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, true);
 
   // Put the outputs in
   for (auto& x : outputs) {
-    set_array_buffer(compute_encoder, x, cnt++);
+    compute_encoder.set_output_array(x, cnt++);
   }
 
   // Put the output shape and strides in
@@ -363,7 +337,7 @@ void Compiled::eval_gpu(
     MTL::Size grid_dims(nthreads, 1, 1);
     MTL::Size group_dims(
         std::min(nthreads, kernel->maxTotalThreadsPerThreadgroup()), 1, 1);
-    compute_encoder->dispatchThreads(grid_dims, group_dims);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
   } else {
     size_t dim0 = ndim > 0 ? shape[ndim - 1] : 1;
     size_t dim1 = ndim > 1 ? shape[ndim - 2] : 1;
@@ -374,7 +348,7 @@ void Compiled::eval_gpu(
     }
     auto group_dims = get_block_dims(dim0, dim1, rest);
     MTL::Size grid_dims = MTL::Size(dim0, dim1, rest);
-    compute_encoder->dispatchThreads(grid_dims, group_dims);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
   }
 }
 

mlx/backend/metal/compiled_preamble.h

@@ -1,9 +0,0 @@
-// Copyright © 2023-24 Apple Inc.
-
-#pragma once
-
-namespace mlx::core::metal {
-
-const char* get_kernel_preamble();
-
-}

mlx/backend/metal/conv.cpp

@@ -7,6 +7,7 @@
 
 #include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/kernels/defines.h"
 #include "mlx/backend/metal/kernels/steel/conv/params.h"
 #include "mlx/backend/metal/matmul.h"
@@ -41,12 +42,12 @@ void explicit_gemm_conv_ND_gpu(
   // Prepare unfolding kernel
   std::ostringstream kname;
   kname << "naive_unfold_nd_" << type_to_name(in_unfolded) << "_" << N;
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, in_unfolded, 1);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_output_array(in_unfolded, 1);
 
   compute_encoder->setBytes(&conv_params, sizeof(conv_params), 2);
 
@@ -59,7 +60,7 @@ void explicit_gemm_conv_ND_gpu(
   MTL::Size grid_dims = MTL::Size(
       conv_params.C, unfolded_shape[1] / conv_params.C, unfolded_shape[0]);
 
-  compute_encoder->dispatchThreads(grid_dims, group_dims);
+  compute_encoder.dispatchThreads(grid_dims, group_dims);
 
   // Reshape weight
   std::vector<int> wt_reshape{implicit_K, implicit_N};
@@ -89,6 +90,90 @@ void explicit_gemm_conv_ND_gpu(
       /*copies = */ copies);
 }
 
+template <int N>
+void explicit_gemm_conv_group_ND_gpu(
+    const Stream& s,
+    metal::Device& d,
+    const array& in,
+    const array& wt,
+    array out,
+    const MLXConvParams<N>& conv_params) {
+  const int groups = conv_params.groups;
+  const int C_per_group = conv_params.C / conv_params.groups;
+  const int O_per_group = conv_params.O / conv_params.groups;
+  // Get gemm shapes
+  const int implicit_M = out.size() / conv_params.O;
+  const int implicit_K = wt.size() / conv_params.O;
+  const int implicit_N = O_per_group;
+
+  int kernel_size = 1;
+  for (int i = 0; i < N; ++i) {
+    kernel_size *= conv_params.wS[i];
+  }
+
+  // Prepare unfolding array
+  std::vector<int> unfolded_shape{implicit_M, implicit_K * groups};
+  array in_unfolded(unfolded_shape, in.dtype(), nullptr, {});
+  in_unfolded.set_data(allocator::malloc_or_wait(in_unfolded.nbytes()));
+
+  // Prepare unfolding kernel
+  std::ostringstream kname;
+  kname << "naive_unfold_transpose_nd_" << type_to_name(in_unfolded) << "_"
+        << N;
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = d.get_kernel(kname.str());
+  compute_encoder->setComputePipelineState(kernel);
+
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_output_array(in_unfolded, 1);
+
+  compute_encoder->setBytes(&conv_params, sizeof(conv_params), 2);
+
+  // Launch unfolding kernel
+  int tgp_x = std::min(conv_params.C, 64);
+  tgp_x = 32 * ((tgp_x + 32 - 1) / 32);
+  int tgp_y = 256 / tgp_x;
+
+  MTL::Size group_dims = MTL::Size(tgp_x, tgp_y, 1);
+  MTL::Size grid_dims = MTL::Size(
+      conv_params.C, unfolded_shape[1] / conv_params.C, unfolded_shape[0]);
+
+  compute_encoder.dispatchThreads(grid_dims, group_dims);
+
+  // Transpose kernel weights so that we can slice them by contiguous chunks
+  // of channel groups.
+  array wt_view(
+      {wt.shape(0), C_per_group, kernel_size}, wt.dtype(), nullptr, {});
+  wt_view.copy_shared_buffer(
+      wt,
+      {wt.strides(0), 1, static_cast<size_t>(C_per_group)},
+      wt.flags(),
+      wt.size());
+
+  // Materialize
+  auto wt_transpose = array(wt_view.shape(), wt_view.dtype(), nullptr, {});
+  copy_gpu(wt_view, wt_transpose, CopyType::General, s);
+
+  // Perform gemm
+  std::vector<array> copies = {in_unfolded, wt_view, wt_transpose};
+  return steel_matmul_conv_groups(
+      s,
+      d,
+      /*a = */ in_unfolded,
+      /*b = */ wt_transpose,
+      /*c = */ out,
+      /*M = */ implicit_M,
+      /*N = */ implicit_N,
+      /*K = */ implicit_K,
+      /*a_cols = */ implicit_K * groups,
+      /*b_cols = */ implicit_K,
+      /*out_cols = */ implicit_N * groups,
+      /*a_transposed = */ false,
+      /*b_transposed = */ true,
+      /* groups = */ groups,
+      /*copies = */ copies);
+}
+
 void conv_1D_gpu(
     const Stream& s,
     metal::Device& d,
@@ -99,6 +184,7 @@ void conv_1D_gpu(
     const std::vector<int>& wt_strides,
     const std::vector<int>& wt_dilation,
     const std::vector<int>& in_dilation,
+    int groups,
     bool flip) {
   // Make conv params
   MLXConvParams<1> conv_params{
@@ -118,11 +204,15 @@ void conv_1D_gpu(
       {wt.strides()[0], wt.strides()[1], wt.strides()[2]},
       /* const size_t out_strides[NDIM + 2] = */
       {out.strides()[0], out.strides()[1], out.strides()[2]},
-      /* const int groups = */ 1,
+      /* const int groups = */ groups,
       /* const bool flip = */ flip};
 
   // Direct to explicit gemm conv
-  return explicit_gemm_conv_ND_gpu(s, d, in, wt, out, conv_params);
+  if (groups > 1) {
+    return explicit_gemm_conv_group_ND_gpu(s, d, in, wt, out, conv_params);
+  } else {
+    return explicit_gemm_conv_ND_gpu(s, d, in, wt, out, conv_params);
+  }
 }
 
 void slow_conv_2D_gpu(
@@ -140,7 +230,7 @@ void slow_conv_2D_gpu(
         << "_tm" << tm << "_tn" << tn;
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
@@ -153,12 +243,12 @@ void slow_conv_2D_gpu(
   MTL::Size group_dims = MTL::Size(bm, bn, 1);
   MTL::Size grid_dims = MTL::Size(grid_dim_x, grid_dim_y, grid_dim_z);
 
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, wt, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
 
   compute_encoder->setBytes(&conv_params, sizeof(MLXConvParams<2>), 3);
-  compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+  compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
 }
 
 void implicit_gemm_conv_2D_gpu(
@@ -168,15 +258,19 @@ void implicit_gemm_conv_2D_gpu(
     const array& wt,
     array out,
     const MLXConvParams<2>& conv_params) {
+  const int groups = conv_params.groups;
+  const int C_per_group = conv_params.C / conv_params.groups;
+  const int O_per_group = conv_params.O / conv_params.groups;
+
   // Deduce implicit gemm size
-  int implicit_M = conv_params.N * conv_params.oS[0] * conv_params.oS[1];
-  int implicit_N = conv_params.O;
-  int implicit_K = conv_params.wS[0] * conv_params.wS[1] * conv_params.C;
+  const int implicit_M = conv_params.N * conv_params.oS[0] * conv_params.oS[1];
+  const int implicit_N = O_per_group;
+  const int implicit_K = conv_params.wS[0] * conv_params.wS[1] * C_per_group;
 
   // Determine block and warp tiles
   int wm = 2, wn = 2;
 
-  int bm = implicit_M >= 8192 && conv_params.C >= 64 ? 64 : 32;
+  int bm = implicit_M >= 8192 && C_per_group >= 64 ? 64 : 32;
   int bn = (bm == 64 || implicit_N >= 64) ? 64 : 32;
   int bk = 16;
 
@@ -192,15 +286,15 @@ void implicit_gemm_conv_2D_gpu(
 
   // Fix small channel specialization
   int n_channel_specialization = 0;
-  int channel_k_iters = ((conv_params.C + bk - 1) / bk);
+  int channel_k_iters = ((C_per_group + bk - 1) / bk);
   int gemm_k_iters = conv_params.wS[0] * conv_params.wS[1] * channel_k_iters;
 
-  if (conv_params.C <= 2) {
+  if (C_per_group <= 2) {
     gemm_k_iters = (implicit_K + bk - 1) / bk;
-    n_channel_specialization = conv_params.C;
-  } else if (conv_params.C <= 4) {
+    n_channel_specialization = C_per_group;
+  } else if (C_per_group <= 4) {
     gemm_k_iters = ((conv_params.wS[0] * conv_params.wS[1] * 4) + bk - 1) / bk;
-    n_channel_specialization = conv_params.C;
+    n_channel_specialization = C_per_group;
   }
 
   bool small_filter = (!n_channel_specialization) &&
@@ -241,8 +335,18 @@ void implicit_gemm_conv_2D_gpu(
         << "_filter_" << (small_filter ? 's' : 'l');
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = get_steel_conv_kernel(
+      d,
+      kname.str(),
+      out,
+      bm,
+      bn,
+      bk,
+      wm,
+      wn,
+      n_channel_specialization,
+      small_filter);
   compute_encoder->setComputePipelineState(kernel);
 
   // Deduce grid launch dimensions
@@ -251,19 +355,19 @@ void implicit_gemm_conv_2D_gpu(
   size_t grid_dim_x = tn * tile;
 
   MTL::Size group_dims = MTL::Size(32, wn, wm);
-  MTL::Size grid_dims = MTL::Size(grid_dim_x, grid_dim_y, 1);
+  MTL::Size grid_dims = MTL::Size(grid_dim_x, grid_dim_y, groups);
 
   // Encode arrays
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, wt, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode params
   compute_encoder->setBytes(&conv_params, sizeof(MLXConvParams<2>), 3);
   compute_encoder->setBytes(&gemm_params, sizeof(ImplicitGemmConv2DParams), 4);
 
   // Launch kernel
-  compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+  compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
 }
 
 void implicit_gemm_conv_2D_general_gpu(
@@ -394,8 +498,9 @@ void implicit_gemm_conv_2D_general_gpu(
         << "_bn" << bn << "_bk" << bk << "_wm" << wm << "_wn" << wn;
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel =
+      get_steel_conv_general_kernel(d, kname.str(), out, bm, bn, bk, wm, wn);
   compute_encoder->setComputePipelineState(kernel);
 
   // Deduce grid launch dimensions
@@ -408,9 +513,9 @@ void implicit_gemm_conv_2D_general_gpu(
   MTL::Size grid_dims = MTL::Size(grid_dim_x, grid_dim_y, grid_dim_z);
 
   // Encode arrays
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, wt, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode params
   compute_encoder->setBytes(&conv_params, sizeof(MLXConvParams<2>), 3);
@@ -423,7 +528,7 @@ void implicit_gemm_conv_2D_general_gpu(
       base_w.data(), sizeof(Conv2DGeneralBaseInfo) * base_w.size(), 7);
 
   // Launch kernel
-  compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+  compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
 }
 
 void winograd_conv_2D_gpu(
@@ -511,12 +616,12 @@ void winograd_conv_2D_gpu(
     std::ostringstream kname;
     kname << "winograd_conv_2d_weight_transform_" << type_to_name(out) << "_bc"
           << bc;
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
-    set_array_buffer(compute_encoder, wt, 0);
-    set_array_buffer(compute_encoder, filt_wg, 1);
+    compute_encoder.set_input_array(wt, 0);
+    compute_encoder.set_output_array(filt_wg, 1);
 
     compute_encoder->setBytes(&C_c, sizeof(int), 2);
     compute_encoder->setBytes(&O_c, sizeof(int), 3);
@@ -524,7 +629,7 @@ void winograd_conv_2D_gpu(
     MTL::Size group_dims = MTL::Size(32, bo, 1);
     MTL::Size grid_dims = MTL::Size(O_c / bo, 1, 1);
 
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+    compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
   }
 
   // Do input transform
@@ -539,12 +644,12 @@ void winograd_conv_2D_gpu(
     std::ostringstream kname;
     kname << "winograd_conv_2d_input_transform_" << type_to_name(out) << "_bc"
           << bc;
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
-    set_array_buffer(compute_encoder, in_padded, 0);
-    set_array_buffer(compute_encoder, inp_wg, 1);
+    compute_encoder.set_input_array(in_padded, 0);
+    compute_encoder.set_output_array(inp_wg, 1);
 
     compute_encoder->setBytes(
         &conv_params_updated, sizeof(MLXConvParams<2>), 2);
@@ -552,7 +657,7 @@ void winograd_conv_2D_gpu(
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(N_tiles_w, N_tiles_h, N_tiles_n);
 
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+    compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
   }
 
   // Do batched gemm
@@ -587,12 +692,12 @@ void winograd_conv_2D_gpu(
     std::ostringstream kname;
     kname << "winograd_conv_2d_output_transform_" << type_to_name(out) << "_bo"
           << bc;
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
-    set_array_buffer(compute_encoder, out_wg, 0);
-    set_array_buffer(compute_encoder, out, 1);
+    compute_encoder.set_input_array(out_wg, 0);
+    compute_encoder.set_output_array(out, 1);
 
     compute_encoder->setBytes(
         &conv_params_updated, sizeof(MLXConvParams<2>), 2);
@@ -600,7 +705,7 @@ void winograd_conv_2D_gpu(
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(N_tiles_w, N_tiles_h, N_tiles_n);
 
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+    compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
   }
 }
 
@@ -614,6 +719,7 @@ void conv_2D_gpu(
     const std::vector<int>& wt_strides,
     const std::vector<int>& wt_dilation,
     const std::vector<int>& in_dilation,
+    const int groups,
     bool flip,
     std::vector<array>& copies) {
   // Make conv params
@@ -629,12 +735,12 @@ void conv_2D_gpu(
       /* const int kdil[NDIM] = */ {wt_dilation[0], wt_dilation[1]},
       /* const int idil[NDIM] = */ {in_dilation[0], in_dilation[1]},
       /* const size_t in_strides[NDIM + 2] = */
-      {in.strides()[0], in.strides()[1], in.strides()[2], in.strides()[3]},
+      {in.strides(0), in.strides(1), in.strides(2), in.strides(3)},
       /* const size_t wt_strides[NDIM + 2] = */
-      {wt.strides()[0], wt.strides()[1], wt.strides()[2], wt.strides()[3]},
+      {wt.strides(0), wt.strides(1), wt.strides(2), wt.strides(3)},
       /* const size_t out_strides[NDIM + 2] = */
-      {out.strides()[0], out.strides()[1], out.strides()[2], out.strides()[3]},
-      /* const int groups = */ 1,
+      {out.strides(0), out.strides(1), out.strides(2), out.strides(3)},
+      /* const int groups = */ groups,
       /* const bool flip = */ flip,
   };
 
@@ -646,6 +752,18 @@ void conv_2D_gpu(
   bool channels_large = (conv_params.C + conv_params.O) >= 512;
   bool channels_med = (conv_params.C + conv_params.O) >= 256;
 
+  if (groups > 1) {
+    const int C_per_group = conv_params.C / groups;
+    const int O_per_group = conv_params.O / groups;
+
+    if (is_idil_one && (C_per_group <= 4 || C_per_group % 16 == 0) &&
+        (O_per_group <= 16 || O_per_group % 16 == 0)) {
+      return implicit_gemm_conv_2D_gpu(s, d, in, wt, out, conv_params);
+    } else {
+      return explicit_gemm_conv_group_ND_gpu(s, d, in, wt, out, conv_params);
+    }
+  }
+
   // Direct to winograd conv
   if (!flip && is_stride_one && is_kdil_one && is_idil_one &&
       conv_params.wS[0] == 3 && conv_params.wS[1] == 3 &&
@@ -670,6 +788,56 @@ void conv_2D_gpu(
   }
 }
 
+void conv_3D_gpu(
+    const Stream& s,
+    metal::Device& d,
+    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,
+    std::vector<array>& copies) {
+  // Make conv params
+  MLXConvParams<3> conv_params{
+      /* const int  N = */ in.shape(0),
+      /* const int  C = */ in.shape(4),
+      /* const int  O = */ wt.shape(0),
+      /* const int iS[NDIM] = */ {in.shape(1), in.shape(2), in.shape(3)},
+      /* const int wS[NDIM] = */ {wt.shape(1), wt.shape(2), wt.shape(3)},
+      /* const int oS[NDIM] = */ {out.shape(1), out.shape(2), out.shape(3)},
+      /* const int str[NDIM] = */ {wt_strides[0], wt_strides[1], wt_strides[2]},
+      /* const int pad[NDIM] = */ {padding[0], padding[1], padding[2]},
+      /* const int kdil[NDIM] = */
+      {wt_dilation[0], wt_dilation[1], wt_dilation[2]},
+      /* const int idil[NDIM] = */
+      {in_dilation[0], in_dilation[1], in_dilation[2]},
+      /* const size_t in_strides[NDIM + 2] = */
+      {in.strides()[0],
+       in.strides()[1],
+       in.strides()[2],
+       in.strides()[3],
+       in.strides()[4]},
+      /* const size_t wt_strides[NDIM + 2] = */
+      {wt.strides()[0],
+       wt.strides()[1],
+       wt.strides()[2],
+       wt.strides()[3],
+       wt.strides()[4]},
+      /* const size_t out_strides[NDIM + 2] = */
+      {out.strides()[0],
+       out.strides()[1],
+       out.strides()[2],
+       out.strides()[3],
+       out.strides()[4]},
+      /* const int groups = */ 1,
+      /* const bool flip = */ flip,
+  };
+  return explicit_gemm_conv_ND_gpu(s, d, in, wt, out, conv_params);
+}
+
 } // namespace
 
 void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -694,8 +862,23 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
     wt = arr_copy;
   }
 
+  // 3D conv
+  if (out.ndim() == 5) {
+    conv_3D_gpu(
+        s,
+        d,
+        in,
+        wt,
+        out,
+        padding_,
+        kernel_strides_,
+        kernel_dilation_,
+        input_dilation_,
+        flip_,
+        copies);
+  }
   // 2D conv
-  if (out.ndim() == 4) {
+  else if (out.ndim() == 4) {
     conv_2D_gpu(
         s,
         d,
@@ -706,6 +889,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
         kernel_strides_,
         kernel_dilation_,
         input_dilation_,
+        groups_,
         flip_,
         copies);
   }
@@ -721,6 +905,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
         kernel_strides_,
         kernel_dilation_,
         input_dilation_,
+        groups_,
         flip_);
   }
   // Throw error

mlx/backend/metal/copy.cpp

@@ -4,12 +4,14 @@
 
 #include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/kernels/defines.h"
+#include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
+constexpr int MAX_COPY_SPECIALIZED_DIMS = 5;
+
 void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s) {
   if (ctype == CopyType::Vector) {
     // If the input is donateable, we are doing a vector copy and the types
@@ -62,36 +64,43 @@ void copy_gpu_inplace(
   auto& strides_out_ = strides[1];
 
   auto& d = metal::device(s.device);
-  std::ostringstream kname;
-  switch (ctype) {
-    case CopyType::Scalar:
-      kname << "scopy";
-      break;
-    case CopyType::Vector:
-      kname << "vcopy";
-      break;
-    case CopyType::General:
-      kname << "gcopy";
-      break;
-    case CopyType::GeneralGeneral:
-      kname << "ggcopy";
-      break;
+  std::string kernel_name;
+  {
+    std::ostringstream kname;
+    switch (ctype) {
+      case CopyType::Scalar:
+        kname << "s";
+        break;
+      case CopyType::Vector:
+        kname << "v";
+        break;
+      case CopyType::General:
+        kname << "g";
+        break;
+      case CopyType::GeneralGeneral:
+        kname << "gg";
+        break;
+    }
+    if ((ctype == CopyType::General || ctype == CopyType::GeneralGeneral) &&
+        shape.size() <= MAX_COPY_SPECIALIZED_DIMS) {
+      kname << shape.size();
+    }
+    kname << "_copy";
+    kname << type_to_name(in) << type_to_name(out);
+    kernel_name = kname.str();
   }
-  kname << type_to_name(in) << type_to_name(out);
-  if ((ctype == CopyType::General || ctype == CopyType::GeneralGeneral) &&
-      shape.size() <= MAX_COPY_SPECIALIZED_DIMS) {
-    kname << "_" << shape.size();
-  }
-  auto kernel = d.get_kernel(kname.str());
-  auto compute_encoder = d.get_command_encoder(s.index);
+
+  auto kernel = get_copy_kernel(d, kernel_name, in, out);
+
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
   bool donate_in = in.data_shared_ptr() == nullptr;
 
   inp_offset *= size_of(in.dtype());
   out_offset *= size_of(out.dtype());
 
-  set_array_buffer(compute_encoder, donate_in ? out : in, inp_offset, 0);
-  set_array_buffer(compute_encoder, out, out_offset, 1);
+  compute_encoder.set_input_array(donate_in ? out : in, 0, inp_offset);
+  compute_encoder.set_output_array(out, 1, out_offset);
 
   if (ctype == CopyType::General || ctype == CopyType::GeneralGeneral) {
     int ndim = shape.size();
@@ -106,7 +115,7 @@ void copy_gpu_inplace(
       set_vector_bytes(compute_encoder, strides_out, ndim, 4);
     }
 
-    if (ndim > MAX_BINARY_SPECIALIZED_DIMS) {
+    if (ndim > MAX_COPY_SPECIALIZED_DIMS) {
       compute_encoder->setBytes(&ndim, sizeof(int), 5);
     }
 
@@ -126,7 +135,7 @@ void copy_gpu_inplace(
 
     auto group_dims = get_block_dims(dim0, dim1, rest);
     MTL::Size grid_dims = MTL::Size(dim0, dim1, rest);
-    compute_encoder->dispatchThreads(grid_dims, group_dims);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
   } else {
     size_t nthreads = out.data_size();
     MTL::Size grid_dims = MTL::Size(nthreads, 1, 1);
@@ -135,7 +144,7 @@ void copy_gpu_inplace(
       thread_group_size = nthreads;
     }
     MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
-    compute_encoder->dispatchThreads(grid_dims, group_dims);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
   }
 }
 

mlx/backend/metal/device.cpp

@@ -1,16 +1,19 @@
-// Copyright © 2023-24 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <dlfcn.h>
 #include <cstdlib>
 #include <filesystem>
 #include <sstream>
 
+#include <sys/sysctl.h>
+
 #define NS_PRIVATE_IMPLEMENTATION
 #define CA_PRIVATE_IMPLEMENTATION
 #define MTL_PRIVATE_IMPLEMENTATION
 
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/metal.h"
+#include "mlx/backend/metal/metal_impl.h"
 #include "mlx/backend/metal/mps/gemm.h"
 #include "mlx/backend/metal/utils.h"
 
@@ -22,9 +25,18 @@ namespace {
 
 // TODO nicer way to set this or possibly expose as an environment variable
 constexpr int MAX_BUFFERS_PER_QUEUE = 12;
+constexpr int MAX_DISPATCHES_PER_ENCODER = 2;
 
 constexpr const char* default_mtllib_path = METAL_PATH;
 
+constexpr auto get_metal_version() {
+#if defined METAL_3_1
+  return MTL::LanguageVersion3_1;
+#else
+  return MTL::LanguageVersion3_0;
+#endif
+}
+
 auto load_device() {
   auto devices = MTL::CopyAllDevices();
   auto device = static_cast<MTL::Device*>(devices->object(0))
@@ -34,7 +46,6 @@ auto load_device() {
   }
   return device;
 }
-
 std::pair<MTL::Library*, NS::Error*> load_library_from_path(
     MTL::Device* device,
     const char* path) {
@@ -113,6 +124,33 @@ MTL::Library* load_library(
 
 } // namespace
 
+void CommandEncoder::dispatchThreadgroups(
+    MTL::Size grid_dims,
+    MTL::Size group_dims) {
+  num_dispatches++;
+  enc->dispatchThreadgroups(grid_dims, group_dims);
+  maybe_split();
+}
+
+void CommandEncoder::dispatchThreads(
+    MTL::Size grid_dims,
+    MTL::Size group_dims) {
+  num_dispatches++;
+  enc->dispatchThreads(grid_dims, group_dims);
+  maybe_split();
+}
+
+void CommandEncoder::maybe_split() {
+  if (num_dispatches > MAX_DISPATCHES_PER_ENCODER && !concurrent) {
+    enc->endEncoding();
+    enc->release();
+    num_dispatches = 0;
+    outputs.clear();
+    enc = cbuf->computeCommandEncoder(MTL::DispatchTypeConcurrent);
+    enc->retain();
+  }
+}
+
 Device::Device() {
   auto pool = new_scoped_memory_pool();
   device_ = load_device();
@@ -127,9 +165,6 @@ Device::~Device() {
   for (auto& b : buffer_map_) {
     b.second.second->release();
   }
-  for (auto& e : encoder_map_) {
-    e.second->release();
-  }
   for (auto& k : kernel_map_) {
     k.second->release();
   }
@@ -166,27 +201,26 @@ void Device::increment_command_buffer_ops(int index) {
 
 MTL::CommandBuffer* Device::get_command_buffer(int index) {
   auto bit = buffer_map_.find(index);
-  return (bit == buffer_map_.end()) ? nullptr : bit->second.second;
-}
+  if (bit == buffer_map_.end()) {
+    auto qit = queue_map_.find(index);
+    if (qit == queue_map_.end()) {
+      throw std::runtime_error(
+          "[metal::Device] Attempting to get command buffer for invalid queue.");
+    }
 
-MTL::CommandBuffer* Device::new_command_buffer(int index) {
-  auto qit = queue_map_.find(index);
-  if (qit == queue_map_.end()) {
-    throw std::runtime_error(
-        "[metal::Device] Attempting to get command buffer for invalid queue.");
+    auto cb = qit->second->commandBufferWithUnretainedReferences();
+
+    if (!cb) {
+      throw std::runtime_error(
+          "[metal::Device] Unable to create new command buffer");
+    }
+
+    // Increment ref count so the buffer is not garbage collected
+    cb->retain();
+
+    bit = buffer_map_.insert({index, {0, cb}}).first;
   }
-
-  auto cb = qit->second->commandBufferWithUnretainedReferences();
-
-  if (!cb) {
-    throw std::runtime_error(
-        "[metal::Device] Unable to create new command buffer");
-  }
-
-  // Increment ref count so the buffer is not garbage collected
-  cb->retain();
-
-  return buffer_map_.insert({index, {0, cb}}).first->second.second;
+  return bit->second.second;
 }
 
 void Device::commit_command_buffer(int index) {
@@ -197,24 +231,17 @@ void Device::commit_command_buffer(int index) {
 }
 
 void Device::end_encoding(int index) {
-  auto eit = encoder_map_.find(index);
-  if (eit != encoder_map_.end()) {
-    eit->second->endEncoding();
-    eit->second->release();
-    encoder_map_.erase(eit);
-  }
+  encoder_map_.erase(index);
 }
 
-MTL::ComputeCommandEncoder* Device::get_command_encoder(int index) {
+CommandEncoder& Device::get_command_encoder(int index) {
   auto eit = encoder_map_.find(index);
   if (eit == encoder_map_.end()) {
     auto cb = get_command_buffer(index);
-    auto compute_encoder = cb->computeCommandEncoder();
-    // Increment ref count so the buffer is not garbage collected
-    compute_encoder->retain();
-    eit = encoder_map_.insert({index, compute_encoder}).first;
+    eit =
+        encoder_map_.emplace(index, std::make_unique<CommandEncoder>(cb)).first;
   }
-  return eit->second;
+  return *(eit->second);
 }
 
 void Device::register_library(
@@ -256,13 +283,16 @@ MTL::Library* Device::get_library_(const std::string& source_string) {
       NS::String::string(source_string.c_str(), NS::ASCIIStringEncoding);
 
   NS::Error* error = nullptr;
-  auto mtl_lib = device_->newLibrary(ns_code, nullptr, &error);
+  auto options = MTL::CompileOptions::alloc()->init();
+  options->setFastMathEnabled(false);
+  options->setLanguageVersion(get_metal_version());
+  auto mtl_lib = device_->newLibrary(ns_code, options, &error);
+  options->release();
 
   // Throw error if unable to compile library
   if (!mtl_lib) {
     std::ostringstream msg;
-    msg << "[metal::Device] Unable to load build metal library from source"
-        << "\n";
+    msg << "[metal::Device] Unable to build metal library from source" << "\n";
     if (error) {
       msg << error->localizedDescription()->utf8String() << "\n";
     }
@@ -281,8 +311,7 @@ MTL::Library* Device::get_library_(const MTL::StitchedLibraryDescriptor* desc) {
   // Throw error if unable to compile library
   if (!mtl_lib) {
     std::ostringstream msg;
-    msg << "[metal::Device] Unable to load build stitched metal library"
-        << "\n";
+    msg << "[metal::Device] Unable to build stitched metal library" << "\n";
     if (error) {
       msg << error->localizedDescription()->utf8String() << "\n";
     }
@@ -340,7 +369,6 @@ MTL::Function* Device::get_function_(
   }
 
   mtl_func_consts->release();
-  desc->release();
 
   return mtl_function;
 }
@@ -509,11 +537,13 @@ MTL::ComputePipelineState* Device::get_kernel(
   // Compile kernel to compute pipeline
   auto mtl_linked_funcs = get_linked_functions_(linked_functions);
   auto kernel = get_kernel_(kname, mtl_function, mtl_linked_funcs);
+
   mtl_function->release();
   mtl_linked_funcs->release();
 
   // Add kernel to cache
   kernel_map_.insert({kname, kernel});
+
   return kernel;
 }
 
@@ -540,11 +570,12 @@ Device& device(mlx::core::Device) {
   return metal_device;
 }
 
-std::shared_ptr<void> new_scoped_memory_pool() {
+std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool() {
   auto dtor = [](void* ptr) {
     static_cast<NS::AutoreleasePool*>(ptr)->release();
   };
-  return std::shared_ptr<void>(NS::AutoreleasePool::alloc()->init(), dtor);
+  return std::unique_ptr<void, std::function<void(void*)>>(
+      NS::AutoreleasePool::alloc()->init(), dtor);
 }
 
 void new_stream(Stream stream) {
@@ -553,4 +584,23 @@ void new_stream(Stream stream) {
   }
 }
 
+std::unordered_map<std::string, std::variant<std::string, size_t>>
+device_info() {
+  auto raw_device = device(default_device()).mtl_device();
+  auto arch = std::string(raw_device->architecture()->name()->utf8String());
+
+  int mib[] = {CTL_HW, HW_MEMSIZE};
+  size_t memsize = 0;
+  size_t length = sizeof(memsize);
+
+  sysctl(mib, 2, &memsize, &length, NULL, 0);
+
+  return {
+      {"architecture", arch},
+      {"max_buffer_length", raw_device->maxBufferLength()},
+      {"max_recommended_working_set_size",
+       raw_device->recommendedMaxWorkingSetSize()},
+      {"memory_size", memsize}};
+}
+
 } // namespace mlx::core::metal

mlx/backend/metal/device.h

@@ -1,4 +1,4 @@
-// Copyright © 2023-24 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #pragma once
 
@@ -7,10 +7,12 @@
 #include <mutex>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 
 #include <dlfcn.h>
 #include <filesystem>
 
+#include "mlx/array.h"
 #include "mlx/device.h"
 
 namespace fs = std::filesystem;
@@ -34,6 +36,84 @@ inline std::string get_colocated_mtllib_path(const std::string& lib_name) {
 using MTLFCList =
     std::vector<std::tuple<const void*, MTL::DataType, NS::UInteger>>;
 
+struct CommandEncoder {
+  CommandEncoder(MTL::CommandBuffer* cbuf) : cbuf(cbuf) {
+    enc = cbuf->computeCommandEncoder(MTL::DispatchTypeConcurrent);
+    enc->retain();
+  };
+  CommandEncoder(const CommandEncoder&) = delete;
+  CommandEncoder& operator=(const CommandEncoder&) = delete;
+
+  struct ConcurrentContext {
+    ConcurrentContext(CommandEncoder& enc) : enc(enc) {
+      enc.concurrent = true;
+    }
+    ~ConcurrentContext() {
+      enc.concurrent = false;
+      enc.outputs.insert(
+          enc.concurrent_outputs.begin(), enc.concurrent_outputs.end());
+      enc.concurrent_outputs.clear();
+    }
+
+   private:
+    CommandEncoder& enc;
+  };
+
+  MTL::ComputeCommandEncoder* operator->() {
+    return enc;
+  }
+
+  void set_input_array(const array& a, int idx, int64_t offset = 0) {
+    auto r_buf =
+        static_cast<MTL::Resource*>(const_cast<void*>(a.buffer().ptr()));
+    if (auto it = outputs.find(r_buf); it != outputs.end()) {
+      // Insert a barrier
+      enc->memoryBarrier(&r_buf, 1);
+
+      // Remove the output
+      outputs.erase(it);
+    }
+    auto a_buf = static_cast<const MTL::Buffer*>(a.buffer().ptr());
+    auto base_offset = a.data<char>() -
+        static_cast<char*>(const_cast<MTL::Buffer*>(a_buf)->contents());
+    base_offset += offset;
+    enc->setBuffer(a_buf, base_offset, idx);
+  }
+
+  void set_output_array(array& a, int idx, int64_t offset = 0) {
+    // Add barriers before adding the output to the output set
+    set_input_array(a, idx, offset);
+    auto buf = static_cast<MTL::Resource*>(a.buffer().ptr());
+    if (concurrent) {
+      concurrent_outputs.insert(buf);
+    } else {
+      outputs.insert(buf);
+    }
+  }
+
+  void dispatchThreadgroups(MTL::Size grid_dims, MTL::Size group_dims);
+  void dispatchThreads(MTL::Size grid_dims, MTL::Size group_dims);
+
+  ConcurrentContext start_concurrent() {
+    return ConcurrentContext(*this);
+  }
+
+  ~CommandEncoder() {
+    enc->endEncoding();
+    enc->release();
+  }
+
+ private:
+  void maybe_split();
+
+  int num_dispatches{0};
+  MTL::CommandBuffer* cbuf;
+  MTL::ComputeCommandEncoder* enc;
+  bool concurrent{false};
+  std::unordered_set<MTL::Resource*> outputs;
+  std::unordered_set<MTL::Resource*> concurrent_outputs;
+};
+
 class Device {
  public:
   Device();
@@ -46,12 +126,11 @@ class Device {
   };
 
   void new_queue(int index);
-  MTL::CommandBuffer* new_command_buffer(int index);
   MTL::CommandBuffer* get_command_buffer(int index);
   int get_command_buffer_ops(int index);
   void increment_command_buffer_ops(int index);
   void commit_command_buffer(int index);
-  MTL::ComputeCommandEncoder* get_command_encoder(int index);
+  CommandEncoder& get_command_encoder(int index);
   void end_encoding(int index);
 
   void register_library(
@@ -132,7 +211,7 @@ class Device {
   MTL::Device* device_;
   std::unordered_map<int32_t, MTL::CommandQueue*> queue_map_;
   std::unordered_map<int32_t, std::pair<int, MTL::CommandBuffer*>> buffer_map_;
-  std::unordered_map<int32_t, MTL::ComputeCommandEncoder*> encoder_map_;
+  std::unordered_map<int32_t, std::unique_ptr<CommandEncoder>> encoder_map_;
   std::unordered_map<std::string, MTL::ComputePipelineState*> kernel_map_;
   std::unordered_map<std::string, MTL::Library*> library_map_;
   std::mutex mtx_;

mlx/backend/metal/event.cpp

@@ -0,0 +1,30 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/event.h"
+#include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/metal_impl.h"
+
+namespace mlx::core {
+
+Event::Event(const Stream& stream) : stream_(stream) {
+  auto dtor = [](void* ptr) {
+    auto p = metal::new_scoped_memory_pool();
+    static_cast<MTL::SharedEvent*>(ptr)->release();
+  };
+  auto p = metal::new_scoped_memory_pool();
+  event_ = std::shared_ptr<void>(
+      metal::device(stream.device).mtl_device()->newSharedEvent(), dtor);
+}
+
+void Event::wait() {
+  if (!static_cast<MTL::SharedEvent*>(raw_event().get())
+           ->waitUntilSignaledValue(value(), -1)) {
+    throw std::runtime_error("[Event::wait] Timed out");
+  }
+}
+
+void Event::signal() {
+  static_cast<MTL::SharedEvent*>(raw_event().get())->setSignaledValue(value());
+}
+
+} // namespace mlx::core

mlx/backend/metal/fft.cpp

@@ -1,12 +1,106 @@
 // Copyright © 2023 Apple Inc.
-
+#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/metal/utils.h"
+#include "mlx/mlx.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 void FFT::eval_gpu(const std::vector<array>& inputs, array& out) {
+  auto& s = out.primitive().stream();
+  auto& d = metal::device(s.device);
+
   auto& in = inputs[0];
-  throw std::runtime_error("[FFT] NYI for Metal backend.");
+
+  if (axes_.size() == 0 || axes_.size() > 1 || inverse_ ||
+      in.dtype() != complex64 || out.dtype() != complex64) {
+    // Could also fallback to CPU implementation here.
+    throw std::runtime_error(
+        "GPU FFT is only implemented for 1D, forward, complex FFTs.");
+  }
+
+  size_t n = in.shape(axes_[0]);
+
+  if (!is_power_of_2(n) || n > 2048 || n < 4) {
+    throw std::runtime_error(
+        "GPU FFT is only implemented for the powers of 2 from 4 -> 2048");
+  }
+
+  // Make sure that the array is contiguous and has stride 1 in the FFT dim
+  std::vector<array> copies;
+  auto check_input = [this, &copies, &s](const array& x) {
+    // TODO: Pass the strides to the kernel so
+    // we can avoid the copy when x is not contiguous.
+    bool no_copy = x.strides()[axes_[0]] == 1 && x.flags().row_contiguous ||
+        x.flags().col_contiguous;
+    if (no_copy) {
+      return x;
+    } else {
+      array x_copy(x.shape(), x.dtype(), nullptr, {});
+      std::vector<size_t> strides;
+      size_t cur_stride = x.shape(axes_[0]);
+      for (int axis = 0; axis < x.ndim(); axis++) {
+        if (axis == axes_[0]) {
+          strides.push_back(1);
+        } else {
+          strides.push_back(cur_stride);
+          cur_stride *= x.shape(axis);
+        }
+      }
+
+      auto flags = x.flags();
+      size_t f_stride = 1;
+      size_t b_stride = 1;
+      flags.col_contiguous = true;
+      flags.row_contiguous = true;
+      for (int i = 0, ri = x.ndim() - 1; i < x.ndim(); ++i, --ri) {
+        flags.col_contiguous &= (strides[i] == f_stride || x.shape(i) == 1);
+        f_stride *= x.shape(i);
+        flags.row_contiguous &= (strides[ri] == b_stride || x.shape(ri) == 1);
+        b_stride *= x.shape(ri);
+      }
+      // This is probably over-conservative
+      flags.contiguous = false;
+
+      x_copy.set_data(
+          allocator::malloc_or_wait(x.nbytes()), x.data_size(), strides, flags);
+      copy_gpu_inplace(x, x_copy, CopyType::GeneralGeneral, s);
+      copies.push_back(x_copy);
+      return x_copy;
+    }
+  };
+  const array& in_contiguous = check_input(inputs[0]);
+
+  // TODO: allow donation here
+  out.set_data(
+      allocator::malloc_or_wait(out.nbytes()),
+      in_contiguous.data_size(),
+      in_contiguous.strides(),
+      in_contiguous.flags());
+
+  // We use n / 4 threads by default since radix-4
+  // is the largest single threaded radix butterfly
+  // we currently implement.
+  size_t m = n / 4;
+  size_t batch = in.size() / in.shape(axes_[0]);
+
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  {
+    std::ostringstream kname;
+    kname << "fft_" << n;
+    auto kernel = d.get_kernel(kname.str());
+
+    bool donated = in.data_shared_ptr() == nullptr;
+    compute_encoder->setComputePipelineState(kernel);
+    compute_encoder.set_input_array(in_contiguous, 0);
+    compute_encoder.set_output_array(out, 1);
+
+    auto group_dims = MTL::Size(1, m, 1);
+    auto grid_dims = MTL::Size(batch, m, 1);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
+  }
+  d.get_command_buffer(s.index)->addCompletedHandler(
+      [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
 }
 
 } // namespace mlx::core

mlx/backend/metal/indexing.cpp

@@ -1,24 +1,35 @@
 // Copyright © 2023-2024 Apple Inc.
-#include <algorithm>
-#include <cassert>
-#include <numeric>
-#include <sstream>
+#include <fmt/format.h>
 
-#include "mlx/backend/common/binary.h"
+#include "mlx/backend/common/compiled.h"
 #include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/kernels/defines.h"
+#include "mlx/backend/metal/jit/includes.h"
+#include "mlx/backend/metal/jit/indexing.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
 
 namespace mlx::core {
 
-namespace {
+constexpr int METAL_MAX_INDEX_ARRAYS = 20;
 
-constexpr int METAL_MAX_INDEX_ARRAYS = 10;
-
-} // namespace
+std::pair<std::string, std::string> make_index_args(
+    const std::string& idx_type,
+    int nidx) {
+  std::ostringstream idx_args;
+  std::ostringstream idx_arr;
+  for (int i = 0; i < nidx; ++i) {
+    idx_args << fmt::format(
+        "const device {0} *idx{1} [[buffer({2})]],", idx_type, i, 20 + i);
+    idx_arr << fmt::format("idx{0}", i);
+    if (i < nidx - 1) {
+      idx_args << "\n";
+      idx_arr << ",";
+    }
+  }
+  return {idx_args.str(), idx_arr.str()};
+}
 
 void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& src = inputs[0];
@@ -42,15 +53,41 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   int idx_ndim = nidx ? inputs[1].ndim() : 0;
   size_t ndim = src.ndim();
 
-  std::ostringstream kname;
+  std::string lib_name;
+  std::string kernel_name;
   std::string idx_type_name = nidx ? type_to_name(inputs[1]) : "";
-  kname << "gather" << type_to_name(src) << idx_type_name << "_" << nidx;
-  if (idx_ndim <= 1) {
-    kname << "_" << idx_ndim;
+  {
+    std::ostringstream kname;
+    kname << "gather" << type_to_name(out) << idx_type_name << "_" << nidx
+          << "_" << idx_ndim;
+    lib_name = kname.str();
+    kernel_name = lib_name;
   }
 
-  auto compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::gather();
+    std::string out_type_str = get_type_string(out.dtype());
+    std::string idx_type_str =
+        nidx ? get_type_string(inputs[1].dtype()) : "bool";
+    auto [idx_args, idx_arr] = make_index_args(idx_type_str, nidx);
+
+    // Index dimension specializations
+    kernel_source << fmt::format(
+        gather_kernels,
+        type_to_name(out) + idx_type_name,
+        out_type_str,
+        idx_type_str,
+        nidx,
+        idx_args,
+        idx_arr,
+        idx_ndim);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = d.get_kernel(kernel_name, lib);
   compute_encoder->setComputePipelineState(kernel);
 
   size_t slice_size = 1;
@@ -81,8 +118,8 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
 
   // Set all the buffers
-  set_array_buffer(compute_encoder, src, 0);
-  set_array_buffer(compute_encoder, out, 1);
+  compute_encoder.set_input_array(src, 0);
+  compute_encoder.set_output_array(out, 1);
 
   // Set source info
   compute_encoder->setBytes(src.shape().data(), ndim * sizeof(int), 2);
@@ -102,12 +139,12 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   compute_encoder->setBytes(&idx_ndim, sizeof(int), 9);
 
   // Set index buffers
-  for (int i = 1; i < nidx + 1; ++i) {
-    set_array_buffer(compute_encoder, inputs[i], 20 + i);
+  for (int i = 0; i < nidx; ++i) {
+    compute_encoder.set_input_array(inputs[i + 1], 20 + i);
   }
 
   // Launch grid
-  compute_encoder->dispatchThreads(grid_dims, group_dims);
+  compute_encoder.dispatchThreads(grid_dims, group_dims);
 }
 
 void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -139,10 +176,6 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& s = stream();
   auto& d = metal::device(s.device);
 
-  // Get kernel name
-  std::ostringstream kname;
-  std::string idx_type_name = nidx ? type_to_name(inputs[1]) : "";
-
   int idx_ndim = nidx ? inputs[1].ndim() : 0;
   bool index_nd1_specialization = (idx_ndim == 1);
 
@@ -159,32 +192,86 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
     index_nd1_specialization &= inputs[i].flags().row_contiguous;
   }
 
-  if (index_nd1_specialization) {
-    kname << "scatter_1d_index" << type_to_name(out) << idx_type_name;
-  } else {
-    kname << "scatter" << type_to_name(out) << idx_type_name;
-  }
+  std::string lib_name;
+  std::string kernel_name;
+  std::string idx_type_name = nidx ? type_to_name(inputs[1]) : "";
+  std::string op_name;
   switch (reduce_type_) {
     case Scatter::None:
-      kname << "_none";
+      op_name = "none";
       break;
     case Scatter::Sum:
-      kname << "_sum";
+      op_name = "sum";
       break;
     case Scatter::Prod:
-      kname << "_prod";
+      op_name = "prod";
       break;
     case Scatter::Max:
-      kname << "_max";
+      op_name = "max";
       break;
     case Scatter::Min:
-      kname << "_min";
+      op_name = "min";
       break;
   }
-  kname << "_" << nidx;
 
-  auto compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  {
+    std::ostringstream kname;
+    if (index_nd1_specialization) {
+      kname << "scatter_1d_index" << type_to_name(out) << idx_type_name;
+    } else {
+      kname << "scatter" << type_to_name(out) << idx_type_name;
+    }
+    kname << "_" << op_name << "_" << nidx;
+    lib_name = kname.str();
+    kernel_name = kname.str();
+  }
+
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::reduce_utils()
+                  << metal::scatter();
+
+    std::string out_type_str = get_type_string(out.dtype());
+    std::string idx_type_str =
+        nidx ? get_type_string(inputs[1].dtype()) : "bool";
+    std::string op_type;
+    switch (reduce_type_) {
+      case Scatter::None:
+        op_type = "None";
+        break;
+      case Scatter::Sum:
+        op_type = "Sum<{0}>";
+        break;
+      case Scatter::Prod:
+        op_type = "Prod<{0}>";
+        break;
+      case Scatter::Max:
+        op_type = "Max<{0}>";
+        break;
+      case Scatter::Min:
+        op_type = "Min<{0}>";
+        break;
+    }
+    if (reduce_type_ != Scatter::None) {
+      op_type = fmt::format(op_type, out_type_str);
+    }
+    auto [idx_args, idx_arr] = make_index_args(idx_type_str, nidx);
+
+    kernel_source << fmt::format(
+        scatter_kernels,
+        type_to_name(out) + idx_type_name + "_" + op_name,
+        out_type_str,
+        idx_type_str,
+        op_type,
+        nidx,
+        idx_args,
+        idx_arr);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = d.get_kernel(kernel_name, lib);
 
   auto& upd = inputs.back();
   size_t nthreads = upd.size();
@@ -192,8 +279,8 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   compute_encoder->setComputePipelineState(kernel);
 
   // Set all the buffers
-  set_array_buffer(compute_encoder, upd, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(upd, 1);
+  compute_encoder.set_output_array(out, 2);
 
   // Set update info
   uint upd_ndim = upd.ndim();
@@ -209,14 +296,14 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
     compute_encoder->setBytes(&upd_size, sizeof(size_t), 5);
 
     // Set index buffers
-    for (int i = 1; i < nidx + 1; ++i) {
-      set_array_buffer(compute_encoder, inputs[i], 20 + i);
+    for (int i = 0; i < nidx; ++i) {
+      compute_encoder.set_input_array(inputs[i + 1], 20 + i);
     }
 
     // Launch grid
     MTL::Size grid_dims = MTL::Size(upd_size, nthreads / upd_size, 1);
     MTL::Size group_dims = get_block_dims(upd_size, nthreads / upd_size, 1);
-    compute_encoder->dispatchThreads(grid_dims, group_dims);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
 
   } else {
     // Collect all idx shapes and strides into one place
@@ -279,14 +366,14 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
     compute_encoder->setBytes(&idx_ndim, sizeof(int), 13);
 
     // Set index buffers
-    for (int i = 1; i < nidx + 1; ++i) {
-      set_array_buffer(compute_encoder, inputs[i], 20 + i);
+    for (int i = 0; i < nidx; ++i) {
+      compute_encoder.set_input_array(inputs[i + 1], 20 + i);
     }
 
     // Launch grid
     MTL::Size grid_dims = MTL::Size(upd_size, nthreads / upd_size, 1);
     MTL::Size group_dims = get_block_dims(upd_size, nthreads / upd_size, 1);
-    compute_encoder->dispatchThreads(grid_dims, group_dims);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
   }
 }
 

mlx/backend/metal/jit/arange.h

@@ -0,0 +1,9 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view arange_kernels = R"(
+template [[host_name("{0}")]] [[kernel]] void arange<{1}>(
+    constant const {1}& start,
+    constant const {1}& step,
+    device {1}* out,
+    uint index [[thread_position_in_grid]]);
+)";

mlx/backend/metal/jit/binary.h

@@ -0,0 +1,87 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view binary_kernels = R"(
+template [[host_name("ss{0}")]] [[kernel]]
+void binary_ss<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("vs{0}")]] [[kernel]]
+void binary_vs<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("sv{0}")]] [[kernel]]
+void binary_sv<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("vv{0}")]] [[kernel]]
+void binary_vv<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("g4{0}")]] [[kernel]] void
+binary_g_nd<{1}, {2}, {3}, 4>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    constant const int shape[4],
+    constant const size_t a_strides[4],
+    constant const size_t b_strides[4],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("g5{0}")]] [[kernel]] void
+binary_g_nd<{1}, {2}, {3}, 5>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    constant const int shape[5],
+    constant const size_t a_strides[5],
+    constant const size_t b_strides[5],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+
+template [[host_name("g1{0}")]] [[kernel]] void
+binary_g_nd1<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    constant const size_t& a_stride,
+    constant const size_t& b_stride,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("g2{0}")]] [[kernel]] void
+binary_g_nd2<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    constant const size_t a_strides[2],
+    constant const size_t b_strides[2],
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]);
+template [[host_name("g3{0}")]] [[kernel]] void
+binary_g_nd3<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    constant const size_t a_strides[3],
+    constant const size_t b_strides[3],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+
+template [[host_name("gn{0}")]] [[kernel]]
+void binary_g<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    constant const int* shape,
+    constant const size_t* a_strides,
+    constant const size_t* b_strides,
+    constant const int& ndim,
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+)";

mlx/backend/metal/jit/binary_two.h

@@ -0,0 +1,98 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view binary_two_kernels = R"(
+template [[host_name("ss{0}")]] [[kernel]]
+void binary_ss<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("vs{0}")]] [[kernel]]
+void binary_vs<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("sv{0}")]] [[kernel]]
+void binary_sv<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("vv{0}")]] [[kernel]]
+void binary_vv<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    uint index [[thread_position_in_grid]]);
+
+template [[host_name("g4{0}")]] [[kernel]] void
+binary_g_nd<{1}, {2}, {3}, 4>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    constant const int shape[4],
+    constant const size_t a_strides[4],
+    constant const size_t b_strides[4],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("g5{0}")]] [[kernel]] void
+binary_g_nd<{1}, {2}, {3}, 5>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    constant const int shape[5],
+    constant const size_t a_strides[5],
+    constant const size_t b_strides[5],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+
+template [[host_name("g1{0}")]] [[kernel]] void
+binary_g_nd1<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    constant const size_t& a_stride,
+    constant const size_t& b_stride,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("g2{0}")]] [[kernel]] void
+binary_g_nd2<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    constant const size_t a_strides[2],
+    constant const size_t b_strides[2],
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]);
+template [[host_name("g3{0}")]] [[kernel]] void
+binary_g_nd3<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    constant const size_t a_strides[3],
+    constant const size_t b_strides[3],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+
+template [[host_name("gn{0}")]] [[kernel]]
+void binary_g<{1}, {2}, {3}>(
+    device const {1}* a,
+    device const {1}* b,
+    device {2}* c,
+    device {2}* d,
+    constant const int* shape,
+    constant const size_t* a_strides,
+    constant const size_t* b_strides,
+    constant const int& ndim,
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+)";

mlx/backend/metal/jit/copy.h

@@ -0,0 +1,100 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view copy_kernels = R"(
+template [[host_name("s_{0}")]] [[kernel]] void copy_s<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    uint index [[thread_position_in_grid]]);
+template [[host_name("v_{0}")]] [[kernel]] void copy_v<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    uint index [[thread_position_in_grid]]);
+
+template [[host_name("g4_{0}")]] [[kernel]] void
+copy_g_nd<{1}, {2}, 4>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("gg4_{0}")]] [[kernel]] void
+copy_gg_nd<{1}, {2}, 4>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
+    uint3 index [[thread_position_in_grid]]);
+template [[host_name("g5_{0}")]] [[kernel]] void
+copy_g_nd<{1}, {2}, 5>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("gg5_{0}")]] [[kernel]] void
+copy_gg_nd<{1}, {2}, 5>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
+    uint3 index [[thread_position_in_grid]]);
+template [[host_name("g1_{0}")]] [[kernel]] void copy_g_nd1<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int64_t& src_stride [[buffer(3)]],
+    uint index [[thread_position_in_grid]]);
+template [[host_name("g2_{0}")]] [[kernel]] void copy_g_nd2<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]);
+template [[host_name("g3_{0}")]] [[kernel]] void copy_g_nd3<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("gg1_{0}")]] [[kernel]] void
+copy_gg_nd1<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int64_t& src_stride [[buffer(3)]],
+    constant const int64_t& dst_stride [[buffer(4)]],
+    uint index [[thread_position_in_grid]]);
+template [[host_name("gg2_{0}")]] [[kernel]] void
+copy_gg_nd2<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
+    uint2 index [[thread_position_in_grid]]);
+template [[host_name("gg3_{0}")]] [[kernel]] void
+copy_gg_nd3<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
+    uint3 index [[thread_position_in_grid]]);
+
+template [[host_name("g_{0}")]] [[kernel]] void copy_g<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int& ndim [[buffer(5)]],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("gg_{0}")]] [[kernel]] void copy_gg<{1}, {2}>(
+    device const {1}* src [[buffer(0)]],
+    device {2}* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
+    constant const int& ndim [[buffer(5)]],
+    uint3 index [[thread_position_in_grid]]);
+)";

mlx/backend/metal/jit/includes.h

@@ -0,0 +1,34 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#pragma once
+
+namespace mlx::core::metal {
+
+const char* utils();
+const char* binary_ops();
+const char* unary_ops();
+const char* ternary_ops();
+const char* reduce_utils();
+const char* gather();
+const char* scatter();
+
+const char* arange();
+const char* unary();
+const char* binary();
+const char* binary_two();
+const char* copy();
+const char* ternary();
+const char* scan();
+const char* softmax();
+const char* sort();
+const char* reduce();
+
+const char* gemm();
+const char* steel_gemm_fused();
+const char* steel_gemm_masked();
+const char* steel_gemm_splitk();
+const char* conv();
+const char* steel_conv();
+const char* steel_conv_general();
+
+} // namespace mlx::core::metal

mlx/backend/metal/jit/indexing.h

@@ -0,0 +1,81 @@
+// Copyright © 2023-2024 Apple Inc.
+
+constexpr std::string_view gather_kernels = R"(
+[[kernel]] void gather{0}_{3}_{6}(
+    const device {1}* src [[buffer(0)]],
+    device {1}* out [[buffer(1)]],
+    const constant int* src_shape [[buffer(2)]],
+    const constant size_t* src_strides [[buffer(3)]],
+    const constant size_t& src_ndim [[buffer(4)]],
+    const constant int* slice_sizes [[buffer(5)]],
+    const constant int* axes [[buffer(6)]],
+    const constant int* idx_shapes [[buffer(7)]],
+    const constant size_t* idx_strides [[buffer(8)]],
+    const constant int& idx_ndim [[buffer(9)]],
+    {4}
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]) {{
+  Indices<{2}, {3}> idxs{{
+    {{ {5} }}, idx_shapes, idx_strides, idx_ndim}};
+
+  return gather_impl<{1}, {2}, {3}, {6}>(
+      src,
+      out,
+      src_shape,
+      src_strides,
+      src_ndim,
+      slice_sizes,
+      axes,
+      idxs,
+      index,
+      grid_dim);
+}}
+)";
+
+constexpr std::string_view scatter_kernels = R"(
+[[kernel]] void scatter_1d_index{0}_{4}(
+    const device {1}* updates [[buffer(1)]],
+    device mlx_atomic<{1}>* out [[buffer(2)]],
+    const constant int* out_shape [[buffer(3)]],
+    const constant size_t* out_strides [[buffer(4)]],
+    const constant size_t& upd_size [[buffer(5)]],
+    {5}
+    uint2 gid [[thread_position_in_grid]]) {{
+  const array<const device {2}*, {4}> idx_buffers = {{ {6} }};
+  return scatter_1d_index_impl<{1}, {2}, {3}, {4}>(
+      updates, out, out_shape, out_strides, upd_size, idx_buffers, gid);
+}}
+
+[[kernel]] void scatter{0}_{4}(
+    const device {1}* updates [[buffer(1)]],
+    device mlx_atomic<{1}>* out [[buffer(2)]],
+    const constant int* upd_shape [[buffer(3)]],
+    const constant size_t* upd_strides [[buffer(4)]],
+    const constant size_t& upd_ndim [[buffer(5)]],
+    const constant size_t& upd_size [[buffer(6)]],
+    const constant int* out_shape [[buffer(7)]],
+    const constant size_t* out_strides [[buffer(8)]],
+    const constant size_t& out_ndim [[buffer(9)]],
+    const constant int* axes [[buffer(10)]],
+    const constant int* idx_shapes [[buffer(11)]],
+    const constant size_t* idx_strides [[buffer(12)]],
+    const constant int& idx_ndim [[buffer(13)]],
+    {5}
+    uint2 gid [[thread_position_in_grid]]) {{
+  Indices<{2}, {4}> idxs{{ {{ {6} }}, idx_shapes, idx_strides, idx_ndim}};
+
+  return scatter_impl<{1}, {2}, {3}, {4}>(
+      updates,
+      out,
+      upd_shape,
+      upd_strides,
+      upd_ndim,
+      upd_size,
+      out_shape,
+      out_strides,
+      out_ndim,
+      axes,
+      idxs,
+      gid);
+}}
+)";

mlx/backend/metal/jit/reduce.h

@@ -0,0 +1,168 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view reduce_init_kernels = R"(
+[[kernel]] void {0}(
+    device {1}* out [[buffer(0)]],
+    uint tid [[thread_position_in_grid]]) {{
+  out[tid] = {2}<{1}>::init;
+}}
+)";
+
+constexpr std::string_view reduce_kernels = R"(
+template [[host_name("all_{0}")]] [[kernel]] void
+all_reduce<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device mlx_atomic<{2}>* out [[buffer(1)]],
+    const device size_t& in_size [[buffer(2)]],
+    uint gid [[thread_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint grid_size [[threads_per_grid]],
+    uint simd_per_group [[simdgroups_per_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+template [[host_name("colGeneral_{0}")]] [[kernel]] void
+col_reduce_general<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device mlx_atomic<{2}>* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& reduction_stride [[buffer(3)]],
+    const constant size_t& out_size [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    threadgroup {2}* local_data [[threadgroup(0)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]],
+    uint3 lsize [[threads_per_threadgroup]]);
+template [[host_name("colSmall_{0}")]] [[kernel]] void
+col_reduce_small<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& reduction_stride [[buffer(3)]],
+    const constant size_t& out_size [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    const constant size_t& non_col_reductions [[buffer(8)]],
+    const constant int* non_col_shapes [[buffer(9)]],
+    const constant size_t* non_col_strides [[buffer(10)]],
+    const constant int& non_col_ndim [[buffer(11)]],
+    uint tid [[thread_position_in_grid]]);
+template [[host_name("rowGeneralSmall_{0}")]] [[kernel]] void
+row_reduce_general_small<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& out_size [[buffer(3)]],
+    const constant size_t& non_row_reductions [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    uint lid [[thread_position_in_grid]]);
+template [[host_name("rowGeneralMed_{0}")]] [[kernel]] void
+row_reduce_general_med<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& out_size [[buffer(3)]],
+    const constant size_t& non_row_reductions [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    uint tid [[threadgroup_position_in_grid]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_per_group [[dispatch_simdgroups_per_threadgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+template [[host_name("rowGeneral_{0}")]] [[kernel]] void
+row_reduce_general<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device mlx_atomic<{2}>* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& out_size [[buffer(3)]],
+    const constant size_t& non_row_reductions [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    uint3 lid [[thread_position_in_threadgroup]],
+    uint3 lsize [[threads_per_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_per_group [[simdgroups_per_threadgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+)";
+
+constexpr std::string_view reduce_non_atomic_kernels = R"(
+template [[host_name("allNoAtomics_{0}")]] [[kernel]] void
+all_reduce_no_atomics<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const device size_t& in_size [[buffer(2)]],
+    uint gid [[thread_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint grid_size [[threads_per_grid]],
+    uint simd_per_group [[simdgroups_per_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint thread_group_id [[threadgroup_position_in_grid]]);
+
+template [[host_name("colGeneralNoAtomics_{0}")]] [[kernel]] void
+  col_reduce_general_no_atomics<{1}, {2}, {3}<{2}>>(
+      const device {1}* in [[buffer(0)]],
+      device {2}* out [[buffer(1)]],
+      const constant size_t& reduction_size [[buffer(2)]],
+      const constant size_t& reduction_stride [[buffer(3)]],
+      const constant size_t& out_size [[buffer(4)]],
+      const constant int* shape [[buffer(5)]],
+      const constant size_t* strides [[buffer(6)]],
+      const constant int& ndim [[buffer(7)]],
+      threadgroup {2}* local_data [[threadgroup(0)]],
+      uint3 tid [[threadgroup_position_in_grid]],
+      uint3 lid [[thread_position_in_threadgroup]],
+      uint3 gid [[thread_position_in_grid]],
+      uint3 lsize [[threads_per_threadgroup]],
+      uint3 gsize [[threads_per_grid]]);
+template [[host_name("colSmall_{0}")]] [[kernel]] void
+col_reduce_small<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& reduction_stride [[buffer(3)]],
+    const constant size_t& out_size [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    const constant size_t& non_col_reductions [[buffer(8)]],
+    const constant int* non_col_shapes [[buffer(9)]],
+    const constant size_t* non_col_strides [[buffer(10)]],
+    const constant int& non_col_ndim [[buffer(11)]],
+    uint tid [[thread_position_in_grid]]);
+template [[host_name("rowGeneralSmall_{0}")]] [[kernel]] void
+row_reduce_general_small<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& out_size [[buffer(3)]],
+    const constant size_t& non_row_reductions [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    uint lid [[thread_position_in_grid]]);
+template [[host_name("rowGeneralNoAtomics_{0}")]] [[kernel]] void
+row_reduce_general_no_atomics<{1}, {2}, {3}<{2}>>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& out_size [[buffer(3)]],
+    const constant size_t& non_row_reductions [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    uint3 lid [[thread_position_in_threadgroup]],
+    uint3 lsize [[threads_per_threadgroup]],
+    uint3 gsize [[threads_per_grid]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_per_group [[simdgroups_per_threadgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+)";

mlx/backend/metal/jit/scan.h

@@ -0,0 +1,26 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view scan_kernels = R"(
+template [[host_name("contig_{0}")]] [[kernel]] void
+contiguous_scan<{1}, {2}, {3}<{2}>, 4, {4}, {5}>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& axis_size [[buffer(2)]],
+    uint gid [[thread_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint lsize [[threads_per_threadgroup]],
+    uint simd_size [[threads_per_simdgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+
+template [[host_name("strided_{0}")]] [[kernel]] void
+strided_scan<{1}, {2}, {3}<{2}>, 4, {4}, {5}>(
+    const device {1}* in [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant size_t& axis_size [[buffer(2)]],
+    const constant size_t& stride [[buffer(3)]],
+    uint2 gid [[thread_position_in_grid]],
+    uint2 lid [[thread_position_in_threadgroup]],
+    uint2 lsize [[threads_per_threadgroup]],
+    uint simd_size [[threads_per_simdgroup]]);
+)";

mlx/backend/metal/jit/softmax.h

@@ -0,0 +1,23 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view softmax_kernels = R"(
+template [[host_name("block_{0}")]] [[kernel]] void
+softmax_single_row<{1}, {2}>(
+    const device {1}* in,
+    device {1}* out,
+    constant int& axis_size,
+    uint gid [[thread_position_in_grid]],
+    uint _lid [[thread_position_in_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+template [[host_name("looped_{0}")]] [[kernel]] void
+softmax_looped<{1}, {2}>(
+    const device {1}* in,
+    device {1}* out,
+    constant int& axis_size,
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint lsize [[threads_per_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+)";

mlx/backend/metal/jit/sort.h

@@ -0,0 +1,81 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view block_sort_kernels = R"(
+template [[host_name("carg_{0}")]] [[kernel]] void
+block_sort<{1}, {2}, true, {3}, {4}>(
+    const device {1}* inp [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant int& size_sorted_axis [[buffer(2)]],
+    const constant int& stride_sorted_axis [[buffer(3)]],
+    const constant int& stride_segment_axis [[buffer(4)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+template [[host_name("ncarg_{0}")]] [[kernel]] void
+block_sort_nc<{1}, {2}, true, {3}, {4}>(
+    const device {1}* inp [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant int& size_sorted_axis [[buffer(2)]],
+    const constant int& stride_sorted_axis [[buffer(3)]],
+    const constant int& nc_dim [[buffer(4)]],
+    const device int* nc_shape [[buffer(5)]],
+    const device size_t* nc_strides [[buffer(6)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+template [[host_name("c_{0}")]] [[kernel]] void
+block_sort<{1}, {2}, false, {3}, {4}>(
+    const device {1}* inp [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant int& size_sorted_axis [[buffer(2)]],
+    const constant int& stride_sorted_axis [[buffer(3)]],
+    const constant int& stride_segment_axis [[buffer(4)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+template [[host_name("nc_{0}")]] [[kernel]] void
+block_sort_nc<{1}, {2}, false, {3}, {4}>(
+    const device {1}* inp [[buffer(0)]],
+    device {2}* out [[buffer(1)]],
+    const constant int& size_sorted_axis [[buffer(2)]],
+    const constant int& stride_sorted_axis [[buffer(3)]],
+    const constant int& nc_dim [[buffer(4)]],
+    const device int* nc_shape [[buffer(5)]],
+    const device size_t* nc_strides [[buffer(6)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";
+
+constexpr std::string_view multiblock_sort_kernels = R"(
+template [[host_name("sort_{0}")]] [[kernel]] void
+mb_block_sort<{1}, {2}, true, {3}, {4}>(
+    const device {1}* inp [[buffer(0)]],
+    device {1}* out_vals [[buffer(1)]],
+    device {2}* out_idxs [[buffer(2)]],
+    const constant int& size_sorted_axis [[buffer(3)]],
+    const constant int& stride_sorted_axis [[buffer(4)]],
+    const constant int& nc_dim [[buffer(5)]],
+    const device int* nc_shape [[buffer(6)]],
+    const device size_t* nc_strides [[buffer(7)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+template [[host_name("partition_{0}")]] [[kernel]] void
+mb_block_partition<{1}, {2}, true, {3}, {4}>(
+    device {2}* block_partitions [[buffer(0)]],
+    const device {1}* dev_vals [[buffer(1)]],
+    const device {2}* dev_idxs [[buffer(2)]],
+    const constant int& size_sorted_axis [[buffer(3)]],
+    const constant int& merge_tiles [[buffer(4)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]],
+    uint3 tgp_dims [[threads_per_threadgroup]]);
+template [[host_name("merge_{0}")]] [[kernel]] void
+mb_block_merge<{1}, {2}, true, {3}, {4}>(
+    const device {2}* block_partitions [[buffer(0)]],
+    const device {1}* dev_vals_in [[buffer(1)]],
+    const device {2}* dev_idxs_in [[buffer(2)]],
+    device {1}* dev_vals_out [[buffer(3)]],
+    device {2}* dev_idxs_out [[buffer(4)]],
+    const constant int& size_sorted_axis [[buffer(5)]],
+    const constant int& merge_tiles [[buffer(6)]],
+    const constant int& num_tiles [[buffer(7)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";

mlx/backend/metal/jit/steel_conv.h

@@ -0,0 +1,32 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view steel_conv_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+implicit_gemm_conv_2d<{itype}, {bm}, {bn}, {bk}, {wm}, {wn}, {n_channels}, {small_filter}>(
+    const device {itype}* A [[buffer(0)]],
+    const device {itype}* B [[buffer(1)]],
+    device {itype}* C [[buffer(2)]],
+    const constant MLXConvParams<2>* params [[buffer(3)]],
+    const constant ImplicitGemmConv2DParams* gemm_params [[buffer(4)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]],
+    uint simd_gid [[simdgroup_index_in_threadgroup]],
+    uint simd_lid [[thread_index_in_simdgroup]]);
+)";
+
+constexpr std::string_view steel_conv_general_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+    implicit_gemm_conv_2d_general<{itype}, {bm}, {bn}, {bk}, {wm}, {wn}>(
+        const device {itype}* A [[buffer(0)]],
+        const device {itype}* B [[buffer(1)]],
+        device {itype}* C [[buffer(2)]],
+        const constant MLXConvParams<2>* params [[buffer(3)]],
+        const constant ImplicitGemmConv2DParams* gemm_params [[buffer(4)]],
+        const constant Conv2DGeneralJumpParams* jump_params [[buffer(5)]],
+        const constant Conv2DGeneralBaseInfo* base_h [[buffer(6)]],
+        const constant Conv2DGeneralBaseInfo* base_w [[buffer(7)]],
+        uint3 tid [[threadgroup_position_in_grid]],
+        uint3 lid [[thread_position_in_threadgroup]],
+        uint simd_gid [[simdgroup_index_in_threadgroup]],
+        uint simd_lid [[thread_index_in_simdgroup]]);
+)";

mlx/backend/metal/jit/steel_gemm.h

@@ -0,0 +1,106 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view steel_gemm_fused_kernels = R"(
+template [[host_name("{name}")]]
+[[kernel]] void gemm<{itype}, {bm}, {bn}, {bk}, {wm}, {wn}, {trans_a}, {trans_b}, float>(
+    const device {itype} *A [[buffer(0)]],
+    const device {itype} *B [[buffer(1)]],
+    const device {itype} *C [[buffer(2), function_constant(use_out_source)]],
+    device {itype} *D [[buffer(3)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    const constant GEMMAddMMParams* addmm_params [[buffer(5), function_constant(use_out_source)]],
+    const constant int* batch_shape [[buffer(6)]],
+    const constant size_t* batch_strides [[buffer(7)]],
+    const constant uint32_t* lhs_indices [[buffer(10), function_constant(do_gather)]],
+    const constant uint32_t* rhs_indices [[buffer(11), function_constant(do_gather)]],
+    const constant uint32_t* C_indices [[buffer(12), function_constant(gather_bias)]],
+    const constant int* operand_shape [[buffer(13), function_constant(do_gather)]],
+    const constant size_t* operand_strides [[buffer(14), function_constant(do_gather)]],
+    const constant packed_int3& operand_batch_ndim [[buffer(15), function_constant(do_gather)]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";
+
+constexpr std::string_view steel_gemm_masked_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+block_masked_gemm<
+    {itype},
+    {outmasktype},
+    {opmasktype},
+    {bm},
+    {bn},
+    {bk},
+    {wm},
+    {wn},
+    {trans_a},
+    {trans_b},
+    {mn_aligned},
+    {k_aligned}>(
+    const device {itype}* A [[buffer(0)]],
+    const device {itype}* B [[buffer(1)]],
+    device {itype}* D [[buffer(3)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    const constant int* batch_shape [[buffer(6)]],
+    const constant size_t* batch_strides [[buffer(7)]],
+    const device {outmasktype}* out_mask [[buffer(10)]],
+    const device {opmasktype}* lhs_mask [[buffer(11)]],
+    const device {opmasktype}* rhs_mask [[buffer(12)]],
+    const constant int* mask_strides [[buffer(13)]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";
+
+constexpr std::string_view steel_gemm_splitk_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+gemm_splitk<
+    {itype},
+    {otype},
+    {bm},
+    {bn},
+    {bk},
+    {wm},
+    {wn},
+    {trans_a},
+    {trans_b},
+    {mn_aligned},
+    {k_aligned}>(
+    const device {itype}* A [[buffer(0)]],
+    const device {itype}* B [[buffer(1)]],
+    device {otype}* C [[buffer(2)]],
+    const constant GEMMSpiltKParams* params [[buffer(3)]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";
+
+constexpr std::string_view steel_gemm_splitk_accum_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+gemm_splitk_accum<{atype}, {otype}>(
+    const device {atype}* C_split [[buffer(0)]],
+    device {otype}* D [[buffer(1)]],
+    const constant int& k_partitions [[buffer(2)]],
+    const constant int& partition_stride [[buffer(3)]],
+    const constant int& ldd [[buffer(4)]],
+    uint2 gid [[thread_position_in_grid]]);
+)";
+
+constexpr std::string_view steel_gemm_splitk_accum_axbpy_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+gemm_splitk_accum_axpby<{atype}, {otype}>(
+    const device {atype}* C_split [[buffer(0)]],
+    device {otype}* D [[buffer(1)]],
+    const constant int& k_partitions [[buffer(2)]],
+    const constant int& partition_stride [[buffer(3)]],
+    const constant int& ldd [[buffer(4)]],
+    const device {otype}* C [[buffer(5)]],
+    const constant int& ldc [[buffer(6)]],
+    const constant int& fdc [[buffer(7)]],
+    const constant float& alpha [[buffer(8)]],
+    const constant float& beta [[buffer(9)]],
+    uint2 gid [[thread_position_in_grid]]);
+)";

mlx/backend/metal/jit/ternary.h

@@ -0,0 +1,80 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view ternary_kernels = R"(
+template [[host_name("v_{0}")]] [[kernel]] void ternary_v<{1}, {2}>(
+    device const bool* a,
+    device const {1}* b,
+    device const {1}* c,
+    device {1}* d,
+    uint index [[thread_position_in_grid]]);
+
+template [[host_name("g_{0}")]] [[kernel]] void ternary_g<{1}, {2}>(
+    device const bool* a,
+    device const {1}* b,
+    device const {1}* c,
+    device {1}* d,
+    constant const int* shape,
+    constant const size_t* a_strides,
+    constant const size_t* b_strides,
+    constant const size_t* c_strides,
+    constant const int& ndim,
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+
+template [[host_name("g1_{0}")]] [[kernel]] void
+ternary_g_nd1<{1}, {2}>(
+    device const bool* a,
+    device const {1}* b,
+    device const {1}* c,
+    device {1}* d,
+    constant const size_t& a_strides,
+    constant const size_t& b_strides,
+    constant const size_t& c_strides,
+    uint index [[thread_position_in_grid]]);
+template [[host_name("g2_{0}")]] [[kernel]] void
+ternary_g_nd2<{1}, {2}>(
+    device const bool* a,
+    device const {1}* b,
+    device const {1}* c,
+    device {1}* d,
+    constant const size_t a_strides[2],
+    constant const size_t b_strides[2],
+    constant const size_t c_strides[2],
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]);
+template [[host_name("g3_{0}")]] [[kernel]] void
+ternary_g_nd3<{1}, {2}>(
+    device const bool* a,
+    device const {1}* b,
+    device const {1}* c,
+    device {1}* d,
+    constant const size_t a_strides[3],
+    constant const size_t b_strides[3],
+    constant const size_t c_strides[3],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("g4_{0}")]] [[kernel]] void
+ternary_g_nd<{1}, {2}, 4>(
+    device const bool* a,
+    device const {1}* b,
+    device const {1}* c,
+    device {1}* d,
+    constant const int shape[4],
+    constant const size_t a_strides[4],
+    constant const size_t b_strides[4],
+    constant const size_t c_strides[4],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+template [[host_name("g5_{0}")]] [[kernel]] void
+ternary_g_nd<{1}, {2}, 5>(
+    device const bool* a,
+    device const {1}* b,
+    device const {1}* c,
+    device {1}* d,
+    constant const int shape[5],
+    constant const size_t a_strides[5],
+    constant const size_t b_strides[5],
+    constant const size_t c_strides[5],
+    uint3 index [[thread_position_in_grid]],
+    uint3 grid_dim [[threads_per_grid]]);
+)";

mlx/backend/metal/jit/unary.h

@@ -0,0 +1,16 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view unary_kernels = R"(
+template [[host_name("v{0}")]] [[kernel]] void unary_v<{1}, {2}>(
+    device const {1}* in,
+    device {1}* out,
+    uint index [[thread_position_in_grid]]);
+
+template [[host_name("g{0}")]] [[kernel]] void unary_g<{1}, {2}>(
+    device const {1}* in,
+    device {1}* out,
+    device const int* in_shape,
+    device const size_t* in_strides,
+    device const int& ndim,
+    uint index [[thread_position_in_grid]]);
+)";

mlx/backend/metal/jit_kernels.cpp

@@ -0,0 +1,486 @@
+// Copyright © 2024 Apple Inc.
+#include <fmt/format.h>
+
+#include "mlx/backend/common/compiled.h"
+#include "mlx/backend/metal/jit/arange.h"
+#include "mlx/backend/metal/jit/binary.h"
+#include "mlx/backend/metal/jit/binary_two.h"
+#include "mlx/backend/metal/jit/copy.h"
+#include "mlx/backend/metal/jit/includes.h"
+#include "mlx/backend/metal/jit/reduce.h"
+#include "mlx/backend/metal/jit/scan.h"
+#include "mlx/backend/metal/jit/softmax.h"
+#include "mlx/backend/metal/jit/sort.h"
+#include "mlx/backend/metal/jit/steel_conv.h"
+#include "mlx/backend/metal/jit/steel_gemm.h"
+#include "mlx/backend/metal/jit/ternary.h"
+#include "mlx/backend/metal/jit/unary.h"
+#include "mlx/backend/metal/kernels.h"
+#include "mlx/backend/metal/utils.h"
+
+using namespace fmt::literals;
+
+namespace mlx::core {
+
+std::string op_name(const array& arr) {
+  std::ostringstream op_t;
+  arr.primitive().print(op_t);
+  return op_t.str();
+}
+
+MTL::ComputePipelineState* get_arange_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& out) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source
+        << metal::utils() << metal::arange()
+        << fmt::format(arange_kernels, lib_name, get_type_string(out.dtype()));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_unary_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::unary_ops() << metal::unary()
+                  << fmt::format(
+                         unary_kernels,
+                         lib_name,
+                         get_type_string(out.dtype()),
+                         op_name(out));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_binary_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(2);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::binary_ops() << metal::binary()
+                  << fmt::format(
+                         binary_kernels,
+                         lib_name,
+                         get_type_string(in.dtype()),
+                         get_type_string(out.dtype()),
+                         op_name(out));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_binary_two_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(2);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::binary_ops()
+                  << metal::binary_two()
+                  << fmt::format(
+                         binary_two_kernels,
+                         lib_name,
+                         get_type_string(in.dtype()),
+                         get_type_string(out.dtype()),
+                         op_name(out));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_ternary_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(kernel_name.find("_") + 1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::ternary_ops() << metal::ternary()
+                  << fmt::format(
+                         ternary_kernels,
+                         lib_name,
+                         get_type_string(out.dtype()),
+                         op_name(out));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_copy_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(kernel_name.find("_") + 1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::copy()
+                  << fmt::format(
+                         copy_kernels,
+                         lib_name,
+                         get_type_string(in.dtype()),
+                         get_type_string(out.dtype()));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_softmax_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    bool precise,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(kernel_name.find("_") + 1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::softmax()
+                  << fmt::format(
+                         softmax_kernels,
+                         lib_name,
+                         get_type_string(out.dtype()),
+                         get_type_string(precise ? float32 : out.dtype()));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_scan_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    bool reverse,
+    bool inclusive,
+    const array& in,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(kernel_name.find("_") + 1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::scan()
+                  << fmt::format(
+                         scan_kernels,
+                         lib_name,
+                         get_type_string(in.dtype()),
+                         get_type_string(out.dtype()),
+                         op_name(out),
+                         inclusive,
+                         reverse);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_sort_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& out,
+    int bn,
+    int tn) {
+  std::string lib_name = kernel_name.substr(kernel_name.find("_") + 1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::sort()
+                  << fmt::format(
+                         block_sort_kernels,
+                         lib_name,
+                         get_type_string(in.dtype()),
+                         get_type_string(out.dtype()),
+                         bn,
+                         tn);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_mb_sort_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& idx,
+    int bn,
+    int tn) {
+  std::string lib_name = kernel_name.substr(kernel_name.find("_") + 1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::sort()
+                  << fmt::format(
+                         multiblock_sort_kernels,
+                         lib_name,
+                         get_type_string(in.dtype()),
+                         get_type_string(idx.dtype()),
+                         bn,
+                         tn);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_reduce_init_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& out) {
+  auto lib = d.get_library(kernel_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::reduce_utils()
+                  << fmt::format(
+                         reduce_init_kernels,
+                         kernel_name,
+                         get_type_string(out.dtype()),
+                         op_name(out));
+    lib = d.get_library(kernel_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_reduce_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& out) {
+  std::string lib_name = kernel_name.substr(kernel_name.find("_") + 1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    bool non_atomic = out.dtype() == int64 || out.dtype() == uint64;
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::reduce_utils() << metal::reduce()
+                  << fmt::format(
+                         non_atomic ? reduce_non_atomic_kernels
+                                    : reduce_kernels,
+                         lib_name,
+                         get_type_string(in.dtype()),
+                         get_type_string(out.dtype()),
+                         op_name(out));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_steel_gemm_fused_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const std::string& hash_name,
+    const metal::MTLFCList& func_consts,
+    const array& out,
+    bool transpose_a,
+    bool transpose_b,
+    int bm,
+    int bn,
+    int bk,
+    int wm,
+    int wn) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::gemm()
+                  << metal::steel_gemm_fused()
+                  << fmt::format(
+                         steel_gemm_fused_kernels,
+                         "name"_a = lib_name,
+                         "itype"_a = get_type_string(out.dtype()),
+                         "bm"_a = bm,
+                         "bn"_a = bn,
+                         "bk"_a = bk,
+                         "wm"_a = wm,
+                         "wn"_a = wn,
+                         "trans_a"_a = transpose_a,
+                         "trans_b"_a = transpose_b);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib, hash_name, func_consts);
+}
+
+MTL::ComputePipelineState* get_steel_gemm_splitk_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& out,
+    bool transpose_a,
+    bool transpose_b,
+    int bm,
+    int bn,
+    int bk,
+    int wm,
+    int wn,
+    bool mn_aligned,
+    bool k_aligned) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::gemm()
+                  << metal::steel_gemm_splitk()
+                  << fmt::format(
+                         steel_gemm_splitk_kernels,
+                         "name"_a = lib_name,
+                         "itype"_a = get_type_string(in.dtype()),
+                         "otype"_a = get_type_string(out.dtype()),
+                         "bm"_a = bm,
+                         "bn"_a = bn,
+                         "bk"_a = bk,
+                         "wm"_a = wm,
+                         "wn"_a = wn,
+                         "trans_a"_a = transpose_a,
+                         "trans_b"_a = transpose_b,
+                         "mn_aligned"_a = mn_aligned,
+                         "k_aligned"_a = k_aligned);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_steel_gemm_splitk_accum_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& in,
+    const array& out,
+    bool axbpy) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::gemm()
+                  << metal::steel_gemm_splitk()
+                  << fmt::format(
+                         axbpy ? steel_gemm_splitk_accum_axbpy_kernels
+                               : steel_gemm_splitk_accum_kernels,
+                         "name"_a = lib_name,
+                         "atype"_a = get_type_string(in.dtype()),
+                         "otype"_a = get_type_string(out.dtype()));
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_steel_gemm_masked_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& out,
+    const std::optional<array>& mask_out,
+    const std::optional<array>& mask_op,
+    bool transpose_a,
+    bool transpose_b,
+    int bm,
+    int bn,
+    int bk,
+    int wm,
+    int wn,
+    bool mn_aligned,
+    bool k_aligned) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    auto out_mask_type = mask_out.has_value()
+        ? get_type_string((*mask_out).dtype())
+        : "nomask_t";
+    auto op_mask_type =
+        mask_op.has_value() ? get_type_string((*mask_op).dtype()) : "nomask_t";
+    kernel_source << metal::utils() << metal::gemm()
+                  << metal::steel_gemm_masked()
+                  << fmt::format(
+                         steel_gemm_masked_kernels,
+                         "name"_a = lib_name,
+                         "itype"_a = get_type_string(out.dtype()),
+                         "outmasktype"_a = out_mask_type,
+                         "opmasktype"_a = op_mask_type,
+                         "bm"_a = bm,
+                         "bn"_a = bn,
+                         "bk"_a = bk,
+                         "wm"_a = wm,
+                         "wn"_a = wn,
+                         "trans_a"_a = transpose_a,
+                         "trans_b"_a = transpose_b,
+                         "mn_aligned"_a = mn_aligned,
+                         "k_aligned"_a = k_aligned);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_steel_conv_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& out,
+    int bm,
+    int bn,
+    int bk,
+    int wm,
+    int wn,
+    int n_channel_specialization,
+    bool small_filter) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::conv() << metal::steel_conv()
+                  << fmt::format(
+                         steel_conv_kernels,
+                         "name"_a = lib_name,
+                         "itype"_a = get_type_string(out.dtype()),
+                         "bm"_a = bm,
+                         "bn"_a = bn,
+                         "bk"_a = bk,
+                         "wm"_a = wm,
+                         "wn"_a = wn,
+                         "n_channels"_a = n_channel_specialization,
+                         "small_filter"_a = small_filter);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+MTL::ComputePipelineState* get_steel_conv_general_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const array& out,
+    int bm,
+    int bn,
+    int bk,
+    int wm,
+    int wn) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    kernel_source << metal::utils() << metal::conv()
+                  << metal::steel_conv_general()
+                  << fmt::format(
+                         steel_conv_general_kernels,
+                         "name"_a = lib_name,
+                         "itype"_a = get_type_string(out.dtype()),
+                         "bm"_a = bm,
+                         "bn"_a = bn,
+                         "bk"_a = bk,
+                         "wm"_a = wm,
+                         "wn"_a = wn);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+} // namespace mlx::core