Move to nanobind v2 (#1316)

* add cholesky inv + tri inv

* always run tri_inv on cpu

* consistent naming

.circleci/config.yml

@@ -31,7 +31,7 @@ jobs:
           name: Install dependencies
           command: |
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install numpy
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
@@ -49,11 +49,6 @@ jobs:
           name: Run Python tests
           command: |
             python3 -m unittest discover python/tests -v
-      # TODO: Reenable when extension api becomes stable
-      # - run:
-      #     name: Build example extension
-      #     command: |
-      #       cd examples/extensions && python3 -m pip install . 
       - run:
           name: Build CPP only
           command: |
@@ -69,18 +64,19 @@ jobs:
         default: "15.2.0"
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
             brew install python@3.8
+            brew install openmpi
             python3.8 -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install numpy
             pip install torch
             pip install tensorflow
@@ -101,11 +97,14 @@ jobs:
             source env/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
-      # TODO: Reenable when extension api becomes stable
+            mpirun -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-      # - run:
+      - run:
-      #     name: Build example extension
+          name: Build example extension
-      #     command: |
+          command: |
-      #       cd examples/extensions && python3.11 -m pip install . 
+            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:
@@ -132,18 +137,19 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: macos.m1.large.gen1
+    resource_class: macos.m1.medium.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
             brew install python@<< parameters.python_version >>
+            brew install openmpi
             python<< parameters.python_version >> -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install --upgrade setuptools
             pip install numpy
             pip install twine
@@ -207,7 +213,7 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
+            pip install git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4
             pip install --upgrade setuptools
             pip install numpy
             pip install auditwheel

.pre-commit-config.yaml

@@ -1,11 +1,11 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v18.1.3
+    rev: v18.1.4
     hooks:
     -   id: clang-format
 # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 24.3.0
+    rev: 24.4.2
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort

ACKNOWLEDGMENTS.md

@@ -7,15 +7,17 @@ with a short description of your contribution(s) below. For example:
 
 MLX was developed with contributions from the following individuals:
 
-- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops.
+- Nripesh Niketan: Added `softsign`, `softmax`, `hardswish`, `logsoftmax` activation functions. Added `dropout3d` ops. Added `LogicalAnd` and `LogicalOR` ops. Added `clip_grad_norm` along with `tree_reduce`.
 - Juarez Bochi: Fixed bug in cross attention.
 - Justin Deschenaux: Sine, Cosine, arange, randint, truncated normal, bernoulli, lion optimizer, Dropout2d, linear and logistic regression python example.
-- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream` and safetensor support.
+- Diogo Da Cruz: Added `tri`, `tril`, `triu`, `tensordot`, `inner`, `outer`, `tile`, `StreamContext`, `stream`, safetensors support, `einsum`, and `einsum_path`.
 - Gabrijel Boduljak: Added `mlx.core.linalg`, implemented `norm` method and `InstanceNorm` layer. Implemented pooling layers and ``Upsample``.
 - Hinrik Snær Guðmundsson: Added `atleast_1d`, `atleast_2d`, `atleast_3d` ops.
 - Luca Arnaboldi: Added `Ceil` and `Floor` ops; implemented pickling, copy and deepcopy for mlx arrays.
 - Brian Keene & Atila Orhon, with Argmax Inc.: Added `fast.scaled_dot_product_attention`
 - AmirHossein Razlighi: Added chaining support for some of the ops in `nn.Module`. 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 0.10.0)
+  set(MLX_VERSION 0.16.2)
 endif()
 
 # --------------------- Processor tests -------------------------
@@ -79,15 +83,17 @@ elseif (MLX_BUILD_METAL)
                   OUTPUT_VARIABLE MACOS_VERSION
                   COMMAND_ERROR_IS_FATAL ANY)
 
-  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
+  if (${MACOS_VERSION} LESS 14.0)
-
-  if (${MACOS_VERSION} GREATER_EQUAL 14.2)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14.2_iOS17.2.zip)
-  elseif (${MACOS_VERSION} GREATER_EQUAL 14.0)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14_iOS17-beta.zip)
-  else()
     message(FATAL_ERROR "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON" )
   endif()
+  message(STATUS "Building with SDK for macOS version ${MACOS_VERSION}")
+
+  set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18-beta.zip)
+  # Get the metal version
+  execute_process(
+    COMMAND zsh "-c" "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal -E -x metal -P - | tail -1 | tr -d '\n'"
+    OUTPUT_VARIABLE MLX_METAL_VERSION
+    COMMAND_ERROR_IS_FATAL ANY)
 
   FetchContent_Declare(
     metal_cpp
@@ -101,17 +107,20 @@ 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()
 
+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 ${ACCELERATE_LIBRARY})
+    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.")
@@ -134,7 +143,7 @@ else()
     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})
+    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)
@@ -149,7 +158,34 @@ else()
     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})
+    target_link_libraries(mlx PUBLIC ${BLAS_LIBRARIES})
+  endif()
+else()
+  set(MLX_BUILD_ACCELERATE OFF)
+endif()
+
+find_package(MPI)
+if (MPI_FOUND)
+  execute_process(
+    COMMAND zsh "-c" "mpirun --version"
+    OUTPUT_VARIABLE MPI_VERSION
+    ERROR_QUIET
+  )
+  if (${MPI_VERSION} MATCHES ".*Open MPI.*")
+    target_include_directories(mlx PRIVATE ${MPI_INCLUDE_PATH})
+  elseif (MPI_VERSION STREQUAL "")
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING
+      "MPI found but mpirun is not available. Building without MPI."
+    )
+  else()
+    set(MPI_FOUND FALSE)
+    message(
+      WARNING
+      "MPI which is not OpenMPI found. Building without MPI."
+    )
+  endif() 
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
@@ -161,6 +197,14 @@ target_include_directories(
   $<INSTALL_INTERFACE:include>
 )
 
+FetchContent_Declare(fmt
+  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+  GIT_TAG 10.2.1 
+  EXCLUDE_FROM_ALL
+)
+FetchContent_MakeAvailable(fmt)
+target_link_libraries(mlx PRIVATE fmt::fmt-header-only)
+
 if (MLX_BUILD_PYTHON_BINDINGS)
   message(STATUS "Building Python bindings.")
   find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)

README.md

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

benchmarks/python/comparative/bench_torch.py

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

benchmarks/python/comparative/compare.py

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

benchmarks/python/compile_bench.py

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

benchmarks/python/conv1d_bench.py

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

benchmarks/python/conv_bench.py

@@ -28,11 +28,11 @@ def bench(f, a, b):
     return (e - s) * 1e-9
 
 
-def make_mx_conv_2D(strides=(1, 1), padding=(0, 0)):
+def make_mx_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
     def mx_conv_2D(a, b):
         ys = []
         for i in range(N_iter_func):
-            y = mx.conv2d(a, b, stride=strides, padding=padding)
+            y = mx.conv2d(a, b, stride=strides, padding=padding, groups=groups)
             ys.append(y)
         mx.eval(ys)
         return ys
@@ -40,12 +40,12 @@ def make_mx_conv_2D(strides=(1, 1), padding=(0, 0)):
     return mx_conv_2D
 
 
-def make_pt_conv_2D(strides=(1, 1), padding=(0, 0)):
+def make_pt_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
     @torch.no_grad()
     def pt_conv_2D(a, b):
         ys = []
         for i in range(N_iter_func):
-            y = torch.conv2d(a, b, stride=strides, padding=padding)
+            y = torch.conv2d(a, b, stride=strides, padding=padding, groups=groups)
             ys.append(y)
         torch.mps.synchronize()
         return ys
@@ -53,11 +53,12 @@ def make_pt_conv_2D(strides=(1, 1), padding=(0, 0)):
     return pt_conv_2D
 
 
-def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
+def bench_shape(N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype):
-
     scale = 1.0 / math.sqrt(kH * kH * C)
     a_np = np.random.uniform(0, 0.5, (N, H, W, C)).astype(np_dtype)
-    b_np = np.random.uniform(-scale, scale, (O, kH, kW, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kH, kW, int(C / groups))).astype(
+        np_dtype
+    )
 
     a_mx = mx.array(a_np)
     b_mx = mx.array(b_np)
@@ -67,15 +68,15 @@ def bench_shape(N, H, W, C, kH, kW, O, strides, padding, np_dtype):
 
     torch.mps.synchronize()
 
-    f_mx = make_mx_conv_2D(strides, padding)
+    f_mx = make_mx_conv_2D(strides, padding, groups)
-    f_pt = make_pt_conv_2D(strides, padding)
+    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, 32, 5, 5, 32, (1, 1), (2, 2), 1),
-        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2)),
+        (4, 32, 32, 64, 5, 5, 64, (1, 1), (2, 2), 1),
-        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2)),
+        (4, 32, 32, 128, 5, 5, 128, (1, 1), (2, 2), 1),
-        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2)),
+        (4, 32, 32, 256, 5, 5, 256, (1, 1), (2, 2), 1),
-        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2)),
+        (4, 32, 32, 512, 5, 5, 512, (1, 1), (2, 2), 1),
-        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2)),
+        (4, 64, 64, 32, 5, 5, 32, (1, 1), (2, 2), 1),
-        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2)),
+        (4, 64, 64, 64, 5, 5, 64, (1, 1), (2, 2), 1),
-        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2)),
+        (4, 64, 64, 128, 5, 5, 128, (1, 1), (2, 2), 1),
-        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2)),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 1),
-        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2)),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 2),
-        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2)),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 16),
-        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2)),
+        (4, 64, 64, 256, 5, 5, 256, (1, 1), (2, 2), 64),
-        (4, 256, 256, 32, 5, 5, 3, (1, 1), (2, 2)),
+        (4, 128, 128, 32, 5, 5, 32, (1, 1), (2, 2), 1),
-        (4, 256, 256, 3, 5, 5, 32, (1, 1), (2, 2)),
+        (4, 128, 128, 64, 5, 5, 64, (1, 1), (2, 2), 1),
-        (4, 128, 128, 64, 5, 5, 3, (1, 1), (2, 2)),
+        (4, 128, 128, 128, 5, 5, 128, (1, 1), (2, 2), 1),
-        (4, 128, 128, 3, 5, 5, 64, (1, 1), (2, 2)),
+        (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%")
+        print(
-        for N, H, W, C, kH, kW, O, strides, padding in shapes:
+            "(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

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

benchmarks/python/hadamard_bench.py

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

benchmarks/python/sdpa_bench.py

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

docs/Doxyfile

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

docs/README.md

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

docs/requirements.txt

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

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

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

docs/src/conf.py

@@ -22,6 +22,7 @@ extensions = [
     "sphinx.ext.autosummary",
     "sphinx.ext.intersphinx",
     "sphinx.ext.napoleon",
+    "breathe",
 ]
 
 python_use_unqualified_type_names = True
@@ -33,6 +34,9 @@ intersphinx_mapping = {
     "numpy": ("https://numpy.org/doc/stable/", None),
 }
 
+breathe_projects = {"mlx": "../build/xml"}
+breathe_default_project = "mlx"
+
 templates_path = ["_templates"]
 html_static_path = ["_static"]
 source_suffix = ".rst"
@@ -79,3 +83,15 @@ def setup(app):
 # -- Options for LaTeX output ------------------------------------------------
 
 latex_documents = [(main_doc, "MLX.tex", "MLX Documentation", author, "manual")]
+latex_elements = {
+    "preamble": r"""
+    \usepackage{enumitem}
+    \setlistdepth{5}
+    \setlist[itemize,1]{label=$\bullet$}
+    \setlist[itemize,2]{label=$\bullet$}
+    \setlist[itemize,3]{label=$\bullet$}
+    \setlist[itemize,4]{label=$\bullet$}
+    \setlist[itemize,5]{label=$\bullet$}
+    \renewlist{itemize}{itemize}{5}
+""",
+}

docs/src/cpp/ops.rst

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

docs/src/dev/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
+        // Make sure the metal library is available
-        // in the same folder as this executable if needed
+        d.register_library("mlx_ext");
-        d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
 
         // Make a kernel from this metal library
         auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
         // Prepare to encode kernel
-        auto compute_encoder = d.get_command_encoder(s.index);
+        auto& compute_encoder = d.get_command_encoder(s.index);
         compute_encoder->setComputePipelineState(kernel);
 
         // 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);
+        compute_encoder.set_input_array(x, 0);
-        set_array_buffer(compute_encoder, y, 1);
+        compute_encoder.set_input_array(y, 1);
 
         // Encode output arrays to kernel
-        set_array_buffer(compute_encoder, out, 2);
+        compute_encoder.set_output_array(out, 2);
 
         // Encode alpha and beta
         compute_encoder->setBytes(&alpha_, sizeof(float), 3);
@@ -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/dev/metal_debugger.rst

@@ -32,10 +32,9 @@ work.
 
     trace_file = "mlx_trace.gputrace"
 
-    if not mx.metal.start_capture(trace_file):
+    # Make sure to run with MTL_CAPTURE_ENABLED=1 and
-      print("Make sure to run with MTL_CAPTURE_ENABLED=1 and "
+    # that the path trace_file does not already exist.
-            f"that the path {trace_file} does not already exist.")
+    mx.metal.start_capture(trace_file)
-      exit(1)
 
     for _ in range(10):
       mx.eval(mx.add(a, b))

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
-
 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/devices_and_streams.rst

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

docs/src/python/distributed.rst

@@ -0,0 +1,19 @@
+.. _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/metal.rst

@@ -7,10 +7,13 @@ Metal
   :toctree: _autosummary
 
   is_available
+  device_info
   get_active_memory
   get_peak_memory
+  reset_peak_memory
   get_cache_memory
   set_memory_limit
   set_cache_limit
+  clear_cache
   start_capture
   stop_capture

docs/src/python/nn.rst

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

docs/src/python/nn/functions.rst

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

docs/src/python/nn/layers.rst

@@ -15,15 +15,21 @@ Layers
    BatchNorm
    Conv1d
    Conv2d
+   Conv3d
    Dropout
    Dropout2d
    Dropout3d
    Embedding
    GELU
+   GLU
    GroupNorm
    GRU
+   HardShrink
+   HardTanh
+   Hardswish
    InstanceNorm
    LayerNorm
+   LeakyReLU
    Linear
    LSTM
    MaxPool1d
@@ -31,16 +37,23 @@ Layers
    Mish
    MultiHeadAttention
    PReLU
+   QuantizedEmbedding
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU6
    RNN
    RoPE
    SELU
    Sequential
    SiLU
    SinusoidalPositionalEncoding
+   Softmin
    Softshrink
+   Softsign
+   Softmax
+   Softplus
    Step
+   Tanh
    Transformer
    Upsample

docs/src/python/ops.rst

@@ -10,6 +10,7 @@ Operations
 
    abs
    add
+   addmm
    all
    allclose
    any
@@ -19,19 +20,27 @@ Operations
    arcsin
    arcsinh
    arctan
+   arctan2
    arctanh
    argmax
    argmin
    argpartition
    argsort
    array_equal
+   as_strided
    atleast_1d
    atleast_2d
    atleast_3d
+   bitwise_and
+   bitwise_or
+   bitwise_xor
+   block_masked_mm
    broadcast_to
    ceil
    clip
    concatenate
+   conj
+   conjugate
    convolve
    conv1d
    conv2d
@@ -42,11 +51,14 @@ Operations
    cummin
    cumprod
    cumsum
+   degrees
    dequantize
    diag
    diagonal
    divide
    divmod
+   einsum
+   einsum_path
    equal
    erf
    erfinv
@@ -58,8 +70,11 @@ Operations
    floor
    floor_divide
    full
+   gather_mm
+   gather_qmm
    greater
    greater_equal
+   hadamard_transform
    identity
    inner
    isclose
@@ -67,6 +82,8 @@ Operations
    isnan
    isneginf
    isposinf
+   issubdtype
+   left_shift
    less
    less_equal
    linspace
@@ -89,18 +106,24 @@ Operations
    minimum
    moveaxis
    multiply
+   nan_to_num
    negative
+   not_equal
    ones
    ones_like
    outer
    partition
    pad
+   power
    prod
    quantize
    quantized_matmul
+   radians
    reciprocal
+   remainder
    repeat
    reshape
+   right_shift
    round
    rsqrt
    save
@@ -131,11 +154,13 @@ Operations
    tensordot
    tile
    topk
+   trace
    transpose
    tri
    tril
    triu
    var
+   view
    where
    zeros
    zeros_like

docs/src/python/optimizers.rst

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

docs/src/python/random.rst

@@ -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/python/tree_utils.rst

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

docs/src/usage/distributed.rst

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

docs/src/usage/lazy_evaluation.rst

@@ -18,7 +18,7 @@ describe below.
 Transforming Compute Graphs
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Lazy evaluation let's us record a compute graph without actually doing any
+Lazy evaluation lets us record a compute graph without actually doing any
 computations. This is useful for function transformations like :func:`grad` and
 :func:`vmap` and graph optimizations.
 

docs/src/usage/numpy.rst

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

examples/cpp/CMakeLists.txt

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

examples/cpp/distributed.cpp

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

examples/cpp/metal_capture.cpp

@@ -11,7 +11,7 @@ int main() {
   // To use Metal debugging and profiling:
   // 1. Build with the MLX_METAL_DEBUG CMake option (i.e. -DMLX_METAL_DEBUG=ON).
   // 2. Run with MTL_CAPTURE_ENABLED=1.
-  assert(metal::start_capture("mlx_trace.gputrace"));
+  metal::start_capture("mlx_trace.gputrace");
 
   // Start at index two because the default GPU and CPU streams have indices
   // zero and one, respectively. This naming matches the label assigned to each

examples/cpp/tutorial.cpp

@@ -89,8 +89,8 @@ void automatic_differentiation() {
   // dfdx is 2 * x
 
   // Get the second derivative by composing grad with grad
-  auto df2dx2 = grad(grad(fn))(x);
+  auto d2fdx2 = grad(grad(fn))(x);
-  // df2dx2 is 2
+  // 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
+  // Make sure the metal library is available
-  // in the same folder as this executable if needed
+  d.register_library("mlx_ext");
-  d.register_library("mlx_ext", metal::get_colocated_mtllib_path);
 
   // Make a kernel from this metal library
   auto kernel = d.get_kernel(kname.str(), "mlx_ext");
 
   // Prepare to encode kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
 
   // 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);
+  compute_encoder.set_input_array(x, 0);
-  set_array_buffer(compute_encoder, y, 1);
+  compute_encoder.set_input_array(y, 1);
 
   // Encode output arrays to kernel
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode alpha and beta
   compute_encoder->setBytes(&alpha_, sizeof(float), 3);
@@ -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/axpby/axpby.metal

@@ -36,8 +36,8 @@ template <typename T>
 }
 
 #define instantiate_axpby(type_name, type)                               \
-  template [[host_name("axpby_general_" #type_name)]] \
+  template [[host_name("axpby_general_" #type_name)]] [[kernel]] void    \
-  [[kernel]] void axpby_general<type>(                \
+  axpby_general<type>(                                                   \
       device const type* x [[buffer(0)]],                                \
       device const type* y [[buffer(1)]],                                \
       device type* out [[buffer(2)]],                                    \
@@ -48,8 +48,8 @@ template <typename T>
       constant const size_t* y_strides [[buffer(7)]],                    \
       constant const int& ndim [[buffer(8)]],                            \
       uint index [[thread_position_in_grid]]);                           \
-  template [[host_name("axpby_contiguous_" #type_name)]] \
+  template [[host_name("axpby_contiguous_" #type_name)]] [[kernel]] void \
-  [[kernel]] void axpby_contiguous<type>(                \
+  axpby_contiguous<type>(                                                \
       device const type* x [[buffer(0)]],                                \
       device const type* y [[buffer(1)]],                                \
       device type* out [[buffer(2)]],                                    \

examples/extensions/mlx_sample_extensions/__init__.py

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

examples/extensions/pyproject.toml

@@ -3,6 +3,6 @@ requires = [
   "setuptools>=42",
   "cmake>=3.24",
   "mlx>=0.9.0",
-  "nanobind@git+https://github.com/wjakob/nanobind.git#egg=4148debcf91f5ccab0c3b8d67b5c3cabd61f407f",
+  "nanobind@git+https://github.com/wjakob/nanobind.git@2f04eac452a6d9142dedb957701bdb20125561e4",
 ]
 build-backend = "setuptools.build_meta"

examples/extensions/requirements.txt

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

@@ -1,5 +1,4 @@
 // Copyright © 2023-2024 Apple Inc.
-
 #include <functional>
 
 #include "mlx/array.h"
@@ -18,6 +17,10 @@ bool in_tracing() {
   return detail::InTracing::in_tracing();
 }
 
+bool retain_graph() {
+  return detail::RetainGraph::retain_graph();
+}
+
 } // namespace
 
 array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
@@ -93,13 +96,17 @@ void array::detach() {
 }
 
 void array::eval() {
-  if (!is_evaled()) {
+  // Ensure the array is ready to be read
+  if (status() == Status::scheduled) {
+    event().wait();
+    set_status(Status::available);
+  } else if (status() == Status::unscheduled) {
     mlx::core::eval({*this});
   }
 }
 
 bool array::is_tracer() const {
-  return array_desc_->is_tracer && in_tracing();
+  return array_desc_->is_tracer && in_tracing() || retain_graph();
 }
 
 void array::set_data(allocator::Buffer buffer, deleter_t d) {
@@ -163,6 +170,40 @@ void array::move_shared_buffer(array other) {
   move_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
+array::~array() {
+  if (array_desc_ == nullptr) {
+    return;
+  }
+
+  // Ignore arrays that might be detached during eval
+  if (status() == array::Status::scheduled) {
+    return;
+  }
+
+  // Break circular reference for non-detached arrays with siblings
+  if (auto n = siblings().size(); n > 0) {
+    bool do_detach = true;
+    // If all siblings have siblings.size() references except
+    // the one we are currently destroying (which has siblings.size() + 1)
+    // then there are no more external references
+    do_detach &= (array_desc_.use_count() == (n + 1));
+    for (auto& s : siblings()) {
+      do_detach &= (s.array_desc_.use_count() == n);
+      if (!do_detach) {
+        break;
+      }
+    }
+    if (do_detach) {
+      for (auto& s : siblings()) {
+        for (auto& ss : s.siblings()) {
+          ss.array_desc_ = nullptr;
+        }
+        s.array_desc_->siblings.clear();
+      }
+    }
+  }
+}
+
 void array::ArrayDesc::init() {
   strides.resize(shape.size());
   size = 1;
@@ -170,13 +211,13 @@ void array::ArrayDesc::init() {
     strides[i] = size;
     size *= shape[i];
   }
-  for (auto& in : inputs) {
+  for (const auto& in : inputs) {
     is_tracer |= in.is_tracer();
   }
 }
 
 array::ArrayDesc::ArrayDesc(std::vector<int> shape, Dtype dtype)
-    : shape(std::move(shape)), dtype(dtype) {
+    : shape(std::move(shape)), dtype(dtype), status(Status::available) {
   init();
 }
 
@@ -187,6 +228,7 @@ array::ArrayDesc::ArrayDesc(
     std::vector<array> inputs)
     : shape(std::move(shape)),
       dtype(dtype),
+      status(Status::unscheduled),
       primitive(std::move(primitive)),
       inputs(std::move(inputs)) {
   init();
@@ -194,7 +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

@@ -9,6 +9,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/dtype.h"
+#include "mlx/event.h"
 
 namespace mlx::core {
 
@@ -72,32 +73,32 @@ class array {
       this->array_desc_ = other.array_desc_;
     }
     return *this;
-  };
+  }
 
   /** The size of the array's datatype in bytes. */
   size_t itemsize() const {
     return size_of(dtype());
-  };
+  }
 
   /** The number of elements in the array. */
   size_t size() const {
     return array_desc_->size;
-  };
+  }
 
   /** The number of bytes in the array. */
   size_t nbytes() const {
     return size() * itemsize();
-  };
+  }
 
   /** The number of dimensions of the array. */
   size_t ndim() const {
     return array_desc_->shape.size();
-  };
+  }
 
   /** The shape of the array as a vector of integers. */
   const std::vector<int>& shape() const {
     return array_desc_->shape;
-  };
+  }
 
   /**
    *  Get the size of the corresponding dimension.
@@ -106,17 +107,26 @@ class array {
    *  bounds checking. */
   int shape(int dim) const {
     return shape().at(dim < 0 ? dim + ndim() : dim);
-  };
+  }
 
   /** The strides of the array. */
   const std::vector<size_t>& strides() const {
     return array_desc_->strides;
-  };
+  }
+
+  /**
+   *  Get the stride of the corresponding dimension.
+   *
+   *  This function supports negative indexing and provides
+   *  bounds checking. */
+  size_t strides(int dim) const {
+    return strides().at(dim < 0 ? dim + ndim() : dim);
+  }
 
   /** Get the arrays data type. */
   Dtype dtype() const {
     return array_desc_->dtype;
-  };
+  }
 
   /** Evaluate the array. */
   void eval();
@@ -150,10 +160,10 @@ class array {
 
     friend bool operator==(const ArrayIterator& a, const ArrayIterator& b) {
       return a.arr.id() == b.arr.id() && a.idx == b.idx;
-    };
+    }
     friend bool operator!=(const ArrayIterator& a, const ArrayIterator& b) {
       return !(a == b);
-    };
+    }
 
    private:
     const array& arr;
@@ -199,7 +209,7 @@ class array {
     allocator::Buffer buffer;
     deleter_t d;
     Data(allocator::Buffer buffer, deleter_t d = allocator::free)
-        : buffer(buffer), d(d){};
+        : buffer(buffer), d(d) {}
     // Not copyable
     Data(const Data& d) = delete;
     Data& operator=(const Data& d) = delete;
@@ -220,22 +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;
@@ -249,24 +259,18 @@ class array {
   /** The array's siblings. */
   const std::vector<array>& siblings() const {
     return array_desc_->siblings;
-  };
+  }
+
+  /** The array's siblings. */
+  std::vector<array>& siblings() {
+    return array_desc_->siblings;
+  }
 
   void set_siblings(std::vector<array> siblings, uint16_t position) {
     array_desc_->siblings = std::move(siblings);
     array_desc_->position = position;
   }
 
-  /** The i-th output of the array's primitive. */
-  const array& output(int i) const {
-    if (i == array_desc_->position) {
-      return *this;
-    } else if (i < array_desc_->position) {
-      return siblings()[i];
-    } else {
-      return siblings()[i + 1];
-    }
-  };
-
   /** The outputs of the array's primitive (i.e. this array and
    * its siblings) in the order the primitive expects. */
   std::vector<array> outputs() const {
@@ -277,7 +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();
@@ -285,19 +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
@@ -308,16 +312,35 @@ class array {
   template <typename T>
   T* data() {
     return static_cast<T*>(array_desc_->data_ptr);
-  };
+  }
 
   template <typename T>
   const T* data() const {
     return static_cast<T*>(array_desc_->data_ptr);
-  };
+  }
 
-  // Check if the array has been evaluated
+  enum Status { unscheduled, scheduled, available };
-  bool is_evaled() const {
+
-    return array_desc_->data != nullptr;
+  bool is_available() const {
+    return status() == Status::available;
+  }
+
+  Status status() const {
+    return array_desc_->status;
+  }
+
+  void set_status(Status s) const {
+    array_desc_->status = s;
+  }
+
+  // Get the array's shared event
+  Event& event() const {
+    return array_desc_->event;
+  }
+
+  // Attach an event to a not yet evaluated array
+  void attach_event(Event e) const {
+    array_desc_->event = std::move(e);
   }
 
   // Mark the array as a tracer array (true) or not.
@@ -358,6 +381,8 @@ class array {
     array_desc_ = other.array_desc_;
   }
 
+  ~array();
+
  private:
   // Initialize the arrays data
   template <typename It>
@@ -370,6 +395,11 @@ class array {
     Dtype dtype;
     std::shared_ptr<Primitive> primitive;
 
+    Status status;
+
+    // An event on the array used for synchronization
+    Event event;
+
     // Indicates an array is being used in a graph transform
     // and should not be detached from the graph
     bool is_tracer{false};
@@ -470,10 +500,11 @@ T array::item() const {
   if (size() != 1) {
     throw std::invalid_argument("item can only be called on arrays of size 1.");
   }
-  if (!is_evaled()) {
+  if (status() == Status::unscheduled) {
     throw std::invalid_argument(
         "item() const can only be called on evaled arrays");
   }
+  const_cast<array*>(this)->eval();
   return *data<T>();
 }
 

mlx/backend/accelerate/conv.cpp

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

mlx/backend/accelerate/matmul.cpp

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

mlx/backend/accelerate/primitives.cpp

@@ -3,8 +3,7 @@
 #include <cassert>
 #include <cmath>
 
-#include <vecLib/vDSP.h>
+#include <Accelerate/Accelerate.h>
-#include <vecLib/vForce.h>
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/binary.h"
@@ -31,11 +30,13 @@ DEFAULT(ArgPartition)
 DEFAULT(ArgReduce)
 DEFAULT(ArgSort)
 DEFAULT(AsStrided)
+DEFAULT(BlockMaskedMM)
 DEFAULT(Broadcast)
 DEFAULT(Ceil)
 DEFAULT(Concatenate)
+DEFAULT(Conjugate)
 DEFAULT(Copy)
-DEFAULT_MULTI(CustomVJP)
+DEFAULT_MULTI(CustomTransforms)
 DEFAULT_MULTI(Depends)
 DEFAULT_MULTI(DivMod)
 DEFAULT(NumberOfElements)
@@ -45,8 +46,11 @@ DEFAULT(ErfInv)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Gather)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
 DEFAULT(Less)
 DEFAULT(LessEqual)
 DEFAULT(Load)
@@ -76,6 +80,7 @@ DEFAULT(StopGradient)
 DEFAULT_MULTI(SVD)
 DEFAULT(Transpose)
 DEFAULT(Inverse)
+DEFAULT(Cholesky)
 
 void Abs::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
@@ -97,7 +102,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -112,7 +117,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vadd((const float*)a, 1, (const float*)b, 1, (float*)o, 1, n);
         });
   } else if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -127,7 +132,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vaddi((const int*)a, 1, (const int*)b, 1, (int*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x + y; });
+    eval(inputs, out);
   }
 }
 
@@ -191,6 +196,26 @@ void ArcTan::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  if (out.dtype() == float32 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    if (a.is_donatable()) {
+      out.copy_shared_buffer(a);
+    } else if (b.is_donatable()) {
+      out.copy_shared_buffer(b);
+    } else {
+      out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    }
+    int size = a.data_size();
+    vvatan2f(out.data<float>(), a.data<float>(), b.data<float>(), &size);
+  } else {
+    eval(inputs, out);
+  }
+}
+
 void ArcTanh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
@@ -262,7 +287,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b = inputs[1];
 
   if (a.dtype() == int32) {
-    binary(
+    binary_op<int>(
         a,
         b,
         out,
@@ -275,7 +300,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdivi((const int*)b, 1, (const int*)a, 1, (int*)o, 1, n);
         });
   } else if (a.dtype() == float32) {
-    binary(
+    binary_op<float>(
         a,
         b,
         out,
@@ -290,7 +315,7 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
           vDSP_vdiv((const float*)b, 1, (const float*)a, 1, (float*)o, 1, n);
         });
   } else {
-    binary(a, b, out, [](auto x, auto y) { return x / y; });
+    eval(inputs, out);
   }
 }
 
@@ -301,12 +326,8 @@ 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(
+    eval(inputs, out);
-        "[exp] Cannot exponentiate elements in array"
-        " with non floating point type.");
   }
 }
 
@@ -368,12 +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(
+    eval(inputs, out);
-        "[log1p] Cannot compute log of elements in array with"
-        " non floating point type.");
   }
 }
 
@@ -383,7 +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,
@@ -398,7 +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);
   }
 }
 
@@ -409,7 +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);
   }
 }
 
@@ -496,7 +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);
   }
 }
 
@@ -522,7 +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,
@@ -540,7 +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,
@@ -552,7 +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 = vmulq_f16(x, vdupq_n_f16(float16_t(1.442695f))); // multiply with log_2(e)
-  x = vmaxq_f16(x, vdupq_n_f16(-14)); // clamp under with -14
+  x = vmaxq_f16(x, vdupq_n_f16(float16_t(-14.f))); // clamp under with -14
-  x = vminq_f16(x, vdupq_n_f16(14)); // clamp over 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 = vdupq_n_f16(float16_t(1.535336188319500e-4f));
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.339887440266574e-3f), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(1.339887440266574e-3f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(9.618437357674640e-3f), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(9.618437357674640e-3f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(5.550332471162809e-2f), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(5.550332471162809e-2f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(2.402264791363012e-1f), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(2.402264791363012e-1f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(6.931472028550421e-1f), x, fpart);
+  x = vfmaq_f16(vdupq_n_f16(float16_t(6.931472028550421e-1f)), x, fpart);
-  x = vfmaq_f16(vdupq_n_f16(1.000000000000000f), 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

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

mlx/backend/common/binary.cpp

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

mlx/backend/common/binary.h

@@ -1,6 +1,8 @@
 // Copyright © 2023 Apple Inc.
 
 #pragma once
+#include <cassert>
+
 #include "mlx/allocator.h"
 #include "mlx/array.h"
 #include "mlx/backend/common/utils.h"

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() &&
+      if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
-          in.is_donatable() &&
+          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

@@ -38,11 +38,15 @@ void slow_conv_1D(
 
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
   const int iH = 1 + in_dilation[0] * (in.shape(1) - 1); // Input spatial dim
+  const int C = in.shape(2); // Input channels
   const int oH = out.shape(1); // Output spatial dim
   const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(2); // In channels
   const int wH = wt.shape(1); // Weight spatial dim
 
+  const int groups = C / wt.shape(2);
+  const int C_per_group = wt.shape(2);
+  const int O_per_group = O / groups;
+
   const size_t in_stride_N = in.strides()[0];
   const size_t in_stride_H = in.strides()[1];
   const size_t in_stride_C = in.strides()[2];
@@ -57,7 +61,8 @@ void slow_conv_1D(
 
   for (int n = 0; n < N; ++n) {
     for (int oh = 0; oh < oH; ++oh) {
-      for (int o = 0; o < O; ++o) {
+      for (int g = 0; g < groups; ++g) {
+        for (int o = g * O_per_group; o < (g + 1) * O_per_group; ++o) {
           const T* filter_wt_ptr = start_wt_ptr + o * wt_stride_O;
           float r = 0.;
 
@@ -70,10 +75,10 @@ void slow_conv_1D(
             auto ih_div = std::div(ih, in_dilation[0]);
 
             if (ih >= 0 && ih < iH && ih_div.rem == 0) {
-            for (int c = 0; c < C; ++c) {
+              for (int c = g * C_per_group; c < (g + 1) * C_per_group; ++c) {
                 r += static_cast<float>(
                          in_ptr[ih_div.quot * in_stride_H + c * in_stride_C]) *
-                  static_cast<float>(wt_ptr[c * wt_stride_C]);
+                    static_cast<float>(wt_ptr[(c % C_per_group) * wt_stride_C]);
               } // c
 
             } // ih check
@@ -81,11 +86,11 @@ void slow_conv_1D(
 
           out_ptr[oh * out_stride_H + o * out_stride_O] = static_cast<T>(r);
         } // o
+      } // g
     } // oh
 
     in_ptr += in_stride_N;
     out_ptr += out_stride_N;
-
   } // n
 }
 
@@ -106,13 +111,17 @@ void slow_conv_2D(
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
   const int iH = 1 + in_dilation[0] * (in.shape(1) - 1); // Input spatial dim
   const int iW = 1 + in_dilation[1] * (in.shape(2) - 1); // Input spatial dim
+  const int C = in.shape(3); // In channels
   const int oH = out.shape(1); // Output spatial dim
   const int oW = out.shape(2); // Output spatial dim
   const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(3); // In channels
   const int wH = wt.shape(1); // Weight spatial dim
   const int wW = wt.shape(2); // Weight spatial dim
 
+  const int groups = C / wt.shape(3);
+  const int C_per_group = wt.shape(3);
+  const int O_per_group = O / groups;
+
   const size_t in_stride_N = in.strides()[0];
   const size_t in_stride_H = in.strides()[1];
   const size_t in_stride_W = in.strides()[2];
@@ -136,7 +145,8 @@ 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) {
+        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) {
@@ -146,16 +156,16 @@ void slow_conv_2D(
                 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* wt_ptr_pt =
-              const T* in_ptr_pt = in_ptr + ih * in_stride_H + iw * in_stride_W;
+                    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) {
+                for (int c = g * C_per_group; c < (g + 1) * C_per_group; ++c) {
-                r += static_cast<float>(in_ptr_pt[0]) *
+                  r += static_cast<float>(in_ptr_pt[c * in_stride_C]) *
-                    static_cast<float>(wt_ptr_pt[0]);
+                      static_cast<float>(
-                in_ptr_pt += in_stride_C;
+                           wt_ptr_pt[(c % C_per_group) * wt_stride_C]);
-                wt_ptr_pt += wt_stride_C;
                 } // c
-
               } // ww
             } // wh
 
@@ -163,6 +173,7 @@ void slow_conv_2D(
             out_ptr += out_stride_O;
             wt_ptr += wt_stride_O;
           } // o
+        } // g
       };
 
   int jump_h = flip ? -wt_dilation[0] : wt_dilation[0];
@@ -214,7 +225,8 @@ 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) {
+        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) {
@@ -234,11 +246,11 @@ void slow_conv_2D(
                   const T* in_ptr_pt =
                       in_ptr + ih_dil * in_stride_H + iw_dil * in_stride_W;
 
-                for (int c = 0; c < C; ++c) {
+                  for (int c = g * C_per_group; c < (g + 1) * C_per_group;
-                  r += static_cast<float>(in_ptr_pt[0]) *
+                       ++c) {
-                      static_cast<float>(wt_ptr_pt[0]);
+                    r += static_cast<float>(in_ptr_pt[c * in_stride_C]) *
-                  in_ptr_pt += in_stride_C;
+                        static_cast<float>(
-                  wt_ptr_pt += wt_stride_C;
+                             wt_ptr_pt[(c % C_per_group) * wt_stride_C]);
                   } // c
 
                 } // ih, iw check
@@ -249,6 +261,7 @@ void slow_conv_2D(
             out_ptr += out_stride_O;
             wt_ptr += wt_stride_O;
           } // o
+        } // g
       };
 
   int oH_border_0 = 0;
@@ -305,6 +318,296 @@ void slow_conv_2D(
   } // n
 }
 
+template <typename T>
+void slow_conv_3D(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    const std::vector<int>& in_dilation,
+    bool flip) {
+  const T* st_wt_ptr = wt.data<T>();
+  const T* st_in_ptr = in.data<T>();
+  T* st_out_ptr = out.data<T>();
+
+  const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
+  const int iD = 1 + in_dilation[0] * (in.shape(1) - 1); // Input spatial dim
+  const int iH = 1 + in_dilation[1] * (in.shape(2) - 1); // Input spatial dim
+  const int iW = 1 + in_dilation[2] * (in.shape(3) - 1); // Input spatial dim
+  const int oD = out.shape(1); // Output spatial dim
+  const int oH = out.shape(2); // Output spatial dim
+  const int oW = out.shape(3); // Output spatial dim
+  const int O = wt.shape(0); // Out channels
+  const int C = wt.shape(4); // In channels
+  const int wD = wt.shape(1); // Weight spatial dim
+  const int wH = wt.shape(2); // Weight spatial dim
+  const int wW = wt.shape(3); // Weight spatial dim
+
+  const size_t in_stride_N = in.strides()[0];
+  const size_t in_stride_D = in.strides()[1];
+  const size_t in_stride_H = in.strides()[2];
+  const size_t in_stride_W = in.strides()[3];
+  const size_t in_stride_C = in.strides()[4];
+
+  const size_t wt_stride_O = wt.strides()[0];
+  const size_t wt_stride_D = wt.strides()[1];
+  const size_t wt_stride_H = wt.strides()[2];
+  const size_t wt_stride_W = wt.strides()[3];
+  const size_t wt_stride_C = wt.strides()[4];
+
+  const size_t out_stride_N = out.strides()[0];
+  const size_t out_stride_D = out.strides()[1];
+  const size_t out_stride_H = out.strides()[2];
+  const size_t out_stride_W = out.strides()[3];
+  const size_t out_stride_O = out.strides()[4];
+
+  bool is_idil_one =
+      in_dilation[0] == 1 && in_dilation[1] == 1 && in_dilation[2] == 1;
+
+  auto pt_conv_no_checks = [&](const T* in_ptr,
+                               const T* wt_ptr,
+                               T* out_ptr,
+                               int od,
+                               int oh,
+                               int ow) {
+    out_ptr += od * out_stride_D + oh * out_stride_H + ow * out_stride_W;
+    int id_base = od * wt_strides[0] - padding[0];
+    int ih_base = oh * wt_strides[1] - padding[1];
+    int iw_base = ow * wt_strides[2] - padding[2];
+
+    for (int o = 0; o < O; ++o) {
+      float r = 0.;
+
+      for (int wd = 0; wd < wD; ++wd) {
+        for (int wh = 0; wh < wH; ++wh) {
+          for (int ww = 0; ww < wW; ++ww) {
+            int wd_flip = flip ? wD - wd - 1 : wd;
+            int wh_flip = flip ? wH - wh - 1 : wh;
+            int ww_flip = flip ? wW - ww - 1 : ww;
+            int id = id_base + wd_flip * wt_dilation[0];
+            int ih = ih_base + wh_flip * wt_dilation[1];
+            int iw = iw_base + ww_flip * wt_dilation[2];
+
+            const T* wt_ptr_pt =
+                wt_ptr + wd * wt_stride_D + wh * wt_stride_H + ww * wt_stride_W;
+            const T* in_ptr_pt =
+                in_ptr + id * in_stride_D + ih * in_stride_H + iw * in_stride_W;
+
+            for (int c = 0; c < C; ++c) {
+              r += static_cast<float>(in_ptr_pt[0]) *
+                  static_cast<float>(wt_ptr_pt[0]);
+              in_ptr_pt += in_stride_C;
+              wt_ptr_pt += wt_stride_C;
+            } // c
+
+          } // ww
+        } // wh
+      } // wd
+
+      out_ptr[0] = static_cast<T>(r);
+      out_ptr += out_stride_O;
+      wt_ptr += wt_stride_O;
+    } // o
+  };
+
+  int jump_d = flip ? -wt_dilation[0] : wt_dilation[0];
+  int jump_h = flip ? -wt_dilation[1] : wt_dilation[1];
+  int jump_w = flip ? -wt_dilation[2] : wt_dilation[2];
+
+  int init_d = (flip ? (wD - 1) * wt_dilation[0] : 0);
+  int init_h = (flip ? (wH - 1) * wt_dilation[1] : 0);
+  int init_w = (flip ? (wW - 1) * wt_dilation[2] : 0);
+
+  int f_wgt_jump_d = std::lcm(in_dilation[0], wt_dilation[0]) / wt_dilation[0];
+  int f_wgt_jump_h = std::lcm(in_dilation[1], wt_dilation[1]) / wt_dilation[1];
+  int f_wgt_jump_w = std::lcm(in_dilation[2], wt_dilation[2]) / wt_dilation[2];
+
+  int f_out_jump_d = std::lcm(in_dilation[0], wt_strides[0]) / wt_strides[0];
+  int f_out_jump_h = std::lcm(in_dilation[1], wt_strides[1]) / wt_strides[1];
+  int f_out_jump_w = std::lcm(in_dilation[2], wt_strides[2]) / wt_strides[2];
+
+  std::vector<int> base_d(f_out_jump_d);
+  std::vector<int> base_h(f_out_jump_h);
+  std::vector<int> base_w(f_out_jump_w);
+
+  for (int i = 0; i < f_out_jump_d; ++i) {
+    int id_loop = i * wt_strides[0] - padding[0] + init_d;
+
+    int wd_base = 0;
+    while (wd_base < wD && id_loop % in_dilation[0] != 0) {
+      wd_base++;
+      id_loop += jump_d;
+    }
+
+    base_d[i] = wd_base;
+  }
+
+  for (int i = 0; i < f_out_jump_h; ++i) {
+    int ih_loop = i * wt_strides[1] - padding[1] + init_h;
+
+    int wh_base = 0;
+    while (wh_base < wH && ih_loop % in_dilation[1] != 0) {
+      wh_base++;
+      ih_loop += jump_h;
+    }
+
+    base_h[i] = wh_base;
+  }
+
+  for (int j = 0; j < f_out_jump_w; ++j) {
+    int iw_loop = j * wt_strides[2] - padding[2] + init_w;
+
+    int ww_base = 0;
+    while (ww_base < wW && iw_loop % in_dilation[2] != 0) {
+      ww_base++;
+      iw_loop += jump_w;
+    }
+
+    base_w[j] = ww_base;
+  }
+
+  auto pt_conv_all_checks = [&](const T* in_ptr,
+                                const T* wt_ptr,
+                                T* out_ptr,
+                                int od,
+                                int oh,
+                                int ow) {
+    out_ptr += od * out_stride_D + oh * out_stride_H + ow * out_stride_W;
+
+    int id_base = od * wt_strides[0] - padding[0];
+    int ih_base = oh * wt_strides[1] - padding[1];
+    int iw_base = ow * wt_strides[2] - padding[2];
+
+    int wd_base = base_d[od % f_out_jump_d];
+    int wh_base = base_h[oh % f_out_jump_h];
+    int ww_base = base_w[ow % f_out_jump_w];
+
+    for (int o = 0; o < O; ++o) {
+      float r = 0.;
+
+      for (int wd = wd_base; wd < wD; wd += f_wgt_jump_d) {
+        for (int wh = wh_base; wh < wH; wh += f_wgt_jump_h) {
+          for (int ww = ww_base; ww < wW; ww += f_wgt_jump_w) {
+            int wd_flip = flip ? wD - wd - 1 : wd;
+            int wh_flip = flip ? wH - wh - 1 : wh;
+            int ww_flip = flip ? wW - ww - 1 : ww;
+            int id = id_base + wd_flip * wt_dilation[0];
+            int ih = ih_base + wh_flip * wt_dilation[1];
+            int iw = iw_base + ww_flip * wt_dilation[2];
+
+            if (id >= 0 && id < iD && ih >= 0 && ih < iH && iw >= 0 &&
+                iw < iW) {
+              const T* wt_ptr_pt = wt_ptr + wd * wt_stride_D +
+                  wh * wt_stride_H + ww * wt_stride_W;
+
+              int id_dil = !is_idil_one ? (id / in_dilation[0]) : id;
+              int ih_dil = !is_idil_one ? (ih / in_dilation[1]) : ih;
+              int iw_dil = !is_idil_one ? (iw / in_dilation[2]) : iw;
+
+              const T* in_ptr_pt = in_ptr + id_dil * in_stride_D +
+                  ih_dil * in_stride_H + iw_dil * in_stride_W;
+
+              for (int c = 0; c < C; ++c) {
+                r += static_cast<float>(in_ptr_pt[0]) *
+                    static_cast<float>(wt_ptr_pt[0]);
+                in_ptr_pt += in_stride_C;
+                wt_ptr_pt += wt_stride_C;
+              } // c
+
+            } // iD, ih, iw check
+          } // ww
+        } // wh
+      } // wd
+
+      out_ptr[0] = static_cast<T>(r);
+      out_ptr += out_stride_O;
+      wt_ptr += wt_stride_O;
+    } // o
+  };
+
+  int oD_border_0 = 0;
+  int oD_border_1 =
+      is_idil_one ? ((padding[0] + wt_strides[0] - 1) / wt_strides[0]) : oD;
+  int oD_border_2 = std::max(
+      oD_border_1, (iD + padding[0] - wD * wt_dilation[0]) / wt_strides[0]);
+  int oD_border_3 = oD;
+
+  int oH_border_0 = 0;
+  int oH_border_1 =
+      is_idil_one ? ((padding[1] + wt_strides[1] - 1) / wt_strides[1]) : oH;
+  int oH_border_2 = std::max(
+      oH_border_1, (iH + padding[1] - wH * wt_dilation[1]) / wt_strides[1]);
+  int oH_border_3 = oH;
+
+  int oW_border_0 = 0;
+  int oW_border_1 =
+      is_idil_one ? ((padding[2] + wt_strides[2] - 1) / wt_strides[2]) : oW;
+  int oW_border_2 = std::max(
+      oW_border_1, (iW + padding[2] - wW * wt_dilation[2]) / wt_strides[2]);
+  int oW_border_3 = oW;
+
+  for (int n = 0; n < N; ++n) {
+    // Case 1: od might put us out of bounds
+    for (int od = oD_border_0; od < oD_border_1; ++od) {
+      for (int oh = 0; oh < oH; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+    } // od
+
+    // Case 2: od in bounds
+    for (int od = oD_border_1; od < oD_border_2; ++od) {
+      // Case 2.1: oh might put us out of bounds
+      for (int oh = oH_border_0; oh < oH_border_1; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+
+      // Case 2.2: oh in bounds
+      for (int oh = oH_border_1; oh < oH_border_2; ++oh) {
+        // Case 2.2.1: ow might put us out of bounds
+        for (int ow = oW_border_0; ow < oW_border_1; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+
+        // Case 2.2.2: ow in bounds
+        for (int ow = oW_border_1; ow < oW_border_2; ++ow) {
+          pt_conv_no_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+
+        // Case 2.2.3: ow might put us out of bounds
+        for (int ow = oW_border_2; ow < oW_border_3; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+
+      // Case 2.3: oh might put us out of bounds
+      for (int oh = oH_border_2; oh < oH_border_3; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+    } // od
+
+    // Case 3: od might put us out of bounds
+    for (int od = oD_border_2; od < oD_border_3; ++od) {
+      for (int oh = 0; oh < oH; ++oh) {
+        for (int ow = 0; ow < oW; ++ow) {
+          pt_conv_all_checks(st_in_ptr, st_wt_ptr, st_out_ptr, od, oh, ow);
+        } // ow
+      } // oh
+    } // od
+
+    st_in_ptr += in_stride_N;
+    st_out_ptr += out_stride_N;
+
+  } // n
+}
+
 void dispatch_slow_conv_1D(
     const array& in,
     const array& wt,
@@ -353,6 +656,30 @@ void dispatch_slow_conv_2D(
   }
 }
 
+void dispatch_slow_conv_3D(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    const std::vector<int>& in_dilation,
+    bool flip) {
+  if (in.dtype() == float32) {
+    return slow_conv_3D<float>(
+        in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+  } else if (in.dtype() == float16) {
+    return slow_conv_3D<float16_t>(
+        in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+  } else if (in.dtype() == bfloat16) {
+    return slow_conv_3D<bfloat16_t>(
+        in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+  } else {
+    throw std::invalid_argument(
+        "[Convolution::eval] got unsupported data type.");
+  }
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Explicit gemm conv
 ///////////////////////////////////////////////////////////////////////////////
@@ -366,11 +693,15 @@ void explicit_gemm_conv_1D_cpu(
     const std::vector<int>& wt_dilation) {
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
   const int iH = in.shape(1); // Input spatial dim
+  const int C = in.shape(2); // Input channels
   const int oH = out.shape(1); // Output spatial dim
   const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(2); // In channels
   const int wH = wt.shape(1); // Weight spatial dim
 
+  const int groups = C / wt.shape(2);
+  const int C_per_group = wt.shape(2);
+  const int O_per_group = O / groups;
+
   auto conv_dtype = float32;
 
   // Pad input
@@ -402,6 +733,11 @@ void explicit_gemm_conv_1D_cpu(
       in_padded.strides()[1],
       in_padded.strides()[2]};
   auto flags = in_padded.flags();
+  if (groups > 1) {
+    // Transpose the last two dimensions for grouped convolutions
+    std::swap(strided_shape[2], strided_shape[3]);
+    std::swap(strided_strides[2], strided_strides[3]);
+  }
 
   array in_strided_view(strided_shape, in_padded.dtype(), nullptr, {});
   in_strided_view.copy_shared_buffer(
@@ -416,7 +752,19 @@ void explicit_gemm_conv_1D_cpu(
   auto gemm_wt = wt;
   auto gemm_out = out;
 
-  if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
+  if (groups > 1) {
+    // Transpose the last two dimensions for grouped convolutions
+    array wt_transpose(
+        {wt.shape(0), wt.shape(2), wt.shape(1)}, wt.dtype(), nullptr, {});
+    wt_transpose.copy_shared_buffer(
+        wt,
+        {wt.strides(0), wt.strides(2), wt.strides(1)},
+        wt.flags(),
+        wt.size(),
+        0);
+    gemm_wt = array(wt_transpose.shape(), float32, nullptr, {});
+    copy(wt_transpose, gemm_wt, CopyType::General);
+  } else if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
@@ -428,21 +776,22 @@ void explicit_gemm_conv_1D_cpu(
     gemm_out.set_data(allocator::malloc_or_wait(gemm_out.nbytes()));
   }
 
+  for (int g = 0; g < groups; ++g) {
     // Perform gemm
     cblas_sgemm(
         CblasRowMajor,
         CblasNoTrans, // no trans A
         CblasTrans, // transB
         strided_reshape[0], // M
-      O, // N
+        O_per_group, // N
-      strided_reshape[1], // K
+        C_per_group * wH, // K
         1.0f, // alpha
-      in_strided.data<float>(),
+        in_strided.data<float>() + g * C_per_group * wH, // A
-      strided_reshape[1], // lda
+        wH * C, // lda
-      gemm_wt.data<float>(),
+        gemm_wt.data<float>() + g * O_per_group * C_per_group * wH, // B
-      strided_reshape[1], // ldb
+        wH * C_per_group, // ldb
         0.0f, // beta
-      gemm_out.data<float>(),
+        gemm_out.data<float>() + g * O_per_group, // C
         O // ldc
     );
 
@@ -451,6 +800,7 @@ void explicit_gemm_conv_1D_cpu(
       copy(gemm_out, out, CopyType::Vector);
     }
   }
+}
 
 void explicit_gemm_conv_2D_cpu(
     const array& in,
@@ -554,6 +904,131 @@ void explicit_gemm_conv_2D_cpu(
   }
 }
 
+void explicit_gemm_conv_ND_cpu(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation) {
+  const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
+  const auto iDim = std::vector<int>(
+      in.shape().begin() + 1, in.shape().end() - 1); // Input spatial dim
+  const auto oDim = std::vector<int>(
+      out.shape().begin() + 1, out.shape().end() - 1); // Output spatial dim
+  const int O = wt.shape(0); // Out channels
+  const int C = wt.shape(-1); // In channels
+  const auto wDim = std::vector<int>(
+      wt.shape().begin() + 1, wt.shape().end() - 1); // Weight spatial dim
+
+  auto conv_dtype = float32;
+
+  // Pad input
+  std::vector<int> padded_shape(in.shape().size());
+  padded_shape.front() = N;
+  for (size_t i = 0; i < iDim.size(); i++) {
+    padded_shape[i + 1] = iDim[i] + 2 * padding[i];
+  }
+  padded_shape.back() = C;
+  array in_padded(padded_shape, conv_dtype, nullptr, {});
+
+  // Fill with zeros
+  copy(array(0, conv_dtype), in_padded, CopyType::Scalar);
+
+  // Pick input slice from padded
+  size_t data_offset = 0;
+  for (size_t i = 0; i < padding.size(); i++) {
+    data_offset += padding[i] * in_padded.strides()[i + 1];
+  }
+  array in_padded_slice(in.shape(), in_padded.dtype(), nullptr, {});
+  in_padded_slice.copy_shared_buffer(
+      in_padded,
+      in_padded.strides(),
+      in_padded.flags(),
+      in_padded_slice.size(),
+      data_offset);
+
+  // Copy input values into the slice
+  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral);
+
+  // Make strided view
+  std::vector<int> strided_shape(oDim.size() + wDim.size() + 2);
+  strided_shape.front() = N;
+  for (size_t i = 0; i < oDim.size(); i++) {
+    strided_shape[i + 1] = oDim[i];
+  }
+  for (size_t i = 0; i < wDim.size(); i++) {
+    strided_shape[i + 1 + oDim.size()] = wDim[i];
+  }
+  strided_shape.back() = C;
+
+  std::vector<size_t> strided_strides(in.shape().size() * 2 - 2);
+  strided_strides[0] = in_padded.strides()[0];
+  for (size_t i = 0; i < wt_strides.size(); i++) {
+    strided_strides[i + 1] = in_padded.strides()[i + 1] * wt_strides[i];
+  }
+  for (size_t i = 1; i < in_padded.strides().size(); i++) {
+    strided_strides[i + wt_strides.size()] = in_padded.strides()[i];
+  }
+
+  auto flags = in_padded.flags();
+
+  array in_strided_view(strided_shape, in_padded.dtype(), nullptr, {});
+  in_strided_view.copy_shared_buffer(
+      in_padded, strided_strides, flags, in_strided_view.size(), 0);
+
+  // Materialize strided view
+  std::vector<int> strided_reshape = {N, C};
+  for (const auto& o : oDim) {
+    strided_reshape[0] *= o;
+  }
+  for (const auto& w : wDim) {
+    strided_reshape[1] *= w;
+  }
+
+  array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
+  copy(in_strided_view, in_strided, CopyType::General);
+
+  // Check wt dtype and prepare
+  auto gemm_wt = wt;
+  auto gemm_out = out;
+
+  if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
+    auto ctype =
+        wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
+    gemm_wt = array(wt.shape(), float32, nullptr, {});
+    copy(wt, gemm_wt, ctype);
+  }
+
+  if (out.dtype() != float32) {
+    gemm_out = array(out.shape(), float32, nullptr, {});
+    gemm_out.set_data(allocator::malloc_or_wait(gemm_out.nbytes()));
+  }
+
+  // Perform gemm
+  cblas_sgemm(
+      CblasRowMajor,
+      CblasNoTrans, // no trans A
+      CblasTrans, // transB
+      strided_reshape[0], // M
+      O, // N
+      strided_reshape[1], // K
+      1.0f, // alpha
+      in_strided.data<float>(),
+      strided_reshape[1], // lda
+      gemm_wt.data<float>(),
+      strided_reshape[1], // ldb
+      0.0f, // beta
+      gemm_out.data<float>(),
+      O // ldc
+  );
+
+  // Copy results if needed
+  if (out.dtype() != float32) {
+    copy(gemm_out, out, CopyType::Vector);
+  }
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Conv routing
 ///////////////////////////////////////////////////////////////////////////////
@@ -589,6 +1064,19 @@ void conv_2D_cpu(
       in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
 }
 
+void conv_3D_cpu(
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    const std::vector<int>& in_dilation,
+    bool flip) {
+  return dispatch_slow_conv_3D(
+      in, wt, out, padding, wt_strides, wt_dilation, in_dilation, flip);
+}
+
 } // namespace
 
 void Convolution::eval(const std::vector<array>& inputs, array& out) {
@@ -597,8 +1085,20 @@ void Convolution::eval(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
   auto& wt = inputs[1];
 
+  // 3D convolution
+  if (in.ndim() == (3 + 2)) {
+    return conv_3D_cpu(
+        in,
+        wt,
+        out,
+        padding_,
+        kernel_strides_,
+        kernel_dilation_,
+        input_dilation_,
+        flip_);
+  }
   // 2D convolution
-  if (in.ndim() == (2 + 2)) {
+  else if (in.ndim() == (2 + 2)) {
     return conv_2D_cpu(
         in,
         wt,

mlx/backend/common/copy.cpp

@@ -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,
+    int64_t i_offset,
-    stride_t o_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,
+    int64_t i_offset,
-    stride_t o_offset) {
+    int64_t o_offset) {
-  switch (src.ndim()) {
+  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 src_offset = i_offset + elem_to_loc(i, new_shape, new_strides[0]);
-    stride_t dst_offset = o_offset + elem_to_loc(i, dst.shape(), o_strides);
+    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 <>
+template void copy_inplace<size_t>(
-void copy_inplace<int64_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) {
+    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);
-  }
-}
 
 } // namespace mlx::core

mlx/backend/common/default_primitives.cpp

@@ -5,7 +5,6 @@
 #else
 #include <cblas.h>
 #endif
-
 #include <cstring>
 
 #include "mlx/array.h"
@@ -34,6 +33,7 @@ DEFAULT(ArcCosh)
 DEFAULT(ArcSin)
 DEFAULT(ArcSinh)
 DEFAULT(ArcTan)
+DEFAULT(ArcTan2)
 DEFAULT(ArcTanh)
 DEFAULT(ArgPartition)
 DEFAULT(ArgReduce)
@@ -41,14 +41,18 @@ DEFAULT(ArgSort)
 DEFAULT(AsType)
 DEFAULT(AsStrided)
 DEFAULT(Broadcast)
+DEFAULT(BlockMaskedMM)
+DEFAULT(GatherMM)
+DEFAULT(GatherQMM)
 DEFAULT_MULTI(DivMod)
 DEFAULT(Ceil)
 DEFAULT(Concatenate)
+DEFAULT(Conjugate)
 DEFAULT(Convolution)
 DEFAULT(Copy)
 DEFAULT(Cos)
 DEFAULT(Cosh)
-DEFAULT_MULTI(CustomVJP)
+DEFAULT_MULTI(CustomTransforms)
 DEFAULT_MULTI(Depends)
 DEFAULT(Divide)
 DEFAULT(NumberOfElements)
@@ -64,6 +68,7 @@ DEFAULT(Full)
 DEFAULT(Gather)
 DEFAULT(Greater)
 DEFAULT(GreaterEqual)
+DEFAULT(Hadamard)
 DEFAULT(Less)
 DEFAULT(LessEqual)
 DEFAULT(Load)
@@ -108,6 +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
@@ -11,24 +10,39 @@
 #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) {
-  // Lapack uses the column-major convention. We take advantage of the following
-  // identity to avoid transposing (see
-  // https://math.stackexchange.com/a/340234):
-  //   (A⁻¹)ᵀ = (Aᵀ)⁻¹
-
-  // The inverse is computed in place, so just copy the input to the output.
-  copy(a, inv, a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-
-  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,
@@ -65,8 +79,7 @@ void inverse_impl(const array& a, array& inv) {
   }
 
   const int lwork = workspace_size;
-    auto scratch =
+  auto scratch = array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
-        array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
 
   // Compute inverse.
   sgetri_(
@@ -84,21 +97,44 @@ void inverse_impl(const array& a, array& inv) {
     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):
+  //   (A⁻¹)ᵀ = (Aᵀ)⁻¹
+
+  // The inverse is computed in place, so just copy the input to the output.
+  copy(a, inv, a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+
+  const int N = a.shape(-1);
+  const size_t num_matrices = a.size() / (N * N);
+
+  for (int i = 0; i < num_matrices; i++) {
+    if (tri) {
+      tri_inv(inv, N, i, upper);
+    } else {
+      general_inv(inv, N, i);
+    }
+  }
 }
 
 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);
+  inverse_impl(inputs[0], output, tri_, upper_);
-}
-
-std::pair<std::vector<array>, std::vector<int>> Inverse::vmap(
-    const std::vector<array>& inputs,
-    const std::vector<int>& axes) {
-  auto ax = axes[0] >= 0 ? 0 : -1;
-  auto a = axes[0] > 0 ? moveaxis(inputs[0], axes[0], 0, stream()) : inputs[0];
-  return {{linalg::inv(a, stream())}, {ax}};
 }
 
 } // namespace mlx::core

mlx/backend/common/make_compiled_preamble.sh

@@ -11,7 +11,7 @@ GCC=$2
 SRCDIR=$3
 CLANG=$4
 
-if [ $CLANG = "TRUE" ]; then
+if [ "$CLANG" = "TRUE" ]; then
   read -r -d '' INCLUDES <<- EOM
   #include <cmath>
   #include <complex>
@@ -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

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

mlx/backend/common/ops.h

@@ -108,133 +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 {
@@ -606,4 +619,39 @@ struct Select {
   }
 };
 
+struct BitwiseAnd {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x & y;
+  }
+};
+
+struct BitwiseOr {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x | y;
+  }
+};
+
+struct BitwiseXor {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x ^ y;
+  }
+};
+
+struct LeftShift {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x << y;
+  }
+};
+
+struct RightShift {
+  template <typename T>
+  T operator()(T x, T y) {
+    return x >> y;
+  }
+};
+
 } // namespace mlx::core::detail

mlx/backend/common/primitives.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <algorithm>
 #include <cassert>
@@ -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) {
+  auto ibytes = size_of(in.dtype());
-    out_strides[ax] = in.strides()[axes_[ax]];
+  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);
 
-  // Conditions for {row/col}_contiguous
+    auto flags = out.flags();
-  // - array must be contiguous (no gaps)
+    flags.contiguous = true;
-  // - 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) {
+    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
-      flags.col_contiguous &= (out_strides[i] == f_stride || out.shape(i) == 1);
+    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
-      f_stride *= out.shape(i);
+    out.move_shared_buffer(tmp, out.strides(), flags, out.size());
-      flags.row_contiguous &=
-          (out_strides[ri] == b_stride || out.shape(ri) == 1);
-      b_stride *= out.shape(ri);
   }
 }
-  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
-}
 
 } // namespace mlx::core

mlx/backend/common/quantized.cpp

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

mlx/backend/common/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()) {
+      if (out.dtype() == int32) {
-        case bool_:
+        // special case since the input type can be 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;
+      } else {
-        case int64:
+        reduction_op<InT, InT>(in, out, axes, 0, op);
-          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;
     }
-    } 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) {
+    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);
-}
 
 std::pair<std::vector<int>, std::vector<size_t>> shapes_without_reduction_axes(
     const array& x,
-    const std::vector<int>& axes) {
+    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);
-}
 
 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];
+  size_t axis_stride = out.strides()[axis];
-  int axis_size = in.shape(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();
+  auto in_remaining_shape = in.shape();
-  remaining_shape.erase(remaining_shape.begin() + axis);
+  in_remaining_shape.erase(in_remaining_shape.begin() + axis);
 
-  auto remaining_strides = in.strides();
+  auto in_remaining_strides = in.strides();
-  remaining_strides.erase(remaining_strides.begin() + axis);
+  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);
+    size_t in_loc = elem_to_loc(i, in_remaining_shape, in_remaining_strides);
-    const T* data_ptr = in.data<T>() + loc;
+    size_t out_loc = elem_to_loc(i, out_remaining_shape, out_remaining_strides);
-    IdxT* idx_ptr = out.data<IdxT>() + loc;
+    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 st_(idx_ptr, out_stride, 0);
-    StridedIterator ed_(idx_ptr, axis_stride, axis_size);
+    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 st(idx_ptr, out_stride, 0);
-    StridedIterator ed(idx_ptr, axis_stride, axis_size);
+    StridedIterator ed(idx_ptr, out_stride, axis_size);
 
-    std::stable_sort(st, ed, [data_ptr, axis_stride](IdxT a, IdxT b) {
+    std::stable_sort(st, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
-      auto v1 = data_ptr[a * axis_stride];
+      auto v1 = data_ptr[a * in_stride];
-      auto v2 = data_ptr[b * axis_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

@@ -1,32 +1,140 @@
+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 ${SRC_FILE} NAME)
   add_custom_command(
-    OUTPUT  compiled_preamble.cpp
+    OUTPUT  jit/${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/${SRC_FILE}.h
-            kernels/unary.h
+            ${ARGN}
-            kernels/binary.h
   )
-
+  add_custom_target(${SRC_NAME} DEPENDS jit/${SRC_NAME}.cpp)
-add_custom_target(
+  add_dependencies(mlx ${SRC_NAME})
-  compiled_preamble
+  target_sources(
-  DEPENDS compiled_preamble.cpp
+    mlx
+    PRIVATE
+    ${CMAKE_CURRENT_BINARY_DIR}/jit/${SRC_NAME}.cpp
   )
+endfunction(make_jit_source)
 
-add_dependencies(mlx compiled_preamble)
+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)
+
+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
@@ -36,10 +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

@@ -140,10 +140,15 @@ void BufferCache::remove_from_list(BufferCache::BufferHolder* to_remove) {
 
 MetalAllocator::MetalAllocator()
     : device_(device(mlx::core::Device::gpu).mtl_device()),
-      buffer_cache_(device_),
+      buffer_cache_(device_) {
-      block_limit_(1.5 * device_->recommendedMaxWorkingSetSize()),
+  auto memsize = std::get<size_t>(device_info()["memory_size"]);
-      gc_limit_(0.95 * device_->recommendedMaxWorkingSetSize()),
+  block_limit_ =
-      max_pool_size_(block_limit_) {}
+      std::min(1.5 * device_->recommendedMaxWorkingSetSize(), 0.95 * memsize);
+  gc_limit_ = std::min(
+      static_cast<size_t>(0.95 * device_->recommendedMaxWorkingSetSize()),
+      block_limit_);
+  max_pool_size_ = block_limit_;
+}
 
 size_t MetalAllocator::set_cache_limit(size_t limit) {
   std::swap(limit, max_pool_size_);
@@ -165,6 +170,15 @@ Buffer MetalAllocator::malloc(size_t size, bool allow_swap /* = false */) {
     return Buffer{nullptr};
   }
 
+  // More helpful message if maximum buffer length is exceeded
+  if (size > device_->maxBufferLength()) {
+    std::ostringstream msg;
+    msg << "Attempting to allocate " << size << " bytes which is greater than"
+        << " the maximum allowed buffer size of " << device_->maxBufferLength()
+        << " bytes.";
+    throw std::runtime_error(msg.str());
+  }
+
   // Align up memory
   if (size > vm_page_size) {
     size = vm_page_size * ((size + vm_page_size - 1) / vm_page_size);
@@ -209,6 +223,11 @@ Buffer MetalAllocator::malloc(size_t size, bool allow_swap /* = false */) {
   return Buffer{static_cast<void*>(buf)};
 }
 
+void MetalAllocator::clear_cache() {
+  std::unique_lock lk(mutex_);
+  buffer_cache_.clear();
+}
+
 void MetalAllocator::free(Buffer buffer) {
   auto buf = static_cast<MTL::Buffer*>(buffer.ptr());
   std::unique_lock lk(mutex_);
@@ -223,8 +242,17 @@ void MetalAllocator::free(Buffer buffer) {
 }
 
 MetalAllocator& allocator() {
-  static MetalAllocator allocator_;
+  // By creating the |allocator_| on heap, the destructor of MetalAllocator will
-  return allocator_;
+  // 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) {
@@ -239,9 +267,15 @@ size_t get_active_memory() {
 size_t get_peak_memory() {
   return allocator().get_peak_memory();
 }
+void reset_peak_memory() {
+  allocator().reset_peak_memory();
+}
 size_t get_cache_memory() {
   return allocator().get_cache_memory();
 }
+void clear_cache() {
+  return allocator().clear_cache();
+}
 
 } // namespace metal
 

mlx/backend/metal/allocator.h

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

mlx/backend/metal/binary.cpp

@@ -0,0 +1,296 @@
+// 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
+         << " [[buffer(" << cnt++ << ")]]," << std::endl;
-         << "    constant const size_t* " << xname << "_strides [[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
+         << xname << "[elem_to_loc(index, output_shape, in_strides + "
-         << "_strides, ndim)];" << std::endl;
+         << 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(
+      in_strides.insert(
-          strides[stride_idx].data(),
+          in_strides.end(),
-          strides[stride_idx].size() * sizeof(size_t),
+          strides[stride_idx].begin(),
-          cnt++);
+          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};
@@ -89,6 +90,90 @@ void explicit_gemm_conv_ND_gpu(
       /*copies = */ copies);
 }
 
+template <int N>
+void explicit_gemm_conv_group_ND_gpu(
+    const Stream& s,
+    metal::Device& d,
+    const array& in,
+    const array& wt,
+    array out,
+    const MLXConvParams<N>& conv_params) {
+  const int groups = conv_params.groups;
+  const int C_per_group = conv_params.C / conv_params.groups;
+  const int O_per_group = conv_params.O / conv_params.groups;
+  // Get gemm shapes
+  const int implicit_M = out.size() / conv_params.O;
+  const int implicit_K = wt.size() / conv_params.O;
+  const int implicit_N = O_per_group;
+
+  int kernel_size = 1;
+  for (int i = 0; i < N; ++i) {
+    kernel_size *= conv_params.wS[i];
+  }
+
+  // Prepare unfolding array
+  std::vector<int> unfolded_shape{implicit_M, implicit_K * groups};
+  array in_unfolded(unfolded_shape, in.dtype(), nullptr, {});
+  in_unfolded.set_data(allocator::malloc_or_wait(in_unfolded.nbytes()));
+
+  // Prepare unfolding kernel
+  std::ostringstream kname;
+  kname << "naive_unfold_transpose_nd_" << type_to_name(in_unfolded) << "_"
+        << N;
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = d.get_kernel(kname.str());
+  compute_encoder->setComputePipelineState(kernel);
+
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_output_array(in_unfolded, 1);
+
+  compute_encoder->setBytes(&conv_params, sizeof(conv_params), 2);
+
+  // Launch unfolding kernel
+  int tgp_x = std::min(conv_params.C, 64);
+  tgp_x = 32 * ((tgp_x + 32 - 1) / 32);
+  int tgp_y = 256 / tgp_x;
+
+  MTL::Size group_dims = MTL::Size(tgp_x, tgp_y, 1);
+  MTL::Size grid_dims = MTL::Size(
+      conv_params.C, unfolded_shape[1] / conv_params.C, unfolded_shape[0]);
+
+  compute_encoder.dispatchThreads(grid_dims, group_dims);
+
+  // Transpose kernel weights so that we can slice them by contiguous chunks
+  // of channel groups.
+  array wt_view(
+      {wt.shape(0), C_per_group, kernel_size}, wt.dtype(), nullptr, {});
+  wt_view.copy_shared_buffer(
+      wt,
+      {wt.strides(0), 1, static_cast<size_t>(C_per_group)},
+      wt.flags(),
+      wt.size());
+
+  // Materialize
+  auto wt_transpose = array(wt_view.shape(), wt_view.dtype(), nullptr, {});
+  copy_gpu(wt_view, wt_transpose, CopyType::General, s);
+
+  // Perform gemm
+  std::vector<array> copies = {in_unfolded, wt_view, wt_transpose};
+  return steel_matmul_conv_groups(
+      s,
+      d,
+      /*a = */ in_unfolded,
+      /*b = */ wt_transpose,
+      /*c = */ out,
+      /*M = */ implicit_M,
+      /*N = */ implicit_N,
+      /*K = */ implicit_K,
+      /*a_cols = */ implicit_K * groups,
+      /*b_cols = */ implicit_K,
+      /*out_cols = */ implicit_N * groups,
+      /*a_transposed = */ false,
+      /*b_transposed = */ true,
+      /* groups = */ groups,
+      /*copies = */ copies);
+}
+
 void conv_1D_gpu(
     const Stream& s,
     metal::Device& d,
@@ -99,6 +184,7 @@ void conv_1D_gpu(
     const std::vector<int>& wt_strides,
     const std::vector<int>& wt_dilation,
     const std::vector<int>& in_dilation,
+    int groups,
     bool flip) {
   // Make conv params
   MLXConvParams<1> conv_params{
@@ -118,12 +204,16 @@ void conv_1D_gpu(
       {wt.strides()[0], wt.strides()[1], wt.strides()[2]},
       /* const size_t out_strides[NDIM + 2] = */
       {out.strides()[0], out.strides()[1], out.strides()[2]},
-      /* const int groups = */ 1,
+      /* const int groups = */ groups,
       /* const bool flip = */ flip};
 
   // Direct to explicit gemm conv
+  if (groups > 1) {
+    return explicit_gemm_conv_group_ND_gpu(s, d, in, wt, out, conv_params);
+  } else {
     return explicit_gemm_conv_ND_gpu(s, d, in, wt, out, conv_params);
   }
+}
 
 void slow_conv_2D_gpu(
     const Stream& s,
@@ -158,7 +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(
@@ -168,15 +258,19 @@ void implicit_gemm_conv_2D_gpu(
     const array& wt,
     array out,
     const MLXConvParams<2>& conv_params) {
+  const int groups = conv_params.groups;
+  const int C_per_group = conv_params.C / conv_params.groups;
+  const int O_per_group = conv_params.O / conv_params.groups;
+
   // Deduce implicit gemm size
-  int implicit_M = conv_params.N * conv_params.oS[0] * conv_params.oS[1];
+  const int implicit_M = conv_params.N * conv_params.oS[0] * conv_params.oS[1];
-  int implicit_N = conv_params.O;
+  const int implicit_N = O_per_group;
-  int implicit_K = conv_params.wS[0] * conv_params.wS[1] * conv_params.C;
+  const int implicit_K = conv_params.wS[0] * conv_params.wS[1] * C_per_group;
 
   // Determine block and warp tiles
   int wm = 2, wn = 2;
 
-  int bm = implicit_M >= 8192 && conv_params.C >= 64 ? 64 : 32;
+  int bm = implicit_M >= 8192 && C_per_group >= 64 ? 64 : 32;
   int bn = (bm == 64 || implicit_N >= 64) ? 64 : 32;
   int bk = 16;
 
@@ -192,15 +286,15 @@ void implicit_gemm_conv_2D_gpu(
 
   // Fix small channel specialization
   int n_channel_specialization = 0;
-  int channel_k_iters = ((conv_params.C + bk - 1) / bk);
+  int channel_k_iters = ((C_per_group + bk - 1) / bk);
   int gemm_k_iters = conv_params.wS[0] * conv_params.wS[1] * channel_k_iters;
 
-  if (conv_params.C <= 2) {
+  if (C_per_group <= 2) {
     gemm_k_iters = (implicit_K + bk - 1) / bk;
-    n_channel_specialization = conv_params.C;
+    n_channel_specialization = C_per_group;
-  } else if (conv_params.C <= 4) {
+  } 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) &&
@@ -242,7 +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
@@ -251,7 +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);
@@ -263,7 +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(
@@ -395,7 +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
@@ -423,7 +528,7 @@ void implicit_gemm_conv_2D_general_gpu(
       base_w.data(), sizeof(Conv2DGeneralBaseInfo) * base_w.size(), 7);
 
   // Launch kernel
-  compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+  compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
 }
 
 void winograd_conv_2D_gpu(
@@ -524,7 +629,7 @@ void winograd_conv_2D_gpu(
     MTL::Size group_dims = MTL::Size(32, bo, 1);
     MTL::Size grid_dims = MTL::Size(O_c / bo, 1, 1);
 
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+    compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
   }
 
   // Do input transform
@@ -552,7 +657,7 @@ void winograd_conv_2D_gpu(
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(N_tiles_w, N_tiles_h, N_tiles_n);
 
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+    compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
   }
 
   // Do batched gemm
@@ -600,7 +705,7 @@ void winograd_conv_2D_gpu(
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(N_tiles_w, N_tiles_h, N_tiles_n);
 
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+    compute_encoder.dispatchThreadgroups(grid_dims, group_dims);
   }
 }
 
@@ -614,6 +719,7 @@ void conv_2D_gpu(
     const std::vector<int>& wt_strides,
     const std::vector<int>& wt_dilation,
     const std::vector<int>& in_dilation,
+    const int groups,
     bool flip,
     std::vector<array>& copies) {
   // Make conv params
@@ -629,12 +735,12 @@ void conv_2D_gpu(
       /* const int kdil[NDIM] = */ {wt_dilation[0], wt_dilation[1]},
       /* const int idil[NDIM] = */ {in_dilation[0], in_dilation[1]},
       /* const size_t in_strides[NDIM + 2] = */
-      {in.strides()[0], in.strides()[1], in.strides()[2], in.strides()[3]},
+      {in.strides(0), in.strides(1), in.strides(2), in.strides(3)},
       /* const size_t wt_strides[NDIM + 2] = */
-      {wt.strides()[0], wt.strides()[1], wt.strides()[2], wt.strides()[3]},
+      {wt.strides(0), wt.strides(1), wt.strides(2), wt.strides(3)},
       /* const size_t out_strides[NDIM + 2] = */
-      {out.strides()[0], out.strides()[1], out.strides()[2], out.strides()[3]},
+      {out.strides(0), out.strides(1), out.strides(2), out.strides(3)},
-      /* const int groups = */ 1,
+      /* const int groups = */ groups,
       /* const bool flip = */ flip,
   };
 
@@ -646,6 +752,18 @@ void conv_2D_gpu(
   bool channels_large = (conv_params.C + conv_params.O) >= 512;
   bool channels_med = (conv_params.C + conv_params.O) >= 256;
 
+  if (groups > 1) {
+    const int C_per_group = conv_params.C / groups;
+    const int O_per_group = conv_params.O / groups;
+
+    if (is_idil_one && (C_per_group <= 4 || C_per_group % 16 == 0) &&
+        (O_per_group <= 16 || O_per_group % 16 == 0)) {
+      return implicit_gemm_conv_2D_gpu(s, d, in, wt, out, conv_params);
+    } else {
+      return explicit_gemm_conv_group_ND_gpu(s, d, in, wt, out, conv_params);
+    }
+  }
+
   // Direct to winograd conv
   if (!flip && is_stride_one && is_kdil_one && is_idil_one &&
       conv_params.wS[0] == 3 && conv_params.wS[1] == 3 &&
@@ -670,6 +788,56 @@ void conv_2D_gpu(
   }
 }
 
+void conv_3D_gpu(
+    const Stream& s,
+    metal::Device& d,
+    const array& in,
+    const array& wt,
+    array out,
+    const std::vector<int>& padding,
+    const std::vector<int>& wt_strides,
+    const std::vector<int>& wt_dilation,
+    const std::vector<int>& in_dilation,
+    bool flip,
+    std::vector<array>& copies) {
+  // Make conv params
+  MLXConvParams<3> conv_params{
+      /* const int  N = */ in.shape(0),
+      /* const int  C = */ in.shape(4),
+      /* const int  O = */ wt.shape(0),
+      /* const int iS[NDIM] = */ {in.shape(1), in.shape(2), in.shape(3)},
+      /* const int wS[NDIM] = */ {wt.shape(1), wt.shape(2), wt.shape(3)},
+      /* const int oS[NDIM] = */ {out.shape(1), out.shape(2), out.shape(3)},
+      /* const int str[NDIM] = */ {wt_strides[0], wt_strides[1], wt_strides[2]},
+      /* const int pad[NDIM] = */ {padding[0], padding[1], padding[2]},
+      /* const int kdil[NDIM] = */
+      {wt_dilation[0], wt_dilation[1], wt_dilation[2]},
+      /* const int idil[NDIM] = */
+      {in_dilation[0], in_dilation[1], in_dilation[2]},
+      /* const size_t in_strides[NDIM + 2] = */
+      {in.strides()[0],
+       in.strides()[1],
+       in.strides()[2],
+       in.strides()[3],
+       in.strides()[4]},
+      /* const size_t wt_strides[NDIM + 2] = */
+      {wt.strides()[0],
+       wt.strides()[1],
+       wt.strides()[2],
+       wt.strides()[3],
+       wt.strides()[4]},
+      /* const size_t out_strides[NDIM + 2] = */
+      {out.strides()[0],
+       out.strides()[1],
+       out.strides()[2],
+       out.strides()[3],
+       out.strides()[4]},
+      /* const int groups = */ 1,
+      /* const bool flip = */ flip,
+  };
+  return explicit_gemm_conv_ND_gpu(s, d, in, wt, out, conv_params);
+}
+
 } // namespace
 
 void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -694,8 +862,23 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
     wt = arr_copy;
   }
 
+  // 3D conv
+  if (out.ndim() == 5) {
+    conv_3D_gpu(
+        s,
+        d,
+        in,
+        wt,
+        out,
+        padding_,
+        kernel_strides_,
+        kernel_dilation_,
+        input_dilation_,
+        flip_,
+        copies);
+  }
   // 2D conv
-  if (out.ndim() == 4) {
+  else if (out.ndim() == 4) {
     conv_2D_gpu(
         s,
         d,
@@ -706,6 +889,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
         kernel_strides_,
         kernel_dilation_,
         input_dilation_,
+        groups_,
         flip_,
         copies);
   }
@@ -721,6 +905,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
         kernel_strides_,
         kernel_dilation_,
         input_dilation_,
+        groups_,
         flip_);
   }
   // Throw error

mlx/backend/metal/copy.cpp

@@ -4,12 +4,14 @@
 
 #include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/kernels/defines.h"
+#include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
+constexpr int MAX_COPY_SPECIALIZED_DIMS = 5;
+
 void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s) {
   if (ctype == CopyType::Vector) {
     // If the input is donateable, we are doing a vector copy and the types
@@ -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::string kernel_name;
+  {
     std::ostringstream kname;
     switch (ctype) {
       case CopyType::Scalar:
-      kname << "scopy";
+        kname << (use_2d ? "s2" : "s");
         break;
       case CopyType::Vector:
-      kname << "vcopy";
+        kname << (use_2d ? "v2" : "v");
         break;
       case CopyType::General:
-      kname << "gcopy";
+        kname << "g";
         break;
       case CopyType::GeneralGeneral:
-      kname << "ggcopy";
+        kname << "gg";
         break;
     }
-  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();
+      kname << shape.size();
     }
-  auto kernel = d.get_kernel(kname.str());
+    kname << "_copy";
+    kname << type_to_name(in) << type_to_name(out);
+    kernel_name = 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

@@ -5,6 +5,8 @@
 #include <filesystem>
 #include <sstream>
 
+#include <sys/sysctl.h>
+
 #define NS_PRIVATE_IMPLEMENTATION
 #define CA_PRIVATE_IMPLEMENTATION
 #define MTL_PRIVATE_IMPLEMENTATION
@@ -12,7 +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;
@@ -23,9 +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))
@@ -35,7 +61,6 @@ auto load_device() {
   }
   return device;
 }
-
 std::pair<MTL::Library*, NS::Error*> load_library_from_path(
     MTL::Device* device,
     const char* path) {
@@ -114,6 +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();
@@ -128,9 +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();
   }
@@ -167,10 +259,7 @@ 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()) {
-}
-
-MTL::CommandBuffer* Device::new_command_buffer(int index) {
     auto qit = queue_map_.find(index);
     if (qit == queue_map_.end()) {
       throw std::runtime_error(
@@ -187,7 +276,9 @@ MTL::CommandBuffer* Device::new_command_buffer(int index) {
     // Increment ref count so the buffer is not garbage collected
     cb->retain();
 
-  return buffer_map_.insert({index, {0, cb}}).first->second.second;
+    bit = buffer_map_.insert({index, {0, cb}}).first;
+  }
+  return bit->second.second;
 }
 
 void Device::commit_command_buffer(int index) {
@@ -198,25 +289,17 @@ void Device::commit_command_buffer(int index) {
 }
 
 void Device::end_encoding(int index) {
-  auto eit = encoder_map_.find(index);
+  encoder_map_.erase(index);
-  if (eit != encoder_map_.end()) {
-    eit->second->endEncoding();
-    eit->second->release();
-    encoder_map_.erase(eit);
-  }
 }
 
 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 =
+    eit =
-        cb->computeCommandEncoder(MTL::DispatchTypeConcurrent);
+        encoder_map_.emplace(index, std::make_unique<CommandEncoder>(cb)).first;
-    // Increment ref count so the buffer is not garbage collected
-    compute_encoder->retain();
-    eit = encoder_map_.emplace(index, CommandEncoder{compute_encoder}).first;
   }
-  return eit->second;
+  return *(eit->second);
 }
 
 void Device::register_library(
@@ -228,13 +311,9 @@ void Device::register_library(
   }
 }
 
-void Device::register_library(
+void Device::register_library(const std::string& lib_name) {
-    const std::string& lib_name,
-    const std::function<std::string(const std::string&)>& lib_path_func) {
   if (auto it = library_map_.find(lib_name); it == library_map_.end()) {
-    std::string new_lib_path = lib_path_func(lib_name);
+    register_library(lib_name, get_colocated_mtllib_path(lib_name));
-    auto new_lib = load_library(device_, lib_name, new_lib_path.c_str());
-    library_map_.insert({lib_name, new_lib});
   }
 }
 
@@ -244,7 +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];
   }
 
@@ -258,13 +337,16 @@ MTL::Library* Device::get_library_(const std::string& source_string) {
       NS::String::string(source_string.c_str(), NS::ASCIIStringEncoding);
 
   NS::Error* error = nullptr;
-  auto mtl_lib = device_->newLibrary(ns_code, nullptr, &error);
+  auto options = MTL::CompileOptions::alloc()->init();
+  options->setFastMathEnabled(false);
+  options->setLanguageVersion(get_metal_version());
+  auto mtl_lib = device_->newLibrary(ns_code, options, &error);
+  options->release();
 
   // Throw error if unable to compile library
   if (!mtl_lib) {
     std::ostringstream msg;
-    msg << "[metal::Device] Unable to load build metal library from source"
+    msg << "[metal::Device] Unable to build metal library from source" << "\n";
-        << "\n";
     if (error) {
       msg << error->localizedDescription()->utf8String() << "\n";
     }
@@ -283,8 +365,7 @@ MTL::Library* Device::get_library_(const MTL::StitchedLibraryDescriptor* desc) {
   // Throw error if unable to compile library
   if (!mtl_lib) {
     std::ostringstream msg;
-    msg << "[metal::Device] Unable to load build stitched metal library"
+    msg << "[metal::Device] Unable to build stitched metal library" << "\n";
-        << "\n";
     if (error) {
       msg << error->localizedDescription()->utf8String() << "\n";
     }
@@ -342,7 +423,6 @@ MTL::Function* Device::get_function_(
   }
 
   mtl_func_consts->release();
-  desc->release();
 
   return mtl_function;
 }
@@ -511,11 +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;
 }
 
@@ -542,11 +624,12 @@ Device& device(mlx::core::Device) {
   return metal_device;
 }
 
-std::shared_ptr<void> new_scoped_memory_pool() {
+std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool() {
   auto dtor = [](void* ptr) {
     static_cast<NS::AutoreleasePool*>(ptr)->release();
   };
-  return std::shared_ptr<void>(NS::AutoreleasePool::alloc()->init(), dtor);
+  return std::unique_ptr<void, std::function<void(void*)>>(
+      NS::AutoreleasePool::alloc()->init(), dtor);
 }
 
 void new_stream(Stream stream) {
@@ -555,4 +638,23 @@ void new_stream(Stream stream) {
   }
 }
 
+std::unordered_map<std::string, std::variant<std::string, size_t>>
+device_info() {
+  auto raw_device = device(default_device()).mtl_device();
+  auto arch = std::string(raw_device->architecture()->name()->utf8String());
+
+  int mib[] = {CTL_HW, HW_MEMSIZE};
+  size_t memsize = 0;
+  size_t length = sizeof(memsize);
+
+  sysctl(mib, 2, &memsize, &length, NULL, 0);
+
+  return {
+      {"architecture", arch},
+      {"max_buffer_length", raw_device->maxBufferLength()},
+      {"max_recommended_working_set_size",
+       raw_device->recommendedMaxWorkingSetSize()},
+      {"memory_size", memsize}};
+}
+
 } // namespace mlx::core::metal

mlx/backend/metal/device.h

@@ -9,36 +9,17 @@
 #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)
+  CommandEncoder(MTL::CommandBuffer* cbuf);
-      : enc(enc), concurrent(false){};
+  CommandEncoder(const CommandEncoder&) = delete;
   CommandEncoder& operator=(const CommandEncoder&) = delete;
 
   struct ConcurrentContext {
@@ -60,41 +41,24 @@ struct CommandEncoder {
     return enc;
   }
 
-  void set_input_array(const array& a, int idx, int offset = 0) {
+  void set_input_array(const array& a, int idx, int64_t offset = 0);
-    auto r_buf =
+  void set_output_array(array& a, int idx, int64_t offset = 0);
-        static_cast<MTL::Resource*>(const_cast<void*>(a.buffer().ptr()));
+  void dispatchThreadgroups(MTL::Size grid_dims, MTL::Size group_dims);
-    if (auto it = outputs.find(r_buf); it != outputs.end()) {
+  void dispatchThreads(MTL::Size grid_dims, MTL::Size group_dims);
-      // 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);
-    }
-  }
 
   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;
 };
@@ -111,7 +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);
@@ -122,10 +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,
+  void register_library(const std::string& lib_name);
-      const std::function<std::string(const std::string&)>& lib_path_func =
-          get_colocated_mtllib_path);
 
   MTL::Library* get_library(const std::string& name);
 
@@ -197,7 +158,7 @@ class Device {
   MTL::Device* device_;
   std::unordered_map<int32_t, MTL::CommandQueue*> queue_map_;
   std::unordered_map<int32_t, std::pair<int, MTL::CommandBuffer*>> buffer_map_;
-  std::unordered_map<int32_t, CommandEncoder> encoder_map_;
+  std::unordered_map<int32_t, std::unique_ptr<CommandEncoder>> encoder_map_;
   std::unordered_map<std::string, MTL::ComputePipelineState*> kernel_map_;
   std::unordered_map<std::string, MTL::Library*> library_map_;
   std::mutex mtx_;

mlx/backend/metal/event.cpp

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