patch (#1320)

Allows opt-in to memory efficient GPU shader at proscribed sequence
length.  Otherwise, utilizes aggregate MLX primitives for best latency.

.circleci/config.yml

@@ -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: |
@@ -69,13 +64,14 @@ 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
@@ -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:
@@ -132,13 +137,14 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
             brew install python@<< parameters.python_version >>
+            brew install openmpi
             python<< parameters.python_version >> -m venv env
             source env/bin/activate
             pip install --upgrade pip

.github/workflows/pull_request.yml

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

ACKNOWLEDGMENTS.md

@@ -10,12 +10,14 @@ MLX was developed with contributions from the following individuals:
 - Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops. Added `clip_grad_norm` along with `tree_reduce`.
 - Juarez Bochi: Fixed bug in cross attention.
 - Justin Deschenaux: Sine, Cosine, arange, randint, truncated normal, bernoulli, lion optimizer, Dropout2d, linear and logistic regression python example.
-- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream` and safetensor support.
+- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream`, safetensors support, `einsum`, and `einsum_path`.
 - Gabrijel Boduljak: Added `mlx.core.linalg`, implemented `norm` method and `InstanceNorm` layer. Implemented pooling layers and ``Upsample``.
 - Hinrik Snær Guðmundsson: Added `atleast_1d`, `atleast_2d`, `atleast_3d` ops.
 - Luca Arnaboldi: Added `Ceil` and `Floor` ops; implemented pickling, copy and deepcopy for mlx arrays.
 - Brian Keene & Atila Orhon, with Argmax Inc.: Added `fast.scaled_dot_product_attention`
 - AmirHossein Razlighi: Added chaining support for some of the ops in `nn.Module`. Comparison works for non array objects in `mlx.core.array`. Exception handling for invalid operations in `mlx.core.array`.
+- Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
+- Paul Paczuski: Improved stability of BCE loss calculation
 
 <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" />

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 0.12.2)
+  set(MLX_VERSION 0.16.3)
 endif()
 
 # --------------------- Processor tests -------------------------
@@ -79,22 +83,21 @@ elseif (MLX_BUILD_METAL)
                   OUTPUT_VARIABLE MACOS_VERSION
                   COMMAND_ERROR_IS_FATAL ANY)
 
-  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
-
-  if (${MACOS_VERSION} GREATER_EQUAL 14.2)
-    set(METAL_CPP_PATCH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/metal.14.2.diff)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14.2_iOS17.2.zip)
-  elseif (${MACOS_VERSION} GREATER_EQUAL 14.0)
-    set(METAL_CPP_PATCH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/metal.14.0.diff)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14_iOS17-beta.zip)
-  else()
+  if (${MACOS_VERSION} LESS 14.0)
     message(FATAL_ERROR "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON" )
   endif()
+  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
+
+  set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18-beta.zip)
+  # Get the metal version
+  execute_process(
+    COMMAND zsh "-c" "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal -E -x metal -P - | tail -1 | tr -d '\n'"
+    OUTPUT_VARIABLE MLX_METAL_VERSION
+    COMMAND_ERROR_IS_FATAL ANY)
 
   FetchContent_Declare(
     metal_cpp
     URL ${METAL_CPP_URL}
-    PATCH_COMMAND /usr/bin/patch -N -i ${METAL_CPP_PATCH} || true
   )
 
   FetchContent_MakeAvailable(metal_cpp)
@@ -104,55 +107,85 @@ elseif (MLX_BUILD_METAL)
     $<INSTALL_INTERFACE:include/metal_cpp>
   )
   target_link_libraries(
-    mlx
+    mlx PUBLIC
     ${METAL_LIB}
     ${FOUNDATION_LIB}
     ${QUARTZ_LIB})
+
+  add_compile_definitions("MLX_METAL_VERSION=${MLX_METAL_VERSION}")
 endif()
 
-find_library(ACCELERATE_LIBRARY Accelerate)
-if (MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
-  message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
-  set(MLX_BUILD_ACCELERATE ON)
-  target_link_libraries(mlx ${ACCELERATE_LIBRARY})
-  add_compile_definitions(ACCELERATE_NEW_LAPACK)
+if (MLX_BUILD_CPU)
+  find_library(ACCELERATE_LIBRARY Accelerate)
+  if (MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
+    message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
+    set(MLX_BUILD_ACCELERATE ON)
+    target_link_libraries(mlx PUBLIC ${ACCELERATE_LIBRARY})
+    add_compile_definitions(ACCELERATE_NEW_LAPACK)
+  else()
+    message(STATUS "Accelerate or arm neon not found, using default backend.")
+    set(MLX_BUILD_ACCELERATE OFF)
+    if(${CMAKE_HOST_APPLE})
+      # The blas shipped in macOS SDK is not supported, search homebrew for
+      # openblas instead.
+      set(BLA_VENDOR OpenBLAS)
+      set(LAPACK_ROOT "${LAPACK_ROOT};$ENV{LAPACK_ROOT};/usr/local/opt/openblas")
+    endif()
+    # Search and link with lapack.
+    find_package(LAPACK REQUIRED)
+    if (NOT LAPACK_FOUND)
+      message(FATAL_ERROR "Must have LAPACK installed")
+    endif()
+    find_path(LAPACK_INCLUDE_DIRS lapacke.h
+      /usr/include
+      /usr/local/include
+      /usr/local/opt/openblas/include)
+    message(STATUS "Lapack lib " ${LAPACK_LIBRARIES})
+    message(STATUS "Lapack include " ${LAPACK_INCLUDE_DIRS})
+    target_include_directories(mlx PRIVATE ${LAPACK_INCLUDE_DIRS})
+    target_link_libraries(mlx PUBLIC ${LAPACK_LIBRARIES})
+    # List blas after lapack otherwise we may accidentally incldue an old version
+    # of lapack.h from the include dirs of blas.
+    find_package(BLAS REQUIRED)
+    if (NOT BLAS_FOUND)
+      message(FATAL_ERROR "Must have BLAS installed")
+    endif()
+    # TODO find a cleaner way to do this
+    find_path(BLAS_INCLUDE_DIRS cblas.h
+      /usr/include
+      /usr/local/include
+      $ENV{BLAS_HOME}/include)
+    message(STATUS "Blas lib " ${BLAS_LIBRARIES})
+    message(STATUS "Blas include " ${BLAS_INCLUDE_DIRS})
+    target_include_directories(mlx PRIVATE ${BLAS_INCLUDE_DIRS})
+    target_link_libraries(mlx PUBLIC ${BLAS_LIBRARIES})
+  endif()
 else()
-  message(STATUS "Accelerate or arm neon not found, using default backend.")
   set(MLX_BUILD_ACCELERATE OFF)
-  if(${CMAKE_HOST_APPLE})
-    # The blas shipped in macOS SDK is not supported, search homebrew for
-    # openblas instead.
-    set(BLA_VENDOR OpenBLAS)
-    set(LAPACK_ROOT "${LAPACK_ROOT};$ENV{LAPACK_ROOT};/usr/local/opt/openblas")
-  endif()
-  # Search and link with lapack.
-  find_package(LAPACK REQUIRED)
-  if (NOT LAPACK_FOUND)
-    message(FATAL_ERROR "Must have LAPACK installed")
-  endif()
-  find_path(LAPACK_INCLUDE_DIRS lapacke.h
-    /usr/include
-    /usr/local/include
-    /usr/local/opt/openblas/include)
-  message(STATUS "Lapack lib " ${LAPACK_LIBRARIES})
-  message(STATUS "Lapack include " ${LAPACK_INCLUDE_DIRS})
-  target_include_directories(mlx PRIVATE ${LAPACK_INCLUDE_DIRS})
-  target_link_libraries(mlx ${LAPACK_LIBRARIES})
-  # List blas after lapack otherwise we may accidentally incldue an old version
-  # of lapack.h from the include dirs of blas.
-  find_package(BLAS REQUIRED)
-  if (NOT BLAS_FOUND)
-    message(FATAL_ERROR "Must have BLAS installed")
-  endif()
-  # TODO find a cleaner way to do this
-  find_path(BLAS_INCLUDE_DIRS cblas.h
-    /usr/include
-    /usr/local/include
-    $ENV{BLAS_HOME}/include)
-  message(STATUS "Blas lib " ${BLAS_LIBRARIES})
-  message(STATUS "Blas include " ${BLAS_INCLUDE_DIRS})
-  target_include_directories(mlx PRIVATE ${BLAS_INCLUDE_DIRS})
-  target_link_libraries(mlx ${BLAS_LIBRARIES})
+endif()
+
+find_package(MPI)
+if (MPI_FOUND)
+  execute_process(
+    COMMAND zsh "-c" "mpirun --version"
+    OUTPUT_VARIABLE MPI_VERSION
+    ERROR_QUIET
+  )
+  if (${MPI_VERSION} MATCHES ".*Open MPI.*")
+    target_include_directories(mlx PRIVATE ${MPI_INCLUDE_PATH})
+  elseif (MPI_VERSION STREQUAL "")
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING
+      "MPI found but mpirun is not available. Building without MPI."
+    )
+  else()
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING
+      "MPI which is not OpenMPI found. Building without MPI."
+    )
+  endif() 
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
@@ -164,6 +197,14 @@ target_include_directories(
   $<INSTALL_INTERFACE:include>
 )
 
+FetchContent_Declare(fmt
+  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+  GIT_TAG 10.2.1 
+  EXCLUDE_FROM_ALL
+)
+FetchContent_MakeAvailable(fmt)
+target_link_libraries(mlx PRIVATE fmt::fmt-header-only)
+
 if (MLX_BUILD_PYTHON_BINDINGS)
   message(STATUS "Building Python bindings.")
   find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)

README.md

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

benchmarks/python/comparative/bench_torch.py

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

benchmarks/python/comparative/compare.py

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

benchmarks/python/compile_bench.py

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

benchmarks/python/conv_bench.py

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

benchmarks/python/einsum_bench.py

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

benchmarks/python/fft_bench.py

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

benchmarks/python/hadamard_bench.py

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

benchmarks/python/sdpa_bench.py

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

cmake/metal.14.0.diff

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

cmake/metal.14.2.diff

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

docs/requirements.txt

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

docs/src/conf.py

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

docs/src/dev/extensions.rst

@@ -1,5 +1,5 @@
-Developer Documentation
-=======================
+Custom Extensions in MLX
+========================
 
 You can extend MLX with custom operations on the CPU or GPU. This guide
 explains how to do that with a simple example.
@@ -486,15 +486,14 @@ below.
         std::ostringstream kname;
         kname << "axpby_" << "general_" << type_to_name(out);
 
-        // Make sure the metal library is available and look for it
-        // in the same folder as this executable if needed
-        d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
+        // Make sure the metal library is available
+        d.register_library("mlx_ext");
 
         // Make a kernel from this metal library
         auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
         // Prepare to encode kernel
-        auto compute_encoder = d.get_command_encoder(s.index);
+        auto& compute_encoder = d.get_command_encoder(s.index);
         compute_encoder->setComputePipelineState(kernel);
 
         // Kernel parameters are registered with buffer indices corresponding to
@@ -503,11 +502,11 @@ 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);
@@ -531,7 +530,7 @@ below.
 
         // Launch the grid with the given number of threads divided among
         // the given threadgroups
-        compute_encoder->dispatchThreads(grid_dims, group_dims);
+        compute_encoder.dispatchThreads(grid_dims, group_dims);
     }
 
 We can now call the :meth:`axpby` operation on both the CPU and the GPU!
@@ -825,7 +824,7 @@ Let's look at a simple script and its results:
 
     print(f"c shape: {c.shape}")
     print(f"c dtype: {c.dtype}")
-    print(f"c correctness: {mx.all(c == 6.0).item()}")
+    print(f"c correct: {mx.all(c == 6.0).item()}")
 
 Output:
 

docs/src/examples/llama-inference.rst

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

docs/src/examples/mlp.rst

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

docs/src/index.rst

@@ -43,6 +43,7 @@ are the CPU and GPU.
    usage/function_transforms
    usage/compile
    usage/numpy
+   usage/distributed
    usage/using_streams
 
 .. toctree::
@@ -69,6 +70,7 @@ are the CPU and GPU.
    python/metal
    python/nn
    python/optimizers
+   python/distributed
    python/tree_utils
 
 .. toctree::

docs/src/install.rst

@@ -70,36 +70,36 @@ To build and install the MLX python library from source, first, clone MLX from
 
    git clone git@github.com:ml-explore/mlx.git mlx && cd mlx
 
-Install `nanobind <https://nanobind.readthedocs.io/en/latest/>`_ with:
-
-.. code-block:: shell
-
-    pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
-
 Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-   env CMAKE_BUILD_PARALLEL_LEVEL="" pip install .
+  CMAKE_BUILD_PARALLEL_LEVEL="" pip install .
 
-For developing use an editable install:
+For developing, install the package with development dependencies, and use an
+editable install:
 
 .. code-block:: shell
 
-  env CMAKE_BUILD_PARALLEL_LEVEL="" pip install -e .
+  CMAKE_BUILD_PARALLEL_LEVEL="" pip install -e ".[dev]"
 
-To make sure the install is working run the tests with:
+Once the development dependencies are installed, you can build faster with:
+
+.. code-block:: shell
+
+ CMAKE_BUILD_PARALLEL_LEVEL="" python setup.py build_ext -j --inplace
+
+Run the tests with:
 
 .. code-block:: shell
 
-  pip install ".[testing]"
   python -m unittest discover python/tests
 
-Optional: Install stubs to enable auto completions and type checking from your IDE:
+Optional: Install stubs to enable auto completions and type checking from your
+IDE:
 
 .. code-block:: shell
 
-  pip install ".[dev]"
   python setup.py generate_stubs
 
 C++ API
@@ -153,11 +153,18 @@ should point to the path to the built metal library.
      - OFF
    * - MLX_BUILD_METAL
      - ON
+   * - MLX_BUILD_CPU
+     - ON
    * - MLX_BUILD_PYTHON_BINDINGS
      - OFF
    * - MLX_METAL_DEBUG
      - OFF
-
+   * - MLX_BUILD_SAFETENSORS
+     - ON
+   * - MLX_BUILD_GGUF
+     - ON
+   * - MLX_METAL_JIT
+     - OFF
 
 .. note::
 
@@ -176,10 +183,37 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
+Binary Size Minimization
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+To produce a smaller binary use the CMake flags ``CMAKE_BUILD_TYPE=MinSizeRel``
+and ``BUILD_SHARED_LIBS=ON``.
+
+The MLX CMake build has several additional options to make smaller binaries.
+For example, if you don't need the CPU backend or support for safetensors and
+GGUF, you can do:
+
+.. code-block:: shell
+
+  cmake .. \
+    -DCMAKE_BUILD_TYPE=MinSizeRel \
+    -DBUILD_SHARED_LIBS=ON \
+    -DMLX_BUILD_CPU=OFF \
+    -DMLX_BUILD_SAFETENSORS=OFF \
+    -DMLX_BUILD_GGUF=OFF \
+    -DMLX_METAL_JIT=ON
+
+THE ``MLX_METAL_JIT`` flag minimizes the size of the MLX Metal library which
+contains pre-built GPU kernels. This substantially reduces the size of the
+Metal library by run-time compiling kernels the first time they are used in MLX
+on a given machine. Note run-time compilation incurs a cold-start cost which can
+be anwywhere from a few hundred millisecond to a few seconds depending on the
+application. Once a kernel is compiled, it will be cached by the system. The
+Metal kernel cache persists accross reboots.
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 
-
 Metal not found
 ~~~~~~~~~~~~~~~
 

docs/src/python/array.rst

@@ -24,6 +24,7 @@ Array
     array.any
     array.argmax
     array.argmin
+    array.conj
     array.cos
     array.cummax
     array.cummin
@@ -57,3 +58,4 @@ Array
     array.transpose
     array.T
     array.var
+    array.view

docs/src/python/distributed.rst

@@ -0,0 +1,19 @@
+.. _distributed:
+
+.. currentmodule:: mlx.core.distributed
+
+Distributed Communication
+==========================
+
+MLX provides a distributed communication package using MPI. The MPI library is
+loaded at runtime; if MPI is available then distributed communication is also
+made available.
+
+.. autosummary::
+   :toctree: _autosummary
+
+    Group
+    is_available
+    init
+    all_sum
+    all_gather

docs/src/python/linalg.rst

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

docs/src/python/nn/functions.rst

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

docs/src/python/nn/layers.rst

@@ -15,15 +15,21 @@ Layers
    BatchNorm
    Conv1d
    Conv2d
+   Conv3d
    Dropout
    Dropout2d
    Dropout3d
    Embedding
    GELU
+   GLU
    GroupNorm
    GRU
+   HardShrink
+   HardTanh
+   Hardswish
    InstanceNorm
    LayerNorm
+   LeakyReLU
    Linear
    LSTM
    MaxPool1d
@@ -35,13 +41,19 @@ Layers
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU6
    RNN
    RoPE
    SELU
    Sequential
    SiLU
    SinusoidalPositionalEncoding
+   Softmin
    Softshrink
+   Softsign
+   Softmax
+   Softplus
    Step
+   Tanh
    Transformer
    Upsample

docs/src/python/ops.rst

@@ -10,6 +10,7 @@ Operations
 
    abs
    add
+   addmm
    all
    allclose
    any
@@ -19,12 +20,14 @@ Operations
    arcsin
    arcsinh
    arctan
+   arctan2
    arctanh
    argmax
    argmin
    argpartition
    argsort
    array_equal
+   as_strided
    atleast_1d
    atleast_2d
    atleast_3d
@@ -32,11 +35,12 @@ Operations
    bitwise_or
    bitwise_xor
    block_masked_mm
-   block_sparse_mm
    broadcast_to
    ceil
    clip
    concatenate
+   conj
+   conjugate
    convolve
    conv1d
    conv2d
@@ -53,6 +57,8 @@ Operations
    diagonal
    divide
    divmod
+   einsum
+   einsum_path
    equal
    erf
    erfinv
@@ -64,8 +70,11 @@ Operations
    floor
    floor_divide
    full
+   gather_mm
+   gather_qmm
    greater
    greater_equal
+   hadamard_transform
    identity
    inner
    isclose
@@ -73,6 +82,7 @@ Operations
    isnan
    isneginf
    isposinf
+   issubdtype
    left_shift
    less
    less_equal
@@ -96,6 +106,7 @@ Operations
    minimum
    moveaxis
    multiply
+   nan_to_num
    negative
    not_equal
    ones
@@ -103,11 +114,13 @@ Operations
    outer
    partition
    pad
+   power
    prod
    quantize
    quantized_matmul
    radians
    reciprocal
+   remainder
    repeat
    reshape
    right_shift
@@ -141,11 +154,13 @@ Operations
    tensordot
    tile
    topk
+   trace
    transpose
    tri
    tril
    triu
    var
+   view
    where
    zeros
    zeros_like

docs/src/python/optimizers.rst

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

docs/src/python/random.rst

@@ -44,3 +44,4 @@ we use a splittable version of Threefry, which is a counter-based PRNG.
    split
    truncated_normal
    uniform
+   laplace

docs/src/python/transforms.rst

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

docs/src/usage/distributed.rst

@@ -0,0 +1,166 @@
+.. _usage_distributed:
+
+Distributed Communication
+=========================
+
+.. currentmodule:: mlx.core.distributed
+
+MLX utilizes `MPI <https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ to
+provide distributed communication operations that allow the computational cost
+of training or inference to be shared across many physical machines. You can
+see a list of the supported operations in the :ref:`API docs<distributed>`.
+
+.. note::
+   A lot of operations may not be supported or not as fast as they should be.
+   We are adding more and tuning the ones we have as we are figuring out the
+   best way to do distributed computing on Macs using MLX.
+
+Getting Started
+---------------
+
+MLX already comes with the ability to "talk" to MPI if it is installed on the
+machine. The minimal distributed program in MLX is as simple as:
+
+.. code:: python
+
+    import mlx.core as mx
+
+    world = mx.distributed.init()
+    x = mx.distributed.all_sum(mx.ones(10))
+    print(world.rank(), x)
+
+The program above sums the array ``mx.ones(10)`` across all
+distributed processes. If simply run with ``python``, however, only one
+process is launched and no distributed communication takes place.
+
+To launch the program in distributed mode we need to use ``mpirun`` or
+``mpiexec`` depending on the MPI installation. The simplest possible way is the
+following:
+
+.. code:: shell
+
+    $ mpirun -np 2 python test.py
+    1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+    0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+
+The above launches two processes on the same (local) machine and we can see
+both standard output streams. The processes send the array of 1s to each other
+and compute the sum which is printed. Launching with ``mpirun -np 4 ...`` would
+print 4 etc.
+
+Installing MPI
+---------------
+
+MPI can be installed with Homebrew, using the Anaconda package manager or
+compiled from source. Most of our testing is done using ``openmpi`` installed
+with the Anaconda package manager as follows:
+
+.. code:: shell
+
+    $ conda install openmpi
+
+Installing with Homebrew may require specifying the location of ``libmpi.dyld``
+so that MLX can find it and load it at runtime. This can simply be achieved by
+passing the ``DYLD_LIBRARY_PATH`` environment variable to ``mpirun``.
+
+.. code:: shell
+
+    $ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python test.py
+
+Setting up Remote Hosts
+-----------------------
+
+MPI can automatically connect to remote hosts and set up the communication over
+the network if the remote hosts can be accessed via ssh. A good checklist to
+debug connectivity issues is the following:
+
+* ``ssh hostname`` works from all machines to all machines without asking for
+  password or host confirmation
+* ``mpirun`` is accessible on all machines. You can call ``mpirun`` using its
+  full path to force all machines to use a specific path.
+* Ensure that the ``hostname`` used by MPI is the one that you have configured
+  in the ``.ssh/config`` files on all machines.
+
+.. note::
+  For an example hostname ``foo.bar.com`` MPI can use only ``foo`` as
+  the hostname passed to ssh if the current hostname matches ``*.bar.com``.
+
+An easy way to pass the host names to MPI is using a host file. A host file
+looks like the following, where ``host1`` and ``host2`` should be the fully
+qualified domain names or IPs for these hosts.
+
+.. code::
+
+    host1 slots=1
+    host2 slots=1
+
+When using MLX, it is very likely that you want to use 1 slot per host, ie one
+process per host.  The hostfile also needs to contain the current
+host if you want to run on the local host. Passing the host file to
+``mpirun`` is simply done using the ``--hostfile`` command line argument.
+
+Training Example
+----------------
+
+In this section we will adapt an MLX training loop to support data parallel
+distributed training. Namely, we will average the gradients across a set of
+hosts before applying them to the model.
+
+Our training loop looks like the following code snippet if we omit the model,
+dataset and optimizer initialization.
+
+.. code:: python
+
+    model = ...
+    optimizer = ...
+    dataset = ...
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        optimizer.update(model, grads)
+        return loss
+
+    for x, y in dataset:
+        loss = step(model, x, y)
+        mx.eval(loss, model.parameters())
+
+All we have to do to average the gradients across machines is perform an
+:func:`all_sum` and divide by the size of the :class:`Group`. Namely we
+have to :func:`mlx.utils.tree_map` the gradients with following function.
+
+.. code:: python
+
+    def all_avg(x):
+        return mx.distributed.all_sum(x) / mx.distributed.init().size()
+
+Putting everything together our training loop step looks as follows with
+everything else remaining the same.
+
+.. code:: python
+
+    from mlx.utils import tree_map
+
+    def all_reduce_grads(grads):
+        N = mx.distributed.init()
+        if N == 1:
+            return grads
+        return tree_map(
+                lambda x: mx.distributed.all_sum(x) / N,
+                grads)
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        grads = all_reduce_grads(grads)  # <--- This line was added
+        optimizer.update(model, grads)
+        return loss
+
+Tuning All Reduce
+-----------------
+
+We are working on improving the performance of all reduce on MLX but for now
+the two main things one can do to extract the most out of distributed training with MLX are:
+
+1. Perform a few large reductions instead of many small ones to improve
+   bandwidth and latency
+2. Pass ``--mca btl_tcp_links 4`` to ``mpirun`` to configure it to use 4 tcp
+   connections between each host to improve bandwidth

docs/src/usage/numpy.rst

@@ -3,7 +3,11 @@
 Conversion to NumPy and Other Frameworks
 ========================================
 
-MLX array implements the `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_.
+MLX array supports conversion between other frameworks with either:  
+
+* The `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_. 
+* `DLPack <https://dmlc.github.io/dlpack/latest/>`_.  
+
 Let's convert an array to NumPy and back.
 
 .. code-block:: python

examples/cpp/CMakeLists.txt

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

examples/cpp/distributed.cpp

@@ -0,0 +1,22 @@
+// Copyright © 2024 Apple Inc.
+
+#include <iostream>
+
+#include "mlx/mlx.h"
+
+using namespace mlx::core;
+
+int main() {
+  if (!distributed::is_available()) {
+    std::cout << "No communication backend found" << std::endl;
+    return 1;
+  }
+
+  auto global_group = distributed::init();
+  std::cout << global_group.rank() << " / " << global_group.size() << std::endl;
+
+  array x = ones({10});
+  array out = distributed::all_sum(x, global_group);
+
+  std::cout << out << std::endl;
+}

examples/cpp/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/README.md

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

examples/extensions/axpby/axpby.cpp

@@ -249,15 +249,14 @@ void Axpby::eval_gpu(
   kname << (contiguous_kernel ? "contiguous_" : "general_");
   kname << type_to_name(out);
 
-  // Make sure the metal library is available and look for it
-  // in the same folder as this executable if needed
-  d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
+  // Make sure the metal library is available
+  d.register_library("mlx_ext");
 
   // Make a kernel from this metal library
   auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
   // Prepare to encode kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
 
   // Kernel parameters are registered with buffer indices corresponding to
@@ -266,11 +265,11 @@ void Axpby::eval_gpu(
   size_t nelem = out.size();
 
   // Encode input arrays to kernel
-  set_array_buffer(compute_encoder, x, 0);
-  set_array_buffer(compute_encoder, y, 1);
+  compute_encoder.set_input_array(x, 0);
+  compute_encoder.set_input_array(y, 1);
 
   // Encode output arrays to kernel
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode alpha and beta
   compute_encoder->setBytes(&alpha_, sizeof(float), 3);
@@ -296,7 +295,7 @@ void Axpby::eval_gpu(
 
   // Launch the grid with the given number of threads divided among
   // the given threadgroups
-  compute_encoder->dispatchThreads(grid_dims, group_dims);
+  compute_encoder.dispatchThreads(grid_dims, group_dims);
 }
 
 #else // Metal is not available

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/requirements.txt

@@ -1,4 +1,4 @@
 setuptools>=42
 cmake>=3.24
-mlx>=0.9.0
-nanobind@git+https://github.com/wjakob/nanobind.git#egg=4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+mlx>=0.16.2
+nanobind==2.0

examples/extensions/test.py

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

mlx/CMakeLists.txt

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

mlx/array.cpp

@@ -17,6 +17,10 @@ bool in_tracing() {
   return detail::InTracing::in_tracing();
 }
 
+bool retain_graph() {
+  return detail::RetainGraph::retain_graph();
+}
+
 } // namespace
 
 array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
@@ -102,7 +106,7 @@ void array::eval() {
 }
 
 bool array::is_tracer() const {
-  return array_desc_->is_tracer && in_tracing();
+  return array_desc_->is_tracer && in_tracing() || retain_graph();
 }
 
 void array::set_data(allocator::Buffer buffer, deleter_t d) {
@@ -171,10 +175,11 @@ array::~array() {
     return;
   }
 
-  // Ignore arrays that will be detached
-  if (status() != array::Status::unscheduled) {
+  // Ignore arrays that might be detached during eval
+  if (status() == array::Status::scheduled) {
     return;
   }
+
   // Break circular reference for non-detached arrays with siblings
   if (auto n = siblings().size(); n > 0) {
     bool do_detach = true;
@@ -206,7 +211,7 @@ void array::ArrayDesc::init() {
     strides[i] = size;
     size *= shape[i];
   }
-  for (auto& in : inputs) {
+  for (const auto& in : inputs) {
     is_tracer |= in.is_tracer();
   }
 }
@@ -231,7 +236,7 @@ array::ArrayDesc::ArrayDesc(
 
 array::ArrayDesc::~ArrayDesc() {
   // When an array description is destroyed it will delete a bunch of arrays
-  // that may also destory their corresponding descriptions and so on and so
+  // that may also destroy their corresponding descriptions and so on and so
   // forth.
   //
   // This calls recursively the destructor and can result in stack overflow, we

mlx/array.h

@@ -73,32 +73,32 @@ class array {
       this->array_desc_ = other.array_desc_;
     }
     return *this;
-  };
+  }
 
   /** The size of the array's datatype in bytes. */
   size_t itemsize() const {
     return size_of(dtype());
-  };
+  }
 
   /** The number of elements in the array. */
   size_t size() const {
     return array_desc_->size;
-  };
+  }
 
   /** The number of bytes in the array. */
   size_t nbytes() const {
     return size() * itemsize();
-  };
+  }
 
   /** The number of dimensions of the array. */
   size_t ndim() const {
     return array_desc_->shape.size();
-  };
+  }
 
   /** The shape of the array as a vector of integers. */
   const std::vector<int>& shape() const {
     return array_desc_->shape;
-  };
+  }
 
   /**
    *  Get the size of the corresponding dimension.
@@ -107,12 +107,12 @@ class array {
    *  bounds checking. */
   int shape(int dim) const {
     return shape().at(dim < 0 ? dim + ndim() : dim);
-  };
+  }
 
   /** The strides of the array. */
   const std::vector<size_t>& strides() const {
     return array_desc_->strides;
-  };
+  }
 
   /**
    *  Get the stride of the corresponding dimension.
@@ -121,12 +121,12 @@ class array {
    *  bounds checking. */
   size_t strides(int dim) const {
     return strides().at(dim < 0 ? dim + ndim() : dim);
-  };
+  }
 
   /** Get the arrays data type. */
   Dtype dtype() const {
     return array_desc_->dtype;
-  };
+  }
 
   /** Evaluate the array. */
   void eval();
@@ -160,10 +160,10 @@ class array {
 
     friend bool operator==(const ArrayIterator& a, const ArrayIterator& b) {
       return a.arr.id() == b.arr.id() && a.idx == b.idx;
-    };
+    }
     friend bool operator!=(const ArrayIterator& a, const ArrayIterator& b) {
       return !(a == b);
-    };
+    }
 
    private:
     const array& arr;
@@ -209,7 +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;
@@ -230,22 +230,22 @@ class array {
   /** The array's primitive. */
   Primitive& primitive() const {
     return *(array_desc_->primitive);
-  };
+  }
 
   /** A shared pointer to the array's primitive. */
   std::shared_ptr<Primitive>& primitive_ptr() const {
     return array_desc_->primitive;
-  };
+  }
 
   /** Check if the array has an attached primitive or is a leaf node. */
   bool has_primitive() const {
     return array_desc_->primitive != nullptr;
-  };
+  }
 
   /** The array's inputs. */
   const std::vector<array>& inputs() const {
     return array_desc_->inputs;
-  };
+  }
 
   std::vector<array>& inputs() {
     return array_desc_->inputs;
@@ -259,12 +259,12 @@ class array {
   /** The array's siblings. */
   const std::vector<array>& siblings() const {
     return array_desc_->siblings;
-  };
+  }
 
   /** The array's siblings. */
   std::vector<array>& siblings() {
     return array_desc_->siblings;
-  };
+  }
 
   void set_siblings(std::vector<array> siblings, uint16_t position) {
     array_desc_->siblings = std::move(siblings);
@@ -281,7 +281,7 @@ class array {
     outputs.push_back(*this);
     outputs.insert(outputs.end(), siblings().begin() + idx, siblings().end());
     return outputs;
-  };
+  }
 
   /** Detach the array from the graph. */
   void detach();
@@ -289,19 +289,19 @@ class array {
   /** Get the Flags bit-field. */
   const Flags& flags() const {
     return array_desc_->flags;
-  };
+  }
 
   /** The size (in elements) of the underlying buffer the array points to. */
   size_t data_size() const {
     return array_desc_->data_size;
-  };
+  }
 
   allocator::Buffer& buffer() {
     return array_desc_->data->buffer;
-  };
+  }
   const allocator::Buffer& buffer() const {
     return array_desc_->data->buffer;
-  };
+  }
 
   // Return a copy of the shared pointer
   // to the array::Data struct
@@ -312,19 +312,20 @@ class array {
   template <typename T>
   T* data() {
     return static_cast<T*>(array_desc_->data_ptr);
-  };
+  }
 
   template <typename T>
   const T* data() const {
     return static_cast<T*>(array_desc_->data_ptr);
-  };
+  }
 
   enum Status { unscheduled, scheduled, available };
 
   bool is_available() const {
     return status() == Status::available;
   }
-  const Status status() const {
+
+  Status status() const {
     return array_desc_->status;
   }
 

mlx/backend/accelerate/conv.cpp

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

mlx/backend/accelerate/matmul.cpp

@@ -2,8 +2,7 @@
 
 #include <cassert>
 
-#include <vecLib/BNNS/bnns.h>
-#include <vecLib/cblas_new.h>
+#include <Accelerate/Accelerate.h>
 
 #include "mlx/backend/accelerate/utils.h"
 #include "mlx/backend/common/copy.h"

mlx/backend/accelerate/primitives.cpp

@@ -3,8 +3,7 @@
 #include <cassert>
 #include <cmath>
 
-#include <vecLib/vDSP.h>
-#include <vecLib/vForce.h>
+#include <Accelerate/Accelerate.h>
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/binary.h"
@@ -32,12 +31,12 @@ DEFAULT(ArgReduce)
 DEFAULT(ArgSort)
 DEFAULT(AsStrided)
 DEFAULT(BlockMaskedMM)
-DEFAULT(BlockSparseMM)
 DEFAULT(Broadcast)
 DEFAULT(Ceil)
 DEFAULT(Concatenate)
+DEFAULT(Conjugate)
 DEFAULT(Copy)
-DEFAULT_MULTI(CustomVJP)
+DEFAULT_MULTI(CustomTransforms)
 DEFAULT_MULTI(Depends)
 DEFAULT_MULTI(DivMod)
 DEFAULT(NumberOfElements)
@@ -47,8 +46,11 @@ DEFAULT(ErfInv)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Gather)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
 DEFAULT(Less)
 DEFAULT(LessEqual)
 DEFAULT(Load)
@@ -78,6 +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);
@@ -99,7 +102,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -114,7 +117,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vadd((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
         });
   } else if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -129,7 +132,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vaddi((const int*)a, 1, (const int*)b, 1, (int*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x + y; });
+    eval(inputs, out);
   }
 }
 
@@ -193,6 +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];
@@ -264,7 +287,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -277,7 +300,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdivi((const int*)b, 1, (const int*)a, 1, (int*)o, 1, n);
         });
   } else if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -292,7 +315,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdiv((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x / y; });
+    eval(inputs, out);
   }
 }
 
@@ -303,12 +326,8 @@ void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
     set_unary_output_data(in, out);
     auto size = in.data_size();
     vvexpf(out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, [](auto x) { return std::exp(x); });
   } else {
-    throw std::invalid_argument(
-        "[exp] Cannot exponentiate elements in array"
-        " with non floating point type.");
+    eval(inputs, out);
   }
 }
 
@@ -370,12 +389,8 @@ void Log1p::eval_cpu(const std::vector<array>& inputs, array& out) {
     auto size = in.data_size();
     vvlog1pf(
         out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
-  } else if (issubdtype(out.dtype(), inexact)) {
-    unary_fp(in, out, [](auto x) { return std::log1p(x); });
   } else {
-    throw std::invalid_argument(
-        "[log1p] Cannot compute log of elements in array with"
-        " non floating point type.");
+    eval(inputs, out);
   }
 }
 
@@ -385,7 +400,7 @@ void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -400,7 +415,7 @@ void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vmul((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x * y; });
+    eval(inputs, out);
   }
 }
 
@@ -411,7 +426,7 @@ void Negative::eval_cpu(const std::vector<array>& inputs, array& out) {
     set_unary_output_data(in, out);
     vDSP_vneg(in.data<float>(), 1, out.data<float>(), 1, in.data_size());
   } else {
-    unary(in, out, [](auto x) { return -x; });
+    eval(inputs, out);
   }
 }
 
@@ -498,7 +513,7 @@ void Square::eval_cpu(const std::vector<array>& inputs, array& out) {
     auto size = in.data_size();
     vDSP_vsq(in.data<float>(), 1, out.data<float>(), 1, size);
   } else {
-    unary(in, out, [](auto x) { return x * x; });
+    eval(inputs, out);
   }
 }
 
@@ -524,7 +539,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -542,7 +557,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vsub((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
         });
   } else if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -554,7 +569,7 @@ void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
         },
         UseDefaultBinaryOp());
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x - y; });
+    eval(inputs, out);
   }
 }
 

mlx/backend/accelerate/reduce.cpp

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

mlx/backend/accelerate/softmax.cpp

@@ -3,7 +3,10 @@
 #include <cassert>
 #include <limits>
 
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 #include <arm_neon.h>
+#endif
+
 #include <simd/math.h>
 #include <simd/vector.h>
 
@@ -53,25 +56,26 @@ inline simd_float16 simd_fast_exp(simd_float16 x) {
   return (*(simd_float16*)&epart) * x;
 }
 
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 /**
  * The ARM neon equivalent of the fast exp above.
  */
 inline float16x8_t neon_fast_exp(float16x8_t x) {
-  x = vmulq_f16(x, vdupq_n_f16(1.442695)); // multiply with log_2(e)
-  x = vmaxq_f16(x, vdupq_n_f16(-14)); // clamp under with -14
-  x = vminq_f16(x, vdupq_n_f16(14)); // clamp over with 14
+  x = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
+  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f))); // clamp under with -14
+  x = vminq_f16(x, vdupq_n_f16(float16_t(14.f))); // clamp over with 14
 
-  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(0.5)));
+  float16x8_t ipart = vrndmq_f16(vaddq_f16(x, vdupq_n_f16(float16_t(0.5f))));
   float16x8_t fpart = vsubq_f16(x, ipart);
 
-  x = vdupq_n_f16(1.535336188319500e-4f);
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(9.618437357674640e-3f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(5.550332471162809e-2f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(2.402264791363012e-1f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(6.931472028550421e-1f), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.000000000000000f), x, fpart);
+  x = vdupq_n_f16(float16_t(1.535336188319500e-4f));
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.000000000000000f)), x, fpart);
 
   // generate 2**ipart in the floating point representation using integer
   // bitshifting
@@ -107,53 +111,6 @@ inline float16_t neon_reduce_add(float16x8_t x) {
   return vget_lane_f16(y, 0);
 }
 
-template <typename T, typename VT>
-struct AccelerateSimdOps {
-  VT init(T a) {
-    return a;
-  }
-
-  VT load(const T* a) {
-    return *(VT*)a;
-  }
-
-  void store(T* dst, VT x) {
-    *(VT*)dst = x;
-  }
-
-  VT max(VT a, VT b) {
-    return simd_max(a, b);
-  };
-
-  VT exp(VT x) {
-    return simd_fast_exp(x);
-  }
-
-  VT add(VT a, VT b) {
-    return a + b;
-  }
-
-  VT sub(VT a, T b) {
-    return a - b;
-  }
-
-  VT mul(VT a, VT b) {
-    return a * b;
-  }
-
-  VT mul(VT a, T b) {
-    return a * b;
-  }
-
-  T reduce_max(VT x) {
-    return simd_reduce_max(x);
-  }
-
-  T reduce_add(VT x) {
-    return simd_reduce_add(x);
-  }
-};
-
 template <typename T, typename VT>
 struct NeonFp16SimdOps {
   VT init(T a) {
@@ -170,7 +127,7 @@ struct NeonFp16SimdOps {
 
   VT max(VT a, VT b) {
     return vmaxq_f16(a, b);
-  };
+  }
 
   VT exp(VT x) {
     return neon_fast_exp(x);
@@ -201,6 +158,55 @@ struct NeonFp16SimdOps {
   }
 };
 
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+template <typename T, typename VT>
+struct AccelerateSimdOps {
+  VT init(T a) {
+    return a;
+  }
+
+  VT load(const T* a) {
+    return *(VT*)a;
+  }
+
+  void store(T* dst, VT x) {
+    *(VT*)dst = x;
+  }
+
+  VT max(VT a, VT b) {
+    return simd_max(a, b);
+  }
+
+  VT exp(VT x) {
+    return simd_fast_exp(x);
+  }
+
+  VT add(VT a, VT b) {
+    return a + b;
+  }
+
+  VT sub(VT a, T b) {
+    return a - b;
+  }
+
+  VT mul(VT a, VT b) {
+    return a * b;
+  }
+
+  VT mul(VT a, T b) {
+    return a * b;
+  }
+
+  T reduce_max(VT x) {
+    return simd_reduce_max(x);
+  }
+
+  T reduce_add(VT x) {
+    return simd_reduce_add(x);
+  }
+};
+
 template <typename T, typename AccT, typename VT, typename Ops, int N>
 void softmax(const array& in, array& out) {
   Ops ops;
@@ -362,12 +368,16 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
             AccelerateSimdOps<float, simd_float16>,
             16>(in, out);
       } else {
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         softmax<
             float16_t,
             float16_t,
             float16x8_t,
             NeonFp16SimdOps<float16_t, float16x8_t>,
             8>(in, out);
+#else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        eval(inputs, out); // Redirect to common backend for consistency
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
       }
       break;
     case bfloat16:

mlx/backend/accelerate/utils.h

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

mlx/backend/common/CMakeLists.txt

@@ -37,16 +37,20 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/common.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/erf.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/masked_mm.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/reduce_utils.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/sort.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/threefry.cpp
@@ -55,6 +59,7 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/svd.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/cholesky.cpp
   ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp
 )
 

mlx/backend/common/binary.cpp

@@ -196,6 +196,20 @@ void LogAddExp::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void LogicalAnd::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalAnd());
+}
+
+void LogicalOr::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2); // LogicalOr requires two input arrays
+  auto& in1 = inputs[0];
+  auto& in2 = inputs[1];
+  binary(in1, in2, out, detail::LogicalOr());
+}
+
 void Maximum::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
@@ -293,4 +307,25 @@ void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
 }
 
+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 +1,304 @@
+// Copyright © 2024 Apple Inc.
+#include <cassert>
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+void AsStrided::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+
+  auto& in = inputs[0];
+
+  if (!in.flags().row_contiguous) {
+    // Just ensuring that inputs[0] came from the ops which would ensure the
+    // input is row contiguous.
+    throw std::runtime_error(
+        "AsStrided must be used with row contiguous arrays only.");
+  }
+
+  // Compute the flags given the shape and strides
+  bool row_contiguous = true, col_contiguous = true;
+  size_t r = 1, c = 1;
+  for (int i = strides_.size() - 1, j = 0; i >= 0; i--, j++) {
+    row_contiguous &= (r == strides_[i]) || (shape_[i] == 1);
+    col_contiguous &= (c == strides_[j]) || (shape_[j] == 1);
+    r *= shape_[i];
+    c *= shape_[j];
+  }
+  auto flags = in.flags();
+  // TODO: Compute the contiguous flag in a better way cause now we are
+  //       unnecessarily strict.
+  flags.contiguous = row_contiguous || col_contiguous;
+  flags.row_contiguous = row_contiguous;
+  flags.col_contiguous = col_contiguous;
+
+  // There is no easy way to compute the actual data size so we use out.size().
+  // The contiguous flag will almost certainly not be set so no code should
+  // rely on data_size anyway.
+  size_t data_size = out.size();
+
+  return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
+}
+
+void Broadcast::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+  std::vector<size_t> strides(out.ndim(), 0);
+  int diff = out.ndim() - in.ndim();
+  for (int i = in.ndim() - 1; i >= 0; --i) {
+    strides[i + diff] = (in.shape()[i] == 1) ? 0 : in.strides()[i];
+  }
+  auto flags = in.flags();
+  if (out.size() > in.size()) {
+    flags.row_contiguous = flags.col_contiguous = false;
+  }
+  out.copy_shared_buffer(in, strides, flags, in.data_size());
+}
+
+void Copy::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  out.copy_shared_buffer(inputs[0]);
+}
+
+void CustomTransforms::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() > outputs.size());
+  for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
+       i++, j++) {
+    outputs[i].copy_shared_buffer(inputs[j]);
+  }
+}
+
+void Depends::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() > outputs.size());
+  for (int i = 0; i < outputs.size(); i++) {
+    outputs[i].copy_shared_buffer(inputs[i]);
+  }
+}
+
+void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  double numel = 1;
+  for (auto ax : axes_) {
+    numel *= inputs[0].shape(ax);
+  }
+
+  if (inverted_) {
+    numel = 1.0 / numel;
+  }
+
+  switch (out.dtype()) {
+    case bool_:
+      *out.data<bool>() = static_cast<bool>(numel);
+      break;
+    case uint8:
+      *out.data<uint8_t>() = static_cast<uint8_t>(numel);
+      break;
+    case uint16:
+      *out.data<uint16_t>() = static_cast<uint16_t>(numel);
+      break;
+    case uint32:
+      *out.data<uint32_t>() = static_cast<uint32_t>(numel);
+      break;
+    case uint64:
+      *out.data<uint64_t>() = static_cast<uint64_t>(numel);
+      break;
+    case int8:
+      *out.data<int8_t>() = static_cast<int8_t>(numel);
+      break;
+    case int16:
+      *out.data<int16_t>() = static_cast<int16_t>(numel);
+      break;
+    case int32:
+      *out.data<int32_t>() = static_cast<int32_t>(numel);
+      break;
+    case int64:
+      *out.data<int64_t>() = static_cast<int64_t>(numel);
+      break;
+    case float16:
+      *out.data<float16_t>() = static_cast<float16_t>(numel);
+      break;
+    case float32:
+      *out.data<float>() = static_cast<float>(numel);
+      break;
+    case bfloat16:
+      *out.data<bfloat16_t>() = static_cast<bfloat16_t>(numel);
+      break;
+    case complex64:
+      *out.data<complex64_t>() = static_cast<complex64_t>(numel);
+      break;
+  }
+}
+
+std::pair<bool, std::vector<size_t>> Reshape::prepare_reshape(
+    const array& in,
+    const array& out) {
+  // Special case for empty arrays or row contiguous arrays
+  if (in.size() == 0 || in.flags().row_contiguous) {
+    return {false, out.strides()};
+  }
+
+  // Special case for scalars
+  if (in.ndim() == 0) {
+    std::vector<size_t> out_strides(out.ndim(), 0);
+    return {false, out_strides};
+  }
+
+  // Firstly let's collapse all the contiguous dimensions of the input
+  auto [shape, _strides] = collapse_contiguous_dims(in);
+  auto& strides = _strides[0];
+
+  // If shapes fit exactly in the contiguous dims then no copy is necessary so
+  // let's check.
+  std::vector<size_t> out_strides;
+  bool copy_necessary = false;
+  int j = 0;
+  for (int i = 0; i < out.ndim(); i++) {
+    int N = out.shape(i);
+    if (j < shape.size() && shape[j] % N == 0) {
+      shape[j] /= N;
+      out_strides.push_back(shape[j] * strides[j]);
+      j += (shape[j] == 1);
+    } else if (N == 1) {
+      // i > 0 because otherwise j < shape.size() && shape[j] % 1 == 0
+      out_strides.push_back(out_strides.back());
+    } else {
+      copy_necessary = true;
+      break;
+    }
+  }
+
+  return {copy_necessary, out_strides};
+}
+
+void Reshape::shared_buffer_reshape(
+    const array& in,
+    const std::vector<size_t>& out_strides,
+    array& out) {
+  auto flags = in.flags();
+  if (flags.row_contiguous) {
+    // For row contiguous reshapes:
+    // - Shallow copy the buffer
+    // - If reshaping into a vector (all singleton dimensions except one) it
+    //    becomes col contiguous again.
+    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
+    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
+  }
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
+}
+
+void Split::eval(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+
+  auto& in = inputs[0];
+
+  auto compute_new_flags = [](const auto& shape,
+                              const auto& strides,
+                              size_t in_data_size,
+                              auto flags) {
+    size_t data_size = 1;
+    size_t f_stride = 1;
+    size_t b_stride = 1;
+    flags.row_contiguous = true;
+    flags.col_contiguous = true;
+    for (int i = 0, ri = shape.size() - 1; ri >= 0; i++, ri--) {
+      flags.col_contiguous &= strides[i] == f_stride || shape[i] == 1;
+      flags.row_contiguous &= strides[ri] == b_stride || shape[ri] == 1;
+      f_stride *= shape[i];
+      b_stride *= shape[ri];
+      if (strides[i] > 0) {
+        data_size *= shape[i];
+      }
+    }
+
+    if (data_size == 1) {
+      // Broadcasted scalar array is contiguous.
+      flags.contiguous = true;
+    } else if (data_size == in_data_size) {
+      // Means we sliced a broadcasted dimension so leave the "no holes" flag
+      // alone.
+    } else {
+      // We sliced something. So either we are row or col contiguous or we
+      // punched a hole.
+      flags.contiguous &= flags.row_contiguous || flags.col_contiguous;
+    }
+
+    return std::pair<decltype(flags), size_t>{flags, data_size};
+  };
+
+  std::vector<int> indices(1, 0);
+  indices.insert(indices.end(), indices_.begin(), indices_.end());
+  for (int i = 0; i < indices.size(); i++) {
+    size_t offset = indices[i] * in.strides()[axis_];
+    auto [new_flags, data_size] = compute_new_flags(
+        outputs[i].shape(), in.strides(), in.data_size(), in.flags());
+    outputs[i].copy_shared_buffer(
+        in, in.strides(), new_flags, data_size, offset);
+  }
+}
+
+std::tuple<int64_t, std::vector<int64_t>> SliceUpdate::prepare_slice(
+    const array& in) {
+  int64_t data_offset = 0;
+  std::vector<int64_t> inp_strides(in.ndim(), 0);
+  for (int i = 0; i < in.ndim(); ++i) {
+    data_offset += start_indices_[i] * in.strides()[i];
+    inp_strides[i] = in.strides()[i] * strides_[i];
+  }
+
+  return std::make_tuple(data_offset, inp_strides);
+}
+
+void StopGradient::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  out.copy_shared_buffer(inputs[0]);
+}
+
+void Transpose::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  std::vector<size_t> out_strides(out.ndim());
+  auto& in = inputs[0];
+  for (int ax = 0; ax < axes_.size(); ++ax) {
+    out_strides[ax] = in.strides()[axes_[ax]];
+  }
+
+  // Conditions for {row/col}_contiguous
+  // - array must be contiguous (no gaps)
+  // - underlying buffer size should have the same size as the array
+  // - cumulative product of shapes is equal to the strides (we can ignore axes
+  //   with size == 1)
+  //   - in the forward direction (column contiguous)
+  //   - in the reverse direction (row contiguous)
+  // - vectors are both row and col contiguous (hence if both row/col are
+  //   true, they stay true)
+  auto flags = in.flags();
+  if (flags.contiguous && in.data_size() == in.size()) {
+    size_t f_stride = 1;
+    size_t b_stride = 1;
+    flags.col_contiguous = true;
+    flags.row_contiguous = true;
+    for (int i = 0, ri = out.ndim() - 1; i < out.ndim(); ++i, --ri) {
+      flags.col_contiguous &= (out_strides[i] == f_stride || out.shape(i) == 1);
+      f_stride *= out.shape(i);
+      flags.row_contiguous &=
+          (out_strides[ri] == b_stride || out.shape(ri) == 1);
+      b_stride *= out.shape(ri);
+    }
+  }
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
+}
+
+} // namespace mlx::core

mlx/backend/common/compiled.cpp

@@ -205,8 +205,8 @@ void compiled_allocate_outputs(
       // - Donatable
       // - Correct size
       // - Not a constant
-      if (in.flags().row_contiguous && in.nbytes() == outputs[o].nbytes() &&
-          in.is_donatable() &&
+      if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
+          in.itemsize() == outputs[o].itemsize() && in.is_donatable() &&
           constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
         if (move_buffers) {
           outputs[o].move_shared_buffer(

mlx/backend/common/conv.cpp

@@ -111,13 +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];
@@ -141,33 +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];
@@ -219,41 +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;
@@ -310,6 +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,
@@ -358,6 +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
 ///////////////////////////////////////////////////////////////////////////////
@@ -582,6 +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
 ///////////////////////////////////////////////////////////////////////////////
@@ -617,6 +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) {
@@ -625,8 +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

@@ -4,6 +4,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/copy.h"
+#include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
@@ -142,29 +143,31 @@ void copy_general(
     const std::vector<int>& data_shape,
     const std::vector<stride_t>& i_strides,
     int64_t i_offset) {
-  switch (src.ndim()) {
+  auto [new_shape, new_strides] = collapse_contiguous_dims(
+      data_shape, std::vector<std::vector<stride_t>>{i_strides});
+  switch (new_shape.size()) {
     case 1:
       copy_general_dim1<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
     case 2:
       copy_general_dim2<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
     case 3:
       copy_general_dim3<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
     case 4:
       copy_general_dim4<SrcT, DstT, stride_t>(
-          src, dst, data_shape, i_strides, i_offset);
+          src, dst, new_shape, new_strides[0], i_offset);
       return;
   }
 
   auto src_ptr = src.data<SrcT>() + i_offset;
   auto dst_ptr = dst.data<DstT>();
   for (size_t i = 0; i < dst.size(); ++i) {
-    stride_t src_elem = elem_to_loc(i, data_shape, i_strides);
+    stride_t src_elem = elem_to_loc(i, new_shape, new_strides[0]);
     dst_ptr[i] = static_cast<DstT>(src_ptr[src_elem]);
   }
 }
@@ -195,10 +198,10 @@ inline void copy_general_general_dims(
     const std::vector<int>& data_shape,
     const std::vector<stride_t>& i_strides,
     const std::vector<stride_t>& o_strides,
-    stride_t i_offset,
-    stride_t o_offset) {
+    int64_t i_offset,
+    int64_t o_offset) {
   if constexpr (D > 1) {
-    int axis = src.ndim() - D;
+    int axis = data_shape.size() - D;
     auto stride_src = i_strides[axis];
     auto stride_dst = o_strides[axis];
     auto N = data_shape[axis];
@@ -209,7 +212,7 @@ inline void copy_general_general_dims(
       o_offset += stride_dst;
     }
   } else {
-    int axis = src.ndim() - 1;
+    int axis = data_shape.size() - 1;
     auto stride_src = i_strides[axis];
     auto stride_dst = o_strides[axis];
     auto N = data_shape[axis];
@@ -230,38 +233,76 @@ void copy_general_general(
     const std::vector<int>& data_shape,
     const std::vector<stride_t>& i_strides,
     const std::vector<stride_t>& o_strides,
-    stride_t i_offset,
-    stride_t o_offset) {
-  switch (src.ndim()) {
+    int64_t i_offset,
+    int64_t o_offset) {
+  auto [new_shape, new_strides] = collapse_contiguous_dims(
+      data_shape, std::vector<std::vector<stride_t>>{i_strides, o_strides});
+  switch (new_shape.size()) {
     case 1:
       copy_general_general_dims<SrcT, DstT, stride_t, 1>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 2:
       copy_general_general_dims<SrcT, DstT, stride_t, 2>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 3:
       copy_general_general_dims<SrcT, DstT, stride_t, 3>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 4:
       copy_general_general_dims<SrcT, DstT, stride_t, 4>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
     case 5:
       copy_general_general_dims<SrcT, DstT, stride_t, 5>(
-          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+          src,
+          dst,
+          new_shape,
+          new_strides[0],
+          new_strides[1],
+          i_offset,
+          o_offset);
       return;
   }
 
   int size = std::accumulate(
-      data_shape.begin() - 5, data_shape.end(), 1, std::multiplies<int>());
+      new_shape.end() - 5, new_shape.end(), 1, std::multiplies<int>());
   for (int i = 0; i < src.size(); i += size) {
-    stride_t src_offset = i_offset + elem_to_loc(i, data_shape, i_strides);
-    stride_t dst_offset = o_offset + elem_to_loc(i, dst.shape(), o_strides);
+    stride_t src_offset = i_offset + elem_to_loc(i, new_shape, new_strides[0]);
+    stride_t dst_offset = o_offset + elem_to_loc(i, new_shape, new_strides[1]);
     copy_general_general_dims<SrcT, DstT, stride_t, 5>(
-        src, dst, data_shape, i_strides, o_strides, src_offset, dst_offset);
+        src,
+        dst,
+        new_shape,
+        new_strides[0],
+        new_strides[1],
+        src_offset,
+        dst_offset);
   }
 }
 
@@ -444,8 +485,17 @@ void copy_inplace(
   }
 }
 
-template <>
-void copy_inplace<int64_t>(
+template void copy_inplace<size_t>(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<size_t>& i_strides,
+    const std::vector<size_t>& o_strides,
+    int64_t i_offset,
+    int64_t o_offset,
+    CopyType ctype);
+
+template void copy_inplace<int64_t>(
     const array& src,
     array& dst,
     const std::vector<int>& data_shape,
@@ -453,24 +503,6 @@ void copy_inplace<int64_t>(
     const std::vector<int64_t>& o_strides,
     int64_t i_offset,
     int64_t o_offset,
-    CopyType ctype) {
-  switch (ctype) {
-    case CopyType::General:
-    case CopyType::GeneralGeneral:
-      return copy_inplace_dispatch(
-          src,
-          dst,
-          ctype,
-          data_shape,
-          i_strides,
-          o_strides,
-          i_offset,
-          o_offset);
-
-    case CopyType::Scalar:
-    case CopyType::Vector:
-      return copy_inplace_dispatch(src, dst, ctype);
-  }
-}
+    CopyType ctype);
 
 } // namespace mlx::core

mlx/backend/common/default_primitives.cpp

@@ -5,7 +5,6 @@
 #else
 #include <cblas.h>
 #endif
-
 #include <cstring>
 
 #include "mlx/array.h"
@@ -34,6 +33,7 @@ DEFAULT(ArcCosh)
 DEFAULT(ArcSin)
 DEFAULT(ArcSinh)
 DEFAULT(ArcTan)
+DEFAULT(ArcTan2)
 DEFAULT(ArcTanh)
 DEFAULT(ArgPartition)
 DEFAULT(ArgReduce)
@@ -42,15 +42,17 @@ DEFAULT(AsType)
 DEFAULT(AsStrided)
 DEFAULT(Broadcast)
 DEFAULT(BlockMaskedMM)
-DEFAULT(BlockSparseMM)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT_MULTI(DivMod)
 DEFAULT(Ceil)
 DEFAULT(Concatenate)
+DEFAULT(Conjugate)
 DEFAULT(Convolution)
 DEFAULT(Copy)
 DEFAULT(Cos)
 DEFAULT(Cosh)
-DEFAULT_MULTI(CustomVJP)
+DEFAULT_MULTI(CustomTransforms)
 DEFAULT_MULTI(Depends)
 DEFAULT(Divide)
 DEFAULT(NumberOfElements)
@@ -66,6 +68,7 @@ DEFAULT(Full)
 DEFAULT(Gather)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
 DEFAULT(Less)
 DEFAULT(LessEqual)
 DEFAULT(Load)
@@ -110,6 +113,7 @@ DEFAULT(Tan)
 DEFAULT(Tanh)
 DEFAULT(Transpose)
 DEFAULT(Inverse)
+DEFAULT(Cholesky)
 
 namespace {
 

mlx/backend/common/hadamard.cpp

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

mlx/backend/common/hadamard.h

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

mlx/backend/common/inverse.cpp

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

mlx/backend/common/make_compiled_preamble.sh

@@ -28,6 +28,7 @@ const char* get_kernel_preamble() {
 return R"preamble(
 $INCLUDES
 $CONTENT
+using namespace mlx::core;
 using namespace mlx::core::detail;
 )preamble";
 }

mlx/backend/common/masked_mm.cpp

@@ -17,24 +17,25 @@ namespace mlx::core {
 
 namespace {
 
-template <typename T>
+template <typename T, typename mask_t>
 inline void mask_matrix(
     T* data,
-    const bool* mask,
+    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 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++) {
-      bool do_mask = mask[i * X_mask_str + j * Y_mask_str];
-      if (!do_mask) {
+      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;
@@ -43,7 +44,11 @@ inline void mask_matrix(
         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++) {
-            data_block[ii * X_data_str + jj * Y_data_str] = T(0.);
+            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;
+            }
           }
         }
       }
@@ -62,36 +67,39 @@ void BlockMaskedMM::eval(const std::vector<array>& inputs, array& out) {
 
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
-  auto& out_mask = inputs[2];
 
-  auto check_transpose = [](const array& arr, bool do_copy) {
-    auto stx = arr.strides()[arr.ndim() - 2];
-    auto sty = arr.strides()[arr.ndim() - 1];
-    if (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 (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);
-    }
-  };
+  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;
-  auto [a_transposed, lda, a] = check_transpose(a_pre, has_op_mask);
-  auto [b_transposed, ldb, b] = check_transpose(b_pre, has_op_mask);
+  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);
@@ -114,27 +122,42 @@ void BlockMaskedMM::eval(const std::vector<array>& inputs, array& out) {
                        int Y,
                        size_t X_data_str,
                        size_t Y_data_str) {
-    const bool* mask_ptr = mask.data<bool>() +
-        elem_to_loc(mask.shape(-1) * mask.shape(-2) * batch_idx,
-                    mask.shape(),
-                    mask.strides());
+    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];
 
-    return mask_matrix(
-        data,
-        mask_ptr,
-        block_size,
-        X,
-        Y,
-        X_data_str,
-        Y_data_str,
-        X_mask_str,
-        Y_mask_str);
+    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 < (a.size() / (M * K)); ++i) {
+  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());
@@ -144,7 +167,7 @@ void BlockMaskedMM::eval(const std::vector<array>& inputs, array& out) {
 
     // Zero out blocks in a and b if needed
     if (has_op_mask) {
-      auto& a_mask = inputs[3];
+      auto& a_mask = inputs[inputs.size() - 2];
       mask_array(
           a_mask,
           ai,
@@ -155,7 +178,7 @@ void BlockMaskedMM::eval(const std::vector<array>& inputs, array& out) {
           a_transposed ? 1 : lda,
           a_transposed ? lda : 1);
 
-      auto& b_mask = inputs[4];
+      auto& b_mask = inputs[inputs.size() - 1];
       mask_array(
           b_mask,
           bi,
@@ -186,14 +209,16 @@ void BlockMaskedMM::eval(const std::vector<array>& inputs, array& out) {
     );
 
     // Zero out blocks in out
-    mask_array(out_mask, ci, block_size_, i, M, N, N, 1);
+    if (has_out_mask) {
+      mask_array(inputs[2], ci, block_size_, i, M, N, N, 1);
+    }
   }
 }
 
-void BlockSparseMM::eval(const std::vector<array>& inputs, array& out) {
+void GatherMM::eval(const std::vector<array>& inputs, array& out) {
   if (out.dtype() != float32) {
     throw std::runtime_error(
-        "[BlockSparseMM::eval] Currently only supports float32.");
+        "[GatherMM::eval] Currently only supports float32.");
   }
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
@@ -277,4 +302,4 @@ void BlockSparseMM::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-} // namespace mlx::core
+} // namespace mlx::core

mlx/backend/common/ops.h

@@ -108,133 +108,146 @@ struct Abs {
   template <typename T>
   T operator()(T x) {
     return std::abs(x);
-  };
+  }
   uint8_t operator()(uint8_t x) {
     return x;
-  };
+  }
   uint16_t operator()(uint16_t x) {
     return x;
-  };
+  }
   uint32_t operator()(uint32_t x) {
     return x;
-  };
+  }
   uint64_t operator()(uint64_t x) {
     return x;
-  };
+  }
   bool operator()(bool x) {
     return x;
-  };
+  }
 };
 
 struct ArcCos {
   template <typename T>
   T operator()(T x) {
     return std::acos(x);
-  };
+  }
 };
 
 struct ArcCosh {
   template <typename T>
   T operator()(T x) {
     return std::acosh(x);
-  };
+  }
 };
 
 struct ArcSin {
   template <typename T>
   T operator()(T x) {
     return std::asin(x);
-  };
+  }
 };
 
 struct ArcSinh {
   template <typename T>
   T operator()(T x) {
     return std::asinh(x);
-  };
+  }
 };
 
 struct ArcTan {
   template <typename T>
   T operator()(T x) {
     return std::atan(x);
-  };
+  }
+};
+
+struct ArcTan2 {
+  template <typename T>
+  T operator()(T y, T x) {
+    return std::atan2(y, x);
+  }
 };
 
 struct ArcTanh {
   template <typename T>
   T operator()(T x) {
     return std::atanh(x);
-  };
+  }
 };
 
 struct Ceil {
   template <typename T>
   T operator()(T x) {
     return std::ceil(x);
-  };
+  }
   int8_t operator()(int8_t x) {
     return x;
-  };
+  }
   int16_t operator()(int16_t x) {
     return x;
-  };
+  }
   int32_t operator()(int32_t x) {
     return x;
-  };
+  }
   int64_t operator()(int64_t x) {
     return x;
-  };
+  }
   uint8_t operator()(uint8_t x) {
     return x;
-  };
+  }
   uint16_t operator()(uint16_t x) {
     return x;
-  };
+  }
   uint32_t operator()(uint32_t x) {
     return x;
-  };
+  }
   uint64_t operator()(uint64_t x) {
     return x;
-  };
+  }
   bool operator()(bool x) {
     return x;
-  };
+  }
+};
+
+struct Conjugate {
+  complex64_t operator()(complex64_t x) {
+    return std::conj(x);
+  }
 };
 
 struct Cos {
   template <typename T>
   T operator()(T x) {
     return std::cos(x);
-  };
+  }
 };
 
 struct Cosh {
   template <typename T>
   T operator()(T x) {
     return std::cosh(x);
-  };
+  }
 };
 
 struct Erf {
   template <typename T>
   T operator()(T x) {
     return static_cast<T>(fast_erf(static_cast<float>(x)));
-  };
+  }
 };
 
 struct ErfInv {
   template <typename T>
   T operator()(T x) {
     return static_cast<T>(fast_erfinv(static_cast<float>(x)));
-  };
+  }
 };
 
 struct Exp {
   template <typename T>
   T operator()(T x) {
     return fast_exp(x);
-  };
+  }
 
   complex64_t operator()(complex64_t x) {
     return std::exp(x);
@@ -245,83 +258,83 @@ struct Expm1 {
   template <typename T>
   T operator()(T x) {
     return expm1(x);
-  };
+  }
 };
 
 struct Floor {
   template <typename T>
   T operator()(T x) {
     return std::floor(x);
-  };
+  }
   int8_t operator()(int8_t x) {
     return x;
-  };
+  }
   int16_t operator()(int16_t x) {
     return x;
-  };
+  }
   int32_t operator()(int32_t x) {
     return x;
-  };
+  }
   int64_t operator()(int64_t x) {
     return x;
-  };
+  }
   uint8_t operator()(uint8_t x) {
     return x;
-  };
+  }
   uint16_t operator()(uint16_t x) {
     return x;
-  };
+  }
   uint32_t operator()(uint32_t x) {
     return x;
-  };
+  }
   uint64_t operator()(uint64_t x) {
     return x;
-  };
+  }
   bool operator()(bool x) {
     return x;
-  };
+  }
 };
 
 struct Log {
   template <typename T>
   T operator()(T x) {
     return std::log(x);
-  };
+  }
 };
 
 struct Log2 {
   template <typename T>
   T operator()(T x) {
     return std::log2(x);
-  };
+  }
 };
 
 struct Log10 {
   template <typename T>
   T operator()(T x) {
     return std::log10(x);
-  };
+  }
 };
 
 struct Log1p {
   template <typename T>
   T operator()(T x) {
     return log1p(x);
-  };
+  }
 };
 
 struct LogicalNot {
   template <typename T>
   T operator()(T x) {
     return !x;
-  };
+  }
 };
 
 struct Negative {
   template <typename T>
   T operator()(T x) {
     return -x;
-  };
+  }
 };
 
 struct Round {
@@ -366,49 +379,49 @@ struct Sin {
   template <typename T>
   T operator()(T x) {
     return std::sin(x);
-  };
+  }
 };
 
 struct Sinh {
   template <typename T>
   T operator()(T x) {
     return std::sinh(x);
-  };
+  }
 };
 
 struct Square {
   template <typename T>
   T operator()(T x) {
     return x * x;
-  };
+  }
 };
 
 struct Sqrt {
   template <typename T>
   T operator()(T x) {
     return std::sqrt(x);
-  };
+  }
 };
 
 struct Rsqrt {
   template <typename T>
   T operator()(T x) {
     return static_cast<decltype(x)>(1.0) / std::sqrt(x);
-  };
+  }
 };
 
 struct Tan {
   template <typename T>
   T operator()(T x) {
     return std::tan(x);
-  };
+  }
 };
 
 struct Tanh {
   template <typename T>
   T operator()(T x) {
     return std::tanh(x);
-  };
+  }
 };
 
 struct Add {
@@ -541,7 +554,7 @@ struct LogAddExp {
         ? maxval
         : static_cast<decltype(x)>(
               maxval + std::log1p(fast_exp(minval - maxval)));
-  };
+  }
 };
 
 struct Multiply {
@@ -589,14 +602,14 @@ struct LogicalAnd {
   template <typename T>
   T operator()(T x, T y) {
     return x && y;
-  };
+  }
 };
 
 struct LogicalOr {
   template <typename T>
   T operator()(T x, T y) {
     return x || y;
-  };
+  }
 };
 
 struct Select {
@@ -610,35 +623,35 @@ struct BitwiseAnd {
   template <typename T>
   T operator()(T x, T y) {
     return x & y;
-  };
+  }
 };
 
 struct BitwiseOr {
   template <typename T>
   T operator()(T x, T y) {
     return x | y;
-  };
+  }
 };
 
 struct BitwiseXor {
   template <typename T>
   T operator()(T x, T y) {
     return x ^ y;
-  };
+  }
 };
 
 struct LeftShift {
   template <typename T>
   T operator()(T x, T y) {
     return x << y;
-  };
+  }
 };
 
 struct RightShift {
   template <typename T>
   T operator()(T x, T y) {
     return x >> y;
-  };
+  }
 };
 
 } // namespace mlx::core::detail

mlx/backend/common/primitives.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <algorithm>
 #include <cassert>
@@ -8,9 +8,9 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/arange.h"
-#include "mlx/backend/common/binary.h"
 #include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/ops.h"
+#include "mlx/backend/common/slicing.h"
 #include "mlx/backend/common/threefry.h"
 #include "mlx/backend/common/unary.h"
 #include "mlx/backend/common/utils.h"
@@ -113,61 +113,6 @@ void AsType::eval(const std::vector<array>& inputs, array& out) {
   copy(in, out, ctype);
 }
 
-void AsStrided::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-
-  auto& in = inputs[0];
-
-  if (!in.flags().row_contiguous) {
-    // Just ensuring that inputs[0] came from the ops which would ensure the
-    // input is row contiguous.
-    throw std::runtime_error(
-        "AsStrided must be used with row contiguous arrays only.");
-  }
-
-  // Compute the flags given the shape and strides
-  bool row_contiguous = true, col_contiguous = true;
-  size_t r = 1, c = 1;
-  for (int i = strides_.size() - 1, j = 0; i >= 0; i--, j++) {
-    row_contiguous &= (r == strides_[i]) || (shape_[i] == 1);
-    col_contiguous &= (c == strides_[j]) || (shape_[j] == 1);
-    r *= shape_[i];
-    c *= shape_[j];
-  }
-  auto flags = in.flags();
-  // TODO: Compute the contiguous flag in a better way cause now we are
-  //       unnecessarily strict.
-  flags.contiguous = row_contiguous || col_contiguous;
-  flags.row_contiguous = row_contiguous;
-  flags.col_contiguous = col_contiguous;
-
-  // There is no easy way to compute the actual data size so we use out.size().
-  // The contiguous flag will almost certainly not be set so no code should
-  // rely on data_size anyway.
-  size_t data_size = out.size();
-
-  return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
-}
-
-void Broadcast::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-  std::vector<size_t> strides(out.ndim(), 0);
-  int diff = out.ndim() - in.ndim();
-  for (int i = in.ndim() - 1; i >= 0; --i) {
-    strides[i + diff] = (in.shape()[i] == 1) ? 0 : in.strides()[i];
-  }
-  auto flags = in.flags();
-  if (out.size() > in.size()) {
-    flags.row_contiguous = flags.col_contiguous = false;
-  }
-  out.copy_shared_buffer(in, strides, flags, in.data_size());
-}
-
 void Ceil::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
@@ -203,9 +148,15 @@ void Concatenate::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void Copy::eval(const std::vector<array>& inputs, array& out) {
+void Conjugate::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  out.copy_shared_buffer(inputs[0]);
+  const auto& in = inputs[0];
+  if (out.dtype() == complex64) {
+    unary_fp(in, out, detail::Conjugate());
+  } else {
+    throw std::invalid_argument(
+        "[conjugate] conjugate must be called on complex input.");
+  }
 }
 
 void Cos::eval(const std::vector<array>& inputs, array& out) {
@@ -232,81 +183,6 @@ void Cosh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void CustomVJP::eval(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() > outputs.size());
-  for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
-       i++, j++) {
-    outputs[i].copy_shared_buffer(inputs[j]);
-  }
-}
-
-void Depends::eval(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() > outputs.size());
-  for (int i = 0; i < outputs.size(); i++) {
-    outputs[i].copy_shared_buffer(inputs[i]);
-  }
-}
-
-void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  double numel = 1;
-  for (auto ax : axes_) {
-    numel *= inputs[0].shape(ax);
-  }
-
-  if (inverted_) {
-    numel = 1.0 / numel;
-  }
-
-  switch (out.dtype()) {
-    case bool_:
-      *out.data<bool>() = static_cast<bool>(numel);
-      break;
-    case uint8:
-      *out.data<uint8_t>() = static_cast<uint8_t>(numel);
-      break;
-    case uint16:
-      *out.data<uint16_t>() = static_cast<uint16_t>(numel);
-      break;
-    case uint32:
-      *out.data<uint32_t>() = static_cast<uint32_t>(numel);
-      break;
-    case uint64:
-      *out.data<uint64_t>() = static_cast<uint64_t>(numel);
-      break;
-    case int8:
-      *out.data<int8_t>() = static_cast<int8_t>(numel);
-      break;
-    case int16:
-      *out.data<int16_t>() = static_cast<int16_t>(numel);
-      break;
-    case int32:
-      *out.data<int32_t>() = static_cast<int32_t>(numel);
-      break;
-    case int64:
-      *out.data<int64_t>() = static_cast<int64_t>(numel);
-      break;
-    case float16:
-      *out.data<float16_t>() = static_cast<float16_t>(numel);
-      break;
-    case float32:
-      *out.data<float>() = static_cast<float>(numel);
-      break;
-    case bfloat16:
-      *out.data<bfloat16_t>() = static_cast<bfloat16_t>(numel);
-      break;
-    case complex64:
-      *out.data<complex64_t>() = static_cast<complex64_t>(numel);
-      break;
-  }
-}
-
 void Erf::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -437,20 +313,6 @@ void LogicalNot::eval(const std::vector<array>& inputs, array& out) {
   unary(in, out, detail::LogicalNot());
 }
 
-void LogicalAnd::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
-  auto& in1 = inputs[0];
-  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalAnd());
-}
-
-void LogicalOr::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 2); // LogicalOr requires two input arrays
-  auto& in1 = inputs[0];
-  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalOr());
-}
-
 void Negative::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
@@ -536,63 +398,6 @@ void RandomBits::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-std::pair<bool, std::vector<size_t>> Reshape::prepare_reshape(
-    const array& in,
-    const array& out) {
-  // Special case for empty arrays or row contiguous arrays
-  if (in.size() == 0 || in.flags().row_contiguous) {
-    return {false, out.strides()};
-  }
-
-  // Special case for scalars
-  if (in.ndim() == 0) {
-    std::vector<size_t> out_strides(out.ndim(), 0);
-    return {false, out_strides};
-  }
-
-  // Firstly let's collapse all the contiguous dimensions of the input
-  auto [shape, _strides] = collapse_contiguous_dims(in);
-  auto& strides = _strides[0];
-
-  // If shapes fit exactly in the contiguous dims then no copy is necessary so
-  // let's check.
-  std::vector<size_t> out_strides;
-  bool copy_necessary = false;
-  int j = 0;
-  for (int i = 0; i < out.ndim(); i++) {
-    int N = out.shape(i);
-    if (j < shape.size() && shape[j] % N == 0) {
-      shape[j] /= N;
-      out_strides.push_back(shape[j] * strides[j]);
-      j += (shape[j] == 1);
-    } else if (N == 1) {
-      // i > 0 because otherwise j < shape.size() && shape[j] % 1 == 0
-      out_strides.push_back(out_strides.back());
-    } else {
-      copy_necessary = true;
-      break;
-    }
-  }
-
-  return {copy_necessary, out_strides};
-}
-
-void Reshape::shared_buffer_reshape(
-    const array& in,
-    const std::vector<size_t>& out_strides,
-    array& out) {
-  auto flags = in.flags();
-  if (flags.row_contiguous) {
-    // For row contiguous reshapes:
-    // - Shallow copy the buffer
-    // - If reshaping into a vector (all singleton dimensions except one) it
-    //    becomes col contiguous again.
-    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
-    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
-  }
-  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
-}
-
 void Reshape::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -600,7 +405,17 @@ void Reshape::eval(const std::vector<array>& inputs, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
 
   if (copy_necessary) {
-    copy(in, out, in.data_size() == 1 ? CopyType::Scalar : CopyType::General);
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    auto out_strides = make_contiguous_strides<size_t>(in.shape());
+    copy_inplace<size_t>(
+        in,
+        out,
+        in.shape(),
+        in.strides(),
+        out_strides,
+        0,
+        0,
+        CopyType::General);
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -663,49 +478,6 @@ void Sinh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-std::tuple<bool, int64_t, std::vector<int64_t>> Slice::prepare_slice(
-    const array& in) {
-  int64_t data_offset = 0;
-  bool copy_needed = false;
-  std::vector<int64_t> inp_strides(in.ndim(), 0);
-  for (int i = 0; i < in.ndim(); ++i) {
-    data_offset += start_indices_[i] * in.strides()[i];
-    inp_strides[i] = in.strides()[i] * strides_[i];
-
-    copy_needed |= strides_[i] < 0;
-  }
-
-  return std::make_tuple(copy_needed, data_offset, inp_strides);
-}
-
-void Slice::shared_buffer_slice(
-    const array& in,
-    const std::vector<size_t>& out_strides,
-    size_t data_offset,
-    array& out) {
-  // Compute row/col contiguity
-  auto [data_size, is_row_contiguous, is_col_contiguous] =
-      check_contiguity(out.shape(), out_strides);
-
-  auto flags = in.flags();
-  flags.row_contiguous = is_row_contiguous;
-  flags.col_contiguous = is_col_contiguous;
-
-  if (data_size == 1) {
-    // Broadcasted scalar array is contiguous.
-    flags.contiguous = true;
-  } else if (data_size == in.data_size()) {
-    // Means we sliced a broadcasted dimension so leave the "no holes" flag
-    // alone.
-  } else {
-    // We sliced something. So either we are row or col contiguous or we
-    // punched a hole.
-    flags.contiguous &= flags.row_contiguous || flags.col_contiguous;
-  }
-
-  out.copy_shared_buffer(in, out_strides, flags, data_size, data_offset);
-}
-
 void Slice::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   if (out.size() == 0) {
@@ -716,7 +488,8 @@ void Slice::eval(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Calculate out strides, initial offset and if copy needs to be made
-  auto [copy_needed, data_offset, inp_strides] = prepare_slice(in);
+  auto [copy_needed, data_offset, inp_strides] =
+      prepare_slice(in, start_indices_, strides_);
 
   // Do copy if needed
   if (copy_needed) {
@@ -737,18 +510,6 @@ 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) {
@@ -786,58 +547,6 @@ 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];
@@ -854,11 +563,6 @@ 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];
@@ -883,38 +587,36 @@ void Tanh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void Transpose::eval(const std::vector<array>& inputs, array& out) {
+void View::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  std::vector<size_t> out_strides(out.ndim());
   auto& in = inputs[0];
-  for (int ax = 0; ax < axes_.size(); ++ax) {
-    out_strides[ax] = in.strides()[axes_[ax]];
-  }
-
-  // Conditions for {row/col}_contiguous
-  // - array must be contiguous (no gaps)
-  // - underlying buffer size should have the same size as the array
-  // - cumulative product of shapes is equal to the strides (we can ignore axes
-  //   with size == 1)
-  //   - in the forward direction (column contiguous)
-  //   - in the reverse direction (row contiguous)
-  // - vectors are both row and col contiguous (hence if both row/col are
-  //   true, they stay true)
-  auto flags = in.flags();
-  if (flags.contiguous && in.data_size() == in.size()) {
-    size_t f_stride = 1;
-    size_t b_stride = 1;
-    flags.col_contiguous = true;
-    flags.row_contiguous = true;
-    for (int i = 0, ri = out.ndim() - 1; i < out.ndim(); ++i, --ri) {
-      flags.col_contiguous &= (out_strides[i] == f_stride || out.shape(i) == 1);
-      f_stride *= out.shape(i);
-      flags.row_contiguous &=
-          (out_strides[ri] == b_stride || out.shape(ri) == 1);
-      b_stride *= out.shape(ri);
+  auto ibytes = size_of(in.dtype());
+  auto obytes = size_of(out.dtype());
+  // Conditions for buffer copying (disjunction):
+  // - type size is the same
+  // - type size is smaller and the last axis is contiguous
+  // - the entire array is row contiguous
+  if (ibytes == obytes || obytes < ibytes && in.strides().back() == 1 ||
+      in.flags().row_contiguous) {
+    auto strides = in.strides();
+    for (int i = 0; i < strides.size() - 1; ++i) {
+      strides[i] *= ibytes;
+      strides[i] /= obytes;
     }
+    out.copy_shared_buffer(
+        in, strides, in.flags(), in.data_size() * obytes / ibytes);
+  } else {
+    auto tmp = array(in.shape(), in.dtype(), nullptr, {});
+    tmp.set_data(allocator::malloc_or_wait(tmp.nbytes()));
+    copy_inplace(in, tmp, CopyType::General);
+
+    auto flags = out.flags();
+    flags.contiguous = true;
+    flags.row_contiguous = true;
+    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
+    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
+    out.move_shared_buffer(tmp, out.strides(), flags, out.size());
   }
-  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
 }
 
 } // namespace mlx::core

mlx/backend/common/quantized.cpp

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

mlx/backend/common/reduce.cpp

@@ -104,48 +104,14 @@ void reduce_dispatch_out(
     }
     case Reduce::Sum: {
       auto op = [](auto y, auto x) { (*y) = (*y) + x; };
-      switch (out.dtype()) {
-        case bool_:
-          reduction_op<InT, bool>(in, out, axes, false, op);
-          break;
-        case uint8:
-          reduction_op<InT, uint8_t>(in, out, axes, 0, op);
-          break;
-        case uint16:
-          reduction_op<InT, uint16_t>(in, out, axes, 0, op);
-          break;
-        case uint32:
-          reduction_op<InT, uint32_t>(in, out, axes, 0, op);
-          break;
-        case uint64:
-          reduction_op<InT, uint64_t>(in, out, axes, 0, op);
-          break;
-        case int8:
-          reduction_op<InT, int8_t>(in, out, axes, 0, op);
-          break;
-        case int16:
-          reduction_op<InT, int16_t>(in, out, axes, 0, op);
-          break;
-        case int32:
-          reduction_op<InT, int32_t>(in, out, axes, 0, op);
-          break;
-        case int64:
-          reduction_op<InT, int64_t>(in, out, axes, 0, op);
-          break;
-        case float16:
-          reduction_op<InT, float16_t>(in, out, axes, 0.0f, op);
-          break;
-        case float32:
-          reduction_op<InT, float>(in, out, axes, 0.0f, op);
-          break;
-        case bfloat16:
-          reduction_op<InT, bfloat16_t>(in, out, axes, 0.0f, op);
-          break;
-        case complex64:
-          reduction_op<InT, complex64_t>(in, out, axes, complex64_t{0.0f}, op);
-          break;
+      if (out.dtype() == int32) {
+        // special case since the input type can be bool
+        reduction_op<InT, int32_t>(in, out, axes, 0, op);
+      } else {
+        reduction_op<InT, InT>(in, out, axes, 0, op);
       }
-    } break;
+      break;
+    }
     case Reduce::Prod: {
       auto op = [](auto y, auto x) { (*y) *= x; };
       reduction_op<InT, InT>(in, out, axes, 1, op);
@@ -168,6 +134,29 @@ void reduce_dispatch_out(
 
 } // namespace
 
+void nd_loop(
+    std::function<void(int)> callback,
+    const std::vector<int>& shape,
+    const std::vector<size_t>& strides) {
+  std::function<void(int, int)> loop_inner;
+  loop_inner = [&](int dim, int offset) {
+    if (dim < shape.size() - 1) {
+      int size = shape[dim];
+      size_t stride = strides[dim];
+      for (int i = 0; i < size; i++) {
+        loop_inner(dim + 1, offset + i * stride);
+      }
+    } else {
+      int size = shape[dim];
+      size_t stride = strides[dim];
+      for (int i = 0; i < size; i++) {
+        callback(offset + i * stride);
+      }
+    }
+  };
+  loop_inner(0, 0);
+}
+
 void Reduce::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];

mlx/backend/common/reduce.h

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

mlx/backend/common/reduce_utils.cpp

@@ -0,0 +1,118 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/common/reduce.h"
+
+namespace mlx::core {
+
+std::pair<std::vector<int>, std::vector<size_t>> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes) {
+  std::vector<int> shape = x.shape();
+  std::vector<size_t> strides = x.strides();
+
+  for (int i = axes.size() - 1; i >= 0; i--) {
+    int a = axes[i];
+    shape.erase(shape.begin() + a);
+    strides.erase(strides.begin() + a);
+  }
+
+  return std::make_pair(shape, strides);
+}
+
+ReductionPlan get_reduction_plan(const array& x, const std::vector<int> axes) {
+  // The data is all there and we are reducing over everything
+  if (x.size() == x.data_size() && axes.size() == x.ndim() &&
+      x.flags().contiguous) {
+    return ContiguousAllReduce;
+  }
+
+  // Row contiguous input so the output is row contiguous
+  if (x.flags().row_contiguous) {
+    // Merge consecutive axes
+    std::vector<int> shape = {x.shape(axes[0])};
+    std::vector<size_t> strides = {x.strides()[axes[0]]};
+    for (int i = 1; i < axes.size(); i++) {
+      if (axes[i] - 1 == axes[i - 1]) {
+        shape.back() *= x.shape(axes[i]);
+        strides.back() = x.strides()[axes[i]];
+      } else {
+        shape.push_back(x.shape(axes[i]));
+        strides.push_back(x.strides()[axes[i]]);
+      }
+    }
+
+    if (strides.back() == 1) {
+      return ReductionPlan(ContiguousReduce, shape, strides);
+    } else if (strides.back() > 1) {
+      return ReductionPlan(ContiguousStridedReduce, shape, strides);
+    }
+  }
+
+  // Let's check if we can optimize our access patterns
+  //
+  // 1. We have a reduction axis with stride 1. Simply call
+  //    GeneralContiguousReduce and be done with it.
+  // 2. We have transpositions and we are not reducing over the axis with
+  //    stride 1. However, we are reducing over an axis where everything is
+  //    contiguous in memory to the right of that axis. We can call strided
+  //    reduce and be done with it.
+  // 2. We have weird transpositions and expands. Copy the strides to the
+  //    output, then call strided reduce.
+
+  // Sort reduction axes by stride in order to merge them and figure out if we
+  // have a contiguous reduction.
+  std::vector<std::pair<int, size_t>> reductions;
+  for (auto a : axes) {
+    reductions.push_back(std::make_pair(x.shape(a), x.strides()[a]));
+  }
+  std::sort(reductions.begin(), reductions.end(), [](auto a, auto b) {
+    return a.second > b.second;
+  });
+  // Extract the two smallest and try to merge them in case the contiguous
+  // reduction can be bigger than just the last axis.
+  for (int i = reductions.size() - 1; i >= 1; i--) {
+    auto a = reductions[i];
+    auto b = reductions[i - 1];
+
+    // b.stride = a.shape * a.stride then a and b are contiguous
+    if (b.second == a.first * a.second) {
+      reductions.erase(reductions.begin() + i);
+      reductions[i - 1] = std::make_pair(a.first * b.first, a.second);
+    }
+  }
+
+  std::vector<int> shape;
+  std::vector<size_t> strides;
+  for (auto r : reductions) {
+    shape.push_back(r.first);
+    strides.push_back(r.second);
+  }
+
+  // We can call the contiguous reduction op for every weird way the input is
+  // structured in the rest of the axes.
+  if (strides.back() == 1) {
+    return ReductionPlan(GeneralContiguousReduce, shape, strides);
+  }
+
+  // Delegate to the general strided reduction op if the axes after
+  // strides.back() are contiguous.
+  if (strides.back() > 1) {
+    int size = 1;
+    for (int i = x.ndim() - 1; i >= 0; i--) {
+      if (axes.back() == i) {
+        continue;
+      }
+      if (x.strides()[i] != size) {
+        break;
+      }
+      size *= x.shape(i);
+    }
+    if (size >= strides.back()) {
+      return ReductionPlan(GeneralStridedReduce, shape, strides);
+    }
+  }
+
+  return ReductionPlan(GeneralReduce, shape, strides);
+}
+
+} // namespace mlx::core

mlx/backend/common/scan.cpp

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

mlx/backend/common/slicing.cpp

@@ -0,0 +1,52 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+std::tuple<bool, int64_t, std::vector<int64_t>> prepare_slice(
+    const array& in,
+    std::vector<int>& start_indices,
+    std::vector<int>& strides) {
+  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 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);
+}
+
+} // namespace mlx::core

mlx/backend/common/slicing.h

@@ -0,0 +1,20 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+
+namespace mlx::core {
+
+std::tuple<bool, int64_t, std::vector<int64_t>> prepare_slice(
+    const array& in,
+    std::vector<int>& start_indices,
+    std::vector<int>& strides);
+
+void shared_buffer_slice(
+    const array& in,
+    const std::vector<size_t>& out_strides,
+    size_t data_offset,
+    array& out);
+
+} // namespace mlx::core

mlx/backend/common/sort.cpp

@@ -113,14 +113,14 @@ void sort(const array& in, array& out, int axis) {
   axis = axis < 0 ? axis + in.ndim() : axis;
   size_t n_rows = in.size() / in.shape(axis);
 
-  auto remaining_shape = in.shape();
+  auto remaining_shape = out.shape();
   remaining_shape.erase(remaining_shape.begin() + axis);
 
-  auto remaining_strides = in.strides();
+  auto remaining_strides = out.strides();
   remaining_strides.erase(remaining_strides.begin() + axis);
 
-  size_t axis_stride = in.strides()[axis];
-  int axis_size = in.shape(axis);
+  size_t axis_stride = out.strides()[axis];
+  int axis_size = out.shape(axis);
 
   // Perform sorting in place
   for (int i = 0; i < n_rows; i++) {
@@ -143,34 +143,42 @@ void argsort(const array& in, array& out, int axis) {
   axis = axis < 0 ? axis + in.ndim() : axis;
   size_t n_rows = in.size() / in.shape(axis);
 
-  auto remaining_shape = in.shape();
-  remaining_shape.erase(remaining_shape.begin() + axis);
+  auto in_remaining_shape = in.shape();
+  in_remaining_shape.erase(in_remaining_shape.begin() + axis);
 
-  auto remaining_strides = in.strides();
-  remaining_strides.erase(remaining_strides.begin() + axis);
+  auto in_remaining_strides = in.strides();
+  in_remaining_strides.erase(in_remaining_strides.begin() + axis);
 
-  size_t axis_stride = in.strides()[axis];
+  auto out_remaining_shape = out.shape();
+  out_remaining_shape.erase(out_remaining_shape.begin() + axis);
+
+  auto out_remaining_strides = out.strides();
+  out_remaining_strides.erase(out_remaining_strides.begin() + axis);
+
+  size_t in_stride = in.strides()[axis];
+  size_t out_stride = out.strides()[axis];
   int axis_size = in.shape(axis);
 
   // Perform sorting
   for (int i = 0; i < n_rows; i++) {
-    size_t loc = elem_to_loc(i, remaining_shape, remaining_strides);
-    const T* data_ptr = in.data<T>() + loc;
-    IdxT* idx_ptr = out.data<IdxT>() + loc;
+    size_t in_loc = elem_to_loc(i, in_remaining_shape, in_remaining_strides);
+    size_t out_loc = elem_to_loc(i, out_remaining_shape, out_remaining_strides);
+    const T* data_ptr = in.data<T>() + in_loc;
+    IdxT* idx_ptr = out.data<IdxT>() + out_loc;
 
-    StridedIterator st_(idx_ptr, axis_stride, 0);
-    StridedIterator ed_(idx_ptr, axis_stride, axis_size);
+    StridedIterator st_(idx_ptr, out_stride, 0);
+    StridedIterator ed_(idx_ptr, out_stride, axis_size);
 
     // Initialize with iota
     std::iota(st_, ed_, IdxT(0));
 
     // Sort according to vals
-    StridedIterator st(idx_ptr, axis_stride, 0);
-    StridedIterator ed(idx_ptr, axis_stride, axis_size);
+    StridedIterator st(idx_ptr, out_stride, 0);
+    StridedIterator ed(idx_ptr, out_stride, axis_size);
 
-    std::stable_sort(st, ed, [data_ptr, axis_stride](IdxT a, IdxT b) {
-      auto v1 = data_ptr[a * axis_stride];
-      auto v2 = data_ptr[b * axis_stride];
+    std::stable_sort(st, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
+      auto v1 = data_ptr[a * in_stride];
+      auto v2 = data_ptr[b * in_stride];
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }

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/common/utils.h

@@ -29,6 +29,15 @@ inline size_t elem_to_loc(int elem, const array& a) {
   return elem_to_loc(elem, a.shape(), a.strides());
 }
 
+template <typename stride_t>
+std::vector<stride_t> make_contiguous_strides(const std::vector<int>& shape) {
+  std::vector<stride_t> strides(shape.size(), 1);
+  for (int i = shape.size() - 1; i > 0; i--) {
+    strides[i - 1] = strides[i] * shape[i];
+  }
+  return strides;
+}
+
 // Collapse dims that are contiguous to possibly route to a better kernel
 // e.g. for x = transpose(array({0, 1, 2, 3, 4, 5, 6, 7}, {2, 2, 2}), {2, 0, 1})
 // should return {{2, 4}, {{1, 2}}}.

mlx/backend/metal/CMakeLists.txt

@@ -1,33 +1,140 @@
-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}
+              ${SRC_FILE}
+              "-DMLX_METAL_VERSION=${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)
+make_jit_source(hadamard)
 
-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(
+    fft
+    kernels/fft/radix.h
+    kernels/fft/readwrite.h
+  )
+  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
+  )
+  make_jit_source(quantized)
+  make_jit_source(gemv_masked)
+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}/hadamard.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cpp
@@ -37,10 +144,13 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/normalization.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/rope.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
   ${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
+  ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
 )
 
 if (NOT MLX_METAL_PATH)

mlx/backend/metal/allocator.cpp

@@ -242,8 +242,17 @@ void MetalAllocator::free(Buffer buffer) {
 }
 
 MetalAllocator& allocator() {
-  static MetalAllocator allocator_;
-  return allocator_;
+  // By creating the |allocator_| on heap, the destructor of MetalAllocator will
+  // not be called on exit and all the buffers will be leaked. This is necessary
+  // because releasing buffers can take more than 30sec when the program holds a
+  // lot of RAM (for example inferencing a LLM), and it would feel frozen to
+  // users when exiting.
+  // TODO(zcbenz): Consider using the `base::NoDestructor` class from Chromium
+  // when applying this pattern to more places, or when introducing sanitizers
+  // to MLX.
+  // https://source.chromium.org/chromium/chromium/src/+/main:base/no_destructor.h
+  static MetalAllocator* allocator_ = new MetalAllocator;
+  return *allocator_;
 }
 
 size_t set_cache_limit(size_t limit) {

mlx/backend/metal/binary.cpp

@@ -0,0 +1,296 @@
+// 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"
+
+#define BINARY_GPU(func)                                              \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    binary_op_gpu(inputs, out, get_primitive_string(this));           \
+  }
+
+#define BINARY_GPU_MULTI(func)                                         \
+  void func::eval_gpu(                                                 \
+      const std::vector<array>& inputs, std::vector<array>& outputs) { \
+    binary_op_gpu(inputs, outputs, get_primitive_string(this));        \
+  }
+
+namespace mlx::core {
+
+constexpr int MAX_BINARY_SPECIALIZED_DIMS = 5;
+
+std::string get_kernel_name(
+    BinaryOpType bopt,
+    const std::string& op,
+    const array& a,
+    bool use_2d,
+    int ndim) {
+  std::ostringstream kname;
+  switch (bopt) {
+    case BinaryOpType::ScalarScalar:
+      kname << "ss";
+      break;
+    case BinaryOpType::ScalarVector:
+      kname << (use_2d ? "sv2" : "sv");
+      break;
+    case BinaryOpType::VectorScalar:
+      kname << (use_2d ? "vs2" : "vs");
+      break;
+    case BinaryOpType::VectorVector:
+      kname << (use_2d ? "vv2" : "vv");
+      break;
+    case BinaryOpType::General:
+      kname << "g";
+      if (ndim <= MAX_BINARY_SPECIALIZED_DIMS) {
+        kname << ndim;
+      } else {
+        kname << "n";
+      }
+      break;
+  }
+  kname << op << type_to_name(a);
+  return kname.str();
+}
+
+void binary_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::string& op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+
+  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];
+
+  bool use_2d = out.data_size() > UINT32_MAX;
+  std::string kernel_name = get_kernel_name(bopt, op, a, use_2d, shape.size());
+  auto& d = metal::device(s.device);
+
+  auto kernel =
+      get_binary_two_kernel(d, kernel_name, a.dtype(), outputs[0].dtype(), op);
+
+  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 or 2D grid of threads
+    size_t nthreads = out.data_size();
+    MTL::Size grid_dims = use_2d
+        ? get_2d_grid_dims(outputs[0].shape(), outputs[0].strides())
+        : 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_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::string& op,
+    const Stream& s) {
+  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);
+  binary_op_gpu_inplace(inputs, outputs, op, s);
+}
+
+void binary_op_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::string& op) {
+  auto& s = outputs[0].primitive().stream();
+  binary_op_gpu(inputs, outputs, op, s);
+}
+
+void binary_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    array& out,
+    const std::string& op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  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];
+
+  bool use_2d = out.data_size() > UINT32_MAX;
+  std::string kernel_name = get_kernel_name(bopt, op, a, use_2d, shape.size());
+  auto& d = metal::device(s.device);
+
+  auto kernel = get_binary_kernel(d, kernel_name, a.dtype(), out.dtype(), op);
+  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 or 2D grid of threads
+
+    size_t nthreads = out.data_size();
+    MTL::Size grid_dims = use_2d ? get_2d_grid_dims(out.shape(), out.strides())
+                                 : 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_gpu(
+    const std::vector<array>& inputs,
+    array& out,
+    const std::string& op,
+    const Stream& s) {
+  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);
+  binary_op_gpu_inplace(inputs, out, op, s);
+}
+
+void binary_op_gpu(
+    const std::vector<array>& inputs,
+    array& out,
+    const std::string& op) {
+  auto& s = out.primitive().stream();
+  binary_op_gpu(inputs, out, op, s);
+}
+
+BINARY_GPU(Add)
+BINARY_GPU(ArcTan2)
+BINARY_GPU(Divide)
+BINARY_GPU_MULTI(DivMod)
+BINARY_GPU(Remainder)
+BINARY_GPU(Equal)
+BINARY_GPU(Greater)
+BINARY_GPU(GreaterEqual)
+BINARY_GPU(Less)
+BINARY_GPU(LessEqual)
+BINARY_GPU(LogicalAnd)
+BINARY_GPU(LogicalOr)
+BINARY_GPU(LogAddExp)
+BINARY_GPU(Maximum)
+BINARY_GPU(Minimum)
+BINARY_GPU(Multiply)
+BINARY_GPU(NotEqual)
+BINARY_GPU(Power)
+BINARY_GPU(Subtract)
+
+void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_op_gpu(inputs, out, get_primitive_string(this));
+      break;
+    case BitwiseBinary::Or:
+      binary_op_gpu(inputs, out, get_primitive_string(this));
+      break;
+    case BitwiseBinary::Xor:
+      binary_op_gpu(inputs, out, get_primitive_string(this));
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_op_gpu(inputs, out, get_primitive_string(this));
+      break;
+    case BitwiseBinary::RightShift:
+      binary_op_gpu(inputs, out, get_primitive_string(this));
+      break;
+  }
+}
+
+} // namespace mlx::core

mlx/backend/metal/binary.h

@@ -0,0 +1,33 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+
+namespace mlx::core {
+
+void binary_op_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::string& op,
+    const Stream& s);
+
+void binary_op_gpu(
+    const std::vector<array>& inputs,
+    array& out,
+    const std::string& op,
+    const Stream& s);
+
+void binary_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::string& op,
+    const Stream& s);
+
+void binary_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    array& out,
+    const std::string& op,
+    const Stream& s);
+
+} // 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"
@@ -56,12 +56,15 @@ inline void build_kernel(
     } else {
       add_indices = true;
       os << "    device const " << get_type_string(x.dtype()) << "* " << xname
-         << " [[buffer(" << cnt++ << ")]]," << std::endl
-         << "    constant const size_t* " << xname << "_strides [[buffer("
-         << cnt++ << ")]]," << std::endl;
+         << " [[buffer(" << cnt++ << ")]]," << std::endl;
     }
   }
 
+  if (add_indices) {
+    os << "    constant const size_t* in_strides [[buffer(" << cnt++
+       << ")]],\n";
+  }
+
   // Add the output arguments
   for (auto& x : outputs) {
     os << "    device " << get_type_string(x.dtype()) << "* "
@@ -110,13 +113,17 @@ inline void build_kernel(
   }
 
   // Read the inputs in tmps
-  for (auto& x : inputs) {
+  int nc_in_count = 0;
+  for (int i = 0; i < inputs.size(); ++i) {
+    auto& x = inputs[i];
     auto& xname = namer.get_name(x);
 
     if (is_constant(x)) {
-      os << "  " << get_type_string(x.dtype()) << " tmp_" << xname << " = ";
+      auto type_str = get_type_string(x.dtype());
+      os << "  auto tmp_" << xname << " = static_cast<"
+         << get_type_string(x.dtype()) << ">(";
       print_constant(os, x);
-      os << ";" << std::endl;
+      os << ");" << std::endl;
     } else if (is_scalar(x)) {
       os << "  " << get_type_string(x.dtype()) << " tmp_" << xname << " = "
          << xname << "[0];" << std::endl;
@@ -124,17 +131,20 @@ inline void build_kernel(
       os << "  " << get_type_string(x.dtype()) << " tmp_" << xname << " = "
          << xname << "[index];" << std::endl;
     } else if (!dynamic_dims) {
+      int offset = nc_in_count * ndim;
       os << "  " << get_type_string(x.dtype()) << " tmp_" << xname << " = "
          << xname << "[";
-      os << "index_0 * " << xname << "_strides[0]";
+      os << "index_0 * " << "in_strides[" << offset << "]";
       for (int i = 1; i < ndim; i++) {
-        os << " + index_" << i << " * " << xname << "_strides[" << i << "]";
+        os << " + index_" << i << " * " << "in_strides[" << offset + i << "]";
       }
       os << "];" << std::endl;
+      nc_in_count++;
     } else {
       os << "  " << get_type_string(x.dtype()) << " tmp_" << xname << " = "
-         << xname << "[elem_to_loc(index, output_shape, " << xname
-         << "_strides, ndim)];" << std::endl;
+         << xname << "[elem_to_loc(index, output_shape, in_strides + "
+         << nc_in_count * ndim << ", ndim)];" << std::endl;
+      nc_in_count++;
     }
   }
 
@@ -190,7 +200,8 @@ 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",
@@ -295,6 +306,7 @@ void Compiled::eval_gpu(
   // Put the inputs in
   int cnt = 0;
   int stride_idx = 1; // idx 0 is the output strides
+  std::vector<size_t> in_strides;
   for (int i = 0; i < inputs.size(); i++) {
     if (constant_ids_.find(inputs_[i].id()) != constant_ids_.end()) {
       continue;
@@ -302,13 +314,17 @@ void Compiled::eval_gpu(
     auto& x = inputs[i];
     compute_encoder.set_input_array(x, cnt++);
     if (!contiguous && !is_scalar(x)) {
-      compute_encoder->setBytes(
-          strides[stride_idx].data(),
-          strides[stride_idx].size() * sizeof(size_t),
-          cnt++);
+      in_strides.insert(
+          in_strides.end(),
+          strides[stride_idx].begin(),
+          strides[stride_idx].end());
       stride_idx++;
     }
   }
+  if (!in_strides.empty()) {
+    compute_encoder->setBytes(
+        in_strides.data(), in_strides.size() * sizeof(size_t), cnt++);
+  }
 
   compiled_allocate_outputs(
       inputs, outputs, inputs_, constant_ids_, contiguous, true);
@@ -336,7 +352,7 @@ 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;
@@ -347,7 +363,7 @@ 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"
@@ -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};
@@ -137,7 +138,7 @@ void explicit_gemm_conv_group_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);
 
   // Transpose kernel weights so that we can slice them by contiguous chunks
   // of channel groups.
@@ -247,7 +248,7 @@ void slow_conv_2D_gpu(
   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(
@@ -257,15 +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;
 
@@ -281,15 +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) &&
@@ -331,7 +336,17 @@ void implicit_gemm_conv_2D_gpu(
 
   // Encode and dispatch kernel
   auto& compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  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
@@ -340,7 +355,7 @@ 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
   compute_encoder.set_input_array(in, 0);
@@ -352,7 +367,7 @@ void implicit_gemm_conv_2D_gpu(
   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(
@@ -484,7 +499,8 @@ void implicit_gemm_conv_2D_general_gpu(
 
   // Encode and dispatch kernel
   auto& compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto kernel =
+      get_steel_conv_general_kernel(d, kname.str(), out, bm, bn, bk, wm, wn);
   compute_encoder->setComputePipelineState(kernel);
 
   // Deduce grid launch dimensions
@@ -512,7 +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(
@@ -613,7 +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
@@ -641,7 +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
@@ -689,7 +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);
   }
 }
 
@@ -703,6 +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
@@ -718,12 +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,
   };
 
@@ -735,6 +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 &&
@@ -759,6 +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) {
@@ -783,8 +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,
@@ -795,6 +889,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
         kernel_strides_,
         kernel_dilation_,
         input_dilation_,
+        groups_,
         flip_,
         copies);
   }

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
@@ -31,9 +33,6 @@ void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s) {
   } else {
     out.set_data(allocator::malloc_or_wait(out.nbytes()));
   }
-  if (out.size() == 0) {
-    return;
-  }
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;
   }
@@ -55,34 +54,46 @@ void copy_gpu_inplace(
     int64_t out_offset,
     CopyType ctype,
     const Stream& s) {
+  if (out.size() == 0) {
+    return;
+  }
+
   // Try to collapse contiguous dims
   auto [shape, strides] = collapse_contiguous_dims(
       data_shape, std::vector{strides_in_pre, strides_out_pre});
   auto& strides_in_ = strides[0];
   auto& strides_out_ = strides[1];
 
+  bool use_2d = out.data_size() > UINT32_MAX;
   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 << (use_2d ? "s2" : "s");
+        break;
+      case CopyType::Vector:
+        kname << (use_2d ? "v2" : "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 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;
@@ -106,7 +117,7 @@ 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);
     }
 
@@ -126,16 +137,17 @@ 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);
+    MTL::Size grid_dims = use_2d ? get_2d_grid_dims(out.shape(), out.strides())
+                                 : 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);
+    compute_encoder.dispatchThreads(grid_dims, group_dims);
   }
 }
 

mlx/backend/metal/device.cpp

@@ -14,7 +14,6 @@
 #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"
 
 namespace fs = std::filesystem;
@@ -25,9 +24,34 @@ 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 (MLX_METAL_VERSION >= 320)
+  return MTL::LanguageVersion3_2;
+#elif (MLX_METAL_VERSION >= 310)
+  return MTL::LanguageVersion3_1;
+#else
+  return MTL::LanguageVersion3_0;
+#endif
+}
+
+std::string get_colocated_mtllib_path(const std::string& lib_name) {
+  Dl_info info;
+  std::string mtllib_path;
+  std::string lib_ext = lib_name + ".metallib";
+
+  int success = dladdr((void*)get_colocated_mtllib_path, &info);
+  if (success) {
+    auto mtllib = fs::path(info.dli_fname).remove_filename() / lib_ext;
+    mtllib_path = mtllib.c_str();
+  }
+
+  return mtllib_path;
+}
+
 auto load_device() {
   auto devices = MTL::CopyAllDevices();
   auto device = static_cast<MTL::Device*>(devices->object(0))
@@ -37,7 +61,6 @@ auto load_device() {
   }
   return device;
 }
-
 std::pair<MTL::Library*, NS::Error*> load_library_from_path(
     MTL::Device* device,
     const char* path) {
@@ -116,6 +139,76 @@ MTL::Library* load_library(
 
 } // namespace
 
+CommandEncoder::CommandEncoder(MTL::CommandBuffer* cbuf) : cbuf(cbuf) {
+  enc = cbuf->computeCommandEncoder(MTL::DispatchTypeConcurrent);
+  enc->retain();
+}
+
+CommandEncoder::~CommandEncoder() {
+  enc->endEncoding();
+  enc->release();
+}
+
+void CommandEncoder::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 CommandEncoder::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 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();
@@ -130,9 +223,6 @@ 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();
   }
@@ -169,27 +259,26 @@ 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) {
@@ -200,25 +289,15 @@ 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);
 }
 
 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(MTL::DispatchTypeConcurrent);
-    // Increment ref count so the buffer is not garbage collected
-    compute_encoder->retain();
-    eit = encoder_map_
-              .emplace(index, std::make_unique<CommandEncoder>(compute_encoder))
-              .first;
+    eit =
+        encoder_map_.emplace(index, std::make_unique<CommandEncoder>(cb)).first;
   }
   return *(eit->second);
 }
@@ -232,13 +311,9 @@ void Device::register_library(
   }
 }
 
-void Device::register_library(
-    const std::string& lib_name,
-    const std::function<std::string(const std::string&)>& lib_path_func) {
+void Device::register_library(const std::string& lib_name) {
   if (auto it = library_map_.find(lib_name); it == library_map_.end()) {
-    std::string new_lib_path = lib_path_func(lib_name);
-    auto new_lib = load_library(device_, lib_name, new_lib_path.c_str());
-    library_map_.insert({lib_name, new_lib});
+    register_library(lib_name, get_colocated_mtllib_path(lib_name));
   }
 }
 
@@ -248,7 +323,7 @@ MTL::Library* Device::get_library_cache_(const std::string& lib_name) {
   if (auto it = library_map_.find(lib_name); it != library_map_.end()) {
     mtl_lib = it->second;
   } else { // Look for metallib alongside library
-    register_library(lib_name);
+    register_library(lib_name, get_colocated_mtllib_path(lib_name));
     mtl_lib = library_map_[lib_name];
   }
 
@@ -262,7 +337,11 @@ 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) {
@@ -344,7 +423,6 @@ MTL::Function* Device::get_function_(
   }
 
   mtl_func_consts->release();
-  desc->release();
 
   return mtl_function;
 }
@@ -513,11 +591,13 @@ 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;
 }
 

mlx/backend/metal/device.h

@@ -9,36 +9,16 @@
 #include <unordered_map>
 #include <unordered_set>
 
-#include <dlfcn.h>
-#include <filesystem>
-
 #include "mlx/array.h"
 #include "mlx/device.h"
 
-namespace fs = std::filesystem;
-
 namespace mlx::core::metal {
 
-inline std::string get_colocated_mtllib_path(const std::string& lib_name) {
-  Dl_info info;
-  std::string mtllib_path;
-  std::string lib_ext = lib_name + ".metallib";
-
-  int success = dladdr((void*)get_colocated_mtllib_path, &info);
-  if (success) {
-    auto mtllib = fs::path(info.dli_fname).remove_filename() / lib_ext;
-    mtllib_path = mtllib.c_str();
-  }
-
-  return mtllib_path;
-}
-
 using MTLFCList =
     std::vector<std::tuple<const void*, MTL::DataType, NS::UInteger>>;
 
 struct CommandEncoder {
-  CommandEncoder(MTL::ComputeCommandEncoder* enc)
-      : enc(enc), concurrent(false) {};
+  CommandEncoder(MTL::CommandBuffer* cbuf);
   CommandEncoder(const CommandEncoder&) = delete;
   CommandEncoder& operator=(const CommandEncoder&) = delete;
 
@@ -61,41 +41,24 @@ struct CommandEncoder {
     return enc;
   }
 
-  void set_input_array(const array& a, int idx, int 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, int 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 set_input_array(const array& a, int idx, int64_t offset = 0);
+  void set_output_array(array& a, int idx, int64_t offset = 0);
+  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();
+
  private:
+  void maybe_split();
+
+  int num_dispatches{0};
+  MTL::CommandBuffer* cbuf;
   MTL::ComputeCommandEncoder* enc;
-  bool concurrent;
+  bool concurrent{false};
   std::unordered_set<MTL::Resource*> outputs;
   std::unordered_set<MTL::Resource*> concurrent_outputs;
 };
@@ -112,7 +75,6 @@ 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);
@@ -123,10 +85,8 @@ class Device {
   void register_library(
       const std::string& lib_name,
       const std::string& lib_path);
-  void register_library(
-      const std::string& lib_name,
-      const std::function<std::string(const std::string&)>& lib_path_func =
-          get_colocated_mtllib_path);
+
+  void register_library(const std::string& lib_name);
 
   MTL::Library* get_library(const std::string& name);
 

mlx/backend/metal/fft.cpp

@@ -1,106 +1,803 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2024 Apple Inc.
+#include <cassert>
+#include <complex>
+#include <map>
+#include <numeric>
+#include <set>
+
+#include "mlx/3rdparty/pocketfft.h"
+#include "mlx/backend/metal/binary.h"
 #include "mlx/backend/metal/copy.h"
+#include "mlx/backend/metal/kernels.h"
+#include "mlx/backend/metal/slicing.h"
+#include "mlx/backend/metal/unary.h"
 #include "mlx/backend/metal/utils.h"
-#include "mlx/mlx.h"
-#include "mlx/primitives.h"
+#include "mlx/utils.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);
+using MTLFC = std::tuple<const void*, MTL::DataType, NS::UInteger>;
 
-  auto& in = inputs[0];
+#define MAX_STOCKHAM_FFT_SIZE 4096
+#define MAX_RADER_FFT_SIZE 2048
+#define MAX_BLUESTEIN_FFT_SIZE 2048
+// Threadgroup memory batching improves throughput for small n
+#define MIN_THREADGROUP_MEM_SIZE 256
+// For strided reads/writes, coalesce at least this many complex64s
+#define MIN_COALESCE_WIDTH 4
 
-  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.");
+inline const std::vector<int> supported_radices() {
+  // Ordered by preference in decomposition.
+  return {13, 11, 8, 7, 6, 5, 4, 3, 2};
+}
+
+std::vector<int> prime_factors(int n) {
+  int z = 2;
+  std::vector<int> factors;
+  while (z * z <= n) {
+    if (n % z == 0) {
+      factors.push_back(z);
+      n /= z;
+    } else {
+      z++;
+    }
+  }
+  if (n > 1) {
+    factors.push_back(n);
+  }
+  return factors;
+}
+
+struct FourStepParams {
+  bool required = false;
+  bool first_step = true;
+  int n1 = 0;
+  int n2 = 0;
+};
+
+// Forward Declaration
+void fft_op(
+    const array& in,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    const FourStepParams four_step_params,
+    bool inplace,
+    const Stream& s);
+
+struct FFTPlan {
+  int n = 0;
+  // Number of steps for each radix in the Stockham decomposition
+  std::vector<int> stockham;
+  // Number of steps for each radix in the Rader decomposition
+  std::vector<int> rader;
+  // Rader factor, 1 if no rader factors
+  int rader_n = 1;
+  int bluestein_n = -1;
+  // Four step FFT
+  bool four_step = false;
+  int n1 = 0;
+  int n2 = 0;
+};
+
+int next_fast_n(int n) {
+  return next_power_of_2(n);
+}
+
+std::vector<int> plan_stockham_fft(int n) {
+  auto radices = supported_radices();
+  std::vector<int> plan(radices.size(), 0);
+  int orig_n = n;
+  if (n == 1) {
+    return plan;
+  }
+  for (int i = 0; i < radices.size(); i++) {
+    int radix = radices[i];
+    // Manually tuned radices for powers of 2
+    if (is_power_of_2(orig_n) && orig_n < 512 && radix > 4) {
+      continue;
+    }
+    while (n % radix == 0) {
+      plan[i] += 1;
+      n /= radix;
+      if (n == 1) {
+        return plan;
+      }
+    }
+  }
+  throw std::runtime_error("Unplannable");
+}
+
+FFTPlan plan_fft(int n) {
+  auto radices = supported_radices();
+  std::set<int> radices_set(radices.begin(), radices.end());
+
+  FFTPlan plan;
+  plan.n = n;
+  plan.rader = std::vector<int>(radices.size(), 0);
+  auto factors = prime_factors(n);
+  int remaining_n = n;
+
+  // Four Step FFT when N is too large for shared mem.
+  if (n > MAX_STOCKHAM_FFT_SIZE && is_power_of_2(n)) {
+    // For power's of two we have a fast, no transpose four step implementation.
+    plan.four_step = true;
+    // Rough heuristic for choosing faster powers of two when we can
+    plan.n2 = n > 65536 ? 1024 : 64;
+    plan.n1 = n / plan.n2;
+    return plan;
+  } else if (n > MAX_STOCKHAM_FFT_SIZE) {
+    // Otherwise we use a multi-upload Bluestein's
+    plan.four_step = true;
+    plan.bluestein_n = next_fast_n(2 * n - 1);
+    return plan;
   }
 
-  size_t n = in.shape(axes_[0]);
+  for (int factor : factors) {
+    // Make sure the factor is a supported radix
+    if (radices_set.find(factor) == radices_set.end()) {
+      // We only support a single Rader factor currently
+      // TODO(alexbarron) investigate weirdness with large
+      // Rader sizes -- possibly a compiler issue?
+      if (plan.rader_n > 1 || n > MAX_RADER_FFT_SIZE) {
+        plan.four_step = n > MAX_BLUESTEIN_FFT_SIZE;
+        plan.bluestein_n = next_fast_n(2 * n - 1);
+        plan.stockham = plan_stockham_fft(plan.bluestein_n);
+        plan.rader = std::vector<int>(radices.size(), 0);
+        return plan;
+      }
+      // See if we can use Rader's algorithm to Stockham decompose n - 1
+      auto rader_factors = prime_factors(factor - 1);
+      int last_factor = -1;
+      for (int rf : rader_factors) {
+        // We don't nest Rader's algorithm so if `factor - 1`
+        // isn't Stockham decomposable we give up and do Bluestein's.
+        if (radices_set.find(rf) == radices_set.end()) {
+          plan.four_step = n > MAX_BLUESTEIN_FFT_SIZE;
+          plan.bluestein_n = next_fast_n(2 * n - 1);
+          plan.stockham = plan_stockham_fft(plan.bluestein_n);
+          plan.rader = std::vector<int>(radices.size(), 0);
+          return plan;
+        }
+      }
+      plan.rader = plan_stockham_fft(factor - 1);
+      plan.rader_n = factor;
+      remaining_n /= factor;
+    }
+  }
 
-  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");
+  plan.stockham = plan_stockham_fft(remaining_n);
+  return plan;
+}
+
+int compute_elems_per_thread(FFTPlan plan) {
+  // Heuristics for selecting an efficient number
+  // of threads to use for a particular mixed-radix FFT.
+  auto n = plan.n;
+
+  std::vector<int> steps;
+  auto radices = supported_radices();
+  steps.insert(steps.end(), plan.stockham.begin(), plan.stockham.end());
+  steps.insert(steps.end(), plan.rader.begin(), plan.rader.end());
+  std::set<int> used_radices;
+  for (int i = 0; i < steps.size(); i++) {
+    int radix = radices[i % radices.size()];
+    if (steps[i] > 0) {
+      used_radices.insert(radix);
+    }
+  }
+
+  // Manual tuning for 7/11/13
+  if (used_radices.find(7) != used_radices.end() &&
+      (used_radices.find(11) != used_radices.end() ||
+       used_radices.find(13) != used_radices.end())) {
+    return 7;
+  } else if (
+      used_radices.find(11) != used_radices.end() &&
+      used_radices.find(13) != used_radices.end()) {
+    return 11;
+  }
+
+  // TODO(alexbarron) Some really weird stuff is going on
+  // for certain `elems_per_thread` on large composite n.
+  // Possibly a compiler issue?
+  if (n == 3159)
+    return 13;
+  if (n == 3645)
+    return 5;
+  if (n == 3969)
+    return 7;
+  if (n == 1982)
+    return 5;
+
+  if (used_radices.size() == 1) {
+    return *(used_radices.begin());
+  }
+  if (used_radices.size() == 2) {
+    if (used_radices.find(11) != used_radices.end() ||
+        used_radices.find(13) != used_radices.end()) {
+      return std::accumulate(used_radices.begin(), used_radices.end(), 0) / 2;
+    }
+    std::vector<int> radix_vec(used_radices.begin(), used_radices.end());
+    return radix_vec[1];
+  }
+  // In all other cases use the second smallest radix.
+  std::vector<int> radix_vec(used_radices.begin(), used_radices.end());
+  return radix_vec[1];
+}
+
+// Rader
+int mod_exp(int x, int y, int n) {
+  int out = 1;
+  while (y) {
+    if (y & 1) {
+      out = out * x % n;
+    }
+    y >>= 1;
+    x = x * x % n;
+  }
+  return out;
+}
+
+int primitive_root(int n) {
+  auto factors = prime_factors(n - 1);
+
+  for (int r = 2; r < n - 1; r++) {
+    bool found = true;
+    for (int factor : factors) {
+      if (mod_exp(r, (n - 1) / factor, n) == 1) {
+        found = false;
+        break;
+      }
+    }
+    if (found) {
+      return r;
+    }
+  }
+  return -1;
+}
+
+std::tuple<array, array, array> compute_raders_constants(
+    int rader_n,
+    const Stream& s) {
+  int proot = primitive_root(rader_n);
+  // Fermat's little theorem
+  int inv = mod_exp(proot, rader_n - 2, rader_n);
+  std::vector<short> g_q(rader_n - 1);
+  std::vector<short> g_minus_q(rader_n - 1);
+  for (int i = 0; i < rader_n - 1; i++) {
+    g_q[i] = mod_exp(proot, i, rader_n);
+    g_minus_q[i] = mod_exp(inv, i, rader_n);
+  }
+  array g_q_arr(g_q.begin(), {rader_n - 1});
+  array g_minus_q_arr(g_minus_q.begin(), {rader_n - 1});
+
+  std::vector<std::complex<float>> b_q(rader_n - 1);
+  for (int i = 0; i < rader_n - 1; i++) {
+    float pi_i = (float)g_minus_q[i] * -2.0 * M_PI / rader_n;
+    b_q[i] = std::exp(std::complex<float>(0, pi_i));
+  }
+
+  array b_q_fft({rader_n - 1}, complex64, nullptr, {});
+  b_q_fft.set_data(allocator::malloc_or_wait(b_q_fft.nbytes()));
+  auto b_q_fft_ptr =
+      reinterpret_cast<std::complex<float>*>(b_q_fft.data<complex64_t>());
+  std::ptrdiff_t item_size = b_q_fft.itemsize();
+  size_t fft_size = rader_n - 1;
+  // This FFT is always small (<4096, batch 1) so save some overhead
+  // and do it on the CPU
+  pocketfft::c2c(
+      /* shape= */ {fft_size},
+      /* stride_in= */ {item_size},
+      /* stride_out= */ {item_size},
+      /* axes= */ {0},
+      /* forward= */ true,
+      /* data_in= */ b_q.data(),
+      /* data_out= */ b_q_fft_ptr,
+      /* scale= */ 1.0f);
+  return std::make_tuple(b_q_fft, g_q_arr, g_minus_q_arr);
+}
+
+// Bluestein
+std::pair<array, array> compute_bluestein_constants(int n, int bluestein_n) {
+  // We need to calculate the Bluestein twiddle factors
+  // in double precision for the overall numerical stability
+  // of Bluestein's FFT algorithm to be acceptable.
+  //
+  // Metal doesn't support float64, so instead we
+  // manually implement the required operations on cpu.
+  //
+  // In numpy:
+  // w_k = np.exp(-1j * np.pi / N * (np.arange(-N + 1, N) ** 2))
+  // w_q = np.fft.fft(1/w_k)
+  // return w_k, w_q
+  int length = 2 * n - 1;
+
+  std::vector<std::complex<float>> w_k_vec(n);
+  std::vector<std::complex<float>> w_q_vec(bluestein_n, 0);
+
+  for (int i = -n + 1; i < n; i++) {
+    double theta = pow(i, 2) * M_PI / (double)n;
+    w_q_vec[i + n - 1] = std::exp(std::complex<double>(0, theta));
+    if (i >= 0) {
+      w_k_vec[i] = std::exp(std::complex<double>(0, -theta));
+    }
+  }
+
+  array w_k({n}, complex64, nullptr, {});
+  w_k.set_data(allocator::malloc_or_wait(w_k.nbytes()));
+  std::copy(w_k_vec.begin(), w_k_vec.end(), w_k.data<complex64_t>());
+
+  array w_q({bluestein_n}, complex64, nullptr, {});
+  w_q.set_data(allocator::malloc_or_wait(w_q.nbytes()));
+  auto w_q_ptr =
+      reinterpret_cast<std::complex<float>*>(w_q.data<complex64_t>());
+
+  std::ptrdiff_t item_size = w_q.itemsize();
+  size_t fft_size = bluestein_n;
+  pocketfft::c2c(
+      /* shape= */ {fft_size},
+      /* stride_in= */ {item_size},
+      /* stride_out= */ {item_size},
+      /* axes= */ {0},
+      /* forward= */ true,
+      /* data_in= */ w_q_vec.data(),
+      /* data_out= */ w_q_ptr,
+      /* scale= */ 1.0f);
+  return std::make_tuple(w_k, w_q);
+}
+
+void multi_upload_bluestein_fft(
+    const array& in,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    FFTPlan& plan,
+    std::vector<array> copies,
+    const Stream& s) {
+  // TODO(alexbarron) Implement fused kernels for mutli upload bluestein's
+  // algorithm
+  int n = inverse ? out.shape(axis) : in.shape(axis);
+  auto [w_k, w_q] = compute_bluestein_constants(n, plan.bluestein_n);
+
+  // Broadcast w_q and w_k to the batch size
+  std::vector<size_t> b_strides(in.ndim(), 0);
+  b_strides[axis] = 1;
+  array w_k_broadcast({}, complex64, nullptr, {});
+  array w_q_broadcast({}, complex64, nullptr, {});
+  w_k_broadcast.copy_shared_buffer(w_k, b_strides, {}, w_k.data_size());
+  w_q_broadcast.copy_shared_buffer(w_q, b_strides, {}, w_q.data_size());
+
+  auto temp_shape = inverse ? out.shape() : in.shape();
+  array temp(temp_shape, complex64, nullptr, {});
+  array temp1(temp_shape, complex64, nullptr, {});
+
+  if (real && !inverse) {
+    // Convert float32->complex64
+    copy_gpu(in, temp, CopyType::General, s);
+  } else if (real && inverse) {
+    int back_offset = n % 2 == 0 ? 2 : 1;
+    auto slice_shape = in.shape();
+    slice_shape[axis] -= back_offset;
+    array slice_temp(slice_shape, complex64, nullptr, {});
+    array conj_temp(in.shape(), complex64, nullptr, {});
+    copies.push_back(slice_temp);
+    copies.push_back(conj_temp);
+
+    std::vector<int> rstarts(in.ndim(), 0);
+    std::vector<int> rstrides(in.ndim(), 1);
+    rstarts[axis] = in.shape(axis) - back_offset;
+    rstrides[axis] = -1;
+    unary_op_gpu({in}, conj_temp, "Conjugate", s);
+    slice_gpu(in, slice_temp, rstarts, rstrides, s);
+    concatenate_gpu({conj_temp, slice_temp}, temp, (int)axis, s);
+  } else if (inverse) {
+    unary_op_gpu({in}, temp, "Conjugate", s);
+  } else {
+    temp.copy_shared_buffer(in);
+  }
+
+  binary_op_gpu({temp, w_k_broadcast}, temp1, "Multiply", s);
+
+  std::vector<std::pair<int, int>> pads;
+  auto padded_shape = out.shape();
+  padded_shape[axis] = plan.bluestein_n;
+  array pad_temp(padded_shape, complex64, nullptr, {});
+  pad_gpu(temp1, array(complex64_t{0.0f, 0.0f}), pad_temp, {(int)axis}, {0}, s);
+
+  array pad_temp1(padded_shape, complex64, nullptr, {});
+  fft_op(
+      pad_temp,
+      pad_temp1,
+      axis,
+      /*inverse=*/false,
+      /*real=*/false,
+      FourStepParams(),
+      /*inplace=*/false,
+      s);
+
+  binary_op_gpu_inplace({pad_temp1, w_q_broadcast}, pad_temp, "Multiply", s);
+
+  fft_op(
+      pad_temp,
+      pad_temp1,
+      axis,
+      /* inverse= */ true,
+      /* real= */ false,
+      FourStepParams(),
+      /*inplace=*/true,
+      s);
+
+  int offset = plan.bluestein_n - (2 * n - 1);
+  std::vector<int> starts(in.ndim(), 0);
+  std::vector<int> strides(in.ndim(), 1);
+  starts[axis] = plan.bluestein_n - offset - n;
+  slice_gpu(pad_temp1, temp, starts, strides, s);
+
+  binary_op_gpu_inplace({temp, w_k_broadcast}, temp1, "Multiply", s);
+
+  if (real && !inverse) {
+    std::vector<int> rstarts(in.ndim(), 0);
+    std::vector<int> rstrides(in.ndim(), 1);
+    slice_gpu(temp1, out, rstarts, strides, s);
+  } else if (real && inverse) {
+    std::vector<size_t> b_strides(in.ndim(), 0);
+    auto inv_n = array({1.0f / n}, {1}, float32);
+    array temp_float(out.shape(), out.dtype(), nullptr, {});
+    copies.push_back(temp_float);
+    copies.push_back(inv_n);
+
+    copy_gpu(temp1, temp_float, CopyType::General, s);
+    binary_op_gpu({temp_float, inv_n}, out, "Multiply", s);
+  } else if (inverse) {
+    auto inv_n = array({1.0f / n}, {1}, complex64);
+    unary_op_gpu({temp1}, temp, "Conjugate", s);
+    binary_op_gpu({temp, inv_n}, out, "Multiply", s);
+    copies.push_back(inv_n);
+  } else {
+    out.copy_shared_buffer(temp1);
+  }
+
+  copies.push_back(w_k);
+  copies.push_back(w_q);
+  copies.push_back(w_k_broadcast);
+  copies.push_back(w_q_broadcast);
+  copies.push_back(temp);
+  copies.push_back(temp1);
+  copies.push_back(pad_temp);
+  copies.push_back(pad_temp1);
+}
+
+void four_step_fft(
+    const array& in,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    FFTPlan& plan,
+    std::vector<array> copies,
+    const Stream& s) {
+  auto& d = metal::device(s.device);
+
+  if (plan.bluestein_n == -1) {
+    // Fast no transpose implementation for powers of 2.
+    FourStepParams four_step_params = {
+        /* required= */ true, /* first_step= */ true, plan.n1, plan.n2};
+    auto temp_shape = (real && inverse) ? out.shape() : in.shape();
+    array temp(temp_shape, complex64, nullptr, {});
+    fft_op(
+        in, temp, axis, inverse, real, four_step_params, /*inplace=*/false, s);
+    four_step_params.first_step = false;
+    fft_op(
+        temp, out, axis, inverse, real, four_step_params, /*inplace=*/false, s);
+    copies.push_back(temp);
+  } else {
+    multi_upload_bluestein_fft(in, out, axis, inverse, real, plan, copies, s);
+  }
+}
+
+void fft_op(
+    const array& in,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    const FourStepParams four_step_params,
+    bool inplace,
+    const Stream& s) {
+  auto& d = metal::device(s.device);
+
+  size_t n = out.dtype() == float32 ? out.shape(axis) : in.shape(axis);
+  if (n == 1) {
+    out.copy_shared_buffer(in);
+    return;
+  }
+
+  if (four_step_params.required) {
+    // Four Step FFT decomposes into two FFTs: n1 on columns, n2 on rows
+    n = four_step_params.first_step ? four_step_params.n1 : four_step_params.n2;
   }
 
   // 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) {
+  auto check_input = [&axis, &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;
+    bool no_copy = x.strides()[axis] == 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]) {
+      size_t cur_stride = x.shape(axis);
+      for (int a = 0; a < x.ndim(); a++) {
+        if (a == axis) {
           strides.push_back(1);
         } else {
           strides.push_back(cur_stride);
-          cur_stride *= x.shape(axis);
+          cur_stride *= x.shape(a);
         }
       }
 
       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;
+      auto [data_size, is_row_contiguous, is_col_contiguous] =
+          check_contiguity(x.shape(), strides);
+
+      flags.col_contiguous = is_row_contiguous;
+      flags.row_contiguous = is_col_contiguous;
+      flags.contiguous = data_size == x_copy.size();
 
       x_copy.set_data(
-          allocator::malloc_or_wait(x.nbytes()), x.data_size(), strides, flags);
+          allocator::malloc_or_wait(x.nbytes()), 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]);
+  const array& in_contiguous = check_input(in);
+
+  // real to complex: n -> (n/2)+1
+  // complex to real: (n/2)+1 -> n
+  auto out_strides = in_contiguous.strides();
+  size_t out_data_size = in_contiguous.data_size();
+  if (in.shape(axis) != out.shape(axis)) {
+    for (int i = 0; i < out_strides.size(); i++) {
+      if (out_strides[i] != 1) {
+        out_strides[i] = out_strides[i] / in.shape(axis) * out.shape(axis);
+      }
+    }
+    out_data_size = out_data_size / in.shape(axis) * out.shape(axis);
+  }
+
+  auto plan = plan_fft(n);
+  if (plan.four_step) {
+    four_step_fft(in, out, axis, inverse, real, plan, copies, s);
+    d.get_command_buffer(s.index)->addCompletedHandler(
+        [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
+    return;
+  }
 
   // TODO: allow donation here
-  out.set_data(
-      allocator::malloc_or_wait(out.nbytes()),
-      in_contiguous.data_size(),
-      in_contiguous.strides(),
-      in_contiguous.flags());
+  if (!inplace) {
+    out.set_data(
+        allocator::malloc_or_wait(out.nbytes()),
+        out_data_size,
+        out_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 radices = supported_radices();
+  int fft_size = plan.bluestein_n > 0 ? plan.bluestein_n : n;
+
+  // Setup function constants
+  bool power_of_2 = is_power_of_2(fft_size);
+
+  auto make_int = [](int* a, int i) {
+    return std::make_tuple(a, MTL::DataType::DataTypeInt, i);
+  };
+  auto make_bool = [](bool* a, int i) {
+    return std::make_tuple(a, MTL::DataType::DataTypeBool, i);
+  };
+
+  std::vector<MTLFC> func_consts = {
+      make_bool(&inverse, 0), make_bool(&power_of_2, 1)};
+
+  // Start of radix/rader step constants
+  int index = 4;
+  for (int i = 0; i < plan.stockham.size(); i++) {
+    func_consts.push_back(make_int(&plan.stockham[i], index));
+    index += 1;
+  }
+  for (int i = 0; i < plan.rader.size(); i++) {
+    func_consts.push_back(make_int(&plan.rader[i], index));
+    index += 1;
+  }
+  int elems_per_thread = compute_elems_per_thread(plan);
+  func_consts.push_back(make_int(&elems_per_thread, 2));
+
+  int rader_m = n / plan.rader_n;
+  func_consts.push_back(make_int(&rader_m, 3));
+
+  // The overall number of FFTs we're going to compute for this input
+  int size = out.dtype() == float32 ? out.size() : in.size();
+  if (real && inverse && four_step_params.required) {
+    size = out.size();
+  }
+  int total_batch_size = size / n;
+  int threads_per_fft = (fft_size + elems_per_thread - 1) / elems_per_thread;
+
+  // We batch among threadgroups for improved efficiency when n is small
+  int threadgroup_batch_size = std::max(MIN_THREADGROUP_MEM_SIZE / fft_size, 1);
+  if (four_step_params.required) {
+    // Require a threadgroup batch size of at least 4 for four step FFT
+    // so we can coalesce the memory accesses.
+    threadgroup_batch_size =
+        std::max(threadgroup_batch_size, MIN_COALESCE_WIDTH);
+  }
+  int threadgroup_mem_size = next_power_of_2(threadgroup_batch_size * fft_size);
+  // FFTs up to 2^20 are currently supported
+  assert(threadgroup_mem_size <= MAX_STOCKHAM_FFT_SIZE);
+
+  // ceil divide
+  int batch_size =
+      (total_batch_size + threadgroup_batch_size - 1) / threadgroup_batch_size;
+
+  if (real && !four_step_params.required) {
+    // We can perform 2 RFFTs at once so the batch size is halved.
+    batch_size = (batch_size + 2 - 1) / 2;
+  }
+  int out_buffer_size = out.size();
 
   auto& compute_encoder = d.get_command_encoder(s.index);
+  auto in_type_str = in.dtype() == float32 ? "float" : "float2";
+  auto out_type_str = out.dtype() == float32 ? "float" : "float2";
+  // Only required by four step
+  int step = -1;
   {
     std::ostringstream kname;
-    kname << "fft_" << n;
-    auto kernel = d.get_kernel(kname.str());
+    std::string inv_string = inverse ? "true" : "false";
+    std::string real_string = real ? "true" : "false";
+    std::string func_name;
+    if (plan.bluestein_n > 0) {
+      kname << "bluestein_fft_mem_" << threadgroup_mem_size << "_"
+            << in_type_str << "_" << out_type_str;
+      func_name = "bluestein_fft";
+    } else if (plan.rader_n > 1) {
+      kname << "rader_fft_mem_" << threadgroup_mem_size << "_" << in_type_str
+            << "_" << out_type_str;
+      func_name = "rader_fft";
+    } else if (four_step_params.required) {
+      step = four_step_params.first_step ? 0 : 1;
+      kname << "four_step_mem_" << threadgroup_mem_size << "_" << in_type_str
+            << "_" << out_type_str << "_" << step << "_" << real_string;
+      func_name = "four_step_fft";
+    } else {
+      kname << "fft_mem_" << threadgroup_mem_size << "_" << in_type_str << "_"
+            << out_type_str;
+      func_name = "fft";
+    }
+    std::string base_name = kname.str();
+    // We use a specialized kernel for each FFT size
+    kname << "_n" << fft_size << "_inv_" << inverse;
+    std::string hash_name = kname.str();
+    auto template_def = func_name == "four_step_fft" ? get_template_definition(
+                                                           base_name,
+                                                           func_name,
+                                                           threadgroup_mem_size,
+                                                           in_type_str,
+                                                           out_type_str,
+                                                           step,
+                                                           real)
+                                                     : get_template_definition(
+                                                           base_name,
+                                                           func_name,
+                                                           threadgroup_mem_size,
+                                                           in_type_str,
+                                                           out_type_str);
+    auto kernel =
+        get_fft_kernel(d, base_name, hash_name, func_consts, template_def);
 
-    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);
+    if (plan.bluestein_n > 0) {
+      // Precomputed twiddle factors for Bluestein's
+      auto [w_k, w_q] = compute_bluestein_constants(n, plan.bluestein_n);
+      copies.push_back(w_q);
+      copies.push_back(w_k);
+
+      compute_encoder.set_input_array(w_q, 2); // w_q
+      compute_encoder.set_input_array(w_k, 3); // w_k
+      compute_encoder->setBytes(&n, sizeof(int), 4);
+      compute_encoder->setBytes(&plan.bluestein_n, sizeof(int), 5);
+      compute_encoder->setBytes(&total_batch_size, sizeof(int), 6);
+    } else if (plan.rader_n > 1) {
+      auto [b_q, g_q, g_minus_q] = compute_raders_constants(plan.rader_n, s);
+      copies.push_back(b_q);
+      copies.push_back(g_q);
+      copies.push_back(g_minus_q);
+
+      compute_encoder.set_input_array(b_q, 2);
+      compute_encoder.set_input_array(g_q, 3);
+      compute_encoder.set_input_array(g_minus_q, 4);
+      compute_encoder->setBytes(&n, sizeof(int), 5);
+      compute_encoder->setBytes(&total_batch_size, sizeof(int), 6);
+      compute_encoder->setBytes(&plan.rader_n, sizeof(int), 7);
+    } else if (four_step_params.required) {
+      compute_encoder->setBytes(&four_step_params.n1, sizeof(int), 2);
+      compute_encoder->setBytes(&four_step_params.n2, sizeof(int), 3);
+      compute_encoder->setBytes(&total_batch_size, sizeof(int), 4);
+    } else {
+      compute_encoder->setBytes(&n, sizeof(int), 2);
+      compute_encoder->setBytes(&total_batch_size, sizeof(int), 3);
+    }
+
+    auto group_dims = MTL::Size(1, threadgroup_batch_size, threads_per_fft);
+    auto grid_dims =
+        MTL::Size(batch_size, threadgroup_batch_size, threads_per_fft);
     compute_encoder->dispatchThreads(grid_dims, group_dims);
   }
   d.get_command_buffer(s.index)->addCompletedHandler(
       [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
 }
 
+void fft_op(
+    const array& in,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    bool inplace,
+    const Stream& s) {
+  fft_op(in, out, axis, inverse, real, FourStepParams(), inplace, s);
+}
+
+void nd_fft_op(
+    const array& in,
+    array& out,
+    const std::vector<size_t>& axes,
+    bool inverse,
+    bool real,
+    const Stream& s) {
+  // Perform ND FFT on GPU as a series of 1D FFTs
+  auto temp_shape = inverse ? in.shape() : out.shape();
+  array temp1(temp_shape, complex64, nullptr, {});
+  array temp2(temp_shape, complex64, nullptr, {});
+  std::vector<array> temp_arrs = {temp1, temp2};
+  for (int i = axes.size() - 1; i >= 0; i--) {
+    int reverse_index = axes.size() - i - 1;
+    // For 5D and above, we don't want to reallocate our two temporary arrays
+    bool inplace = reverse_index >= 3 && i != 0;
+    // Opposite order for fft vs ifft
+    int index = inverse ? reverse_index : i;
+    size_t axis = axes[index];
+    // Mirror np.fft.(i)rfftn and perform a real transform
+    // only on the final axis.
+    bool step_real = (real && index == axes.size() - 1);
+    int step_shape = inverse ? out.shape(axis) : in.shape(axis);
+    const array& in_arr = i == axes.size() - 1 ? in : temp_arrs[1 - i % 2];
+    array& out_arr = i == 0 ? out : temp_arrs[i % 2];
+    fft_op(in_arr, out_arr, axis, inverse, step_real, inplace, s);
+  }
+
+  auto& d = metal::device(s.device);
+  d.get_command_buffer(s.index)->addCompletedHandler(
+      [temp_arrs](MTL::CommandBuffer*) mutable { temp_arrs.clear(); });
+}
+
+void FFT::eval_gpu(const std::vector<array>& inputs, array& out) {
+  auto& s = stream();
+  auto& in = inputs[0];
+
+  if (axes_.size() > 1) {
+    nd_fft_op(in, out, axes_, inverse_, real_, s);
+  } else {
+    fft_op(in, out, axes_[0], inverse_, real_, /*inplace=*/false, s);
+  }
+}
+
 } // namespace mlx::core

mlx/backend/metal/hadamard.cpp

@@ -0,0 +1,203 @@
+// Copyright © 2024 Apple Inc.
+
+#include <map>
+
+#include "mlx/backend/common/compiled.h"
+#include "mlx/backend/common/hadamard.h"
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/jit/includes.h"
+#include "mlx/backend/metal/kernels.h"
+#include "mlx/backend/metal/utils.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+constexpr int MAX_HADAMARD_THREADS_PER_GROUP = 256;
+constexpr int MAX_HADAMARD_BYTES = 32768; // 32KB
+
+std::string gen_hadamard_codelet(int m) {
+  // Generate a O(m^2) hadamard codelet for a given M
+  // using the hadamard matrices above
+  //
+  // e.g. m = 2
+  // METAL_FUNC void hadamard_m(thread float *x) {
+  //   float tmp[2];
+  //   tmp[0] = + x[0] + x[1];
+  //   tmp[1] = + x[0] - x[1];
+  //   for (int i = 0; i < 2; i++) { x[i] = tmp[i]; }
+  // }
+  //
+  auto h_matrices = hadamard_matrices();
+  auto& matrix = h_matrices[m];
+
+  std::ostringstream source;
+  source << "METAL_FUNC void hadamard_radix_m(thread float *x) {" << std::endl;
+  if (m == 1) {
+    source << "}" << std::endl;
+    return source.str();
+  }
+  source << "  float tmp[" << m << "];" << std::endl;
+  auto start = 1;
+  auto end = matrix.find('\n', start);
+
+  int index = 0;
+  while (end != std::string_view::npos) {
+    source << "  tmp[" << index << "] = ";
+    auto row = matrix.substr(start, end - start);
+    for (int i = 0; i < row.length(); i++) {
+      source << " " << row[i] << " x[" << i << "]";
+    }
+    source << ";" << std::endl;
+    start = end + 1;
+    end = matrix.find('\n', start);
+    index++;
+  }
+  source << "  for (int i = 0; i < " << m << "; i++) { x[i] = tmp[i]; }"
+         << std::endl;
+  source << "}" << std::endl;
+  return source.str();
+}
+
+void launch_hadamard(
+    const array& in,
+    array& out,
+    int batch_size,
+    int threads_per,
+    const std::string kernel_name,
+    float scale,
+    const Stream& s) {
+  auto& d = metal::device(s.device);
+
+  const auto& lib_name = kernel_name.substr(1);
+  auto lib = d.get_library(lib_name);
+  auto kernel = d.get_kernel(kernel_name, lib);
+  assert(threads_per <= kernel->maxTotalThreadsPerThreadgroup());
+
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder->setComputePipelineState(kernel);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_output_array(out, 1);
+  compute_encoder->setBytes(&scale, sizeof(float), 2);
+
+  MTL::Size group_dims = MTL::Size(1, threads_per, 1);
+  MTL::Size grid_dims = MTL::Size(batch_size, threads_per, 1);
+  compute_encoder->dispatchThreads(grid_dims, group_dims);
+}
+
+void Hadamard::eval_gpu(const std::vector<array>& inputs, array& out) {
+  auto& s = stream();
+
+  auto& in = inputs[0];
+
+  std::vector<array> copies;
+  // Only support the last axis for now
+  int axis = in.ndim() - 1;
+  auto check_input = [&copies, &s](const array& x) {
+    // TODO(alexbarron) pass strides to kernel to relax this constraint
+    bool no_copy = x.flags().row_contiguous;
+    if (no_copy) {
+      return x;
+    } else {
+      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
+      copy_gpu(x, copies.back(), CopyType::General, s);
+      return copies.back();
+    }
+  };
+  const array& in_contiguous = check_input(in);
+
+  if (in_contiguous.is_donatable()) {
+    out.move_shared_buffer(in_contiguous);
+  } else {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  }
+
+  auto [n, m] = decompose_hadamard(in.shape(axis));
+
+  if (n * (int)size_of(in.dtype()) > MAX_HADAMARD_BYTES) {
+    throw std::invalid_argument(
+        "[hadamard] For n = m*2^k, 2^k > 8192 for FP32 or 2^k > 16384 for FP16/BF16 NYI");
+  }
+
+  int max_radix = std::min(n, 16);
+  // Use read_width 2 for m = 28 to avoid register spilling
+  int read_width = (n == 2 || m == 28) ? 2 : 4;
+
+  std::ostringstream kname;
+  kname << "hadamard_" << n * m << "_" << type_to_name(out);
+  auto kernel_name = kname.str();
+  auto& d = metal::device(s.device);
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    auto codelet = gen_hadamard_codelet(m);
+    kernel_source << metal::utils() << codelet << metal::hadamard();
+    kernel_source << get_template_definition(
+        "n" + kernel_name,
+        "hadamard_n",
+        get_type_string(in.dtype()),
+        n,
+        max_radix,
+        read_width);
+    kernel_source << get_template_definition(
+        "m" + kernel_name,
+        "hadamard_m",
+        get_type_string(in.dtype()),
+        n,
+        m,
+        read_width);
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+
+  int batch_size = in.size() / n;
+  int threads_per = n / max_radix;
+
+  if (m > 1) {
+    // When m is greater than 1, we decompose the
+    // computation into two uploads to the GPU:
+    //
+    // e.g. len(x) = 12*4 = 48, m = 12, n = 4
+    //
+    // y = h48 @ x
+    //
+    // Upload 1:
+    // tmp = a.reshape(12, 4) @ h4
+    //
+    // Upload 2:
+    // y = h12 @ tmp
+    array temp(in.shape(), in.dtype(), nullptr, {});
+    temp.set_data(allocator::malloc_or_wait(temp.nbytes()));
+    copies.push_back(temp);
+
+    launch_hadamard(
+        in_contiguous,
+        temp,
+        batch_size,
+        threads_per,
+        "n" + kernel_name,
+        1.0,
+        s);
+
+    // Metal sometimes reports 256 max threads per group for hadamard_m kernel
+    threads_per = std::min(n / read_width, MAX_HADAMARD_THREADS_PER_GROUP);
+    batch_size = in.size() / m / read_width / threads_per;
+    launch_hadamard(
+        temp, out, batch_size, threads_per, "m" + kernel_name, scale_, s);
+  } else {
+    launch_hadamard(
+        in_contiguous,
+        out,
+        batch_size,
+        threads_per,
+        "n" + kernel_name,
+        scale_,
+        s);
+  }
+
+  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 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(kname.str());
+  auto kernel = d.get_kernel(kernel_name, lib);
   compute_encoder->setComputePipelineState(kernel);
 
   size_t slice_size = 1;
@@ -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) {
-    compute_encoder.set_input_array(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;
+
+  {
+    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(kname.str());
+  auto kernel = d.get_kernel(kernel_name, lib);
 
   auto& upd = inputs.back();
   size_t nthreads = upd.size();
@@ -206,17 +293,28 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
         out.shape().data(), out.shape().size() * sizeof(int), 3);
     compute_encoder->setBytes(
         out.strides().data(), out.strides().size() * sizeof(size_t), 4);
-    compute_encoder->setBytes(&upd_size, sizeof(size_t), 5);
+
+    size_t out_ndim = out.ndim();
+    compute_encoder->setBytes(&out_ndim, sizeof(out_ndim), 5);
+    if (upd_ndim <= 1) {
+      // Placeholder so Metal doesn't compalain
+      int shape_ = 0;
+      compute_encoder->setBytes(&shape_, sizeof(int), 6);
+    } else {
+      compute_encoder->setBytes(upd.shape().data(), upd_ndim * sizeof(int), 6);
+    }
+    compute_encoder->setBytes(&upd_ndim, sizeof(size_t), 7);
+    compute_encoder->setBytes(&upd_size, sizeof(size_t), 8);
 
     // Set index buffers
-    for (int i = 1; i < nidx + 1; ++i) {
-      compute_encoder.set_input_array(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 +377,14 @@ 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) {
-      compute_encoder.set_input_array(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/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/gemv_masked.h

@@ -0,0 +1,25 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view gemv_masked_kernel = R"(
+template [[host_name("{name}")]] [[kernel]] void
+gemv_{trans}masked<{itype}, {outm_t}, {opm_t}, {bm}, {bn}, {sm}, {sn}, {tm}, {tn}, {nc}>(
+    const device {itype}* mat [[buffer(0)]],
+    const device {itype}* in_vec [[buffer(1)]],
+    device {itype}* out_vec [[buffer(3)]],
+    const constant int& in_vec_size [[buffer(4)]],
+    const constant int& out_vec_size [[buffer(5)]],
+    const constant int& marix_ld [[buffer(6)]],
+    const constant int& batch_ndim [[buffer(9)]],
+    const constant int* batch_shape [[buffer(10)]],
+    const constant size_t* vector_batch_stride [[buffer(11)]],
+    const constant size_t* matrix_batch_stride [[buffer(12)]],
+    const device {outm_t}* out_mask [[buffer(20)]],
+    const device {opm_t}* mat_mask [[buffer(21)]],
+    const device {opm_t}* vec_mask [[buffer(22)]],
+    const constant int* mask_strides [[buffer(23)]],
+    const constant size_t* mask_batch_strides [[buffer(24)]],
+    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/includes.h

@@ -0,0 +1,38 @@
+// 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* fft();
+const char* hadamard();
+const char* quantized();
+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();
+const char* gemv_masked();
+
+} // namespace mlx::core::metal

mlx/backend/metal/jit/indexing.h

@@ -0,0 +1,93 @@
+// 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& out_ndim [[buffer(5)]],
+    const constant int* upd_shape [[buffer(6)]],
+    const constant size_t& upd_ndim [[buffer(7)]],
+    const constant size_t& upd_size [[buffer(8)]],
+    {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,
+      out_ndim,
+      upd_shape,
+      upd_ndim,
+      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/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_kernels.cpp

@@ -0,0 +1,642 @@
+// Copyright © 2024 Apple Inc.
+#include <map>
+
+#include "mlx/backend/common/compiled.h"
+#include "mlx/backend/metal/jit/arange.h"
+#include "mlx/backend/metal/jit/copy.h"
+#include "mlx/backend/metal/jit/gemv_masked.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/steel_conv.h"
+#include "mlx/backend/metal/jit/steel_gemm.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,
+    Dtype out_type,
+    const std::string op) {
+  std::string lib_name = kernel_name.substr(1);
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    auto u_def = get_template_definition(
+        "v" + lib_name, "unary_v", get_type_string(out_type), op);
+    auto u2_def = get_template_definition(
+        "v2" + lib_name, "unary_v2", get_type_string(out_type), op);
+    auto g_def = get_template_definition(
+        "g" + lib_name, "unary_g", get_type_string(out_type), op);
+    kernel_source << metal::utils() << metal::unary_ops() << metal::unary()
+                  << u_def << u2_def << g_def;
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+void add_binary_kernels(
+    const std::string lib_name,
+    Dtype in_type,
+    Dtype out_type,
+    const std::string op,
+    std::ostringstream& kernel_source) {
+  const std::map<std::string, std::string> kernel_types = {
+      {"ss", "binary_ss"},
+      {"vs", "binary_vs"},
+      {"sv", "binary_sv"},
+      {"vv", "binary_vv"},
+      {"vs2", "binary_vs2"},
+      {"sv2", "binary_sv2"},
+      {"vv2", "binary_vv2"},
+      {"g1", "binary_g_nd1"},
+      {"g2", "binary_g_nd2"},
+      {"g3", "binary_g_nd3"},
+      {"g4", "binary_g_nd"},
+      {"g5", "binary_g_nd"},
+      {"gn", "binary_g"},
+  };
+  for (auto [name, func] : kernel_types) {
+    std::string template_def;
+    if (name == "g4" || name == "g5") {
+      int dim = std::stoi(name.substr(1));
+      template_def = get_template_definition(
+          name + lib_name,
+          func,
+          get_type_string(in_type),
+          get_type_string(out_type),
+          op,
+          dim);
+    } else {
+      template_def = get_template_definition(
+          name + lib_name,
+          func,
+          get_type_string(in_type),
+          get_type_string(out_type),
+          op);
+    }
+    kernel_source << template_def;
+  }
+}
+
+MTL::ComputePipelineState* get_binary_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    Dtype in_type,
+    Dtype out_type,
+    const std::string op) {
+  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();
+    add_binary_kernels(lib_name, in_type, out_type, op, kernel_source);
+    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,
+    Dtype in_type,
+    Dtype out_type,
+    const std::string op) {
+  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();
+    add_binary_kernels(lib_name, in_type, out_type, op, kernel_source);
+    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,
+    Dtype type,
+    const std::string op) {
+  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;
+    const std::map<std::string, std::string> kernel_types = {
+        {"v", "ternary_v"},
+        {"v2", "ternary_v2"},
+        {"g", "ternary_g"},
+        {"g1", "ternary_g_nd1"},
+        {"g2", "ternary_g_nd2"},
+        {"g3", "ternary_g_nd3"},
+        {"g4", "ternary_g_nd"},
+        {"g5", "ternary_g_nd"},
+    };
+    kernel_source << metal::utils() << metal::ternary_ops() << metal::ternary();
+    for (auto [name, func] : kernel_types) {
+      std::string template_def;
+      if (name == "g4" || name == "g5") {
+        int dim = std::stoi(name.substr(1));
+        template_def = get_template_definition(
+            name + "_" + lib_name, func, get_type_string(type), op, dim);
+      } else {
+        template_def = get_template_definition(
+            name + "_" + lib_name, func, get_type_string(type), op);
+      }
+      kernel_source << template_def;
+    }
+    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 std::string& reduce_type,
+    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::string op_name = "Cum" + reduce_type;
+    op_name[3] = toupper(op_name[3]);
+    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,
+                         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;
+    auto in_type = get_type_string(in.dtype());
+    auto out_type = get_type_string(out.dtype());
+    kernel_source << metal::utils() << metal::sort();
+    for (bool is_argsort : {true, false}) {
+      std::string bool_string = is_argsort ? "true" : "false";
+      std::string func_string = is_argsort ? "carg_" : "c_";
+      kernel_source << get_template_definition(
+          func_string + lib_name,
+          "block_sort",
+          in_type,
+          out_type,
+          bool_string,
+          bn,
+          tn);
+      kernel_source << get_template_definition(
+          "n" + func_string + lib_name,
+          "block_sort_nc",
+          in_type,
+          out_type,
+          bool_string,
+          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();
+    std::vector<std::pair<std::string, std::string>> kernel_types = {
+        {"sort_", "mb_block_sort"},
+        {"partition_", "mb_block_partition"},
+        {"merge_", "mb_block_merge"}};
+    for (auto [name, func] : kernel_types) {
+      kernel_source << get_template_definition(
+          name + lib_name,
+          func,
+          get_type_string(in.dtype()),
+          get_type_string(idx.dtype()),
+          "true",
+          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 std::string& op_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::string op_type = op_name;
+    op_type[0] = std::toupper(op_name[0]);
+    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_type);
+    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_gemv_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_mat,
+    int bm,
+    int bn,
+    int sm,
+    int sn,
+    int tm,
+    int tn,
+    bool contiguous) {
+  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::gemv_masked()
+                  << fmt::format(
+                         gemv_masked_kernel,
+                         "name"_a = lib_name,
+                         "itype"_a = get_type_string(out.dtype()),
+                         "outm_t"_a = out_mask_type,
+                         "opm_t"_a = op_mask_type,
+                         "bm"_a = bm,
+                         "bn"_a = bn,
+                         "sm"_a = sm,
+                         "sn"_a = sn,
+                         "tm"_a = tm,
+                         "tn"_a = tn,
+                         "trans"_a = transpose_mat ? "t_" : "",
+                         "nc"_a = contiguous ? "0" : "1");
+    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);
+}
+
+MTL::ComputePipelineState* get_fft_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const std::string& hash_name,
+    const metal::MTLFCList& func_consts,
+    const std::string& template_def) {
+  const auto& lib_name = kernel_name;
+  auto lib = d.get_library(lib_name);
+  if (lib == nullptr) {
+    std::ostringstream kernel_source;
+    std::string kernel_string;
+    kernel_source << metal::fft() << template_def;
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib, hash_name, func_consts);
+}
+
+MTL::ComputePipelineState* get_quantized_kernel(
+    metal::Device& d,
+    const std::string& kernel_name,
+    const std::string& template_def) {
+  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::quantized()
+                  << template_def;
+    lib = d.get_library(lib_name, kernel_source.str());
+  }
+  return d.get_kernel(kernel_name, lib);
+}
+
+} // namespace mlx::core