bumpity bump (#987)

* rebase

* nit

* fix eval in vmap

.circleci/config.yml

@@ -31,8 +31,7 @@ jobs:
           name: Install dependencies
           command: |
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
-            pip install pybind11-stubgen
+            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
             pip install numpy
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
@@ -44,7 +43,8 @@ jobs:
       - run:
           name: Generate package stubs
           command: |
-            python3 setup.py generate_stubs
+            echo "stubs"
+            python setup.py generate_stubs 
       - run:
           name: Run Python tests
           command: |
@@ -63,21 +63,24 @@ jobs:
           command: ./build/tests/tests
 
   mac_build_and_test:
+    parameters:
+      xcode_version:
+        type: string
+        default: "15.2.0"
     macos:
-      xcode: "15.2.0"
+      xcode: << parameters.xcode_version >>
     resource_class: macos.m1.large.gen1
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
-            brew install python@3.9
-            python3.9 -m venv env
+            brew install python@3.8
+            python3.8 -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
-            pip install pybind11-stubgen
+            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
             pip install numpy
             pip install torch
             pip install tensorflow
@@ -91,13 +94,13 @@ jobs:
           name: Generate package stubs
           command: |
             source env/bin/activate
-            python setup.py generate_stubs
+            python setup.py generate_stubs 
       - run:
           name: Run Python tests
           command: |
             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 python3.9 -m xmlrunner discover -v python/tests -o test-results/gpu
+            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
       # TODO: Reenable when extension api becomes stable
       # - run:
       #     name: Build example extension
@@ -140,9 +143,8 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
+            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
             pip install --upgrade setuptools
-            pip install pybind11-stubgen
             pip install numpy
             pip install twine
             pip install build
@@ -157,7 +159,7 @@ jobs:
           name: Generate package stubs
           command: |
             source env/bin/activate
-            python setup.py generate_stubs
+            python setup.py generate_stubs 
       - run:
           name: Build Python package
           command: |
@@ -205,9 +207,8 @@ jobs:
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install --upgrade pybind11[global]
+            pip install git+https://github.com/wjakob/nanobind.git@4148debcf91f5ccab0c3b8d67b5c3cabd61f407f
             pip install --upgrade setuptools
-            pip install pybind11-stubgen
             pip install numpy
             pip install auditwheel
             pip install patchelf
@@ -215,7 +216,7 @@ jobs:
             << parameters.extra_env >> \
               CMAKE_BUILD_PARALLEL_LEVEL="" \
               pip install . -v
-            python setup.py generate_stubs
+            python setup.py generate_stubs 
             << parameters.extra_env >> \
               CMAKE_BUILD_PARALLEL_LEVEL="" \
               python -m build --wheel
@@ -235,7 +236,10 @@ workflows:
         - not: << pipeline.parameters.weekly_build >>
         - not: << pipeline.parameters.test_release >>
     jobs:
-      - mac_build_and_test
+      - mac_build_and_test:
+          matrix:
+            parameters:
+              xcode_version: ["15.0.0", "15.2.0"]
       - linux_build_and_test
 
   build_pypi_release:
@@ -254,7 +258,7 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["14.3.1", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0"]
               build_env: ["PYPI_RELEASE=1"]
   prb:
     when:
@@ -268,6 +272,9 @@ workflows:
           context: pr-approval
       - mac_build_and_test:
           requires: [ hold ]
+          matrix:
+            parameters:
+              xcode_version: ["15.0.0", "15.2.0"]
       - linux_build_and_test:
           requires: [ hold ]
   nightly_build:
@@ -280,7 +287,7 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["14.3.1", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0"]
   weekly_build:
     when:
       and:
@@ -291,7 +298,7 @@ workflows:
           matrix:
             parameters:
               python_version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
-              xcode_version: ["14.3.1", "15.2.0"]
+              xcode_version: ["15.0.0", "15.2.0"]
               build_env: ["DEV_RELEASE=1"]
   linux_test_release:
     when:

.pre-commit-config.yaml

@@ -1,11 +1,11 @@
 repos:
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v17.0.6
+    rev: v18.1.3
     hooks:
     -   id: clang-format
 # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 24.2.0
+    rev: 24.3.0
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort

ACKNOWLEDGMENTS.md

@@ -15,6 +15,8 @@ MLX was developed with contributions from the following individuals:
 - Hinrik Snær Guðmundsson: Added `atleast_1d`, `atleast_2d`, `atleast_3d` ops.
 - Luca Arnaboldi: Added `Ceil` and `Floor` ops; implemented pickling, copy and deepcopy for mlx arrays.
 - Brian Keene & Atila Orhon, with Argmax Inc.: Added `fast.scaled_dot_product_attention`
+- AmirHossein Razlighi: Added chaining support for some of the ops in `nn.Module`. Comparison works for non array objects in `mlx.core.array`. Exception handling for invalid operations in `mlx.core.array`.
+
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>

CMakeLists.txt

@@ -15,31 +15,33 @@ 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_METAL_DEBUG "Enhance metal debug workflow" OFF)
+option(MLX_ENABLE_X64_MAC "Enable building for x64 macOS" OFF)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
 
 if(NOT MLX_VERSION)
-  set(MLX_VERSION 0.5.1)
+  set(MLX_VERSION 0.10.0)
 endif()
 
 # --------------------- Processor tests -------------------------
 
-message(STATUS "Building MLX for ${CMAKE_HOST_SYSTEM_PROCESSOR} processor on ${CMAKE_SYSTEM_NAME}")
+message(STATUS "Building MLX for ${CMAKE_SYSTEM_PROCESSOR} processor on ${CMAKE_SYSTEM_NAME}")
 
 set(MLX_BUILD_ARM OFF)
 
 if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
-  if (${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "x86_64" AND ${CMAKE_HOST_APPLE})
-    message(FATAL_ERROR
-      "Building for x86_64 on macOS is not supported."
-      " If you are on an Apple silicon system, check the build"
-      " documentation for possible fixes: "
-      "https://ml-explore.github.io/mlx/build/html/install.html#build-from-source")
-  elseif (${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "x86_64")
-    message(WARNING
-      "Building for x86_64 on macOS is not supported."
-      " If you are on an Apple silicon system, "
-      " make sure you are building for arm64.")
-  elseif(${CMAKE_HOST_SYSTEM_PROCESSOR} MATCHES "arm64")
+  if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
+    if(NOT MLX_ENABLE_X64_MAC)
+      message(FATAL_ERROR
+        "Building for x86_64 on macOS is not supported."
+        " If you are on an Apple silicon system, check the build"
+        " documentation for possible fixes: "
+        "https://ml-explore.github.io/mlx/build/html/install.html#build-from-source")
+    else()
+      message(WARNING "Building for x86_64 arch is not officially supported.")
+    endif()
+    set(MLX_BUILD_METAL OFF)
+  elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm64")
     set(MLX_BUILD_ARM ON)
   endif()
 
@@ -64,8 +66,14 @@ endif()
 if (MLX_BUILD_METAL AND NOT METAL_LIB)
   message(STATUS "Metal not found. Unable to build GPU")
   set(MLX_BUILD_METAL OFF)
+  set(MLX_METAL_DEBUG OFF)
 elseif (MLX_BUILD_METAL)
   message(STATUS "Building METAL sources")
+
+  if (MLX_METAL_DEBUG)
+    add_compile_definitions(MLX_METAL_DEBUG)
+  endif()
+
   # Throw an error if xcrun not found
   execute_process(COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
                   OUTPUT_VARIABLE MACOS_VERSION
@@ -77,10 +85,8 @@ elseif (MLX_BUILD_METAL)
     set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14.2_iOS17.2.zip)
   elseif (${MACOS_VERSION} GREATER_EQUAL 14.0)
     set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS14_iOS17-beta.zip)
-  elseif (${MACOS_VERSION} GREATER_EQUAL 13.3)
-    set(METAL_CPP_URL https://developer.apple.com/metal/cpp/files/metal-cpp_macOS13.3_iOS16.4.zip)
   else()
-    message(FATAL_ERROR "MLX requires macOS >= 13.4 to be built with MLX_BUILD_METAL=ON" )
+    message(FATAL_ERROR "MLX requires macOS SDK >= 14.0 to be built with MLX_BUILD_METAL=ON" )
   endif()
 
   FetchContent_Declare(
@@ -110,7 +116,27 @@ if (MLX_BUILD_ARM AND ACCELERATE_LIBRARY)
 else()
   message(STATUS "Accelerate or arm neon not found, using default backend.")
   set(MLX_BUILD_ACCELERATE OFF)
-  #set(BLA_VENDOR Generic)
+  if(${CMAKE_HOST_APPLE})
+    # The blas shipped in macOS SDK is not supported, search homebrew for
+    # openblas instead.
+    set(BLA_VENDOR OpenBLAS)
+    set(LAPACK_ROOT "${LAPACK_ROOT};$ENV{LAPACK_ROOT};/usr/local/opt/openblas")
+  endif()
+  # Search and link with lapack.
+  find_package(LAPACK REQUIRED)
+  if (NOT LAPACK_FOUND)
+    message(FATAL_ERROR "Must have LAPACK installed")
+  endif()
+  find_path(LAPACK_INCLUDE_DIRS lapacke.h
+    /usr/include
+    /usr/local/include
+    /usr/local/opt/openblas/include)
+  message(STATUS "Lapack lib " ${LAPACK_LIBRARIES})
+  message(STATUS "Lapack include " ${LAPACK_INCLUDE_DIRS})
+  target_include_directories(mlx PRIVATE ${LAPACK_INCLUDE_DIRS})
+  target_link_libraries(mlx ${LAPACK_LIBRARIES})
+  # List blas after lapack otherwise we may accidentally incldue an old version
+  # of lapack.h from the include dirs of blas.
   find_package(BLAS REQUIRED)
   if (NOT BLAS_FOUND)
     message(FATAL_ERROR "Must have BLAS installed")
@@ -124,17 +150,6 @@ else()
   message(STATUS "Blas include " ${BLAS_INCLUDE_DIRS})
   target_include_directories(mlx PRIVATE ${BLAS_INCLUDE_DIRS})
   target_link_libraries(mlx ${BLAS_LIBRARIES})
-  find_package(LAPACK REQUIRED)
-  if (NOT LAPACK_FOUND)
-      message(FATAL_ERROR "Must have LAPACK installed")
-  endif()
-  find_path(LAPACK_INCLUDE_DIRS lapacke.h
-    /usr/include
-    /usr/local/include)
-  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})
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
@@ -148,8 +163,12 @@ target_include_directories(
 
 if (MLX_BUILD_PYTHON_BINDINGS)
   message(STATUS "Building Python bindings.")
-  find_package(Python COMPONENTS Interpreter Development)
-  find_package(pybind11 CONFIG REQUIRED)
+  find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)
+  execute_process(
+    COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
+    OUTPUT_STRIP_TRAILING_WHITESPACE OUTPUT_VARIABLE NB_DIR)
+  list(APPEND CMAKE_PREFIX_PATH "${NB_DIR}")
+  find_package(nanobind CONFIG REQUIRED)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/python/src)
 endif()
 

benchmarks/cpp/time_utils.h

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

benchmarks/python/layer_norm_bench.py

@@ -0,0 +1,41 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import mlx.core as mx
+import mlx.nn as nn
+from time_utils import time_fn
+
+
+def layer_norm(x, w, b, eps):
+    ot = x.dtype
+    x = x.astype(mx.float32)
+    mu = mx.mean(x, -1, keepdims=True)
+    v = mx.var(x, -1, keepdims=True)
+    return (x - mu) * mx.rsqrt(v + eps) * w + b
+
+
+def time_layer_norm():
+    f1 = lambda x, w, b, y: (layer_norm(x, w, b, 1e-5) * y).sum()
+    f2 = lambda x, w, b, y: (mx.fast.layer_norm(x, w, b, 1e-5) * y).sum()
+    g1 = mx.grad(f1, argnums=(0, 1, 2))
+    g2 = mx.grad(f2, argnums=(0, 1, 2))
+
+    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    b = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    mx.eval(x, w, b, y)
+
+    def layer_norm_loop(g, x, w, b):
+        gx, gw, gb = x, w, b
+        for _ in range(32):
+            gx, gw, gb = g(gx, gw, gb, y)
+        return gx, gw, gb
+
+    time_fn(layer_norm_loop, g1, x, w, b)
+    time_fn(layer_norm_loop, g2, x, w, b)
+    time_fn(layer_norm_loop, mx.compile(g1), x, w, b)
+    time_fn(layer_norm_loop, mx.compile(g2), x, w, b)
+
+
+if __name__ == "__main__":
+    time_layer_norm()

benchmarks/python/rms_norm_bench.py

@@ -0,0 +1,39 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import mlx.core as mx
+import mlx.nn as nn
+from time_utils import time_fn
+
+
+def rms_norm(x, w, eps):
+    ot = x.dtype
+    x = x.astype(mx.float32)
+    n = mx.rsqrt(x.square().mean(-1, keepdims=True) + eps)
+    return (x * n).astype(ot) * w
+
+
+def time_rms_norm():
+    f1 = lambda x, w, y: (rms_norm(x, w, 1e-5) * y).sum()
+    f2 = lambda x, w, y: (mx.fast.rms_norm(x, w, 1e-5) * y).sum()
+    g1 = mx.grad(f1, argnums=(0, 1))
+    g2 = mx.grad(f2, argnums=(0, 1))
+
+    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
+    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    mx.eval(x, w, y)
+
+    def rms_norm_loop(g, x, w):
+        gx, gw = x, w
+        for _ in range(32):
+            gx, gw = g(gx, gw, y)
+        return gx, gw
+
+    time_fn(rms_norm_loop, g1, x, w)
+    time_fn(rms_norm_loop, g2, x, w)
+    time_fn(rms_norm_loop, mx.compile(g1), x, w)
+    time_fn(rms_norm_loop, mx.compile(g2), x, w)
+
+
+if __name__ == "__main__":
+    time_rms_norm()

benchmarks/python/rope_bench.py

@@ -6,21 +6,21 @@ from time_utils import time_fn
 
 
 def time_rope():
-    rope = nn.RoPE(4096)
+    rope = nn.RoPE(64)
 
     # vec
-    x = mx.random.uniform(shape=(1, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(1, 32, 1, 128)).astype(mx.float16)
     mx.eval(x)
 
     def rope_vec(x):
         for _ in range(32):
-            x = rope(x)
+            x = rope(x, offset=100)
         return x
 
     time_fn(rope_vec, x)
 
     # matrix
-    x = mx.random.uniform(shape=(1024, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(1, 32, 1024, 128)).astype(mx.float16)
     mx.eval(x)
 
     def rope_mat(x):

docs/src/conf.py

@@ -29,16 +29,17 @@ autosummary_generate = True
 autosummary_filename_map = {"mlx.core.Stream": "stream_class"}
 
 intersphinx_mapping = {
-    "https://docs.python.org/3": None,
-    "https://numpy.org/doc/stable/": None,
+    "python": ("https://docs.python.org/3", None),
+    "numpy": ("https://numpy.org/doc/stable/", None),
 }
 
 templates_path = ["_templates"]
 html_static_path = ["_static"]
 source_suffix = ".rst"
-master_doc = "index"
+main_doc = "index"
 highlight_language = "python"
 pygments_style = "sphinx"
+add_module_names = False
 
 # -- Options for HTML output -------------------------------------------------
 
@@ -59,3 +60,22 @@ html_theme_options = {
 # -- Options for HTMLHelp output ---------------------------------------------
 
 htmlhelp_basename = "mlx_doc"
+
+
+def setup(app):
+    from sphinx.util import inspect
+
+    wrapped_isfunc = inspect.isfunction
+
+    def isfunc(obj):
+        type_name = str(type(obj))
+        if "nanobind.nb_method" in type_name or "nanobind.nb_func" in type_name:
+            return True
+        return wrapped_isfunc(obj)
+
+    inspect.isfunction = isfunc
+
+
+# -- Options for LaTeX output ------------------------------------------------
+
+latex_documents = [(main_doc, "MLX.tex", "MLX Documentation", author, "manual")]

docs/src/dev/extensions.rst

@@ -1,24 +1,16 @@
 Developer Documentation
 =======================
 
-MLX provides a open and flexible backend to which users may add operations 
-and specialized implementations without much hassle. While the library supplies
-efficient operations that can be used and composed for any number of 
-applications, there may arise cases where new functionalities or highly 
-optimized implementations are needed. For such cases, you may design and 
-implement your own operations that link to and build on top of :mod:`mlx.core`.
-We will introduce the inner-workings of MLX and go over a simple example to 
-learn the steps involved in adding new operations to MLX with your own CPU 
-and GPU implementations. 
+You can extend MLX with custom operations on the CPU or GPU. This guide
+explains how to do that with a simple example.
 
 Introducing the Example
 -----------------------
 
-Let's say that you would like an operation that takes in two arrays, 
-``x`` and ``y``, scales them both by some coefficients ``alpha`` and ``beta``
-respectively, and then adds them together to get the result 
-``z = alpha * x + beta * y``. Well, you can very easily do that by just 
-writing out a function as follows:
+Let's say you would like an operation that takes in two arrays, ``x`` and
+``y``, scales them both by coefficients ``alpha`` and ``beta`` respectively,
+and then adds them together to get the result ``z = alpha * x + beta * y``.
+You can do that in MLX directly:
 
 .. code-block:: python
 
@@ -27,44 +19,35 @@ writing out a function as follows:
     def simple_axpby(x: mx.array, y: mx.array, alpha: float, beta: float) -> mx.array:
         return alpha * x + beta * y
 
-This function performs that operation while leaving the implementations and 
-differentiation to MLX. 
+This function performs that operation while leaving the implementation and
+function transformations to MLX.
 
-However, you work with vector math libraries often and realize that the 
-``axpby`` routine defines the same operation ``Y = (alpha * X) + (beta * Y)``. 
-You would really like the part of your applications that does this operation 
-on the CPU to be very fast - so you decide that you want it to rely on the 
-``axpby`` routine provided by the Accelerate_ framework. Continuing to impose 
-our assumptions on to you, let's also assume that you want to learn how to add 
-your own implementation for the gradients of your new operation while going 
-over the ins-and-outs of the MLX framework. 
+However you may need to customize the underlying implementation, perhaps to
+make it faster or for custom differentiation. In this tutorial we will go
+through adding custom extensions. It will cover:
 
-Well, what a coincidence! You are in the right place. Over the course of this 
-example, we will learn:
-
-* The structure of the MLX library from the frontend API to the backend implementations.
-* How to implement your own CPU backend that redirects to Accelerate_ when appropriate (and a fallback if needed).
-* How to implement your own GPU implementation using metal.
-* How to add your own ``vjp`` and ``jvp``.
-* How to build your implementations, link them to MLX, and bind them to python.
+* The structure of the MLX library.
+* Implementing a CPU operation that redirects to Accelerate_ when appropriate.
+* Implementing a GPU operation using metal.
+* Adding the ``vjp`` and ``jvp`` function transformation.
+* Building a custom extension and binding it to python.
 
 Operations and Primitives
 -------------------------
 
-In one sentence, operations in MLX build the computation graph, and primitives 
-provide the rules for evaluation and transformations of said graph. Let's start 
-by discussing operations in more detail. 
+Operations in MLX build the computation graph. Primitives provide the rules for
+evaluating and transforming the graph. Let's start by discussing operations in
+more detail.
 
 Operations
 ^^^^^^^^^^^
 
-Operations are the frontend functions that operate on arrays. They are defined 
-in the C++ API (:ref:`cpp_ops`) and then we provide bindings to these 
-operations in the Python API (:ref:`ops`). 
+Operations are the front-end functions that operate on arrays. They are defined
+in the C++ API (:ref:`cpp_ops`), and the Python API (:ref:`ops`) binds them.
 
-We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and ``y``,
-and two scalars, ``alpha`` and ``beta``. This is how we would define it in the 
-C++ API:
+We would like an operation, :meth:`axpby` that takes in two arrays ``x`` and
+``y``, and two scalars, ``alpha`` and ``beta``. This is how to define it in
+C++:
 
 .. code-block:: C++
 
@@ -83,10 +66,7 @@ C++ API:
         StreamOrDevice s = {} // Stream on which to schedule the operation
     );
 
-
-This operation itself can call other operations within it if needed. So, the 
-simplest way to go about implementing this operation would be do so in terms 
-of existing operations. 
+The simplest way to this operation is in terms of existing operations:
 
 .. code-block:: C++
 
@@ -100,25 +80,23 @@ of existing operations.
         // Scale x and y on the provided stream
         auto ax = multiply(array(alpha), x, s);
         auto by = multiply(array(beta), y, s);
-        
+
         // Add and return
         return add(ax, by, s);
     }
 
-However, as we discussed earlier, this is not our goal. The operations themselves 
-do not contain the implementations that act on the data, nor do they contain the
-rules of transformations. Rather, they are an easy to use interface that build 
-on top of the building blocks we call :class:`Primitive`. 
+The operations themselves do not contain the implementations that act on the
+data, nor do they contain the rules of transformations. Rather, they are an
+easy to use interface that use :class:`Primitive` building blocks.
 
 Primitives
 ^^^^^^^^^^^
 
-A :class:`Primitive` is part of the computation graph of an :class:`array`. It 
-defines how to create an output given a set of input :class:`array` . Further,
-a :class:`Primitive` is a class that contains rules on how it is evaluated 
-on the CPU or GPU, and how it acts under transformations such as ``vjp`` and 
-``jvp``. These words on their own can be a bit abstract, so lets take a step 
-back and go to our example to give ourselves a more concrete image. 
+A :class:`Primitive` is part of the computation graph of an :class:`array`. It
+defines how to create outputs arrays given a input arrays. Further, a
+:class:`Primitive` has methods to run on the CPU or GPU and for function
+transformations such as ``vjp`` and ``jvp``.  Lets go back to our example to be
+more concrete:
 
 .. code-block:: C++
 
@@ -134,11 +112,15 @@ back and go to our example to give ourselves a more concrete image.
         * To avoid unnecessary allocations, the evaluation function
         * is responsible for allocating space for the array.
         */
-        void eval_cpu(const std::vector<array>& inputs, array& out) override;
-        void eval_gpu(const std::vector<array>& inputs, array& out) override;
+        void eval_cpu(
+            const std::vector<array>& inputs,
+            std::vector<array>& outputs) override;
+        void eval_gpu(
+            const std::vector<array>& inputs,
+            std::vector<array>& outputs) override;
 
         /** The Jacobian-vector product. */
-        array jvp(
+        std::vector<array> jvp(
             const std::vector<array>& primals,
             const std::vector<array>& tangents,
             const std::vector<int>& argnums) override;
@@ -147,7 +129,8 @@ back and go to our example to give ourselves a more concrete image.
         std::vector<array> vjp(
             const std::vector<array>& primals,
             const array& cotan,
-            const std::vector<int>& argnums) override;
+            const std::vector<int>& argnums,
+            const std::vector<array>& outputs) override;
 
         /**
         * The primitive must know how to vectorize itself across
@@ -155,7 +138,7 @@ back and go to our example to give ourselves a more concrete image.
         * representing the vectorized computation and the axis which
         * corresponds to the output vectorized dimension.
         */
-        std::pair<array, int> vmap(
+        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) override;
 
@@ -175,22 +158,22 @@ back and go to our example to give ourselves a more concrete image.
         void eval(const std::vector<array>& inputs, array& out);
     };
 
-The :class:`Axpby` class derives from the base :class:`Primitive` class and 
-follows the above demonstrated interface. :class:`Axpby` treats ``alpha`` and 
-``beta`` as parameters. It then provides implementations of how the array ``out`` 
-is produced given ``inputs`` through :meth:`Axpby::eval_cpu` and 
-:meth:`Axpby::eval_gpu`. Further, it provides rules of transformations in 
-:meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and :meth:`Axpby::vmap`. 
+The :class:`Axpby` class derives from the base :class:`Primitive` class. The
+:class:`Axpby` treats ``alpha`` and ``beta`` as parameters. It then provides
+implementations of how the output array is produced given the inputs through
+:meth:`Axpby::eval_cpu` and :meth:`Axpby::eval_gpu`. It also provides rules
+of transformations in :meth:`Axpby::jvp`, :meth:`Axpby::vjp`, and
+:meth:`Axpby::vmap`.
 
-Using the Primitives
-^^^^^^^^^^^^^^^^^^^^^
+Using the Primitive
+^^^^^^^^^^^^^^^^^^^
 
-Operations can use this :class:`Primitive` to add a new :class:`array` to 
-the computation graph. An :class:`array` can be constructed by providing its 
-data type, shape, the :class:`Primitive` that computes it, and the 
-:class:`array` inputs that are passed to the primitive.
+Operations can use this :class:`Primitive` to add a new :class:`array` to the
+computation graph. An :class:`array` can be constructed by providing its data
+type, shape, the :class:`Primitive` that computes it, and the :class:`array`
+inputs that are passed to the primitive.
 
-Let's re-implement our operation now in terms of our :class:`Axpby` primitive.
+Let's reimplement our operation now in terms of our :class:`Axpby` primitive.
 
 .. code-block:: C++
 
@@ -223,7 +206,7 @@ Let's re-implement our operation now in terms of our :class:`Axpby` primitive.
             /* const std::vector<int>& shape = */ out_shape,
             /* Dtype dtype = */ out_dtype,
             /* std::unique_ptr<Primitive> primitive = */
-            std::make_unique<Axpby>(to_stream(s), alpha, beta),
+            std::make_shared<Axpby>(to_stream(s), alpha, beta),
             /* const std::vector<array>& inputs = */ broadcasted_inputs);
     }
 
@@ -238,27 +221,26 @@ This operation now handles the following:
 Implementing the Primitive
 --------------------------
 
-No computation happens when we call the operation alone. In effect, the 
-operation only builds the computation graph. When we evaluate the output 
-array, MLX schedules the execution of the computation graph, and calls
-:meth:`Axpby::eval_cpu` or :meth:`Axpby::eval_gpu` depending on the 
-stream/device specified by the user. 
+No computation happens when we call the operation alone. The operation only
+builds the computation graph. When we evaluate the output array, MLX schedules
+the execution of the computation graph, and calls :meth:`Axpby::eval_cpu` or
+:meth:`Axpby::eval_gpu` depending on the stream/device specified by the user.
 
 .. warning::
     When :meth:`Primitive::eval_cpu` or :meth:`Primitive::eval_gpu` are called,
     no memory has been allocated for the output array. It falls on the implementation
-    of these functions to allocate memory as needed
+    of these functions to allocate memory as needed.
 
-Implementing the CPU Backend
+Implementing the CPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Let's start by trying to implement a naive and generic version of 
-:meth:`Axpby::eval_cpu`. We declared this as a private member function of 
-:class:`Axpby` earlier called :meth:`Axpby::eval`. 
+Let's start by implementing a naive and generic version of
+:meth:`Axpby::eval_cpu`. We declared this as a private member function of
+:class:`Axpby` earlier called :meth:`Axpby::eval`.
 
-Our naive method will go over each element of the output array, find the 
-corresponding input elements of ``x`` and ``y`` and perform the operation 
-pointwise. This is captured in the templated function :meth:`axpby_impl`. 
+Our naive method will go over each element of the output array, find the
+corresponding input elements of ``x`` and ``y`` and perform the operation
+point-wise. This is captured in the templated function :meth:`axpby_impl`.
 
 .. code-block:: C++
 
@@ -296,19 +278,19 @@ pointwise. This is captured in the templated function :meth:`axpby_impl`.
         }
     }
 
-Now, we would like our implementation to be able to do this pointwise operation 
-for all incoming floating point arrays. Accordingly, we add dispatches for 
-``float32``, ``float16``, ``bfloat16`` and ``complex64``. We throw an error 
-if we encounter an unexpected type.
+Our implementation should work for all incoming floating point arrays.
+Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
+``complex64``. We throw an error if we encounter an unexpected type.
 
 .. code-block:: C++
 
     /** Fall back implementation for evaluation on CPU */
-    void Axpby::eval(const std::vector<array>& inputs, array& out) {
-        // Check the inputs (registered in the op while constructing the out array)
-        assert(inputs.size() == 2);
+    void Axpby::eval(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Dispatch to the correct dtype
         if (out.dtype() == float32) {
@@ -321,28 +303,26 @@ if we encounter an unexpected type.
             return axpby_impl<complex64_t>(x, y, out, alpha_, beta_);
         } else {
             throw std::runtime_error(
-                "Axpby is only supported for floating point types.");
+                "[Axpby] Only supports floating point types.");
         }
     }
 
-We have a fallback implementation! Now, to do what we are really here to do. 
-Remember we wanted to use the ``axpby`` routine provided by the Accelerate_
-framework? Well, there are 3 complications to keep in mind:
+This is good as a fallback implementation. We can use the ``axpby`` routine
+provided by the Accelerate_ framework for a faster implementation in certain
+cases:
 
 #.  Accelerate does not provide implementations of ``axpby`` for half precision
-    floats. We can only direct to it for ``float32`` types 
-#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all elements
-    have fixed strides between them. Possibly due to broadcasts and transposes, 
-    we aren't guaranteed that the inputs fit this requirement. We can 
-    only direct to Accelerate if both ``x`` and ``y`` are row contiguous or 
-    column contiguous. 
-#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` inplace. 
-    MLX expects to write out the answer to a new array. We must copy the elements 
-    of ``y`` into the output array and use that as an input to ``axpby``
+    floats. We can only use it for ``float32`` types.
+#.  Accelerate assumes the inputs ``x`` and ``y`` are contiguous and all
+    elements have fixed strides between them. We only direct to Accelerate
+    if both ``x`` and ``y`` are row contiguous or column contiguous.
+#.  Accelerate performs the routine ``Y = (alpha * X) + (beta * Y)`` in-place.
+    MLX expects to write the output to a new array. We must copy the elements
+    of ``y`` into the output and use that as an input to ``axpby``.
 
-Let's write out an implementation that uses Accelerate in the right conditions. 
-It must simply allocate data for the output, copy elements of ``y`` into it, 
-and then call the :meth:`catlas_saxpby` from accelerate. 
+Let's write an implementation that uses Accelerate in the right conditions.
+It allocates data for the output, copies ``y`` into it, and then calls the
+:func:`catlas_saxpby` from accelerate.
 
 .. code-block:: C++
 
@@ -356,17 +336,7 @@ and then call the :meth:`catlas_saxpby` from accelerate.
         // Accelerate library provides catlas_saxpby which does
         // Y = (alpha * X) + (beta * Y) in place
         // To use it, we first copy the data in y over to the output array
-
-        // This specialization requires both x and y be contiguous in the same mode
-        // i.e: corresponding linear indices in both point to corresponding elements
-        // The data in the output array is allocated to match the strides in y
-        // such that x, y, and out are contiguous in the same mode and
-        // no transposition is needed
-        out.set_data(
-            allocator::malloc_or_wait(y.data_size() * out.itemsize()),
-            y.data_size(),
-            y.strides(),
-            y.flags());
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
         // We then copy over the elements using the contiguous vector specialization
         copy_inplace(y, out, CopyType::Vector);
@@ -389,18 +359,20 @@ and then call the :meth:`catlas_saxpby` from accelerate.
             /* INCY = */ 1);
     }
 
-Great! But what about the inputs that do not fit the criteria for accelerate?
-Luckily, we can always just direct back to :meth:`Axpby::eval`.
-
-With this in mind, lets finally implement our :meth:`Axpby::eval_cpu`.
+For inputs that do not fit the criteria for accelerate, we fall back to
+:meth:`Axpby::eval`. With this in mind, let's finish our
+:meth:`Axpby::eval_cpu`.
 
 .. code-block:: C++
 
     /** Evaluate primitive on CPU using accelerate specializations */
-    void Axpby::eval_cpu(const std::vector<array>& inputs, array& out) {
+    void Axpby::eval_cpu(
+      const std::vector<array>& inputs,
+      const std::vector<array>& outputs) {
         assert(inputs.size() == 2);
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Accelerate specialization for contiguous single precision float arrays
         if (out.dtype() == float32 &&
@@ -410,35 +382,33 @@ With this in mind, lets finally implement our :meth:`Axpby::eval_cpu`.
             return;
         }
 
-        // Fall back to common backend if specializations are not available
-        eval(inputs, out);
+        // Fall back to common back-end if specializations are not available
+        eval(inputs, outputs);
     }
 
-We have now hit a milestone! Just this much is enough to run the operation 
-:meth:`axpby` on a CPU stream! 
+Just this much is enough to run the operation :meth:`axpby` on a CPU stream! If
+you do not plan on running the operation on the GPU or using transforms on
+computation graphs that contain :class:`Axpby`, you can stop implementing the
+primitive here and enjoy the speed-ups you get from the Accelerate library.
 
-If you do not plan on running the operation on the GPU or using transforms on 
-computation graphs that contain :class:`Axpby`, you can stop implementing the 
-primitive here and enjoy the speed-ups you get from the Accelerate library. 
-
-Implementing the GPU Backend
+Implementing the GPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Apple silicon devices address their GPUs using the Metal_ shading language, and 
-all GPU kernels in MLX are written using metal. 
+Apple silicon devices address their GPUs using the Metal_ shading language, and
+GPU kernels in MLX are written using Metal.
 
 .. note::
 
-    Here are some helpful resources if you are new to metal!
+    Here are some helpful resources if you are new to Metal:
 
     * A walkthrough of the metal compute pipeline: `Metal Example`_
     * Documentation for metal shading language: `Metal Specification`_
     * Using metal from C++: `Metal-cpp`_
 
-Let's keep the GPU algorithm simple. We will launch exactly as many threads 
-as there are elements in the output. Each thread will pick the element it needs 
-from ``x`` and ``y``, do the pointwise operation, and then update its assigned 
-element in the output. 
+Let's keep the GPU kernel simple. We will launch exactly as many threads as
+there are elements in the output. Each thread will pick the element it needs
+from ``x`` and ``y``, do the point-wise operation, and update its assigned
+element in the output.
 
 .. code-block:: C++
 
@@ -457,15 +427,14 @@ element in the output.
         // Convert linear indices to offsets in array
         auto x_offset = elem_to_loc(index, shape, x_strides, ndim);
         auto y_offset = elem_to_loc(index, shape, y_strides, ndim);
-        
+
         // Do the operation and update the output
-        out[index] = 
+        out[index] =
             static_cast<T>(alpha) * x[x_offset] + static_cast<T>(beta) * y[y_offset];
     }
 
 We then need to instantiate this template for all floating point types and give
-each instantiation a unique host name so we can identify the right kernel for 
-each data type. 
+each instantiation a unique host name so we can identify it.
 
 .. code-block:: C++
 
@@ -488,29 +457,21 @@ each data type.
     instantiate_axpby(bfloat16, bfloat16_t);
     instantiate_axpby(complex64, complex64_t);
 
-This kernel will be compiled into a metal library ``mlx_ext.metallib`` as we 
-will see later in :ref:`Building with CMake`. In the following example, we 
-assume that the library ``mlx_ext.metallib`` will always be co-located with 
-the executable/ shared-library calling the :meth:`register_library` function. 
-The :meth:`register_library` function takes the library's name and potential 
-path (or in this case, a function that can produce the path of the metal 
-library) and tries to load that library if it hasn't already been registered 
-by the relevant static :class:`mlx::core::metal::Device` object. This is why, 
-it is important to package your C++ library with the metal library. We will 
-go over this process in more detail later. 
-
-The logic to determine the kernel, set the inputs, resolve the grid dimensions 
-and dispatch it to the GPU are contained in :meth:`Axpby::eval_gpu` as shown 
+The logic to determine the kernel, set the inputs, resolve the grid dimensions,
+and dispatch to the GPU are contained in :meth:`Axpby::eval_gpu` as shown
 below.
 
 .. code-block:: C++
 
     /** Evaluate primitive on GPU */
-    void Axpby::eval_gpu(const std::vector<array>& inputs, array& out) {
+    void Axpby::eval_gpu(
+      const std::vector<array>& inputs,
+      std::vector<array>& outputs) {
         // Prepare inputs
         assert(inputs.size() == 2);
         auto& x = inputs[0];
         auto& y = inputs[1];
+        auto& out = outputs[0];
 
         // Each primitive carries the stream it should execute on
         // and each stream carries its device identifiers
@@ -518,10 +479,10 @@ below.
         // We get the needed metal device using the stream
         auto& d = metal::device(s.device);
 
-        // Allocate output memory 
+        // Allocate output memory
         out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
-        // Resolve name of kernel (corresponds to axpby.metal)
+        // Resolve name of kernel
         std::ostringstream kname;
         kname << "axpby_" << "general_" << type_to_name(out);
 
@@ -552,7 +513,7 @@ below.
         compute_encoder->setBytes(&alpha_, sizeof(float), 3);
         compute_encoder->setBytes(&beta_, sizeof(float), 4);
 
-        // Encode shape, strides and ndim 
+        // Encode shape, strides and ndim
         compute_encoder->setBytes(x.shape().data(), ndim * sizeof(int), 5);
         compute_encoder->setBytes(x.strides().data(), ndim * sizeof(size_t), 6);
         compute_encoder->setBytes(y.strides().data(), ndim * sizeof(size_t), 7);
@@ -575,28 +536,25 @@ below.
 
 We can now call the :meth:`axpby` operation on both the CPU and the GPU!
 
-A few things to note about MLX and metal before moving on. MLX keeps track 
-of the active ``compute_encoder``. We rely on :meth:`d.get_command_encoder` 
-to give us the active metal compute command encoder instead of building a 
-new one and calling :meth:`compute_encoder->end_encoding` at the end. 
-MLX keeps adding kernels (compute pipelines) to the active command encoder 
-until some specified limit is hit or the compute encoder needs to be flushed 
-for synchronization. MLX also handles enqueuing and committing the associated 
-command buffers as needed. We suggest taking a deeper dive into 
-:class:`metal::Device` if you would like to study this routine further.
+A few things to note about MLX and Metal before moving on. MLX keeps track of
+the active ``command_buffer`` and the ``MTLCommandBuffer`` to which it is
+associated. We rely on :meth:`d.get_command_encoder` to give us the active
+metal compute command encoder instead of building a new one and calling
+:meth:`compute_encoder->end_encoding` at the end. MLX adds kernels (compute
+pipelines) to the active command buffer until some specified limit is hit or
+the command buffer needs to be flushed for synchronization.
 
 Primitive Transforms
 ^^^^^^^^^^^^^^^^^^^^^
 
-Now that we have come this far, let's also learn how to add implementations to 
-transformations in a :class:`Primitive`. These transformations can be built on 
-top of our operations, including the one we just defined now. Which then gives 
-us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
+Next, let's add implementations for transformations in a :class:`Primitive`.
+These transformations can be built on top of other operations, including the
+one we just defined:
 
 .. code-block:: C++
 
     /** The Jacobian-vector product. */
-    array Axpby::jvp(
+    std::vector<array> Axpby::jvp(
             const std::vector<array>& primals,
             const std::vector<array>& tangents,
             const std::vector<int>& argnums) {
@@ -611,12 +569,12 @@ us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
         if (argnums.size() > 1) {
             auto scale = argnums[0] == 0 ? alpha_ : beta_;
             auto scale_arr = array(scale, tangents[0].dtype());
-            return multiply(scale_arr, tangents[0], stream());
+            return {multiply(scale_arr, tangents[0], stream())};
         }
         // If, argnums = {0, 1}, we take contributions from both
         // which gives us jvp = tangent_x * alpha + tangent_y * beta
         else {
-            return axpby(tangents[0], tangents[1], alpha_, beta_, stream());
+            return {axpby(tangents[0], tangents[1], alpha_, beta_, stream())};
         }
     }
 
@@ -625,34 +583,35 @@ us the following :meth:`Axpby::jvp` and :meth:`Axpby::vjp` implementations.
     /** The vector-Jacobian product. */
     std::vector<array> Axpby::vjp(
             const std::vector<array>& primals,
-            const array& cotan,
-            const std::vector<int>& argnums) {
+            const std::vector<array>& cotangents,
+            const std::vector<int>& argnums,
+            const std::vector<int>& /* unused */) {
         // Reverse mode diff
         std::vector<array> vjps;
         for (auto arg : argnums) {
             auto scale = arg == 0 ? alpha_ : beta_;
-            auto scale_arr = array(scale, cotan.dtype());
-            vjps.push_back(multiply(scale_arr, cotan, stream()));
+            auto scale_arr = array(scale, cotangents[0].dtype());
+            vjps.push_back(multiply(scale_arr, cotangents[0], stream()));
         }
         return vjps;
     }
 
-Finally, you need not have a transformation fully defined to start using your 
-own :class:`Primitive`.
+Note, a transformation does not need to be fully defined to start using
+the :class:`Primitive`.
 
 .. code-block:: C++
 
     /** Vectorize primitive along given axis */
-    std::pair<array, int> Axpby::vmap(
+    std::pair<std::vector<array>, std::vector<int>> Axpby::vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) {
-        throw std::runtime_error("Axpby has no vmap implementation.");
+        throw std::runtime_error("[Axpby] vmap not implemented.");
     }
 
 Building and Binding
 --------------------
 
-Let's look at the overall directory structure first. 
+Let's look at the overall directory structure first.
 
 | extensions
 | ├── axpby
@@ -666,40 +625,39 @@ Let's look at the overall directory structure first.
 | └── setup.py
 
 * ``extensions/axpby/`` defines the C++ extension library
-* ``extensions/mlx_sample_extensions`` sets out the structure for the 
-  associated python package
-* ``extensions/bindings.cpp`` provides python bindings for our operation
-* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and 
-  python bindings
+* ``extensions/mlx_sample_extensions`` sets out the structure for the
+  associated Python package
+* ``extensions/bindings.cpp`` provides Python bindings for our operation
+* ``extensions/CMakeLists.txt`` holds CMake rules to build the library and
+  Python bindings
 * ``extensions/setup.py`` holds the ``setuptools`` rules to build and install
-  the python package
+  the Python package
 
 Binding to Python
 ^^^^^^^^^^^^^^^^^^
 
-We use PyBind11_ to build a Python API for the C++ library. Since bindings for
+We use nanobind_ to build a Python API for the C++ library. Since bindings for
 components such as :class:`mlx.core.array`, :class:`mlx.core.stream`, etc. are
-already provided, adding our :meth:`axpby` is simple!
+already provided, adding our :meth:`axpby` is simple.
 
 .. code-block:: C++
 
-    PYBIND11_MODULE(mlx_sample_extensions, m) {
-        m.doc() = "Sample C++ and metal extensions for MLX";
+   NB_MODULE(_ext, m) {
+        m.doc() = "Sample extension for MLX";
 
         m.def(
             "axpby",
             &axpby,
             "x"_a,
             "y"_a,
-            py::pos_only(),
             "alpha"_a,
             "beta"_a,
-            py::kw_only(),
-            "stream"_a = py::none(),
-            R"pbdoc(
+            nb::kw_only(),
+            "stream"_a = nb::none(),
+            R"(
                 Scale and sum two vectors element-wise
                 ``z = alpha * x + beta * y``
-                
+
                 Follows numpy style broadcasting between ``x`` and ``y``
                 Inputs are upcasted to floats if needed
 
@@ -711,17 +669,17 @@ already provided, adding our :meth:`axpby` is simple!
 
                 Returns:
                     array: ``alpha * x + beta * y``
-            )pbdoc");
+            )");
     }
 
-Most of the complexity in the above example comes from additional bells and 
+Most of the complexity in the above example comes from additional bells and
 whistles such as the literal names and doc-strings.
 
 .. warning::
 
-    :mod:`mlx.core` needs to be imported before importing 
-    :mod:`mlx_sample_extensions` as defined by the pybind11 module above to 
-    ensure that the casters for :mod:`mlx.core` components like 
+    :mod:`mlx.core` must be imported before importing
+    :mod:`mlx_sample_extensions` as defined by the nanobind module above to
+    ensure that the casters for :mod:`mlx.core` components like
     :class:`mlx.core.array` are available.
 
 .. _Building with CMake:
@@ -729,8 +687,8 @@ whistles such as the literal names and doc-strings.
 Building with CMake
 ^^^^^^^^^^^^^^^^^^^^
 
-Building the C++ extension library itself is simple, it only requires that you 
-``find_package(MLX CONFIG)`` and then link it to your library. 
+Building the C++ extension library only requires that you ``find_package(MLX
+CONFIG)`` and then link it to your library.
 
 .. code-block:: cmake
 
@@ -752,12 +710,12 @@ Building the C++ extension library itself is simple, it only requires that you
     # Link to mlx
     target_link_libraries(mlx_ext PUBLIC mlx)
 
-We also need to build the attached metal library. For convenience, we provide a 
-:meth:`mlx_build_metallib` function that builds a ``.metallib`` target given 
-sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and 
-automatically imported with MLX package). 
+We also need to build the attached Metal library. For convenience, we provide a
+:meth:`mlx_build_metallib` function that builds a ``.metallib`` target given
+sources, headers, destinations, etc. (defined in ``cmake/extension.cmake`` and
+automatically imported with MLX package).
 
-Here is what that looks like in practice!
+Here is what that looks like in practice:
 
 .. code-block:: cmake
 
@@ -779,27 +737,29 @@ Here is what that looks like in practice!
 
     endif()
 
-Finally, we build the Pybind11_ bindings
+Finally, we build the nanobind_ bindings
 
 .. code-block:: cmake
 
-    pybind11_add_module(
-        mlx_sample_extensions
-        ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
+    nanobind_add_module(
+      _ext
+      NB_STATIC STABLE_ABI LTO NOMINSIZE
+      NB_DOMAIN mlx
+      ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
     )
-    target_link_libraries(mlx_sample_extensions PRIVATE mlx_ext)
+    target_link_libraries(_ext PRIVATE mlx_ext)
 
     if(BUILD_SHARED_LIBS)
-        target_link_options(mlx_sample_extensions PRIVATE -Wl,-rpath,@loader_path)
+      target_link_options(_ext PRIVATE -Wl,-rpath,@loader_path)
     endif()
 
 Building with ``setuptools``
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Once we have set out the CMake build rules as described above, we can use the
-build utilities defined in :mod:`mlx.extension` for a simple build process. 
+build utilities defined in :mod:`mlx.extension`:
 
-.. code-block:: python 
+.. code-block:: python
 
     from mlx import extension
     from setuptools import setup
@@ -809,48 +769,50 @@ build utilities defined in :mod:`mlx.extension` for a simple build process.
             name="mlx_sample_extensions",
             version="0.0.0",
             description="Sample C++ and Metal extensions for MLX primitives.",
-            ext_modules=[extension.CMakeExtension("mlx_sample_extensions")],
+            ext_modules=[extension.CMakeExtension("mlx_sample_extensions._ext")],
             cmdclass={"build_ext": extension.CMakeBuild},
-            packages = ["mlx_sample_extensions"],
-            package_dir = {"": "mlx_sample_extensions"},
-            package_data = {"mlx_sample_extensions" : ["*.so", "*.dylib", "*.metallib"]},
+            packages=["mlx_sample_extensions"],
+            package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
+            extras_require={"dev":[]},
             zip_safe=False,
-            python_requires=">=3.7",
+            python_requires=">=3.8",
         )
 
 .. note::
     We treat ``extensions/mlx_sample_extensions`` as the package directory
     even though it only contains a ``__init__.py`` to ensure the following:
-    
-    * :mod:`mlx.core` is always imported before importing  :mod:`mlx_sample_extensions`
-    * The C++ extension library and the metal library are co-located with the python 
-      bindings and copied together if the package is installed 
 
-You can build inplace for development using
+    * :mod:`mlx.core` must be imported before importing :mod:`_ext`
+    * The C++ extension library and the metal library are co-located with the python
+      bindings and copied together if the package is installed
+
+To build the package, first install the build dependencies with ``pip install
+-r requirements.txt``.  You can then build inplace for development using
 ``python setup.py build_ext -j8 --inplace`` (in ``extensions/``)
 
-This will result in a directory structure as follows:
+This results in the directory structure:
 
 | extensions
 | ├── mlx_sample_extensions
 | │   ├── __init__.py
 | │   ├── libmlx_ext.dylib # C++ extension library
 | │   ├── mlx_ext.metallib # Metal library
-| │   └── mlx_sample_extensions.cpython-3x-darwin.so # Python Binding
+| │   └── _ext.cpython-3x-darwin.so # Python Binding
 | ...
 
-When you try to install using the command ``python -m pip install .`` 
-(in ``extensions/``), the package will be installed with the same structure as 
-``extensions/mlx_sample_extensions`` and the C++ and metal library will be 
-copied along with the python binding since they are specified as ``package_data``.
+When you try to install using the command ``python -m pip install .`` (in
+``extensions/``), the package will be installed with the same structure as
+``extensions/mlx_sample_extensions`` and the C++ and Metal library will be
+copied along with the Python binding since they are specified as
+``package_data``.
 
 Usage
 -----
 
-After installing the extension as described above, you should be able to simply 
-import the python package and play with it as you would any other MLX operation!
+After installing the extension as described above, you should be able to simply
+import the Python package and play with it as you would any other MLX operation.
 
-Let's looks at a simple script and it's results!
+Let's look at a simple script and its results:
 
 .. code-block:: python
 
@@ -874,12 +836,12 @@ Output:
     c correctness: True
 
 Results
-^^^^^^^^^^^^^^^^
+^^^^^^^
 
-Let's run a quick benchmark and see how our new ``axpby`` operation compares 
-with the naive :meth:`simple_axpby` we defined at first on the CPU. 
+Let's run a quick benchmark and see how our new ``axpby`` operation compares
+with the naive :meth:`simple_axpby` we first defined on the CPU.
 
-.. code-block:: python 
+.. code-block:: python
 
     import mlx.core as mx
     from mlx_sample_extensions import axpby
@@ -898,7 +860,7 @@ with the naive :meth:`simple_axpby` we defined at first on the CPU.
     alpha = 4.0
     beta = 2.0
 
-    mx.eval((x, y))
+    mx.eval(x, y)
 
     def bench(f):
         # Warm up
@@ -919,30 +881,23 @@ with the naive :meth:`simple_axpby` we defined at first on the CPU.
 
     print(f"Simple axpby: {simple_time:.3f} s | Custom axpby: {custom_time:.3f} s")
 
-Results:
-
-.. code-block::
-
-    Simple axpby: 0.114 s | Custom axpby: 0.109 s
-
-We see some modest improvements right away! 
+The results are ``Simple axpby: 0.114 s | Custom axpby: 0.109 s``. We see
+modest improvements right away!
 
 This operation is now good to be used to build other operations, in
 :class:`mlx.nn.Module` calls, and also as a part of graph transformations like
-:meth:`grad`!
+:meth:`grad`.
 
 Scripts
 -------
 
 .. admonition:: Download the code
 
-   The full example code is available in `mlx <code>`_.
-
-.. code: `https://github.com/ml-explore/mlx/tree/main/examples/extensions/`_
+   The full example code is available in `mlx <https://github.com/ml-explore/mlx/tree/main/examples/extensions/>`_.
 
 .. _Accelerate: https://developer.apple.com/documentation/accelerate/blas?language=objc
 .. _Metal: https://developer.apple.com/documentation/metal?language=objc
 .. _Metal-cpp: https://developer.apple.com/metal/cpp/
 .. _`Metal Specification`: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
 .. _`Metal Example`: https://developer.apple.com/documentation/metal/performing_calculations_on_a_gpu?language=objc
-.. _PyBind11: https://pybind11.readthedocs.io/en/stable/
+.. _nanobind: https://nanobind.readthedocs.io/en/latest/

docs/src/dev/metal_debugger.rst

@@ -0,0 +1,69 @@
+Metal Debugger
+==============
+
+.. currentmodule:: mlx.core
+
+Profiling is a key step for performance optimization. You can build MLX with
+the ``MLX_METAL_DEBUG`` option to improve the Metal debugging and
+optimization workflow. The ``MLX_METAL_DEBUG`` debug option:
+
+* Records source during Metal compilation, for later inspection while
+  debugging.
+* Labels Metal objects such as command queues, improving capture readability.
+
+To build with debugging enabled in Python prepend
+``CMAKE_ARGS="-DMLX_METAL_DEBUG=ON"`` to the build call.
+
+The :func:`metal.start_capture` function initiates a capture of all MLX GPU
+work.
+
+.. note::
+
+   To capture a GPU trace you must run the application with
+   ``MTL_CAPTURE_ENABLED=1``.
+
+.. code-block:: python
+
+    import mlx.core as mx
+
+    a = mx.random.uniform(shape=(512, 512))
+    b = mx.random.uniform(shape=(512, 512))
+    mx.eval(a, b)
+
+    trace_file = "mlx_trace.gputrace"
+
+    if not mx.metal.start_capture(trace_file):
+      print("Make sure to run with MTL_CAPTURE_ENABLED=1 and "
+            f"that the path {trace_file} does not already exist.")
+      exit(1)
+
+    for _ in range(10):
+      mx.eval(mx.add(a, b))
+
+    mx.metal.stop_capture()
+
+You can open and replay the GPU trace in Xcode. The ``Dependencies`` view
+has a great overview of all operations. Checkout the `Metal debugger
+documentation`_ for more information.
+
+.. image:: ../_static/metal_debugger/capture.png
+    :class: dark-light
+
+Xcode Workflow
+--------------
+
+You can skip saving to a path by running within Xcode. First, generate an
+Xcode project using CMake.
+
+.. code-block::
+
+    mkdir build && cd build
+    cmake .. -DMLX_METAL_DEBUG=ON -G Xcode
+    open mlx.xcodeproj
+
+Select the ``metal_capture`` example schema and run.
+
+.. image:: ../_static/metal_debugger/schema.png
+    :class: dark-light
+
+.. _`Metal debugger documentation`: https://developer.apple.com/documentation/xcode/metal-debugger

docs/src/index.rst

@@ -58,10 +58,12 @@ are the CPU and GPU.
    :maxdepth: 1
 
    python/array
+   python/data_types
    python/devices_and_streams
    python/ops
    python/random
    python/transforms
+   python/fast
    python/fft
    python/linalg
    python/metal
@@ -80,3 +82,4 @@ are the CPU and GPU.
    :maxdepth: 1
 
    dev/extensions
+   dev/metal_debugger

docs/src/install.rst

@@ -15,10 +15,10 @@ To install from PyPI you must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
 - Using a native Python >= 3.8
-- macOS >= 13.3
+- macOS >= 13.5
 
 .. note::
-    MLX is only available on devices running macOS >= 13.3 
+    MLX is only available on devices running macOS >= 13.5
     It is highly recommended to use macOS 14 (Sonoma)
 
 
@@ -54,7 +54,7 @@ Build Requirements
 
 - A C++ compiler with C++17 support (e.g. Clang >= 5.0)
 - `cmake <https://cmake.org/>`_ -- version 3.24 or later, and ``make``
-- Xcode >= 14.3 (Xcode >= 15.0 for macOS 14 and above)
+- Xcode >= 15.0 and macOS SDK >= 14.0
 
 .. note::
    Ensure your shell environment is native ``arm``, not ``x86`` via Rosetta. If
@@ -70,16 +70,13 @@ To build and install the MLX python library from source, first, clone MLX from
 
    git clone git@github.com:ml-explore/mlx.git mlx && cd mlx
 
-Make sure that you have `pybind11 <https://pybind11.readthedocs.io/en/stable/index.html>`_
-installed. You can install ``pybind11`` with ``pip``, ``brew`` or ``conda`` as follows:
+Install `nanobind <https://nanobind.readthedocs.io/en/latest/>`_ with:
 
 .. code-block:: shell
 
-    pip install "pybind11[global]"
-    conda install pybind11
-    brew install pybind11
+    pip install git+https://github.com/wjakob/nanobind.git
 
-Then simply build and install it using pip:
+Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
@@ -158,6 +155,8 @@ should point to the path to the built metal library.
      - ON
    * - MLX_BUILD_PYTHON_BINDINGS
      - OFF
+   * - MLX_METAL_DEBUG
+     - OFF
 
 
 .. note::

docs/src/python/array.rst

@@ -10,27 +10,38 @@ Array
 
     array
     array.astype
+    array.at
     array.item
     array.tolist
     array.dtype
+    array.itemsize
+    array.nbytes
     array.ndim
     array.shape
     array.size
-    Dtype
     array.abs
     array.all
     array.any
     array.argmax
     array.argmin
     array.cos
-    array.dtype
+    array.cummax
+    array.cummin
+    array.cumprod
+    array.cumsum
+    array.diag
+    array.diagonal
     array.exp
+    array.flatten
     array.log
+    array.log10
     array.log1p
+    array.log2
     array.logsumexp
     array.max
     array.mean
     array.min
+    array.moveaxis
     array.prod
     array.reciprocal
     array.reshape
@@ -40,6 +51,8 @@ Array
     array.split
     array.sqrt
     array.square
+    array.squeeze
+    array.swapaxes
     array.sum
     array.transpose
     array.T

docs/src/python/data_types.rst

@@ -1,7 +1,5 @@
 .. _data_types:
 
-:orphan:
-
 Data Types
 ==========
 
@@ -44,9 +42,27 @@ The default floating point type is ``float32`` and the default integer type is
    * - ``int64``
      - 8 
      - 64-bit signed integer 
+   * - ``bfloat16``
+     - 2 
+     - 16-bit brain float (e8, m7)
    * - ``float16``
      - 2 
-     - 16-bit float, only available with `ARM C language extensions <https://developer.arm.com/documentation/101028/0012/3--C-language-extensions?lang=en>`_
+     - 16-bit IEEE float (e5, m10)
    * - ``float32``
      - 4 
      - 32-bit float
+   * - ``complex64``
+     - 8 
+     - 64-bit complex float
+
+
+Data type are aranged in a hierarchy. See the :obj:`DtypeCategory` object
+documentation for more information. Use :func:`issubdtype` to determine if one
+``dtype`` (or category) is a subtype of another category.
+
+.. autosummary::
+   :toctree: _autosummary
+
+   Dtype
+   DtypeCategory
+   issubdtype

docs/src/python/fast.rst

@@ -0,0 +1,14 @@
+.. _fast:
+
+Fast
+====
+
+.. currentmodule:: mlx.core.fast
+
+.. autosummary:: 
+  :toctree: _autosummary
+
+  rms_norm
+  layer_norm
+  rope
+  scaled_dot_product_attention

docs/src/python/metal.rst

@@ -3,7 +3,7 @@ Metal
 
 .. currentmodule:: mlx.core.metal
 
-.. autosummary:: 
+.. autosummary::
   :toctree: _autosummary
 
   is_available
@@ -12,3 +12,5 @@ Metal
   get_cache_memory
   set_memory_limit
   set_cache_limit
+  start_capture
+  stop_capture

docs/src/python/nn/layers.rst

@@ -21,9 +21,11 @@ Layers
    Embedding
    GELU
    GroupNorm
+   GRU
    InstanceNorm
    LayerNorm
    Linear
+   LSTM
    MaxPool1d
    MaxPool2d
    Mish
@@ -32,6 +34,7 @@ Layers
    QuantizedLinear
    RMSNorm
    ReLU
+   RNN
    RoPE
    SELU
    Sequential

docs/src/python/nn/module.rst

@@ -30,6 +30,7 @@ Module
       Module.named_modules
       Module.parameters
       Module.save_weights
+      Module.set_dtype
       Module.train
       Module.trainable_parameters
       Module.unfreeze

docs/src/python/ops.rst

@@ -5,13 +5,13 @@ Operations
 
 .. currentmodule:: mlx.core
 
-.. autosummary:: 
+.. autosummary::
   :toctree: _autosummary
 
    abs
    add
    all
-   allclose 
+   allclose
    any
    arange
    arccos
@@ -38,6 +38,10 @@ Operations
    conv_general
    cos
    cosh
+   cummax
+   cummin
+   cumprod
+   cumsum
    dequantize
    diag
    diagonal
@@ -47,6 +51,7 @@ Operations
    erf
    erfinv
    exp
+   expm1
    expand_dims
    eye
    flatten
@@ -58,10 +63,10 @@ Operations
    identity
    inner
    isclose
-   isnan
-   isposinf
-   isneginf
    isinf
+   isnan
+   isneginf
+   isposinf
    less
    less_equal
    linspace
@@ -79,6 +84,7 @@ Operations
    max
    maximum
    mean
+   meshgrid
    min
    minimum
    moveaxis
@@ -113,6 +119,7 @@ Operations
    square
    squeeze
    stack
+   std
    stop_gradient
    subtract
    sum

docs/src/python/random.rst

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

docs/src/usage/function_transforms.rst

@@ -40,7 +40,7 @@ getting higher order derivatives.
 
 Any of the MLX function transformations can be composed in any order to any
 depth. See the following sections for more information on :ref:`automatic
-differentiaion <auto diff>` and :ref:`automatic vectorization <vmap>`.
+differentiation <auto diff>` and :ref:`automatic vectorization <vmap>`.
 For more information on :func:`compile` see the :ref:`compile documentation <compile>`.
 
 

docs/src/usage/saving_and_loading.rst

@@ -49,7 +49,7 @@ it will be added. You can load the array with:
 
 .. code-block:: shell
 
-   >>> mx.load("array.npy", a)
+   >>> mx.load("array.npy")
    array([1], dtype=float32)
 
 Here's an example of saving several arrays to a single file:

examples/cpp/CMakeLists.txt

@@ -8,3 +8,4 @@ endfunction(build_example)
 build_example(tutorial.cpp)
 build_example(linear_regression.cpp)
 build_example(logistic_regression.cpp)
+build_example(metal_capture.cpp)

examples/cpp/metal_capture.cpp

@@ -0,0 +1,31 @@
+// Copyright © 2024 Apple Inc.
+
+#include <cassert>
+#include <iostream>
+
+#include "mlx/mlx.h"
+
+using namespace mlx::core;
+
+int main() {
+  // To use Metal debugging and profiling:
+  // 1. Build with the MLX_METAL_DEBUG CMake option (i.e. -DMLX_METAL_DEBUG=ON).
+  // 2. Run with MTL_CAPTURE_ENABLED=1.
+  assert(metal::start_capture("mlx_trace.gputrace"));
+
+  // Start at index two because the default GPU and CPU streams have indices
+  // zero and one, respectively. This naming matches the label assigned to each
+  // stream's command queue.
+  auto s2 = new_stream(Device::gpu);
+  auto s3 = new_stream(Device::gpu);
+
+  auto a = arange(1.f, 10.f, 1.f, float32, s2);
+  auto b = arange(1.f, 10.f, 1.f, float32, s3);
+  auto x = add(a, a, s2);
+  auto y = add(b, b, s3);
+
+  // The multiply will happen on the default stream.
+  std::cout << multiply(x, y) << std::endl;
+
+  metal::stop_capture();
+}

examples/extensions/CMakeLists.txt

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

examples/extensions/README.md

@@ -0,0 +1,18 @@
+
+## Build the extensions
+
+```
+pip install -e .
+```
+
+For faster builds during development, you can also pre-install the requirements:
+
+```
+pip install -r requirements.txt
+```
+
+And then run:
+
+```
+python setup.py build_ext -j8 --inplace
+```

examples/extensions/axpby/axpby.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <iostream>
@@ -43,7 +43,7 @@ array axpby(
   auto promoted_dtype = promote_types(x.dtype(), y.dtype());
 
   // Upcast to float32 for non-floating point inputs x and y
-  auto out_dtype = is_floating_point(promoted_dtype)
+  auto out_dtype = issubdtype(promoted_dtype, float32)
       ? promoted_dtype
       : promote_types(promoted_dtype, float32);
 
@@ -61,7 +61,7 @@ array axpby(
       /* const std::vector<int>& shape = */ out_shape,
       /* Dtype dtype = */ out_dtype,
       /* std::unique_ptr<Primitive> primitive = */
-      std::make_unique<Axpby>(to_stream(s), alpha, beta),
+      std::make_shared<Axpby>(to_stream(s), alpha, beta),
       /* const std::vector<array>& inputs = */ broadcasted_inputs);
 }
 
@@ -106,12 +106,12 @@ void axpby_impl(
 /** Fall back implementation for evaluation on CPU */
 void Axpby::eval(
     const std::vector<array>& inputs,
-    std::vector<array>& out_arr) {
-  auto out = out_arr[0];
+    std::vector<array>& outputs) {
   // Check the inputs (registered in the op while constructing the out array)
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Dispatch to the correct dtype
   if (out.dtype() == float32) {
@@ -150,11 +150,7 @@ void axpby_impl_accelerate(
   // The data in the output array is allocated to match the strides in y
   // such that x, y, and out are contiguous in the same mode and
   // no transposition is needed
-  out.set_data(
-      allocator::malloc_or_wait(y.data_size() * out.itemsize()),
-      y.data_size(),
-      y.strides(),
-      y.flags());
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
   // We then copy over the elements using the contiguous vector specialization
   copy_inplace(y, out, CopyType::Vector);
@@ -180,11 +176,11 @@ void axpby_impl_accelerate(
 /** Evaluate primitive on CPU using accelerate specializations */
 void Axpby::eval_cpu(
     const std::vector<array>& inputs,
-    std::vector<array>& outarr) {
-  auto out = outarr[0];
+    std::vector<array>& outputs) {
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Accelerate specialization for contiguous single precision float arrays
   if (out.dtype() == float32 &&
@@ -195,7 +191,7 @@ void Axpby::eval_cpu(
   }
 
   // Fall back to common backend if specializations are not available
-  eval(inputs, outarr);
+  eval(inputs, outputs);
 }
 
 #else // Accelerate not available
@@ -203,8 +199,8 @@ void Axpby::eval_cpu(
 /** Evaluate primitive on CPU falling back to common backend */
 void Axpby::eval_cpu(
     const std::vector<array>& inputs,
-    std::vector<array>& out) {
-  eval(inputs, out);
+    const std::vector<array>& outputs) {
+  eval(inputs, outputs);
 }
 
 #endif
@@ -218,12 +214,12 @@ void Axpby::eval_cpu(
 /** Evaluate primitive on GPU */
 void Axpby::eval_gpu(
     const std::vector<array>& inputs,
-    std::vector<array>& outarr) {
+    std::vector<array>& outputs) {
   // Prepare inputs
-  auto out = outarr[0];
   assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
+  auto& out = outputs[0];
 
   // Each primitive carries the stream it should execute on
   // and each stream carries its device identifiers
@@ -372,4 +368,4 @@ bool Axpby::is_equivalent(const Primitive& other) const {
   return alpha_ == r_other.alpha_ && beta_ == r_other.beta_;
 }
 
-} // namespace mlx::core
+} // namespace mlx::core

examples/extensions/axpby/axpby.h

@@ -42,9 +42,9 @@ class Axpby : public Primitive {
    * To avoid unnecessary allocations, the evaluation function
    * is responsible for allocating space for the array.
    */
-  void eval_cpu(const std::vector<array>& inputs, std::vector<array>& out)
+  void eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs)
       override;
-  void eval_gpu(const std::vector<array>& inputs, std::vector<array>& out)
+  void eval_gpu(const std::vector<array>& inputs, std::vector<array>& outputs)
       override;
 
   /** The Jacobian-vector product. */
@@ -83,7 +83,7 @@ class Axpby : public Primitive {
   float beta_;
 
   /** Fall back implementation for evaluation on CPU */
-  void eval(const std::vector<array>& inputs, std::vector<array>& out);
+  void eval(const std::vector<array>& inputs, std::vector<array>& outputs);
 };
 
-} // namespace mlx::core
+} // namespace mlx::core

examples/extensions/bindings.cpp

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

examples/extensions/pyproject.toml

@@ -1,3 +1,8 @@
 [build-system]
-requires = ["setuptools>=42", "pybind11>=2.10", "cmake>=3.24", "mlx @ git+https://github.com/mlx-explore/mlx@main"]
-build-backend = "setuptools.build_meta"
+requires = [
+  "setuptools>=42",
+  "cmake>=3.24",
+  "mlx>=0.9.0",
+  "nanobind@git+https://github.com/wjakob/nanobind.git#egg=4148debcf91f5ccab0c3b8d67b5c3cabd61f407f",
+]
+build-backend = "setuptools.build_meta"

examples/extensions/requirements.txt

@@ -0,0 +1,4 @@
+setuptools>=42
+cmake>=3.24
+mlx>=0.9.0
+nanobind@git+https://github.com/wjakob/nanobind.git#egg=4148debcf91f5ccab0c3b8d67b5c3cabd61f407f

examples/extensions/setup.py

@@ -1,4 +1,4 @@
-# Copyright © 2023 Apple Inc.
+# Copyright © 2023-2024 Apple Inc.
 
 from setuptools import setup
 
@@ -9,11 +9,11 @@ if __name__ == "__main__":
         name="mlx_sample_extensions",
         version="0.0.0",
         description="Sample C++ and Metal extensions for MLX primitives.",
-        ext_modules=[extension.CMakeExtension("mlx_sample_extensions")],
+        ext_modules=[extension.CMakeExtension("mlx_sample_extensions._ext")],
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
-        package_dir={"": "."},
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
+        extras_require={"dev": []},
         zip_safe=False,
         python_requires=">=3.8",
     )

mlx/array.cpp

@@ -12,16 +12,6 @@ namespace mlx::core {
 
 namespace {
 
-std::pair<size_t, std::vector<size_t>> cum_prod(const std::vector<int>& shape) {
-  std::vector<size_t> strides(shape.size());
-  size_t cum_prod = 1;
-  for (int i = shape.size() - 1; i >= 0; --i) {
-    strides[i] = cum_prod;
-    cum_prod *= shape[i];
-  }
-  return {cum_prod, strides};
-}
-
 /** Return true if we are currently performing a function transformation in
  * order to keep the graph when evaluating tracer arrays. */
 bool in_tracing() {
@@ -36,22 +26,11 @@ array::array(const std::complex<float>& val, Dtype dtype /* = complex64 */)
   init(&cval);
 }
 
-array::array(
-    const std::vector<int>& shape,
-    Dtype dtype,
-    std::shared_ptr<Primitive> primitive,
-    const std::vector<array>& inputs)
-    : array_desc_(std::make_shared<ArrayDesc>(
-          shape,
-          dtype,
-          std::move(primitive),
-          inputs)) {}
-
 array::array(
     std::vector<int> shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
-    std::vector<array>&& inputs)
+    std::vector<array> inputs)
     : array_desc_(std::make_shared<ArrayDesc>(
           std::move(shape),
           dtype,
@@ -59,15 +38,16 @@ array::array(
           std::move(inputs))) {}
 
 std::vector<array> array::make_arrays(
-    const std::vector<std::vector<int>>& shapes,
+    std::vector<std::vector<int>> shapes,
     const std::vector<Dtype>& dtypes,
-    std::shared_ptr<Primitive> primitive,
+    const std::shared_ptr<Primitive>& primitive,
     const std::vector<array>& inputs) {
   std::vector<array> outputs;
-  for (int i = 0; i < shapes.size(); ++i) {
-    outputs.push_back(array(shapes[i], dtypes[i], primitive, inputs));
+  for (size_t i = 0; i < shapes.size(); ++i) {
+    outputs.emplace_back(std::move(shapes[i]), dtypes[i], primitive, inputs);
   }
-  for (int i = 0; i < outputs.size(); ++i) {
+  // For each node in |outputs|, its siblings are the other nodes.
+  for (size_t i = 0; i < outputs.size(); ++i) {
     auto siblings = outputs;
     siblings.erase(siblings.begin() + i);
     outputs[i].set_siblings(std::move(siblings), i);
@@ -92,10 +72,10 @@ array::array(std::initializer_list<int> data, Dtype dtype)
 /* Build an array from a shared buffer */
 array::array(
     allocator::Buffer data,
-    const std::vector<int>& shape,
+    std::vector<int> shape,
     Dtype dtype,
     deleter_t deleter)
-    : array_desc_(std::make_shared<ArrayDesc>(shape, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   set_data(data, deleter);
 }
 
@@ -104,18 +84,18 @@ void array::detach() {
     s.array_desc_->inputs.clear();
     s.array_desc_->siblings.clear();
     s.array_desc_->position = 0;
-    s.array_desc_->depth = 0;
     s.array_desc_->primitive = nullptr;
   }
   array_desc_->inputs.clear();
   array_desc_->siblings.clear();
   array_desc_->position = 0;
-  array_desc_->depth = 0;
   array_desc_->primitive = nullptr;
 }
 
 void array::eval() {
-  mlx::core::eval({*this});
+  if (!is_evaled()) {
+    mlx::core::eval({*this});
+  }
 }
 
 bool array::is_tracer() const {
@@ -164,51 +144,82 @@ void array::copy_shared_buffer(const array& other) {
   copy_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
-void array::move_shared_buffer(array other) {
+void array::move_shared_buffer(
+    array other,
+    const std::vector<size_t>& strides,
+    Flags flags,
+    size_t data_size,
+    size_t offset /* = 0 */) {
   array_desc_->data = std::move(other.array_desc_->data);
-  array_desc_->strides = other.strides();
-  array_desc_->flags = other.flags();
-  array_desc_->data_size = other.data_size();
-  array_desc_->data_ptr = other.array_desc_->data_ptr;
+  array_desc_->strides = strides;
+  array_desc_->flags = flags;
+  array_desc_->data_size = data_size;
+  auto char_offset = sizeof(char) * itemsize() * offset;
+  array_desc_->data_ptr = static_cast<void*>(
+      static_cast<char*>(other.array_desc_->data_ptr) + char_offset);
 }
 
-array::ArrayDesc::ArrayDesc(const std::vector<int>& shape, Dtype dtype)
-    : shape(shape), dtype(dtype) {
-  std::tie(size, strides) = cum_prod(shape);
+void array::move_shared_buffer(array other) {
+  move_shared_buffer(other, other.strides(), other.flags(), other.data_size());
 }
 
-array::ArrayDesc::ArrayDesc(
-    const std::vector<int>& shape,
-    Dtype dtype,
-    std::shared_ptr<Primitive> primitive,
-    const std::vector<array>& inputs)
-    : shape(shape),
-      dtype(dtype),
-      primitive(std::move(primitive)),
-      inputs(inputs) {
-  std::tie(size, strides) = cum_prod(this->shape);
-  for (auto& in : this->inputs) {
-    is_tracer |= in.is_tracer();
-    depth = std::max(in.graph_depth(), depth);
+void array::ArrayDesc::init() {
+  strides.resize(shape.size());
+  size = 1;
+  for (int i = shape.size() - 1; i >= 0; --i) {
+    strides[i] = size;
+    size *= shape[i];
   }
-  depth++;
+  for (auto& in : inputs) {
+    is_tracer |= in.is_tracer();
+  }
+}
+
+array::ArrayDesc::ArrayDesc(std::vector<int> shape, Dtype dtype)
+    : shape(std::move(shape)), dtype(dtype) {
+  init();
 }
 
 array::ArrayDesc::ArrayDesc(
-    std::vector<int>&& shape,
+    std::vector<int> shape,
     Dtype dtype,
     std::shared_ptr<Primitive> primitive,
-    std::vector<array>&& inputs)
+    std::vector<array> inputs)
     : shape(std::move(shape)),
       dtype(dtype),
       primitive(std::move(primitive)),
       inputs(std::move(inputs)) {
-  std::tie(size, strides) = cum_prod(this->shape);
-  for (auto& in : this->inputs) {
-    is_tracer |= in.is_tracer();
-    depth = std::max(in.graph_depth(), depth);
+  init();
+}
+
+array::ArrayDesc::~ArrayDesc() {
+  // When an array description is destroyed it will delete a bunch of arrays
+  // that may also destory their corresponding descriptions and so on and so
+  // forth.
+  //
+  // This calls recursively the destructor and can result in stack overflow, we
+  // instead put them in a vector and destroy them one at a time resulting in a
+  // max stack depth of 2.
+  std::vector<std::shared_ptr<ArrayDesc>> for_deletion;
+
+  for (array& a : inputs) {
+    if (a.array_desc_.use_count() == 1) {
+      for_deletion.push_back(std::move(a.array_desc_));
+    }
+  }
+
+  while (!for_deletion.empty()) {
+    // top is going to be deleted at the end of the block *after* the arrays
+    // with inputs have been moved into the vector
+    auto top = std::move(for_deletion.back());
+    for_deletion.pop_back();
+
+    for (array& a : top->inputs) {
+      if (a.array_desc_.use_count() == 1) {
+        for_deletion.push_back(std::move(a.array_desc_));
+      }
+    }
   }
-  depth++;
 }
 
 array::ArrayIterator::ArrayIterator(const array& arr, int idx)

mlx/array.h

@@ -1,5 +1,6 @@
 // Copyright © 2023 Apple Inc.
 #pragma once
+
 #include <algorithm>
 #include <cstdint>
 #include <functional>
@@ -31,7 +32,7 @@ class array {
   template <typename It>
   array(
       It data,
-      const std::vector<int>& shape,
+      std::vector<int> shape,
       Dtype dtype =
           TypeToDtype<typename std::iterator_traits<It>::value_type>());
 
@@ -47,13 +48,13 @@ class array {
   template <typename T>
   array(
       std::initializer_list<T> data,
-      const std::vector<int>& shape,
+      std::vector<int> shape,
       Dtype dtype = TypeToDtype<T>());
 
   /* Build an array from a buffer */
   array(
       allocator::Buffer data,
-      const std::vector<int>& shape,
+      std::vector<int> shape,
       Dtype dtype,
       deleter_t deleter = allocator::free);
 
@@ -172,22 +173,16 @@ class array {
    * API may change.
    */
 
-  array(
-      const std::vector<int>& shape,
-      Dtype dtype,
-      std::shared_ptr<Primitive> primitive,
-      const std::vector<array>& inputs);
-
   array(
       std::vector<int> shape,
       Dtype dtype,
       std::shared_ptr<Primitive> primitive,
-      std::vector<array>&& inputs);
+      std::vector<array> inputs);
 
   static std::vector<array> make_arrays(
-      const std::vector<std::vector<int>>& shapes,
+      std::vector<std::vector<int>> shapes,
       const std::vector<Dtype>& dtypes,
-      std::shared_ptr<Primitive> primitive,
+      const std::shared_ptr<Primitive>& primitive,
       const std::vector<array>& inputs);
 
   /** A unique identifier for an array. */
@@ -261,6 +256,17 @@ class array {
     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 {
@@ -273,11 +279,6 @@ class array {
     return outputs;
   };
 
-  /** The depth of the array in the graph. Evaluated arrays have depth 0. */
-  uint16_t graph_depth() const {
-    return array_desc_->depth;
-  }
-
   /** Detach the array from the graph. */
   void detach();
 
@@ -344,6 +345,13 @@ class array {
 
   void copy_shared_buffer(const array& other);
 
+  void move_shared_buffer(
+      array other,
+      const std::vector<size_t>& strides,
+      Flags flags,
+      size_t data_size,
+      size_t offset = 0);
+
   void move_shared_buffer(array other);
 
   void overwrite_descriptor(const array& other) {
@@ -360,7 +368,7 @@ class array {
     std::vector<size_t> strides;
     size_t size;
     Dtype dtype;
-    std::shared_ptr<Primitive> primitive{nullptr};
+    std::shared_ptr<Primitive> primitive;
 
     // Indicates an array is being used in a graph transform
     // and should not be detached from the graph
@@ -368,7 +376,7 @@ class array {
 
     // This is a shared pointer so that *different* arrays
     // can share the underlying data buffer.
-    std::shared_ptr<Data> data{nullptr};
+    std::shared_ptr<Data> data;
 
     // Properly offset data pointer
     void* data_ptr{nullptr};
@@ -388,29 +396,26 @@ class array {
     // The arrays position in the output list
     uint32_t position{0};
 
-    // The depth of the array in the graph.
-    uint16_t depth{0};
-
-    explicit ArrayDesc(const std::vector<int>& shape, Dtype dtype);
+    explicit ArrayDesc(std::vector<int> shape, Dtype dtype);
 
     explicit ArrayDesc(
-        const std::vector<int>& shape,
+        std::vector<int> shape,
         Dtype dtype,
         std::shared_ptr<Primitive> primitive,
-        const std::vector<array>& inputs);
+        std::vector<array> inputs);
 
-    explicit ArrayDesc(
-        std::vector<int>&& shape,
-        Dtype dtype,
-        std::shared_ptr<Primitive> primitive,
-        std::vector<array>&& inputs);
+    ~ArrayDesc();
+
+   private:
+    // Initialize size, strides, and other metadata
+    void init();
   };
 
   // The ArrayDesc contains the details of the materialized array including the
   // shape, strides, the data type. It also includes
   // the primitive which knows how to compute the array's data from its inputs
   // and the list of array's inputs for the primitive.
-  std::shared_ptr<ArrayDesc> array_desc_{nullptr};
+  std::shared_ptr<ArrayDesc> array_desc_;
 };
 
 template <typename T>
@@ -422,9 +427,9 @@ array::array(T val, Dtype dtype /* = TypeToDtype<T>() */)
 template <typename It>
 array::array(
   It data,
-  const std::vector<int>& shape,
+  std::vector<int> shape,
   Dtype dtype /* = TypeToDtype<typename std::iterator_traits<It>::value_type>() */) :
-    array_desc_(std::make_shared<ArrayDesc>(shape, dtype)) {
+    array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   init(data);
 }
 
@@ -441,9 +446,9 @@ array::array(
 template <typename T>
 array::array(
     std::initializer_list<T> data,
-    const std::vector<int>& shape,
+    std::vector<int> shape,
     Dtype dtype /* = TypeToDtype<T>() */)
-    : array_desc_(std::make_shared<ArrayDesc>(shape, dtype)) {
+    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
   if (data.size() != size()) {
     throw std::invalid_argument(
         "Data size and provided shape mismatch in array construction.");
@@ -518,4 +523,15 @@ void array::init(It src) {
   }
 }
 
+/* Utilities for determining whether a template parameter is array. */
+template <typename T>
+inline constexpr bool is_array_v =
+    std::is_same_v<std::remove_cv_t<std::remove_reference_t<T>>, array>;
+
+template <typename... T>
+inline constexpr bool is_arrays_v = (is_array_v<T> && ...);
+
+template <typename... T>
+using enable_for_arrays_t = typename std::enable_if_t<is_arrays_v<T...>>;
+
 } // namespace mlx::core

mlx/backend/accelerate/primitives.cpp

@@ -38,6 +38,7 @@ DEFAULT(Copy)
 DEFAULT_MULTI(CustomVJP)
 DEFAULT_MULTI(Depends)
 DEFAULT_MULTI(DivMod)
+DEFAULT(NumberOfElements)
 DEFAULT(Equal)
 DEFAULT(Erf)
 DEFAULT(ErfInv)
@@ -68,10 +69,13 @@ DEFAULT(Select)
 DEFAULT(Sigmoid)
 DEFAULT(Sign)
 DEFAULT(Slice)
+DEFAULT(SliceUpdate)
 DEFAULT_MULTI(Split)
 DEFAULT(Sort)
 DEFAULT(StopGradient)
+DEFAULT_MULTI(SVD)
 DEFAULT(Transpose)
+DEFAULT(Inverse)
 
 void Abs::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
@@ -297,7 +301,7 @@ 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 (is_floating_point(out.dtype())) {
+  } else if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, [](auto x) { return std::exp(x); });
   } else {
     throw std::invalid_argument(
@@ -306,6 +310,19 @@ void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void Expm1::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (out.dtype() == float32 && in.flags().contiguous) {
+    set_unary_output_data(in, out);
+    auto size = in.data_size();
+    vvexpm1f(
+        out.data<float>(), in.data<float>(), reinterpret_cast<int*>(&size));
+  } else {
+    eval(inputs, out);
+  }
+}
+
 void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
@@ -351,7 +368,7 @@ 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 (is_floating_point(out.dtype())) {
+  } else if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, [](auto x) { return std::log1p(x); });
   } else {
     throw std::invalid_argument(

mlx/backend/accelerate/reduce.cpp

@@ -10,78 +10,65 @@
 
 namespace mlx::core {
 
-template <typename T, typename VT, int N>
-void _vectorized_strided_sum(const T* x, T* accum, int size, size_t stride) {
-  for (int i = 0; i < size; i++) {
-    size_t s = stride;
-    T* a = accum;
-    while (s >= N) {
-      VT val = (*(VT*)x);
-      *(VT*)a += val;
-      x += N;
-      a += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      *a++ += *x++;
-    }
-  }
-}
+namespace {
 
-// TODO: Add proper templates for the strided reduce algorithm so we don't have
-// to write max/min/sum etc.
-template <typename T, typename VT, int N>
-void _vectorized_strided_max(const T* x, T* accum, int size, size_t stride) {
-  for (int i = 0; i < size; i++) {
-    size_t s = stride;
-    T* a = accum;
-    while (s >= N) {
-      *(VT*)a = simd_max((*(VT*)x), (*(VT*)a));
-      x += N;
-      a += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      *a = std::max(*a, *x);
-      a++;
-      x++;
-    }
+template <typename T, typename VT>
+struct MinReduction {
+  T operator()(const T& a, const T& b) {
+    return std::min(a, b);
   }
-}
 
-template <typename T, typename VT, int N>
-void _vectorized_strided_min(const T* x, T* accum, int size, size_t stride) {
-  for (int i = 0; i < size; i++) {
-    size_t s = stride;
-    T* a = accum;
-    while (s >= N) {
-      *(VT*)a = simd_min((*(VT*)x), (*(VT*)a));
-      x += N;
-      a += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      *a = std::min(*a, *x);
-      a++;
-      x++;
-    }
+  VT operator()(VT a, VT b) {
+    return simd_min(a, b);
   }
-}
+};
 
-template <typename T, typename VT, int N>
-void _vectorized_sum(const T* x, T* accum, int size) {
-  VT _sum = {0};
-  while (size >= N) {
-    _sum += (*(VT*)x);
-    x += N;
-    size -= N;
+template <typename T, typename VT>
+struct MaxReduction {
+  T operator()(const T& a, const T& b) {
+    return std::max(a, b);
   }
-  T sum = _sum[0];
-  for (int i = 1; i < N; i++) {
-    sum += _sum[i];
+
+  VT operator()(VT a, VT b) {
+    return simd_max(a, b);
   }
-  *accum += sum;
-}
+};
+
+template <typename T, typename VT>
+struct SumReduction {
+  T operator()(const T& a, const T& b) {
+    return a + b;
+  }
+
+  VT operator()(VT a, VT b) {
+    return a + b;
+  }
+};
+
+template <typename T, typename VT, int N, typename Reduction>
+struct StridedReduce {
+  void operator()(const T* x, T* accum, int size, size_t stride) {
+    Reduction op;
+
+    for (int i = 0; i < size; i++) {
+      size_t s = stride;
+      T* a = accum;
+      while (s >= N) {
+        *(VT*)a = op((*(VT*)x), (*(VT*)a));
+        x += N;
+        a += N;
+        s -= N;
+      }
+      while (s-- > 0) {
+        *a = op(*a, *x);
+        a++;
+        x++;
+      }
+    }
+  }
+};
+
+} // namespace
 
 void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
@@ -94,10 +81,11 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
           out,
           axes_,
           0,
-          [](const auto* x, auto* accum, int size, size_t stride) {
-            _vectorized_strided_sum<float, simd_float16, 16>(
-                (const float*)x, (float*)accum, size, stride);
-          },
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              SumReduction<float, simd_float16>>(),
           [](const auto* x, auto* accum, int size) {
             float acc;
             vDSP_sve((const float*)x, 1, &acc, size);
@@ -111,10 +99,11 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
           out,
           axes_,
           -std::numeric_limits<float>::infinity(),
-          [](const auto* x, auto* accum, int size, size_t stride) {
-            _vectorized_strided_max<float, simd_float16, 16>(
-                (const float*)x, (float*)accum, size, stride);
-          },
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              MaxReduction<float, simd_float16>>(),
           [](const auto* x, auto* accum, int size) {
             float max;
             vDSP_maxv((const float*)x, 1, &max, size);
@@ -128,10 +117,11 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
           out,
           axes_,
           std::numeric_limits<float>::infinity(),
-          [](const auto* x, auto* accum, int size, size_t stride) {
-            _vectorized_strided_min<float, simd_float16, 16>(
-                (const float*)x, (float*)accum, size, stride);
-          },
+          StridedReduce<
+              float,
+              simd_float16,
+              16,
+              MinReduction<float, simd_float16>>(),
           [](const auto* x, auto* accum, int size) {
             float min;
             vDSP_minv((const float*)x, 1, &min, size);

mlx/backend/accelerate/softmax.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <limits>
@@ -201,7 +201,7 @@ struct NeonFp16SimdOps {
   }
 };
 
-template <typename T, typename VT, typename Ops, int N>
+template <typename T, typename AccT, typename VT, typename Ops, int N>
 void softmax(const array& in, array& out) {
   Ops ops;
 
@@ -218,13 +218,21 @@ void softmax(const array& in, array& out) {
     VT vmaximum = ops.init(-std::numeric_limits<float>::infinity());
     size_t s = M;
     while (s >= N) {
-      vmaximum = ops.max(ops.load(current_in_ptr), vmaximum);
+      VT vals;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vals = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vals[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vmaximum = ops.max(vals, vmaximum);
       current_in_ptr += N;
       s -= N;
     }
-    T maximum = ops.reduce_max(vmaximum);
+    AccT maximum = ops.reduce_max(vmaximum);
     while (s-- > 0) {
-      maximum = std::max(maximum, *current_in_ptr);
+      maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
       current_in_ptr++;
     }
 
@@ -234,18 +242,29 @@ void softmax(const array& in, array& out) {
     current_in_ptr = in_ptr;
     s = M;
     while (s >= N) {
-      VT vexp = ops.exp(ops.sub(*(VT*)current_in_ptr, maximum));
-      ops.store(current_out_ptr, vexp);
-      *(VT*)current_out_ptr = vexp;
+      VT vexp;
+      if constexpr (std::is_same<T, AccT>::value) {
+        vexp = ops.load(current_in_ptr);
+      } else {
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+      }
+      vexp = ops.exp(ops.sub(vexp, maximum));
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, vexp);
+      }
       vnormalizer = ops.add(vnormalizer, vexp);
       current_in_ptr += N;
       current_out_ptr += N;
       s -= N;
     }
-    T normalizer = ops.reduce_add(vnormalizer);
+    AccT normalizer = ops.reduce_add(vnormalizer);
     while (s-- > 0) {
-      T _exp = std::exp(*current_in_ptr - maximum);
-      *current_out_ptr = _exp;
+      AccT _exp = std::exp(*current_in_ptr - maximum);
+      if (std::is_same<T, AccT>::value) {
+        *current_out_ptr = _exp;
+      }
       normalizer += _exp;
       current_in_ptr++;
       current_out_ptr++;
@@ -254,14 +273,33 @@ void softmax(const array& in, array& out) {
 
     // Normalize
     current_out_ptr = out_ptr;
+    current_in_ptr = in_ptr;
     s = M;
     while (s >= N) {
-      ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
+      if constexpr (std::is_same<T, AccT>::value) {
+        ops.store(current_out_ptr, ops.mul(*(VT*)current_out_ptr, normalizer));
+      } else {
+        VT vexp;
+        for (int i = 0; i < N; ++i) {
+          vexp[i] = static_cast<AccT>(current_in_ptr[i]);
+        }
+        vexp = ops.mul(ops.exp(ops.sub(vexp, maximum)), normalizer);
+        for (int i = 0; i < N; ++i) {
+          current_out_ptr[i] = vexp[i];
+        }
+        current_in_ptr += N;
+      }
       current_out_ptr += N;
       s -= N;
     }
     while (s-- > 0) {
-      *current_out_ptr *= normalizer;
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr *= normalizer;
+      } else {
+        AccT _exp = std::exp(*current_in_ptr - maximum);
+        *current_out_ptr = static_cast<T>(_exp * normalizer);
+        current_in_ptr++;
+      }
       current_out_ptr++;
     }
   }
@@ -308,15 +346,29 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
           "Softmax is defined only for floating point types");
       break;
     case float32:
-      softmax<float, simd_float16, AccelerateSimdOps<float, simd_float16>, 16>(
-          in, out);
+      softmax<
+          float,
+          float,
+          simd_float16,
+          AccelerateSimdOps<float, simd_float16>,
+          16>(in, out);
       break;
     case float16:
-      softmax<
-          float16_t,
-          float16x8_t,
-          NeonFp16SimdOps<float16_t, float16x8_t>,
-          8>(in, out);
+      if (precise_) {
+        softmax<
+            float16_t,
+            float,
+            simd_float16,
+            AccelerateSimdOps<float, simd_float16>,
+            16>(in, out);
+      } else {
+        softmax<
+            float16_t,
+            float16_t,
+            float16x8_t,
+            NeonFp16SimdOps<float16_t, float16x8_t>,
+            8>(in, out);
+      }
       break;
     case bfloat16:
       eval(inputs, out);

mlx/backend/common/CMakeLists.txt

@@ -44,7 +44,6 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
-  ${CMAKE_CURRENT_SOURCE_DIR}/rope.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp
@@ -53,6 +52,8 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/qrf.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/svd.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
   ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp
 )
 

mlx/backend/common/binary.cpp

@@ -179,18 +179,16 @@ void LogAddExp::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  if (is_floating_point(out.dtype())) {
-    if (out.dtype() == float32) {
-      binary_op<float>(a, b, out, detail::LogAddExp());
-    } else if (out.dtype() == float16) {
-      binary_op<float16_t>(a, b, out, detail::LogAddExp());
-    } else if (out.dtype() == bfloat16) {
-      binary_op<bfloat16_t>(a, b, out, detail::LogAddExp());
-    } else {
-      std::ostringstream err;
-      err << "[logaddexp] Does not support " << out.dtype();
-      throw std::invalid_argument(err.str());
-    }
+  if (out.dtype() == float32) {
+    binary_op<float>(a, b, out, detail::LogAddExp());
+  } else if (out.dtype() == float16) {
+    binary_op<float16_t>(a, b, out, detail::LogAddExp());
+  } else if (out.dtype() == bfloat16) {
+    binary_op<bfloat16_t>(a, b, out, detail::LogAddExp());
+  } else if (issubdtype(out.dtype(), inexact)) {
+    std::ostringstream err;
+    err << "[logaddexp] Does not support " << out.dtype();
+    throw std::invalid_argument(err.str());
   } else {
     throw std::invalid_argument(
         "[logaddexp] Cannot compute logaddexp for arrays with"

mlx/backend/common/compiled.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2023-2024 Apple Inc.
 
 #include "mlx/backend/common/compiled.h"
+#include "mlx/graph_utils.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
 
@@ -81,13 +82,27 @@ std::string build_lib_name(
     const std::vector<array>& outputs,
     const std::vector<array>& tape,
     const std::unordered_set<uintptr_t>& constant_ids) {
+  NodeNamer namer;
   std::ostringstream os;
   std::ostringstream constant_hasher;
 
+  // Fill the input names. This is not really necessary, I just like having A,
+  // B, C, ... as the inputs.
+  for (auto& x : inputs) {
+    namer.get_name(x);
+  }
+
   // The primitives describing the tape. For unary and binary primitives this
   // must be enough to describe the full computation.
   for (auto& a : tape) {
+    // name and type of output
+    os << namer.get_name(a) << kindof(a.dtype()) << a.itemsize();
+    // computation performed
     a.primitive().print(os);
+    // name of inputs to the function
+    for (auto& inp : a.inputs()) {
+      os << namer.get_name(inp);
+    }
   }
   os << "_";
 
@@ -111,4 +126,102 @@ std::string build_lib_name(
   return os.str();
 }
 
+bool compiled_check_contiguity(
+    const std::vector<array>& inputs,
+    const std::vector<int>& shape) {
+  bool contiguous = true;
+  bool all_contig = true;
+  bool all_row_contig = true;
+  bool all_col_contig = true;
+  int non_scalar_inputs = 0;
+  for (const auto& x : inputs) {
+    if (is_scalar(x)) {
+      continue;
+    }
+    non_scalar_inputs++;
+    bool shape_eq = x.shape() == shape;
+    all_contig &= (x.flags().contiguous && shape_eq);
+    all_row_contig &= (x.flags().row_contiguous && shape_eq);
+    all_col_contig &= (x.flags().col_contiguous && shape_eq);
+  }
+  if (non_scalar_inputs > 1 && !all_row_contig && !all_col_contig) {
+    contiguous = false;
+  } else if (non_scalar_inputs == 1 && !all_contig) {
+    contiguous = false;
+  } else if (non_scalar_inputs == 0 && !shape.empty()) {
+    contiguous = false;
+  }
+  return contiguous;
+}
+
+void compiled_allocate_outputs(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const std::vector<array>& inputs_,
+    const std::unordered_set<uintptr_t>& constant_ids_,
+    bool contiguous,
+    bool move_buffers /* = false */) {
+  if (contiguous) {
+    int o = 0;
+    std::vector<size_t> strides;
+    size_t data_size;
+    array::Flags flags;
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+      auto& in = inputs[i];
+      // Conditions for donation
+      // - Correct size
+      // - Not a scalar
+      // - Donatable
+      // - Not a constant
+      if (in.itemsize() == outputs[o].itemsize() && !is_scalar(in) &&
+          in.is_donatable() &&
+          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+        if (move_buffers) {
+          outputs[o++].move_shared_buffer(in);
+        } else {
+          outputs[o++].copy_shared_buffer(in);
+        }
+      }
+      // Get representative input flags to properly set non-donated outputs
+      if (strides.empty() && in.size() == outputs[0].size()) {
+        strides = in.strides();
+        flags = in.flags();
+        data_size = in.data_size();
+      }
+    }
+    for (; o < outputs.size(); ++o) {
+      outputs[o].set_data(
+          allocator::malloc_or_wait(data_size * outputs[o].itemsize()),
+          data_size,
+          strides,
+          flags);
+    }
+  } else {
+    int o = 0;
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+      auto& in = inputs[i];
+      // Conditions for donation
+      // - Row contiguous
+      // - Donatable
+      // - Correct size
+      // - Not a constant
+      if (in.flags().row_contiguous && in.nbytes() == outputs[o].nbytes() &&
+          in.is_donatable() &&
+          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
+        if (move_buffers) {
+          outputs[o].move_shared_buffer(
+              in, outputs[o].strides(), in.flags(), in.data_size());
+        } else {
+          outputs[o].copy_shared_buffer(
+              in, outputs[o].strides(), in.flags(), in.data_size());
+        }
+        o++;
+      }
+    }
+    for (; o < outputs.size(); ++o) {
+      outputs[o].set_data(allocator::malloc_or_wait(outputs[o].nbytes()));
+    }
+  }
+}
+
 } // namespace mlx::core

mlx/backend/common/compiled.h

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

mlx/backend/common/compiled_cpu.cpp

@@ -52,8 +52,25 @@ void* compile(
     return nullptr;
   }
 
+  std::string kernel_file_name;
+
+  // Deal with long kernel names. Maximum length for files on macOS is 255
+  // characters. Clip file name with a little extra room and append a 16
+  // character hash.
+  constexpr int max_file_name_length = 245;
+  if (kernel_name.size() > max_file_name_length) {
+    std::ostringstream file_name;
+    file_name
+        << std::string_view(kernel_name).substr(0, max_file_name_length - 16);
+    auto file_id = std::hash<std::string>{}(kernel_name);
+    file_name << "_" << std::hex << std::setw(16) << file_id << std::dec;
+    kernel_file_name = file_name.str();
+  } else {
+    kernel_file_name = kernel_name;
+  }
+
   std::ostringstream shared_lib_name;
-  shared_lib_name << "lib" << kernel_name << ".so";
+  shared_lib_name << "lib" << kernel_file_name << ".so";
   auto shared_lib_path = get_temp_file(shared_lib_name.str());
   bool lib_exists = false;
   {
@@ -64,7 +81,7 @@ void* compile(
   if (!lib_exists) {
     // Open source file and write source code to it
     std::ostringstream source_file_name;
-    source_file_name << kernel_name << ".cpp";
+    source_file_name << kernel_file_name << ".cpp";
     auto source_file_path = get_temp_file(source_file_name.str());
 
     std::ofstream source_file(source_file_path);
@@ -248,28 +265,7 @@ void Compiled::eval_cpu(
 
   // Figure out which kernel we are using
   auto& shape = outputs[0].shape();
-  bool contiguous = true;
-  {
-    bool all_contig = true;
-    bool all_row_contig = true;
-    bool all_col_contig = true;
-    int non_scalar_inputs = 0;
-    for (auto& x : inputs) {
-      if (is_scalar(x)) {
-        continue;
-      }
-      non_scalar_inputs++;
-      bool shape_eq = x.shape() == shape;
-      all_contig &= (x.flags().contiguous && shape_eq);
-      all_row_contig &= (x.flags().row_contiguous && shape_eq);
-      all_col_contig &= (x.flags().col_contiguous && shape_eq);
-    }
-    if (non_scalar_inputs > 1 && !all_row_contig && !all_col_contig) {
-      contiguous = false;
-    } else if (non_scalar_inputs == 1 && !all_contig) {
-      contiguous = false;
-    }
-  }
+  bool contiguous = compiled_check_contiguity(inputs, shape);
 
   // Handle all broadcasting and collect function input arguments
   std::vector<void*> args;
@@ -342,56 +338,8 @@ void Compiled::eval_cpu(
     fn_ptr = compile(kernel_name, kernel.str());
   }
 
-  // Allocate space for the outputs possibly with input donation
-  if (contiguous) {
-    int o = 0;
-    std::vector<size_t> strides;
-    size_t data_size;
-    array::Flags flags;
-    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
-      auto& in = inputs[i];
-      // Conditions for donation
-      // - Contiguous
-      // - Donatable
-      // - Correct size
-      // - Not a constant
-      if (in.flags().contiguous && !is_scalar(in) && in.is_donatable() &&
-          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        outputs[o++].copy_shared_buffer(in);
-      }
-      // Get representative input flags to properly set non-donated outputs
-      if (strides.empty() && in.size() == outputs[0].size()) {
-        strides = in.strides();
-        flags = in.flags();
-        data_size = in.data_size();
-      }
-    }
-    for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(
-          allocator::malloc_or_wait(data_size * outputs[o].itemsize()),
-          data_size,
-          strides,
-          flags);
-    }
-  } else {
-    int o = 0;
-    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
-      auto& in = inputs[i];
-      // Conditions for donation
-      // - Row contiguous
-      // - Donatable
-      // - Correct size
-      // - Not a constant
-      if (in.flags().row_contiguous && in.nbytes() == outputs[o].nbytes() &&
-          in.is_donatable() &&
-          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        outputs[o++].copy_shared_buffer(in);
-      }
-    }
-    for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(allocator::malloc_or_wait(outputs[o].nbytes()));
-    }
-  }
+  compiled_allocate_outputs(
+      inputs, outputs, inputs_, constant_ids_, contiguous, false);
 
   for (auto& x : outputs) {
     args.push_back(x.data<void>());

mlx/backend/common/copy.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <numeric>
 
@@ -25,121 +25,196 @@ void copy_vector(const array& src, array& dst) {
   std::copy(src_ptr, src_ptr + src.data_size(), dst_ptr);
 }
 
-template <typename SrcT, typename DstT>
-void copy_general_dim1(const array& src, array& dst) {
+template <typename SrcT, typename DstT, typename stride_t>
+void copy_general_dim1(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    int64_t i_offset) {
   const SrcT* src_ptr = src.data<SrcT>();
   DstT* dst_ptr = dst.data<DstT>();
-  size_t src_idx = 0;
-  size_t dst_idx = 0;
-  for (size_t i = 0; i < src.shape()[0]; ++i) {
+  stride_t src_idx = i_offset;
+  stride_t dst_idx = 0;
+  for (int i = 0; i < data_shape[0]; ++i) {
     dst_ptr[dst_idx++] = static_cast<DstT>(src_ptr[src_idx]);
-    src_idx += src.strides()[0];
+    src_idx += i_strides[0];
   }
 }
 
 template <typename SrcT, typename DstT>
-void copy_general_dim2(const array& src, array& dst) {
+inline void copy_general_dim1(const array& src, array& dst) {
+  return copy_general_dim1<SrcT, DstT, size_t>(
+      src, dst, src.shape(), src.strides(), 0);
+}
+
+template <typename SrcT, typename DstT, typename stride_t>
+void copy_general_dim2(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    int64_t i_offset) {
   const SrcT* src_ptr = src.data<SrcT>();
   DstT* dst_ptr = dst.data<DstT>();
-  size_t src_idx = 0;
-  size_t dst_idx = 0;
-  for (size_t i = 0; i < src.shape()[0]; ++i) {
-    for (size_t j = 0; j < src.shape()[1]; ++j) {
+  stride_t src_idx = i_offset;
+  stride_t dst_idx = 0;
+  for (int i = 0; i < data_shape[0]; ++i) {
+    for (int j = 0; j < data_shape[1]; ++j) {
       dst_ptr[dst_idx++] = static_cast<DstT>(src_ptr[src_idx]);
-      src_idx += src.strides()[1];
+      src_idx += i_strides[1];
     }
-    src_idx += src.strides()[0] - src.strides()[1] * src.shape()[1];
+    src_idx += i_strides[0] - i_strides[1] * data_shape[1];
   }
 }
 
 template <typename SrcT, typename DstT>
-void copy_general_dim3(const array& src, array& dst) {
+inline void copy_general_dim2(const array& src, array& dst) {
+  return copy_general_dim2<SrcT, DstT, size_t>(
+      src, dst, src.shape(), src.strides(), 0);
+}
+
+template <typename SrcT, typename DstT, typename stride_t>
+void copy_general_dim3(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    int64_t i_offset) {
   const SrcT* src_ptr = src.data<SrcT>();
   DstT* dst_ptr = dst.data<DstT>();
-  size_t src_idx = 0;
-  size_t dst_idx = 0;
-  for (size_t i = 0; i < src.shape()[0]; ++i) {
-    for (size_t j = 0; j < src.shape()[1]; ++j) {
-      for (size_t k = 0; k < src.shape()[2]; ++k) {
+  stride_t src_idx = i_offset;
+  stride_t dst_idx = 0;
+  for (int i = 0; i < data_shape[0]; ++i) {
+    for (int j = 0; j < data_shape[1]; ++j) {
+      for (int k = 0; k < data_shape[2]; ++k) {
         dst_ptr[dst_idx++] = static_cast<DstT>(src_ptr[src_idx]);
-        src_idx += src.strides()[2];
+        src_idx += i_strides[2];
       }
-      src_idx += src.strides()[1] - src.strides()[2] * src.shape()[2];
+      src_idx += i_strides[1] - i_strides[2] * data_shape[2];
     }
-    src_idx += src.strides()[0] - src.strides()[1] * src.shape()[1];
+    src_idx += i_strides[0] - i_strides[1] * data_shape[1];
   }
 }
 
 template <typename SrcT, typename DstT>
-void copy_general_dim4(const array& src, array& dst) {
+inline void copy_general_dim3(const array& src, array& dst) {
+  return copy_general_dim3<SrcT, DstT, size_t>(
+      src, dst, src.shape(), src.strides(), 0);
+}
+
+template <typename SrcT, typename DstT, typename stride_t>
+void copy_general_dim4(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    int64_t i_offset) {
   const SrcT* src_ptr = src.data<SrcT>();
   DstT* dst_ptr = dst.data<DstT>();
-  size_t src_idx = 0;
-  size_t dst_idx = 0;
-  for (size_t i = 0; i < src.shape()[0]; ++i) {
-    for (size_t j = 0; j < src.shape()[1]; ++j) {
-      for (size_t k = 0; k < src.shape()[2]; ++k) {
-        for (size_t ii = 0; ii < src.shape()[3]; ++ii) {
+  stride_t src_idx = i_offset;
+  stride_t dst_idx = 0;
+  for (int i = 0; i < data_shape[0]; ++i) {
+    for (int j = 0; j < data_shape[1]; ++j) {
+      for (int k = 0; k < data_shape[2]; ++k) {
+        for (int ii = 0; ii < data_shape[3]; ++ii) {
           dst_ptr[dst_idx++] = static_cast<DstT>(src_ptr[src_idx]);
-          src_idx += src.strides()[3];
+          src_idx += i_strides[3];
         }
-        src_idx += src.strides()[2] - src.strides()[3] * src.shape()[3];
+        src_idx += i_strides[2] - i_strides[3] * data_shape[3];
       }
-      src_idx += src.strides()[1] - src.strides()[2] * src.shape()[2];
+      src_idx += i_strides[1] - i_strides[2] * data_shape[2];
     }
-    src_idx += src.strides()[0] - src.strides()[1] * src.shape()[1];
+    src_idx += i_strides[0] - i_strides[1] * data_shape[1];
   }
 }
 
 template <typename SrcT, typename DstT>
-void copy_general(const array& src, array& dst) {
+inline void copy_general_dim4(const array& src, array& dst) {
+  return copy_general_dim4<SrcT, DstT, size_t>(
+      src, dst, src.shape(), src.strides(), 0);
+}
+
+template <typename SrcT, typename DstT, typename stride_t>
+void copy_general(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    int64_t i_offset) {
   switch (src.ndim()) {
     case 1:
-      copy_general_dim1<SrcT, DstT>(src, dst);
+      copy_general_dim1<SrcT, DstT, stride_t>(
+          src, dst, data_shape, i_strides, i_offset);
       return;
     case 2:
-      copy_general_dim2<SrcT, DstT>(src, dst);
+      copy_general_dim2<SrcT, DstT, stride_t>(
+          src, dst, data_shape, i_strides, i_offset);
       return;
     case 3:
-      copy_general_dim3<SrcT, DstT>(src, dst);
+      copy_general_dim3<SrcT, DstT, stride_t>(
+          src, dst, data_shape, i_strides, i_offset);
       return;
     case 4:
-      copy_general_dim4<SrcT, DstT>(src, dst);
+      copy_general_dim4<SrcT, DstT, stride_t>(
+          src, dst, data_shape, i_strides, i_offset);
       return;
   }
 
-  auto src_ptr = src.data<SrcT>();
+  auto src_ptr = src.data<SrcT>() + i_offset;
   auto dst_ptr = dst.data<DstT>();
   for (size_t i = 0; i < dst.size(); ++i) {
-    size_t src_elem = elem_to_loc(i, src.shape(), src.strides());
+    stride_t src_elem = elem_to_loc(i, data_shape, i_strides);
     dst_ptr[i] = static_cast<DstT>(src_ptr[src_elem]);
   }
 }
 
-template <typename SrcT, typename DstT, int D>
+template <typename SrcT, typename DstT>
+inline void copy_general(const array& src, array& dst) {
+  return copy_general<SrcT, DstT, size_t>(
+      src, dst, src.shape(), src.strides(), 0);
+}
+
+template <typename SrcT, typename DstT, typename stride_t>
+inline void copy_general(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    const std::vector<stride_t>& o_strides,
+    int64_t i_offset,
+    int64_t o_offset) {
+  return copy_general<SrcT, DstT, stride_t>(
+      src, dst, data_shape, i_strides, i_offset);
+}
+
+template <typename SrcT, typename DstT, typename stride_t, int D>
 inline void copy_general_general_dims(
     const array& src,
     array& dst,
-    size_t offset_src,
-    size_t offset_dst) {
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    const std::vector<stride_t>& o_strides,
+    stride_t i_offset,
+    stride_t o_offset) {
   if constexpr (D > 1) {
     int axis = src.ndim() - D;
-    auto stride_src = src.strides()[axis];
-    auto stride_dst = dst.strides()[axis];
-    auto N = src.shape(axis);
+    auto stride_src = i_strides[axis];
+    auto stride_dst = o_strides[axis];
+    auto N = data_shape[axis];
     for (int i = 0; i < N; i++) {
-      copy_general_general_dims<SrcT, DstT, D - 1>(
-          src, dst, offset_src, offset_dst);
-      offset_src += stride_src;
-      offset_dst += stride_dst;
+      copy_general_general_dims<SrcT, DstT, stride_t, D - 1>(
+          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
+      i_offset += stride_src;
+      o_offset += stride_dst;
     }
   } else {
     int axis = src.ndim() - 1;
-    auto stride_src = src.strides()[axis];
-    auto stride_dst = dst.strides()[axis];
-    auto N = src.shape(axis);
-    const SrcT* src_ptr = src.data<SrcT>() + offset_src;
-    DstT* dst_ptr = dst.data<DstT>() + offset_dst;
+    auto stride_src = i_strides[axis];
+    auto stride_dst = o_strides[axis];
+    auto N = data_shape[axis];
+    const SrcT* src_ptr = src.data<SrcT>() + i_offset;
+    DstT* dst_ptr = dst.data<DstT>() + o_offset;
     for (int i = 0; i < N; i++) {
       *dst_ptr = static_cast<DstT>(*src_ptr);
       src_ptr += stride_src;
@@ -148,37 +223,56 @@ inline void copy_general_general_dims(
   }
 }
 
-template <typename SrcT, typename DstT>
-void copy_general_general(const array& src, array& dst) {
+template <typename SrcT, typename DstT, typename stride_t>
+void copy_general_general(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    const std::vector<stride_t>& o_strides,
+    stride_t i_offset,
+    stride_t o_offset) {
   switch (src.ndim()) {
     case 1:
-      copy_general_general_dims<SrcT, DstT, 1>(src, dst, 0, 0);
+      copy_general_general_dims<SrcT, DstT, stride_t, 1>(
+          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
       return;
     case 2:
-      copy_general_general_dims<SrcT, DstT, 2>(src, dst, 0, 0);
+      copy_general_general_dims<SrcT, DstT, stride_t, 2>(
+          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
       return;
     case 3:
-      copy_general_general_dims<SrcT, DstT, 3>(src, dst, 0, 0);
+      copy_general_general_dims<SrcT, DstT, stride_t, 3>(
+          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
       return;
     case 4:
-      copy_general_general_dims<SrcT, DstT, 4>(src, dst, 0, 0);
+      copy_general_general_dims<SrcT, DstT, stride_t, 4>(
+          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
       return;
     case 5:
-      copy_general_general_dims<SrcT, DstT, 5>(src, dst, 0, 0);
+      copy_general_general_dims<SrcT, DstT, stride_t, 5>(
+          src, dst, data_shape, i_strides, o_strides, i_offset, o_offset);
       return;
   }
 
   int size = std::accumulate(
-      src.shape().begin() - 5, src.shape().end(), 1, std::multiplies<int>());
+      data_shape.begin() - 5, data_shape.end(), 1, std::multiplies<int>());
   for (int i = 0; i < src.size(); i += size) {
-    size_t offset_src = elem_to_loc(i, src.shape(), src.strides());
-    size_t offset_dst = elem_to_loc(i, dst.shape(), dst.strides());
-    copy_general_general_dims<SrcT, DstT, 5>(src, dst, offset_src, offset_dst);
+    stride_t src_offset = i_offset + elem_to_loc(i, data_shape, i_strides);
+    stride_t dst_offset = o_offset + elem_to_loc(i, dst.shape(), o_strides);
+    copy_general_general_dims<SrcT, DstT, stride_t, 5>(
+        src, dst, data_shape, i_strides, o_strides, src_offset, dst_offset);
   }
 }
 
 template <typename SrcT, typename DstT>
-void copy(const array& src, array& dst, CopyType ctype) {
+inline void copy_general_general(const array& src, array& dst) {
+  return copy_general_general<SrcT, DstT, size_t>(
+      src, dst, src.shape(), src.strides(), dst.strides(), 0, 0);
+}
+
+template <typename SrcT, typename DstT, typename... Args>
+void copy(const array& src, array& dst, CopyType ctype, Args&&... args) {
   switch (ctype) {
     case CopyType::Scalar:
       copy_single<SrcT, DstT>(src, dst);
@@ -187,54 +281,103 @@ void copy(const array& src, array& dst, CopyType ctype) {
       copy_vector<SrcT, DstT>(src, dst);
       return;
     case CopyType::General:
-      copy_general<SrcT, DstT>(src, dst);
+      copy_general<SrcT, DstT>(src, dst, std::forward<Args>(args)...);
       return;
     case CopyType::GeneralGeneral:
-      copy_general_general<SrcT, DstT>(src, dst);
+      copy_general_general<SrcT, DstT>(src, dst, std::forward<Args>(args)...);
   }
 }
 
-template <typename SrcT>
-void copy(const array& src, array& dst, CopyType ctype) {
+template <typename SrcT, typename... Args>
+void copy(const array& src, array& dst, CopyType ctype, Args&&... args) {
   switch (dst.dtype()) {
     case bool_:
-      copy<SrcT, bool>(src, dst, ctype);
+      copy<SrcT, bool>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint8:
-      copy<SrcT, uint8_t>(src, dst, ctype);
+      copy<SrcT, uint8_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint16:
-      copy<SrcT, uint16_t>(src, dst, ctype);
+      copy<SrcT, uint16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint32:
-      copy<SrcT, uint32_t>(src, dst, ctype);
+      copy<SrcT, uint32_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case uint64:
-      copy<SrcT, uint64_t>(src, dst, ctype);
+      copy<SrcT, uint64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int8:
-      copy<SrcT, int8_t>(src, dst, ctype);
+      copy<SrcT, int8_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int16:
-      copy<SrcT, int16_t>(src, dst, ctype);
+      copy<SrcT, int16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int32:
-      copy<SrcT, int32_t>(src, dst, ctype);
+      copy<SrcT, int32_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case int64:
-      copy<SrcT, int64_t>(src, dst, ctype);
+      copy<SrcT, int64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case float16:
-      copy<SrcT, float16_t>(src, dst, ctype);
+      copy<SrcT, float16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case float32:
-      copy<SrcT, float>(src, dst, ctype);
+      copy<SrcT, float>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case bfloat16:
-      copy<SrcT, bfloat16_t>(src, dst, ctype);
+      copy<SrcT, bfloat16_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
     case complex64:
-      copy<SrcT, complex64_t>(src, dst, ctype);
+      copy<SrcT, complex64_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+  }
+}
+
+template <typename... Args>
+inline void copy_inplace_dispatch(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Args&&... args) {
+  switch (src.dtype()) {
+    case bool_:
+      copy<bool>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case uint8:
+      copy<uint8_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case uint16:
+      copy<uint16_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case uint32:
+      copy<uint32_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case uint64:
+      copy<uint64_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case int8:
+      copy<int8_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case int16:
+      copy<int16_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case int32:
+      copy<int32_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case int64:
+      copy<int64_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case float16:
+      copy<float16_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case float32:
+      copy<float>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case bfloat16:
+      copy<bfloat16_t>(src, dst, ctype, std::forward<Args>(args)...);
+      break;
+    case complex64:
+      copy<complex64_t>(src, dst, ctype, std::forward<Args>(args)...);
       break;
   }
 }
@@ -242,47 +385,7 @@ void copy(const array& src, array& dst, CopyType ctype) {
 } // namespace
 
 void copy_inplace(const array& src, array& dst, CopyType ctype) {
-  switch (src.dtype()) {
-    case bool_:
-      copy<bool>(src, dst, ctype);
-      break;
-    case uint8:
-      copy<uint8_t>(src, dst, ctype);
-      break;
-    case uint16:
-      copy<uint16_t>(src, dst, ctype);
-      break;
-    case uint32:
-      copy<uint32_t>(src, dst, ctype);
-      break;
-    case uint64:
-      copy<uint64_t>(src, dst, ctype);
-      break;
-    case int8:
-      copy<int8_t>(src, dst, ctype);
-      break;
-    case int16:
-      copy<int16_t>(src, dst, ctype);
-      break;
-    case int32:
-      copy<int32_t>(src, dst, ctype);
-      break;
-    case int64:
-      copy<int64_t>(src, dst, ctype);
-      break;
-    case float16:
-      copy<float16_t>(src, dst, ctype);
-      break;
-    case float32:
-      copy<float>(src, dst, ctype);
-      break;
-    case bfloat16:
-      copy<bfloat16_t>(src, dst, ctype);
-      break;
-    case complex64:
-      copy<complex64_t>(src, dst, ctype);
-      break;
-  }
+  return copy_inplace_dispatch(src, dst, ctype);
 }
 
 void copy(const array& src, array& dst, CopyType ctype) {
@@ -312,4 +415,62 @@ void copy(const array& src, array& dst, CopyType ctype) {
   copy_inplace(src, dst, ctype);
 }
 
+template <typename stride_t>
+void copy_inplace(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    const std::vector<stride_t>& o_strides,
+    int64_t i_offset,
+    int64_t o_offset,
+    CopyType ctype) {
+  switch (ctype) {
+    case CopyType::General:
+    case CopyType::GeneralGeneral:
+      return copy_inplace_dispatch(
+          src,
+          dst,
+          ctype,
+          data_shape,
+          i_strides,
+          o_strides,
+          i_offset,
+          o_offset);
+
+    case CopyType::Scalar:
+    case CopyType::Vector:
+      return copy_inplace_dispatch(src, dst, ctype);
+  }
+}
+
+template <>
+void copy_inplace<int64_t>(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<int64_t>& i_strides,
+    const std::vector<int64_t>& o_strides,
+    int64_t i_offset,
+    int64_t o_offset,
+    CopyType ctype) {
+  switch (ctype) {
+    case CopyType::General:
+    case CopyType::GeneralGeneral:
+      return copy_inplace_dispatch(
+          src,
+          dst,
+          ctype,
+          data_shape,
+          i_strides,
+          o_strides,
+          i_offset,
+          o_offset);
+
+    case CopyType::Scalar:
+    case CopyType::Vector:
+      return copy_inplace_dispatch(src, dst, ctype);
+  }
+}
+
 } // namespace mlx::core

mlx/backend/common/copy.h

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #pragma once
 
@@ -26,4 +26,15 @@ enum class CopyType {
 void copy(const array& src, array& dst, CopyType ctype);
 void copy_inplace(const array& src, array& dst, CopyType ctype);
 
+template <typename stride_t>
+void copy_inplace(
+    const array& src,
+    array& dst,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    const std::vector<stride_t>& o_strides,
+    int64_t i_offset,
+    int64_t o_offset,
+    CopyType ctype);
+
 } // namespace mlx::core

mlx/backend/common/default_primitives.cpp

@@ -51,11 +51,13 @@ DEFAULT(Cosh)
 DEFAULT_MULTI(CustomVJP)
 DEFAULT_MULTI(Depends)
 DEFAULT(Divide)
+DEFAULT(NumberOfElements)
 DEFAULT(Remainder)
 DEFAULT(Equal)
 DEFAULT(Erf)
 DEFAULT(ErfInv)
 DEFAULT(Exp)
+DEFAULT(Expm1)
 DEFAULT(FFT)
 DEFAULT(Floor)
 DEFAULT(Full)
@@ -93,6 +95,7 @@ DEFAULT(Sign)
 DEFAULT(Sin)
 DEFAULT(Sinh)
 DEFAULT(Slice)
+DEFAULT(SliceUpdate)
 DEFAULT(Softmax)
 DEFAULT(Sort)
 DEFAULT_MULTI(Split)
@@ -100,9 +103,11 @@ DEFAULT(Square)
 DEFAULT(Sqrt)
 DEFAULT(StopGradient)
 DEFAULT(Subtract)
+DEFAULT_MULTI(SVD)
 DEFAULT(Tan)
 DEFAULT(Tanh)
 DEFAULT(Transpose)
+DEFAULT(Inverse)
 
 namespace {
 

mlx/backend/common/inverse.cpp

@@ -0,0 +1,104 @@
+// 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 {
+
+void inverse_impl(const array& a, array& inv) {
+  // 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,
+        /* n = */ &N,
+        /* a = */ inv.data<float>() + N * N * i,
+        /* lda = */ &N,
+        /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "inverse_impl: LU factorization failed with error code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    static const int lwork_query = -1;
+    float workspace_size = 0;
+
+    // Compute workspace size.
+    sgetri_(
+        /* m = */ &N,
+        /* a = */ nullptr,
+        /* lda = */ &N,
+        /* ipiv = */ nullptr,
+        /* work = */ &workspace_size,
+        /* lwork = */ &lwork_query,
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "inverse_impl: LU workspace calculation failed with error code "
+         << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    const int lwork = workspace_size;
+    auto scratch =
+        array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
+
+    // Compute inverse.
+    sgetri_(
+        /* m = */ &N,
+        /* a = */ inv.data<float>() + N * N * i,
+        /* lda = */ &N,
+        /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
+        /* work = */ static_cast<float*>(scratch.buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "inverse_impl: inversion failed with error code " << info;
+      throw std::runtime_error(ss.str());
+    }
+  }
+}
+
+void Inverse::eval(const std::vector<array>& inputs, array& output) {
+  if (inputs[0].dtype() != float32) {
+    throw std::runtime_error("[Inverse::eval] only supports float32.");
+  }
+  inverse_impl(inputs[0], output);
+}
+
+std::pair<std::vector<array>, std::vector<int>> Inverse::vmap(
+    const std::vector<array>& inputs,
+    const std::vector<int>& axes) {
+  auto ax = axes[0] >= 0 ? 0 : -1;
+  auto a = axes[0] > 0 ? moveaxis(inputs[0], axes[0], 0, stream()) : inputs[0];
+  return {{linalg::inv(a, stream())}, {ax}};
+}
+
+} // namespace mlx::core

mlx/backend/common/lapack_helper.h

@@ -0,0 +1,23 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#ifdef ACCELERATE_NEW_LAPACK
+#include <Accelerate/Accelerate.h>
+#else
+#include <lapack.h>
+#endif
+
+#if defined(LAPACK_GLOBAL) || defined(LAPACK_NAME)
+
+// This is to work around a change in the function signatures of lapack >= 3.9.1
+// where functions taking char* also include a strlen argument, see a similar
+// change in OpenCV:
+// https://github.com/opencv/opencv/blob/1eb061f89de0fb85c4c75a2deeb0f61a961a63ad/cmake/OpenCVFindLAPACK.cmake#L57
+#define MLX_LAPACK_FUNC(f) LAPACK_##f
+
+#else
+
+#define MLX_LAPACK_FUNC(f) f##_
+
+#endif

mlx/backend/common/ops.h

@@ -241,6 +241,13 @@ struct Exp {
   }
 };
 
+struct Expm1 {
+  template <typename T>
+  T operator()(T x) {
+    return expm1(x);
+  };
+};
+
 struct Floor {
   template <typename T>
   T operator()(T x) {

mlx/backend/common/primitives.cpp

@@ -22,7 +22,7 @@ namespace mlx::core {
 void Abs::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  if (is_unsigned(in.dtype())) {
+  if (issubdtype(in.dtype(), unsignedinteger)) {
     // No-op for unsigned types
     out.copy_shared_buffer(in);
   } else {
@@ -37,7 +37,7 @@ void Arange::eval(const std::vector<array>& inputs, array& out) {
 void ArcCos::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::ArcCos());
   } else {
     throw std::invalid_argument(
@@ -49,7 +49,7 @@ void ArcCos::eval(const std::vector<array>& inputs, array& out) {
 void ArcCosh::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::ArcCosh());
   } else {
     throw std::invalid_argument(
@@ -61,7 +61,7 @@ void ArcCosh::eval(const std::vector<array>& inputs, array& out) {
 void ArcSin::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::ArcSin());
   } else {
     throw std::invalid_argument(
@@ -73,7 +73,7 @@ void ArcSin::eval(const std::vector<array>& inputs, array& out) {
 void ArcSinh::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::ArcSinh());
   } else {
     throw std::invalid_argument(
@@ -85,7 +85,7 @@ void ArcSinh::eval(const std::vector<array>& inputs, array& out) {
 void ArcTan::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::ArcTan());
   } else {
     throw std::invalid_argument(
@@ -97,7 +97,7 @@ void ArcTan::eval(const std::vector<array>& inputs, array& out) {
 void ArcTanh::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::ArcTanh());
   } else {
     throw std::invalid_argument(
@@ -171,7 +171,7 @@ void Broadcast::eval(const std::vector<array>& inputs, array& out) {
 void Ceil::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  if (not is_integral(in.dtype())) {
+  if (issubdtype(in.dtype(), inexact)) {
     unary_fp(in, out, detail::Ceil());
   } else {
     // No-op integer types
@@ -211,7 +211,7 @@ void Copy::eval(const std::vector<array>& inputs, array& out) {
 void Cos::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Cos());
   } else {
     throw std::invalid_argument(
@@ -223,7 +223,7 @@ void Cos::eval(const std::vector<array>& inputs, array& out) {
 void Cosh::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Cosh());
   } else {
     throw std::invalid_argument(
@@ -251,6 +251,62 @@ void Depends::eval(
   }
 }
 
+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];
@@ -294,7 +350,7 @@ void ErfInv::eval(const std::vector<array>& inputs, array& out) {
 void Exp::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Exp());
   } else {
     throw std::invalid_argument(
@@ -303,10 +359,22 @@ void Exp::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
+void Expm1::eval(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 1);
+  const auto& in = inputs[0];
+  if (issubdtype(out.dtype(), inexact)) {
+    unary_fp(in, out, detail::Expm1());
+  } else {
+    throw std::invalid_argument(
+        "[expm1] Cannot exponentiate elements in array"
+        " with non floating point type.");
+  }
+}
+
 void Floor::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  if (not is_integral(in.dtype())) {
+  if (issubdtype(in.dtype(), inexact)) {
     unary_fp(in, out, detail::Floor());
   } else {
     // No-op integer types
@@ -332,7 +400,7 @@ void Full::eval(const std::vector<array>& inputs, array& out) {
 void Log::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     switch (base_) {
       case Base::e:
         unary_fp(in, out, detail::Log());
@@ -354,7 +422,7 @@ void Log::eval(const std::vector<array>& inputs, array& out) {
 void Log1p::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Log1p());
   } else {
     throw std::invalid_argument(
@@ -468,27 +536,80 @@ void RandomBits::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void Reshape::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  if (in.flags().row_contiguous) {
+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 flags = in.flags();
     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());
-  } else {
+  }
+  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];
+
+  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
+
+  if (copy_necessary) {
     copy(in, out, in.data_size() == 1 ? CopyType::Scalar : CopyType::General);
+  } else {
+    shared_buffer_reshape(in, out_strides, out);
   }
 }
 
 void Round::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  if (not is_integral(in.dtype())) {
+  if (issubdtype(in.dtype(), inexact)) {
     unary_fp(in, out, detail::Round());
   } else {
     // No-op integer types
@@ -499,7 +620,7 @@ void Round::eval(const std::vector<array>& inputs, array& out) {
 void Sigmoid::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Sigmoid());
   } else {
     throw std::invalid_argument(
@@ -521,7 +642,7 @@ void Sign::eval(const std::vector<array>& inputs, array& out) {
 void Sin::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Sin());
   } else {
     throw std::invalid_argument(
@@ -533,7 +654,7 @@ void Sin::eval(const std::vector<array>& inputs, array& out) {
 void Sinh::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Sinh());
   } else {
     throw std::invalid_argument(
@@ -542,36 +663,33 @@ void Sinh::eval(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void Slice::eval(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-  auto& in = inputs[0];
-  auto strides = in.strides();
-  auto flags = in.flags();
-  size_t data_offset = 0;
+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];
-    strides[i] *= 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
-  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 = out.ndim() - 1; ri >= 0; i++, ri--) {
-    flags.col_contiguous &= strides[i] == f_stride || out.shape(i) == 1;
-    flags.row_contiguous &= strides[ri] == b_stride || out.shape(ri) == 1;
-    f_stride *= out.shape(i);
-    b_stride *= out.shape(ri);
-    if (strides[i] > 0) {
-      data_size *= out.shape(i);
-    }
-  }
+  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.
@@ -585,7 +703,87 @@ void Slice::eval(const std::vector<array>& inputs, array& out) {
     flags.contiguous &= flags.row_contiguous || flags.col_contiguous;
   }
 
-  out.copy_shared_buffer(in, strides, flags, data_size, data_offset);
+  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) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  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);
+
+  // Do copy if needed
+  if (copy_needed) {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    std::vector<int64_t> ostrides{out.strides().begin(), out.strides().end()};
+    copy_inplace<int64_t>(
+        /* const array& src = */ in,
+        /* array& dst = */ out,
+        /* const std::vector<int>& data_shape = */ out.shape(),
+        /* const std::vector<stride_t>& i_strides = */ inp_strides,
+        /* const std::vector<stride_t>& o_strides = */ ostrides,
+        /* int64_t i_offset = */ data_offset,
+        /* int64_t o_offset = */ 0,
+        /* CopyType ctype = */ CopyType::General);
+  } else {
+    std::vector<size_t> ostrides{inp_strides.begin(), inp_strides.end()};
+    shared_buffer_slice(in, ostrides, data_offset, 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) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  auto& in = inputs[0];
+  auto& upd = inputs[1];
+
+  if (upd.size() == 0) {
+    out.copy_shared_buffer(in);
+    return;
+  }
+
+  // Check if materialization is needed
+  auto ctype = in.flags().contiguous && in.size() == in.data_size()
+      ? CopyType::Vector
+      : CopyType::General;
+  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype);
+
+  // Calculate out strides, initial offset and if copy needs to be made
+  auto [data_offset, out_strides] = prepare_slice(out);
+
+  // Do copy
+  std::vector<int64_t> upd_strides{upd.strides().begin(), upd.strides().end()};
+  copy_inplace<int64_t>(
+      /* const array& src = */ upd,
+      /* array& dst = */ out,
+      /* const std::vector<int>& data_shape = */ upd.shape(),
+      /* const std::vector<stride_t>& i_strides = */ upd_strides,
+      /* const std::vector<stride_t>& o_strides = */ out_strides,
+      /* int64_t i_offset = */ 0,
+      /* int64_t o_offset = */ data_offset,
+      /* CopyType ctype = */ CopyType::GeneralGeneral);
 }
 
 void Split::eval(
@@ -664,7 +862,7 @@ void StopGradient::eval(const std::vector<array>& inputs, array& out) {
 void Tan::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Tan());
   } else {
     throw std::invalid_argument(
@@ -676,7 +874,7 @@ void Tan::eval(const std::vector<array>& inputs, array& out) {
 void Tanh::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (is_floating_point(out.dtype())) {
+  if (issubdtype(out.dtype(), inexact)) {
     unary_fp(in, out, detail::Tanh());
   } else {
     throw std::invalid_argument(

mlx/backend/common/reduce.h

@@ -6,8 +6,6 @@
 
 namespace mlx::core {
 
-namespace {
-
 enum ReductionOpType {
   // Self-explanatory. Read everything and produce 1 output.
   ContiguousAllReduce,
@@ -38,6 +36,21 @@ enum ReductionOpType {
   GeneralReduce
 };
 
+struct ReductionPlan {
+  ReductionOpType type;
+  std::vector<int> shape;
+  std::vector<size_t> strides;
+
+  ReductionPlan(
+      ReductionOpType type_,
+      std::vector<int> shape_,
+      std::vector<size_t> strides_)
+      : type(type_), shape(std::move(shape_)), strides(std::move(strides_)) {}
+  ReductionPlan(ReductionOpType type_) : type(type_) {}
+};
+
+namespace {
+
 // Helper for the ndimensional strided loop
 // Should this be in utils?
 inline void nd_loop(
@@ -110,19 +123,6 @@ struct DefaultContiguousReduce {
   }
 };
 
-struct ReductionPlan {
-  ReductionOpType type;
-  std::vector<int> shape;
-  std::vector<size_t> strides;
-
-  ReductionPlan(
-      ReductionOpType type_,
-      std::vector<int> shape_,
-      std::vector<size_t> strides_)
-      : type(type_), shape(std::move(shape_)), strides(std::move(strides_)) {}
-  ReductionPlan(ReductionOpType type_) : type(type_) {}
-};
-
 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() &&

mlx/backend/common/rope.cpp

@@ -1,13 +0,0 @@
-// Copyright © 2023-2024 Apple Inc.
-
-#include "mlx/fast_primitives.h"
-
-namespace mlx::core::fast {
-
-void RoPE::eval_cpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  throw std::runtime_error("NYI");
-}
-
-} // namespace mlx::core::fast

mlx/backend/common/scan.cpp

@@ -222,7 +222,7 @@ void scan_dispatch(
     }
     case Scan::Min: {
       auto op = [](U* o, const U* y, const T* x) { *o = (*x < *y) ? *x : *y; };
-      auto init = (is_floating_point(input.dtype()))
+      auto init = (issubdtype(input.dtype(), floating))
           ? static_cast<U>(std::numeric_limits<float>::infinity())
           : std::numeric_limits<U>::max();
       auto opcs = DefaultContiguousScan<T, U, decltype(op)>(op, init);
@@ -232,7 +232,7 @@ void scan_dispatch(
     }
     case Scan::Max: {
       auto op = [](U* o, const U* y, const T* x) { *o = (*x < *y) ? *y : *x; };
-      auto init = (is_floating_point(input.dtype()))
+      auto init = (issubdtype(input.dtype(), floating))
           ? static_cast<U>(-std::numeric_limits<float>::infinity())
           : std::numeric_limits<U>::max();
       auto opcs = DefaultContiguousScan<T, U, decltype(op)>(op, init);

mlx/backend/common/softmax.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <cassert>
 #include <cmath>
@@ -10,7 +10,7 @@ namespace mlx::core {
 
 namespace {
 
-template <typename T>
+template <typename T, typename AccT>
 void softmax(const array& in, array& out) {
   const T* in_ptr = in.data<T>();
   T* out_ptr = out.data<T>();
@@ -22,26 +22,36 @@ void softmax(const array& in, array& out) {
   for (int i = 0; i < M; i++, in_ptr += N, out_ptr += N) {
     // Find the maximum
     current_in_ptr = in_ptr;
-    T maximum = *current_in_ptr;
+    AccT maximum = *current_in_ptr;
     for (int j = 0; j < N; j++, current_in_ptr++) {
-      maximum = (maximum < *current_in_ptr) ? *current_in_ptr : maximum;
+      maximum = (maximum < *current_in_ptr) ? static_cast<AccT>(*current_in_ptr)
+                                            : maximum;
     }
 
     // Compute the normalizer and the exponentials
-    T normalizer = 0;
+    AccT normalizer = 0;
     current_out_ptr = out_ptr;
     current_in_ptr = in_ptr;
     for (int j = 0; j < N; j++, current_out_ptr++, current_in_ptr++) {
-      T expv = std::exp(*current_in_ptr - maximum);
+      AccT expv = std::exp(*current_in_ptr - maximum);
       normalizer += expv;
-      *current_out_ptr = expv;
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr = expv;
+      }
     }
     normalizer = 1 / normalizer;
 
     // Normalize
+    current_in_ptr = in_ptr;
     current_out_ptr = out_ptr;
     for (int j = 0; j < N; j++, current_out_ptr++) {
-      *current_out_ptr *= normalizer;
+      if constexpr (std::is_same<T, AccT>::value) {
+        *current_out_ptr *= normalizer;
+      } else {
+        auto v = std::exp(*current_in_ptr - maximum);
+        *current_out_ptr = static_cast<T>(v * normalizer);
+        current_in_ptr++;
+      }
     }
   }
 }
@@ -67,11 +77,15 @@ void Softmax::eval(const std::vector<array>& inputs, array& out) {
     }
   };
   array in = check_input(std::move(inputs[0]));
-  out.set_data(
-      allocator::malloc_or_wait(in.data_size() * in.itemsize()),
-      in.data_size(),
-      in.strides(),
-      in.flags());
+  if (in.is_donatable()) {
+    out.copy_shared_buffer(in);
+  } else {
+    out.set_data(
+        allocator::malloc_or_wait(in.data_size() * in.itemsize()),
+        in.data_size(),
+        in.strides(),
+        in.flags());
+  }
 
   switch (in.dtype()) {
     case bool_:
@@ -87,13 +101,21 @@ void Softmax::eval(const std::vector<array>& inputs, array& out) {
           "Softmax is defined only for floating point types");
       break;
     case float32:
-      softmax<float>(in, out);
+      softmax<float, float>(in, out);
       break;
     case float16:
-      softmax<float16_t>(in, out);
+      if (precise_) {
+        softmax<float16_t, float>(in, out);
+      } else {
+        softmax<float16_t, float16_t>(in, out);
+      }
       break;
     case bfloat16:
-      softmax<bfloat16_t>(in, out);
+      if (precise_) {
+        softmax<bfloat16_t, float>(in, out);
+      } else {
+        softmax<bfloat16_t, bfloat16_t>(in, out);
+      }
       break;
     case complex64:
       throw std::invalid_argument(

mlx/backend/common/svd.cpp

@@ -0,0 +1,156 @@
+// Copyright © 2024 Apple Inc.
+
+#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 {
+
+void svd_impl(const array& a, array& u, array& s, array& vt) {
+  // Lapack uses the column-major convention. To avoid having to transpose
+  // the input and then transpose the outputs, we swap the indices/sizes of the
+  // matrices and take advantage of the following identity (see
+  // https://math.stackexchange.com/a/30077)
+  //    A = UΣVᵀ
+  //    Aᵀ = VΣUᵀ
+  // As a result some of the indices/sizes are swapped as noted above.
+
+  // Rows and cols of the original matrix in row-major order.
+  const int M = a.shape(-2);
+  const int N = a.shape(-1);
+  const int K = std::min(M, N);
+
+  // A of shape M x N. The leading dimension is N since lapack receives Aᵀ.
+  const int lda = N;
+  // U of shape M x M. (N x N in lapack).
+  const int ldu = N;
+  // Vᵀ of shape N x N. (M x M in lapack).
+  const int ldvt = M;
+
+  size_t num_matrices = a.size() / (M * N);
+
+  // lapack clobbers the input, so we have to make a copy.
+  array in(a.shape(), float32, nullptr, {});
+  copy(a, in, a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+
+  // Allocate outputs.
+  u.set_data(allocator::malloc_or_wait(u.nbytes()));
+  s.set_data(allocator::malloc_or_wait(s.nbytes()));
+  vt.set_data(allocator::malloc_or_wait(vt.nbytes()));
+
+  static constexpr auto job_u = "V";
+  static constexpr auto job_vt = "V";
+  static constexpr auto range = "A";
+
+  // Will contain the number of singular values after the call has returned.
+  int ns = 0;
+  float workspace_dimension = 0;
+
+  // Will contain the indices of eigenvectors that failed to converge (not used
+  // here but required by lapack).
+  auto iwork = array::Data{allocator::malloc_or_wait(sizeof(int) * 12 * K)};
+
+  static const int lwork_query = -1;
+
+  static const int ignored_int = 0;
+  static const float ignored_float = 0;
+
+  int info;
+
+  // Compute workspace size.
+  MLX_LAPACK_FUNC(sgesvdx)
+  (
+      /* jobu = */ job_u,
+      /* jobvt = */ job_vt,
+      /* range = */ range,
+      // M and N are swapped since lapack expects column-major.
+      /* m = */ &N,
+      /* n = */ &M,
+      /* a = */ nullptr,
+      /* lda = */ &lda,
+      /* vl = */ &ignored_float,
+      /* vu = */ &ignored_float,
+      /* il = */ &ignored_int,
+      /* iu = */ &ignored_int,
+      /* ns = */ &ns,
+      /* s = */ nullptr,
+      /* u = */ nullptr,
+      /* ldu = */ &ldu,
+      /* vt = */ nullptr,
+      /* ldvt = */ &ldvt,
+      /* work = */ &workspace_dimension,
+      /* lwork = */ &lwork_query,
+      /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
+      /* info = */ &info);
+
+  if (info != 0) {
+    std::stringstream ss;
+    ss << "svd_impl: sgesvdx_ workspace calculation failed with code " << info;
+    throw std::runtime_error(ss.str());
+  }
+
+  const int lwork = workspace_dimension;
+  auto scratch = array::Data{allocator::malloc_or_wait(sizeof(float) * lwork)};
+
+  // Loop over matrices.
+  for (int i = 0; i < num_matrices; i++) {
+    MLX_LAPACK_FUNC(sgesvdx)
+    (
+        /* jobu = */ job_u,
+        /* jobvt = */ job_vt,
+        /* range = */ range,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ in.data<float>() + M * N * i,
+        /* lda = */ &lda,
+        /* vl = */ &ignored_float,
+        /* vu = */ &ignored_float,
+        /* il = */ &ignored_int,
+        /* iu = */ &ignored_int,
+        /* ns = */ &ns,
+        /* s = */ s.data<float>() + K * i,
+        // According to the identity above, lapack will write Vᵀᵀ as U.
+        /* u = */ vt.data<float>() + N * N * i,
+        /* ldu = */ &ldu,
+        // According to the identity above, lapack will write Uᵀ as Vᵀ.
+        /* vt = */ u.data<float>() + M * M * i,
+        /* ldvt = */ &ldvt,
+        /* work = */ static_cast<float*>(scratch.buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "svd_impl: sgesvdx_ failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    if (ns != K) {
+      std::stringstream ss;
+      ss << "svd_impl: expected " << K << " singular values, but " << ns
+         << " were computed.";
+      throw std::runtime_error(ss.str());
+    }
+  }
+}
+
+void SVD::eval(const std::vector<array>& inputs, std::vector<array>& outputs) {
+  if (!(inputs[0].dtype() == float32)) {
+    throw std::runtime_error("[SVD::eval] only supports float32.");
+  }
+  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

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #pragma once
 
@@ -8,11 +8,12 @@
 
 namespace mlx::core {
 
-inline size_t elem_to_loc(
+template <typename stride_t>
+inline stride_t elem_to_loc(
     int elem,
     const std::vector<int>& shape,
-    const std::vector<size_t>& strides) {
-  size_t loc = 0;
+    const std::vector<stride_t>& strides) {
+  stride_t loc = 0;
   for (int i = shape.size() - 1; i >= 0; --i) {
     auto q_and_r = ldiv(elem, shape[i]);
     loc += q_and_r.rem * strides[i];
@@ -28,4 +29,93 @@ inline size_t elem_to_loc(int elem, const array& a) {
   return elem_to_loc(elem, a.shape(), a.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}}}.
+//
+// When multiple arrays are passed they should all have the same shape. The
+// collapsed axes are also the same so one shape is returned.
+template <typename stride_t>
+inline std::tuple<std::vector<int>, std::vector<std::vector<stride_t>>>
+collapse_contiguous_dims(
+    const std::vector<int>& shape,
+    const std::vector<std::vector<stride_t>> strides) {
+  // Make a vector that has axes separated with -1. Collapse all axes between
+  // -1.
+  std::vector<int> to_collapse;
+  if (shape.size() > 0) {
+    to_collapse.push_back(0);
+    for (int i = 1; i < shape.size(); i++) {
+      bool contiguous = true;
+      for (const std::vector<stride_t>& st : strides) {
+        if (st[i] * shape[i] != st[i - 1]) {
+          contiguous = false;
+        }
+        if (!contiguous) {
+          break;
+        }
+      }
+      if (!contiguous) {
+        to_collapse.push_back(-1);
+      }
+      to_collapse.push_back(i);
+    }
+    to_collapse.push_back(-1);
+  }
+
+  std::vector<int> out_shape;
+  std::vector<std::vector<stride_t>> out_strides(strides.size());
+  for (int i = 0; i < to_collapse.size(); i++) {
+    int current_shape = shape[to_collapse[i]];
+    while (to_collapse[++i] != -1) {
+      current_shape *= shape[to_collapse[i]];
+    }
+    out_shape.push_back(current_shape);
+    for (int j = 0; j < strides.size(); j++) {
+      const std::vector<stride_t>& st = strides[j];
+      out_strides[j].push_back(st[to_collapse[i - 1]]);
+    }
+  }
+
+  return std::make_tuple(out_shape, out_strides);
+}
+
+inline std::tuple<std::vector<int>, std::vector<std::vector<size_t>>>
+collapse_contiguous_dims(const std::vector<array>& xs) {
+  std::vector<std::vector<size_t>> strides;
+  for (auto& x : xs) {
+    strides.emplace_back(x.strides());
+  }
+  return collapse_contiguous_dims(xs[0].shape(), strides);
+}
+
+template <typename... Arrays, typename = enable_for_arrays_t<Arrays...>>
+inline auto collapse_contiguous_dims(Arrays&&... xs) {
+  return collapse_contiguous_dims(
+      std::vector<array>{std::forward<Arrays>(xs)...});
+}
+
+template <typename stride_t>
+inline auto check_contiguity(
+    const std::vector<int>& shape,
+    const std::vector<stride_t>& strides) {
+  size_t data_size = 1;
+  size_t f_stride = 1;
+  size_t b_stride = 1;
+  bool is_row_contiguous = true;
+  bool is_col_contiguous = true;
+
+  for (int i = 0, ri = shape.size() - 1; ri >= 0; i++, ri--) {
+    is_row_contiguous &= strides[i] == f_stride || shape[i] == 1;
+    is_col_contiguous &= strides[ri] == b_stride || shape[ri] == 1;
+    f_stride *= shape[i];
+    b_stride *= shape[ri];
+    if (strides[i] > 0) {
+      data_size *= shape[i];
+    }
+  }
+
+  return std::make_tuple(data_size, is_row_contiguous, is_col_contiguous);
+}
+
 } // namespace mlx::core

mlx/backend/metal/CMakeLists.txt

@@ -33,6 +33,7 @@ target_sources(
   ${CMAKE_CURRENT_SOURCE_DIR}/metal.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/normalization.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/rope.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp

mlx/backend/metal/allocator.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2023-2024 Apple Inc.
 #include "mlx/backend/metal/allocator.h"
 #include "mlx/backend/metal/metal.h"
+#include "mlx/backend/metal/metal_impl.h"
 
 #include <mach/vm_page_size.h>
 #include <unistd.h>

mlx/backend/metal/compiled.cpp

@@ -3,6 +3,7 @@
 #include <sstream>
 
 #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/utils.h"
@@ -228,14 +229,7 @@ void Compiled::eval_gpu(
 
   // Figure out which kernel we are using
   auto& output_shape = outputs[0].shape();
-  bool contiguous = true;
-  for (auto& x : inputs) {
-    if ((!x.flags().row_contiguous || x.shape() != output_shape) &&
-        !is_scalar(x)) {
-      contiguous = false;
-      break;
-    }
-  }
+  bool contiguous = compiled_check_contiguity(inputs, output_shape);
 
   // Collapse contiguous dims to route to a faster kernel if possible. Also
   // handle all broadcasting.
@@ -295,7 +289,7 @@ void Compiled::eval_gpu(
     }
   }
   auto kernel = d.get_kernel(kernel_name, lib);
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
 
   // Put the inputs in
@@ -306,7 +300,7 @@ void Compiled::eval_gpu(
       continue;
     }
     auto& x = inputs[i];
-    set_array_buffer(compute_encoder, x, cnt++);
+    compute_encoder.set_input_array(x, cnt++);
     if (!contiguous && !is_scalar(x)) {
       compute_encoder->setBytes(
           strides[stride_idx].data(),
@@ -316,30 +310,12 @@ void Compiled::eval_gpu(
     }
   }
 
-  // Allocate space for the outputs possibly with input donation
-  {
-    int o = 0;
-    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
-      auto& in = inputs[i];
-      // Conditions for donation
-      // - Row contiguous
-      // - Donatable
-      // - Correct size
-      // - Not a constant
-      if (in.flags().row_contiguous && in.nbytes() == outputs[o].nbytes() &&
-          in.is_donatable() &&
-          constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        outputs[o++].move_shared_buffer(in);
-      }
-    }
-    for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(allocator::malloc_or_wait(outputs[o].nbytes()));
-    }
-  }
+  compiled_allocate_outputs(
+      inputs, outputs, inputs_, constant_ids_, contiguous, true);
 
   // Put the outputs in
   for (auto& x : outputs) {
-    set_array_buffer(compute_encoder, x, cnt++);
+    compute_encoder.set_output_array(x, cnt++);
   }
 
   // Put the output shape and strides in

mlx/backend/metal/conv.cpp

@@ -28,10 +28,12 @@ void explicit_gemm_conv_ND_gpu(
     const array& wt,
     array out,
     const MLXConvParams<N>& conv_params) {
+  // Get gemm shapes
+  int implicit_M = out.size() / conv_params.O;
+  int implicit_K = wt.size() / conv_params.O;
+  int implicit_N = conv_params.O;
   // Prepare unfolding array
-  std::vector<int> unfolded_shape = {
-      static_cast<int>(out.size() / conv_params.O),
-      static_cast<int>(wt.size() / conv_params.O)};
+  std::vector<int> unfolded_shape{implicit_M, implicit_K};
   array in_unfolded(unfolded_shape, in.dtype(), nullptr, {});
 
   in_unfolded.set_data(allocator::malloc_or_wait(in_unfolded.nbytes()));
@@ -39,12 +41,12 @@ void explicit_gemm_conv_ND_gpu(
   // Prepare unfolding kernel
   std::ostringstream kname;
   kname << "naive_unfold_nd_" << type_to_name(in_unfolded) << "_" << N;
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, in_unfolded, 1);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_output_array(in_unfolded, 1);
 
   compute_encoder->setBytes(&conv_params, sizeof(conv_params), 2);
 
@@ -59,20 +61,29 @@ void explicit_gemm_conv_ND_gpu(
 
   compute_encoder->dispatchThreads(grid_dims, group_dims);
 
+  // Reshape weight
+  std::vector<int> wt_reshape{implicit_K, implicit_N};
+  std::vector<size_t> wt_restride{1, static_cast<size_t>(implicit_K)};
+  array wt_reshaped(wt_reshape, wt.dtype(), nullptr, {});
+  auto wt_flags = wt.flags();
+  wt_flags.row_contiguous = false;
+  wt_flags.col_contiguous = true;
+  wt_reshaped.copy_shared_buffer(wt, wt_restride, wt_flags, wt.data_size());
+
   // Perform gemm
-  std::vector<array> copies;
+  std::vector<array> copies = {in_unfolded, wt_reshaped};
   return steel_matmul(
       s,
       d,
       /*a = */ in_unfolded,
-      /*b = */ wt,
+      /*b = */ wt_reshaped,
       /*c = */ out,
-      /*M = */ unfolded_shape[0],
-      /*N = */ conv_params.O,
-      /*K = */ unfolded_shape[1],
+      /*M = */ implicit_M,
+      /*N = */ implicit_N,
+      /*K = */ implicit_K,
       /*batch_size_out = */ 1,
-      /*a_cols = */ unfolded_shape[1],
-      /*b_cols = */ unfolded_shape[1],
+      /*a_cols = */ implicit_K,
+      /*b_cols = */ implicit_K,
       /*a_transposed = */ false,
       /*b_transposed = */ true,
       /*copies = */ copies);
@@ -129,7 +140,7 @@ void slow_conv_2D_gpu(
         << "_tm" << tm << "_tn" << tn;
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
@@ -142,9 +153,9 @@ void slow_conv_2D_gpu(
   MTL::Size group_dims = MTL::Size(bm, bn, 1);
   MTL::Size grid_dims = MTL::Size(grid_dim_x, grid_dim_y, grid_dim_z);
 
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, wt, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
 
   compute_encoder->setBytes(&conv_params, sizeof(MLXConvParams<2>), 3);
   compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
@@ -230,7 +241,7 @@ void implicit_gemm_conv_2D_gpu(
         << "_filter_" << (small_filter ? 's' : 'l');
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
@@ -243,9 +254,9 @@ void implicit_gemm_conv_2D_gpu(
   MTL::Size grid_dims = MTL::Size(grid_dim_x, grid_dim_y, 1);
 
   // Encode arrays
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, wt, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode params
   compute_encoder->setBytes(&conv_params, sizeof(MLXConvParams<2>), 3);
@@ -383,7 +394,7 @@ void implicit_gemm_conv_2D_general_gpu(
         << "_bn" << bn << "_bk" << bk << "_wm" << wm << "_wn" << wn;
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
@@ -397,9 +408,9 @@ void implicit_gemm_conv_2D_general_gpu(
   MTL::Size grid_dims = MTL::Size(grid_dim_x, grid_dim_y, grid_dim_z);
 
   // Encode arrays
-  set_array_buffer(compute_encoder, in, 0);
-  set_array_buffer(compute_encoder, wt, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_input_array(wt, 1);
+  compute_encoder.set_output_array(out, 2);
 
   // Encode params
   compute_encoder->setBytes(&conv_params, sizeof(MLXConvParams<2>), 3);
@@ -500,12 +511,12 @@ void winograd_conv_2D_gpu(
     std::ostringstream kname;
     kname << "winograd_conv_2d_weight_transform_" << type_to_name(out) << "_bc"
           << bc;
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
-    set_array_buffer(compute_encoder, wt, 0);
-    set_array_buffer(compute_encoder, filt_wg, 1);
+    compute_encoder.set_input_array(wt, 0);
+    compute_encoder.set_output_array(filt_wg, 1);
 
     compute_encoder->setBytes(&C_c, sizeof(int), 2);
     compute_encoder->setBytes(&O_c, sizeof(int), 3);
@@ -528,12 +539,12 @@ void winograd_conv_2D_gpu(
     std::ostringstream kname;
     kname << "winograd_conv_2d_input_transform_" << type_to_name(out) << "_bc"
           << bc;
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
-    set_array_buffer(compute_encoder, in_padded, 0);
-    set_array_buffer(compute_encoder, inp_wg, 1);
+    compute_encoder.set_input_array(in_padded, 0);
+    compute_encoder.set_output_array(inp_wg, 1);
 
     compute_encoder->setBytes(
         &conv_params_updated, sizeof(MLXConvParams<2>), 2);
@@ -576,12 +587,12 @@ void winograd_conv_2D_gpu(
     std::ostringstream kname;
     kname << "winograd_conv_2d_output_transform_" << type_to_name(out) << "_bo"
           << bc;
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
-    set_array_buffer(compute_encoder, out_wg, 0);
-    set_array_buffer(compute_encoder, out, 1);
+    compute_encoder.set_input_array(out_wg, 0);
+    compute_encoder.set_output_array(out, 1);
 
     compute_encoder->setBytes(
         &conv_params_updated, sizeof(MLXConvParams<2>), 2);

mlx/backend/metal/copy.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <sstream>
 
@@ -12,8 +12,15 @@ namespace mlx::core {
 
 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
+    // have the same size, then the input buffer can hold the output.
     if (in.is_donatable() && in.itemsize() == out.itemsize()) {
       out.move_shared_buffer(in);
+      // If the output has the same type as the input then there is nothing to
+      // copy, just use the buffer.
+      if (in.dtype() == out.dtype()) {
+        return;
+      }
     } else {
       out.set_data(
           allocator::malloc_or_wait(in.data_size() * out.itemsize()),
@@ -37,15 +44,22 @@ void copy_gpu(const array& in, array& out, CopyType ctype) {
   copy_gpu(in, out, ctype, out.primitive().stream());
 }
 
+template <typename stride_t>
 void copy_gpu_inplace(
     const array& in,
     array& out,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& strides_in_pre,
+    const std::vector<stride_t>& strides_out_pre,
+    int64_t inp_offset,
+    int64_t out_offset,
     CopyType ctype,
     const Stream& s) {
   // Try to collapse contiguous dims
-  auto [shape, strides] = collapse_contiguous_dims(in, out);
-  auto& strides_in = strides[0];
-  auto& strides_out = strides[1];
+  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];
 
   auto& d = metal::device(s.device);
   std::ostringstream kname;
@@ -69,42 +83,47 @@ void copy_gpu_inplace(
     kname << "_" << shape.size();
   }
   auto kernel = d.get_kernel(kname.str());
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
   bool donate_in = in.data_shared_ptr() == nullptr;
-  set_array_buffer(compute_encoder, donate_in ? out : in, 0);
-  set_array_buffer(compute_encoder, out, 1);
+
+  inp_offset *= size_of(in.dtype());
+  out_offset *= size_of(out.dtype());
+
+  compute_encoder.set_input_array(donate_in ? out : in, 0, inp_offset);
+  compute_encoder.set_output_array(out, 1, out_offset);
 
   if (ctype == CopyType::General || ctype == CopyType::GeneralGeneral) {
-    size_t ndim = shape.size();
+    int ndim = shape.size();
+    std::vector<int64_t> strides_in{strides_in_.begin(), strides_in_.end()};
+    std::vector<int64_t> strides_out{strides_out_.begin(), strides_out_.end()};
+
     if (ndim > 3) {
-      compute_encoder->setBytes(shape.data(), ndim * sizeof(int), 2);
-      compute_encoder->setBytes(strides_in.data(), ndim * sizeof(size_t), 3);
-      if (ctype == CopyType::GeneralGeneral) {
-        compute_encoder->setBytes(strides_out.data(), ndim * sizeof(size_t), 4);
-      }
-    } else {
-      // The shape is implicit in the grid for <= 3D
-      compute_encoder->setBytes(strides_in.data(), ndim * sizeof(size_t), 2);
-      if (ctype == CopyType::GeneralGeneral) {
-        compute_encoder->setBytes(strides_out.data(), ndim * sizeof(size_t), 3);
-      }
+      set_vector_bytes(compute_encoder, shape, ndim, 2);
+    }
+    set_vector_bytes(compute_encoder, strides_in, ndim, 3);
+    if (ctype == CopyType::GeneralGeneral) {
+      set_vector_bytes(compute_encoder, strides_out, ndim, 4);
     }
 
     if (ndim > MAX_BINARY_SPECIALIZED_DIMS) {
-      compute_encoder->setBytes(
-          &ndim, sizeof(int), (ctype == CopyType::GeneralGeneral) ? 5 : 4);
+      compute_encoder->setBytes(&ndim, sizeof(int), 5);
     }
 
     int dim0 = ndim > 0 ? shape[ndim - 1] : 1;
     int dim1 = ndim > 1 ? shape[ndim - 2] : 1;
-    int rest = in.size() / (dim0 * dim1);
+
+    size_t data_size = 1;
+    for (auto& s : shape)
+      data_size *= s;
+    int rest = data_size / (dim0 * dim1);
 
     // NB assuming thread_group_size is a power of 2 larger than 32 x 32
     NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
     if (thread_group_size != 1024) {
       throw std::runtime_error("[Metal::copy] 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);
@@ -120,4 +139,25 @@ void copy_gpu_inplace(
   }
 }
 
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    CopyType ctype,
+    const Stream& s) {
+  return copy_gpu_inplace(
+      in, out, in.shape(), in.strides(), out.strides(), 0, 0, ctype, s);
+}
+
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    const std::vector<int64_t>& istride,
+    int64_t ioffset,
+    CopyType ctype,
+    const Stream& s) {
+  std::vector<int64_t> ostrides{out.strides().begin(), out.strides().end()};
+  return copy_gpu_inplace(
+      in, out, in.shape(), istride, ostrides, ioffset, 0, ctype, s);
+}
+
 } // namespace mlx::core

mlx/backend/metal/copy.h

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #pragma once
 
@@ -7,12 +7,34 @@
 
 namespace mlx::core {
 
+// Generic copy inplace
+template <typename stride_t>
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    const std::vector<int>& data_shape,
+    const std::vector<stride_t>& i_strides,
+    const std::vector<stride_t>& o_strides,
+    int64_t i_offset,
+    int64_t o_offset,
+    CopyType ctype,
+    const Stream& s);
+
 void copy_gpu(const array& src, array& out, CopyType ctype, const Stream& s);
 void copy_gpu(const array& src, array& out, CopyType ctype);
+
 void copy_gpu_inplace(
     const array& src,
     array& out,
     CopyType ctype,
     const Stream& s);
 
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    const std::vector<int64_t>& istride,
+    int64_t ioffset,
+    CopyType ctype,
+    const Stream& s);
+
 } // namespace mlx::core

mlx/backend/metal/device.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023-24 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <dlfcn.h>
 #include <cstdlib>
@@ -11,7 +11,9 @@
 
 #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;
 
@@ -20,9 +22,9 @@ namespace mlx::core::metal {
 namespace {
 
 // TODO nicer way to set this or possibly expose as an environment variable
-static constexpr int MAX_BUFFERS_PER_QUEUE = 12;
+constexpr int MAX_BUFFERS_PER_QUEUE = 12;
 
-static constexpr const char* default_mtllib_path = METAL_PATH;
+constexpr const char* default_mtllib_path = METAL_PATH;
 
 auto load_device() {
   auto devices = MTL::CopyAllDevices();
@@ -145,6 +147,7 @@ void Device::new_queue(int index) {
   // We lock this as a critical section for safety
   const std::lock_guard<std::mutex> lock(mtx_);
   auto q = device_->newCommandQueue(MAX_BUFFERS_PER_QUEUE);
+  debug_set_stream_queue_label(q, index);
   if (!q) {
     throw std::runtime_error(
         "[metal::Device] Failed to make new command queue.");
@@ -203,14 +206,15 @@ void Device::end_encoding(int index) {
   }
 }
 
-MTL::ComputeCommandEncoder* Device::get_command_encoder(int index) {
+CommandEncoder& Device::get_command_encoder(int index) {
   auto eit = encoder_map_.find(index);
   if (eit == encoder_map_.end()) {
     auto cb = get_command_buffer(index);
-    auto compute_encoder = cb->computeCommandEncoder();
+    auto compute_encoder =
+        cb->computeCommandEncoder(MTL::DispatchTypeConcurrent);
     // Increment ref count so the buffer is not garbage collected
     compute_encoder->retain();
-    eit = encoder_map_.insert({index, compute_encoder}).first;
+    eit = encoder_map_.emplace(index, CommandEncoder{compute_encoder}).first;
   }
   return eit->second;
 }

mlx/backend/metal/device.h

@@ -1,4 +1,4 @@
-// Copyright © 2023-24 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #pragma once
 
@@ -7,10 +7,12 @@
 #include <mutex>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 
 #include <dlfcn.h>
 #include <filesystem>
 
+#include "mlx/array.h"
 #include "mlx/device.h"
 
 namespace fs = std::filesystem;
@@ -34,6 +36,69 @@ inline std::string get_colocated_mtllib_path(const std::string& lib_name) {
 using MTLFCList =
     std::vector<std::tuple<const void*, MTL::DataType, NS::UInteger>>;
 
+struct CommandEncoder {
+  CommandEncoder(MTL::ComputeCommandEncoder* enc)
+      : enc(enc), concurrent(false){};
+  CommandEncoder& operator=(const CommandEncoder&) = delete;
+
+  struct ConcurrentContext {
+    ConcurrentContext(CommandEncoder& enc) : enc(enc) {
+      enc.concurrent = true;
+    }
+    ~ConcurrentContext() {
+      enc.concurrent = false;
+      enc.outputs.insert(
+          enc.concurrent_outputs.begin(), enc.concurrent_outputs.end());
+      enc.concurrent_outputs.clear();
+    }
+
+   private:
+    CommandEncoder& enc;
+  };
+
+  MTL::ComputeCommandEncoder* operator->() {
+    return enc;
+  }
+
+  void set_input_array(const array& a, int idx, int offset = 0) {
+    auto r_buf =
+        static_cast<MTL::Resource*>(const_cast<void*>(a.buffer().ptr()));
+    if (auto it = outputs.find(r_buf); it != outputs.end()) {
+      // Insert a barrier
+      enc->memoryBarrier(&r_buf, 1);
+
+      // Remove the output
+      outputs.erase(it);
+    }
+    auto a_buf = static_cast<const MTL::Buffer*>(a.buffer().ptr());
+    auto base_offset = a.data<char>() -
+        static_cast<char*>(const_cast<MTL::Buffer*>(a_buf)->contents());
+    base_offset += offset;
+    enc->setBuffer(a_buf, base_offset, idx);
+  }
+
+  void set_output_array(array& a, int idx, int offset = 0) {
+    // Add barriers before adding the output to the output set
+    set_input_array(a, idx, offset);
+    auto buf = static_cast<MTL::Resource*>(a.buffer().ptr());
+    if (concurrent) {
+      concurrent_outputs.insert(buf);
+    } else {
+      outputs.insert(buf);
+    }
+  }
+
+  ConcurrentContext start_concurrent() {
+    return ConcurrentContext(*this);
+  }
+
+ private:
+  MTL::ComputeCommandEncoder* enc;
+  bool concurrent;
+  std::unordered_set<MTL::Resource*> outputs;
+  std::unordered_set<MTL::Resource*> concurrent_outputs;
+};
+
 class Device {
  public:
   Device();
@@ -51,7 +116,7 @@ class Device {
   int get_command_buffer_ops(int index);
   void increment_command_buffer_ops(int index);
   void commit_command_buffer(int index);
-  MTL::ComputeCommandEncoder* get_command_encoder(int index);
+  CommandEncoder& get_command_encoder(int index);
   void end_encoding(int index);
 
   void register_library(
@@ -132,7 +197,7 @@ class Device {
   MTL::Device* device_;
   std::unordered_map<int32_t, MTL::CommandQueue*> queue_map_;
   std::unordered_map<int32_t, std::pair<int, MTL::CommandBuffer*>> buffer_map_;
-  std::unordered_map<int32_t, MTL::ComputeCommandEncoder*> encoder_map_;
+  std::unordered_map<int32_t, 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/indexing.cpp

@@ -16,7 +16,7 @@ namespace mlx::core {
 
 namespace {
 
-static constexpr int METAL_MAX_INDEX_ARRAYS = 10;
+constexpr int METAL_MAX_INDEX_ARRAYS = 10;
 
 } // namespace
 
@@ -49,7 +49,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
     kname << "_" << idx_ndim;
   }
 
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
@@ -81,8 +81,8 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
 
   // Set all the buffers
-  set_array_buffer(compute_encoder, src, 0);
-  set_array_buffer(compute_encoder, out, 1);
+  compute_encoder.set_input_array(src, 0);
+  compute_encoder.set_output_array(out, 1);
 
   // Set source info
   compute_encoder->setBytes(src.shape().data(), ndim * sizeof(int), 2);
@@ -103,7 +103,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   // Set index buffers
   for (int i = 1; i < nidx + 1; ++i) {
-    set_array_buffer(compute_encoder, inputs[i], 20 + i);
+    compute_encoder.set_input_array(inputs[i], 20 + i);
   }
 
   // Launch grid
@@ -183,7 +183,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
   kname << "_" << nidx;
 
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
 
   auto& upd = inputs.back();
@@ -192,8 +192,8 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   compute_encoder->setComputePipelineState(kernel);
 
   // Set all the buffers
-  set_array_buffer(compute_encoder, upd, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_input_array(upd, 1);
+  compute_encoder.set_output_array(out, 2);
 
   // Set update info
   uint upd_ndim = upd.ndim();
@@ -201,19 +201,16 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   for (int i = idx_ndim; i < upd.ndim(); ++i) {
     upd_size *= upd.shape(i);
   }
-
   if (index_nd1_specialization) {
-    bool upd_col_contiguous = upd.flags().col_contiguous;
     compute_encoder->setBytes(
         out.shape().data(), out.shape().size() * sizeof(int), 3);
     compute_encoder->setBytes(
         out.strides().data(), out.strides().size() * sizeof(size_t), 4);
     compute_encoder->setBytes(&upd_size, sizeof(size_t), 5);
-    compute_encoder->setBytes(&upd_col_contiguous, sizeof(bool), 6);
 
     // Set index buffers
     for (int i = 1; i < nidx + 1; ++i) {
-      set_array_buffer(compute_encoder, inputs[i], 20 + i);
+      compute_encoder.set_input_array(inputs[i], 20 + i);
     }
 
     // Launch grid
@@ -283,7 +280,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
 
     // Set index buffers
     for (int i = 1; i < nidx + 1; ++i) {
-      set_array_buffer(compute_encoder, inputs[i], 20 + i);
+      compute_encoder.set_input_array(inputs[i], 20 + i);
     }
 
     // Launch grid

mlx/backend/metal/kernels/CMakeLists.txt

@@ -7,6 +7,7 @@ set(
   ${CMAKE_CURRENT_SOURCE_DIR}/complex.h
   ${CMAKE_CURRENT_SOURCE_DIR}/defines.h
   ${CMAKE_CURRENT_SOURCE_DIR}/erf.h
+  ${CMAKE_CURRENT_SOURCE_DIR}/expm1f.h
   ${CMAKE_CURRENT_SOURCE_DIR}/indexing.h
   ${CMAKE_CURRENT_SOURCE_DIR}/unary.h
   ${CMAKE_CURRENT_SOURCE_DIR}/utils.h
@@ -23,6 +24,8 @@ set(
   "gemv"
   "quantized"
   "random"
+  "rms_norm"
+  "layer_norm"
   "rope"
   "scan"
   "scaled_dot_product_attention"
@@ -35,11 +38,17 @@ set(
 )
 
 function(build_kernel_base TARGET SRCFILE DEPS)
+  set(METAL_FLAGS -Wall -Wextra -fno-fast-math)
+  if(MLX_METAL_DEBUG)
+    set(METAL_FLAGS ${METAL_FLAGS}
+        -gline-tables-only
+        -frecord-sources)
+  endif()
   add_custom_command(
-    COMMAND xcrun -sdk macosx metal -Wall -Wextra
-                  -fno-fast-math
-                  -c ${SRCFILE} 
-                  -I${PROJECT_SOURCE_DIR} 
+    COMMAND xcrun -sdk macosx metal
+                  ${METAL_FLAGS}
+                  -c ${SRCFILE}
+                  -I${PROJECT_SOURCE_DIR}
                   -o ${TARGET}.air
     DEPENDS ${SRCFILE} ${DEPS}
     OUTPUT ${TARGET}.air

mlx/backend/metal/kernels/copy.metal

@@ -1,29 +1,29 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include "mlx/backend/metal/kernels/bf16.h"
 #include "mlx/backend/metal/kernels/utils.h"
 
 template <typename T, typename U>
 [[kernel]] void copy_s(
-    device const T* src,
-    device U* dst,
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
     uint index [[thread_position_in_grid]]) {
   dst[index] = static_cast<U>(src[0]);
 }
 
 template <typename T, typename U>
 [[kernel]] void copy_v(
-    device const T* src,
-    device U* dst,
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
     uint index [[thread_position_in_grid]]) {
   dst[index] = static_cast<U>(src[index]);
 }
 
 template <typename T, typename U>
 [[kernel]] void copy_g_nd1(
-    device const T* src,
-    device U* dst,
-    constant const size_t& src_stride,
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int64_t& src_stride [[buffer(3)]],
     uint index [[thread_position_in_grid]]) {
   auto src_idx = elem_to_loc_1(index, src_stride);
   dst[index] = static_cast<U>(src[src_idx]);
@@ -31,61 +31,61 @@ template <typename T, typename U>
 
 template <typename T, typename U>
 [[kernel]] void copy_g_nd2(
-    device const T* src,
-    device U* dst,
-    constant const size_t src_strides[2],
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
   auto src_idx = elem_to_loc_2(index, src_strides);
-  size_t dst_idx = index.x + (size_t)grid_dim.x * index.y;
+  int64_t dst_idx = index.x + (int64_t)grid_dim.x * index.y;
   dst[dst_idx] = static_cast<U>(src[src_idx]);
 }
 
 template <typename T, typename U>
 [[kernel]] void copy_g_nd3(
-    device const T* src,
-    device U* dst,
-    constant const size_t src_strides[3],
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
     uint3 index [[thread_position_in_grid]],
     uint3 grid_dim [[threads_per_grid]]) {
   auto src_idx = elem_to_loc_3(index, src_strides);
-  size_t dst_idx = index.x + (size_t)grid_dim.x * (index.y + (size_t)grid_dim.y * index.z);
+  int64_t dst_idx = index.x + (int64_t)grid_dim.x * (index.y + (int64_t)grid_dim.y * index.z);
   dst[dst_idx] = static_cast<U>(src[src_idx]);
 }
 
 template <typename T, typename U, int DIM>
 [[kernel]] void copy_g_nd(
-    device const T* src,
-    device U* dst,
-    constant const int src_shape[DIM],
-    constant const size_t src_strides[DIM],
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
     uint3 index [[thread_position_in_grid]],
     uint3 grid_dim [[threads_per_grid]]) {
   auto src_idx = elem_to_loc_nd<DIM>(index, src_shape, src_strides);
-  size_t dst_idx = index.x + (size_t)grid_dim.x * (index.y + (size_t)grid_dim.y * index.z);
+  int64_t dst_idx = index.x + (int64_t)grid_dim.x * (index.y + (int64_t)grid_dim.y * index.z);
   dst[dst_idx] = static_cast<U>(src[src_idx]);
 }
 
 template <typename T, typename U>
 [[kernel]] void copy_g(
-    device const T* src,
-    device U* dst,
-    constant const int* src_shape,
-    constant const size_t* src_strides,
-    constant const int& ndim,
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int& ndim [[buffer(5)]],
     uint3 index [[thread_position_in_grid]],
     uint3 grid_dim [[threads_per_grid]]) {
   auto src_idx = elem_to_loc(index, src_shape, src_strides, ndim);
-  size_t dst_idx = index.x + (size_t)grid_dim.x * (index.y + (size_t)grid_dim.y * index.z);
+  int64_t dst_idx = index.x + (int64_t)grid_dim.x * (index.y + (int64_t)grid_dim.y * index.z);
   dst[dst_idx] = static_cast<U>(src[src_idx]);
 }
 
 template <typename T, typename U>
 [[kernel]] void copy_gg_nd1(
-    device const T* src,
-    device U* dst,
-    constant const size_t& src_stride,
-    constant const size_t& dst_stride,
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int64_t& src_stride [[buffer(3)]],
+    constant const int64_t& dst_stride [[buffer(4)]],
     uint index [[thread_position_in_grid]]) {
   auto src_idx = elem_to_loc_1(index, src_stride);
   auto dst_idx = elem_to_loc_1(index, dst_stride);
@@ -94,10 +94,10 @@ template <typename T, typename U>
 
 template <typename T, typename U>
 [[kernel]] void copy_gg_nd2(
-    device const T* src,
-    device U* dst,
-    constant const size_t src_strides[2],
-    constant const size_t dst_strides[2],
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
     uint2 index [[thread_position_in_grid]]) {
   auto src_idx = elem_to_loc_2(index, src_strides);
   auto dst_idx = elem_to_loc_2(index, dst_strides);
@@ -106,10 +106,10 @@ template <typename T, typename U>
 
 template <typename T, typename U>
 [[kernel]] void copy_gg_nd3(
-    device const T* src,
-    device U* dst,
-    constant const size_t src_strides[3],
-    constant const size_t dst_strides[3],
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
     uint3 index [[thread_position_in_grid]]) {
   auto src_idx = elem_to_loc_3(index, src_strides);
   auto dst_idx = elem_to_loc_3(index, dst_strides);
@@ -118,11 +118,11 @@ template <typename T, typename U>
 
 template <typename T, typename U, int DIM>
 [[kernel]] void copy_gg_nd(
-    device const T* src,
-    device U* dst,
-    constant const int src_shape[DIM],
-    constant const size_t src_strides[DIM],
-    constant const size_t dst_strides[DIM],
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
     uint3 index [[thread_position_in_grid]]) {
   auto src_idx = elem_to_loc_nd<DIM>(index, src_shape, src_strides);
   auto dst_idx = elem_to_loc_nd<DIM>(index, src_shape, dst_strides);
@@ -131,12 +131,12 @@ template <typename T, typename U, int DIM>
 
 template <typename T, typename U>
 [[kernel]] void copy_gg(
-    device const T* src,
-    device U* dst,
-    constant const int* src_shape,
-    constant const size_t* src_strides,
-    constant const size_t* dst_strides,
-    constant const int& ndim,
+    device const T* src [[buffer(0)]],
+    device U* dst [[buffer(1)]],
+    constant const int* src_shape [[buffer(2)]],
+    constant const int64_t* src_strides [[buffer(3)]],
+    constant const int64_t* dst_strides [[buffer(4)]],
+    constant const int& ndim [[buffer(5)]],
     uint3 index [[thread_position_in_grid]]) {
   auto src_idx = elem_to_loc(index, src_shape, src_strides, ndim);
   auto dst_idx = elem_to_loc(index, src_shape, dst_strides, ndim);
@@ -146,70 +146,70 @@ template <typename T, typename U>
 #define instantiate_copy(name, itype, otype, ctype) \
   template [[host_name(name)]] \
   [[kernel]] void copy_##ctype<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
       uint index [[thread_position_in_grid]]);
 
 #define instantiate_copy_g_dim(name, itype, otype, dims) \
   template [[host_name(name "_" #dims)]] \
   [[kernel]] void copy_g_nd<itype, otype, dims>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const int src_shape[dims], \
-      constant const size_t src_strides[dims], \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int* src_shape [[buffer(2)]], \
+      constant const int64_t* src_strides [[buffer(3)]], \
       uint3 index [[thread_position_in_grid]], \
       uint3 grid_dim [[threads_per_grid]]); \
   template [[host_name("g" name "_" #dims)]] \
   [[kernel]] void copy_gg_nd<itype, otype, dims>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const int src_shape[dims], \
-      constant const size_t src_strides[dims], \
-      constant const size_t dst_strides[dims], \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int* src_shape [[buffer(2)]], \
+      constant const int64_t* src_strides [[buffer(3)]], \
+      constant const int64_t* dst_strides [[buffer(4)]], \
       uint3 index [[thread_position_in_grid]]);
 
 
 #define instantiate_copy_g_nd(name, itype, otype) \
   template [[host_name(name "_1")]] \
   [[kernel]] void copy_g_nd1<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const size_t& src_stride, \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int64_t& src_stride [[buffer(3)]], \
       uint index [[thread_position_in_grid]]); \
   template [[host_name(name "_2")]] \
   [[kernel]] void copy_g_nd2<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const size_t src_strides[2], \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int64_t* src_strides [[buffer(3)]], \
       uint2 index [[thread_position_in_grid]], \
       uint2 grid_dim [[threads_per_grid]]); \
   template [[host_name(name "_3")]] \
   [[kernel]] void copy_g_nd3<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const size_t src_strides[3], \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int64_t* src_strides [[buffer(3)]], \
       uint3 index [[thread_position_in_grid]], \
       uint3 grid_dim [[threads_per_grid]]); \
   template [[host_name("g" name "_1")]] \
   [[kernel]] void copy_gg_nd1<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const size_t& src_stride, \
-      constant const size_t& dst_stride, \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int64_t& src_stride [[buffer(3)]], \
+      constant const int64_t& dst_stride [[buffer(4)]], \
       uint index [[thread_position_in_grid]]); \
   template [[host_name("g" name "_2")]] \
   [[kernel]] void copy_gg_nd2<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const size_t src_strides[2], \
-      constant const size_t dst_strides[2], \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int64_t* src_strides [[buffer(3)]], \
+      constant const int64_t* dst_strides [[buffer(4)]], \
       uint2 index [[thread_position_in_grid]]); \
   template [[host_name("g" name "_3")]] \
   [[kernel]] void copy_gg_nd3<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const size_t src_strides[3], \
-      constant const size_t dst_strides[3], \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int64_t* src_strides [[buffer(3)]], \
+      constant const int64_t* dst_strides [[buffer(4)]], \
       uint3 index [[thread_position_in_grid]]); \
   instantiate_copy_g_dim(name, itype, otype, 4) \
   instantiate_copy_g_dim(name, itype, otype, 5)
@@ -218,21 +218,21 @@ template <typename T, typename U>
 #define instantiate_copy_g(name, itype, otype) \
   template [[host_name(name)]] \
   [[kernel]] void copy_g<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const int* src_shape, \
-      constant const size_t* src_strides, \
-      constant const int& ndim, \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int* src_shape [[buffer(2)]], \
+      constant const int64_t* src_strides  [[buffer(3)]], \
+      constant const int& ndim [[buffer(5)]], \
       uint3 index [[thread_position_in_grid]], \
       uint3 grid_dim [[threads_per_grid]]); \
   template [[host_name("g" name)]] \
   [[kernel]] void copy_gg<itype, otype>( \
-      device const itype* src, \
-      device otype* dst, \
-      constant const int* src_shape, \
-      constant const size_t* src_strides, \
-      constant const size_t* dst_strides, \
-      constant const int& ndim, \
+      device const itype* src [[buffer(0)]], \
+      device otype* dst [[buffer(1)]], \
+      constant const int* src_shape [[buffer(2)]], \
+      constant const int64_t* src_strides [[buffer(3)]], \
+      constant const int64_t* dst_strides [[buffer(4)]], \
+      constant const int& ndim [[buffer(5)]], \
       uint3 index [[thread_position_in_grid]]);
 
 #define instantiate_copy_all(tname, itype, otype) \

mlx/backend/metal/kernels/defines.h

@@ -14,3 +14,5 @@ static MTL_CONST constexpr int MAX_REDUCE_SPECIALIZED_DIMS = 4;
 static MTL_CONST constexpr int REDUCE_N_READS = 16;
 static MTL_CONST constexpr int SOFTMAX_N_READS = 4;
 static MTL_CONST constexpr int SOFTMAX_LOOPED_LIMIT = 4096;
+static MTL_CONST constexpr int RMS_N_READS = 4;
+static MTL_CONST constexpr int RMS_LOOPED_LIMIT = 4096;

mlx/backend/metal/kernels/expm1f.h

@@ -0,0 +1,89 @@
+// Copyright © 2023 Apple Inc.
+
+#pragma once
+
+#include <metal_math>
+
+// Original license copied below:
+//  Copyright (c) 2015-2023 Norbert Juffa
+//  All rights reserved.
+//
+//  Redistribution and use in source and binary forms, with or without
+//  modification, are permitted provided that the following conditions
+//  are met:
+//
+//  1. Redistributions of source code must retain the above copyright
+//     notice, this list of conditions and the following disclaimer.
+//
+//  2. Redistributions in binary form must reproduce the above copyright
+//     notice, this list of conditions and the following disclaimer in the
+//     documentation and/or other materials provided with the distribution.
+//
+//  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+//  "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+//  LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+//  A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+//  HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+//  SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+//  LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+//  DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+//  THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+//  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+//  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+/* Compute exponential base e minus 1. Maximum ulp error = 0.997458
+
+   i = rint(a/log(2)), f = a-i*log(2). Then expm1(a) = 2**i * (expm1(f)+1) - 1.
+   Compute r = expm1(f). Then expm1(a)= 2 * (0.5 * 2**i * r + 0.5 * 2**i - 0.5).
+   With t = 0.5*2**i, expm1(a) = 2*(r * t + t-0.5). However, for best accuracy,
+   when i == 1, expm1(a)= 2*(r + 0.5), and when i == 0, expm1(a) = r.
+
+   NOTE: Scale factor b is only applied if i < 0 or i > 1 (should be power of 2)
+*/
+float expm1f_scaled_unchecked(float a, float b) {
+  float f, j, r, s, t, u, v, x, y;
+  int i;
+
+  // exp(a) = 2**i * exp(f); i = rintf (a / log(2))
+  j = fma(1.442695f, a, 12582912.f); // 0x1.715476p0, 0x1.8p23
+  j = j - 12582912.0f; // 0x1.8p23
+  i = (int)j;
+  f = fma(j, -6.93145752e-1f, a);
+
+  // approximate r = exp(f)-1 on interval [-log(2)/2, +log(2)/2]
+  s = f * f;
+  if (a == 0.0f)
+    s = a; // ensure -0 is passed through
+  // err = 0.997458  ulp1 = 11081805
+  r = 1.97350979e-4f; // 0x1.9de000p-13
+  r = fma(r, f, 1.39309070e-3f); // 0x1.6d30bcp-10
+  r = fma(r, f, 8.33343994e-3f); // 0x1.1111f6p-7
+  r = fma(r, f, 4.16668020e-2f); // 0x1.55559ep-5
+  r = fma(r, f, 1.66666716e-1f); // 0x1.55555cp-3
+  r = fma(r, f, 4.99999970e-1f); // 0x1.fffffep-2
+  u = (j == 1) ? (f + 0.5f) : f;
+  v = fma(r, s, u);
+  s = 0.5f * b;
+  t = ldexp(s, i);
+  y = t - s;
+  x = (t - y) - s; // double-float canonicalization of difference
+  r = fma(v, t, x) + y;
+  r = r + r;
+  if (j == 0)
+    r = v;
+  if (j == 1)
+    r = v + v;
+  return r;
+}
+
+/* Compute exponential base e minus 1. max ulp err = 0.99746 */
+float expm1f(float a) {
+  float r;
+
+  r = expm1f_scaled_unchecked(a, 1.0f);
+  /* handle severe overflow and underflow */
+  if (abs(a - 1.0f) > 88.0f) {
+    r = fma(r, r, -1.0f);
+  }
+  return r;
+}

mlx/backend/metal/kernels/gemv.metal

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #include <metal_stdlib>
 #include <metal_simdgroup>
@@ -22,7 +22,8 @@ template <
   const int BM, /* Threadgroup rows (in threads) */
   const int BN, /* Threadgroup cols (in threads) */
   const int TM, /* Thread rows (in elements) */
-  const int TN > /* Thread cols (in elements) */ 
+  const int TN , /* Thread cols (in elements) */ 
+  const bool kDoAxpby> /* Do out = alpha * out + beta * bias */
 struct GEMVKernel {
 
   static_assert(BN == SIMD_SIZE, "gemv block must have a width of SIMD_SIZE");
@@ -48,11 +49,16 @@ struct GEMVKernel {
   MLX_MTL_CONST short tgp_mem_size = BN * TN * 2;
 
   static METAL_FUNC void run(
-      const device T* mat,
-      const device T* in_vec,
-      device T* out_vec, 
-      const constant int& in_vec_size [[buffer(3)]],
-      const constant int& out_vec_size [[buffer(4)]],
+      const device T* mat [[buffer(0)]],
+      const device T* in_vec [[buffer(1)]],
+      const device T* bias [[buffer(2)]],
+      device T* out_vec [[buffer(3)]], 
+      const constant int& in_vec_size [[buffer(4)]],
+      const constant int& out_vec_size [[buffer(5)]],
+      const constant int& marix_ld [[buffer(6)]],
+      const constant float& alpha [[buffer(7)]],
+      const constant float& beta [[buffer(8)]],
+      const constant int& bias_stride [[buffer(14)]],
       threadgroup T* tgp_memory [[threadgroup(0)]],
       uint3 tid [[threadgroup_position_in_grid]],
       uint3 lid [[thread_position_in_threadgroup]],
@@ -81,7 +87,7 @@ struct GEMVKernel {
     out_row = out_row + TM <= out_vec_size ? out_row : out_vec_size - TM;
 
     // Advance matrix
-    mat += out_row * in_vec_size;
+    mat += out_row * marix_ld;
 
     // Loop over in_vec in blocks of BN * TN
     for(int bn = simd_lid * TN; bn < in_vec_size; bn += BN * TN) {
@@ -124,14 +130,14 @@ struct GEMVKernel {
         if(bn + TN <= in_vec_size) {
           #pragma clang loop unroll(full)
           for(int tn = 0; tn < TN; tn++) {
-            inter[tn] = mat[tm * in_vec_size + bn + tn];
+            inter[tn] = mat[tm * marix_ld + bn + tn];
           }
 
         } else { // Edgecase
           #pragma clang loop unroll(full)
           for(int tn = 0; tn < TN; tn++) {
             int col_idx = (bn + tn) < in_vec_size ? (bn + tn) : (in_vec_size - 1);
-            inter[tn] = mat[tm * in_vec_size + col_idx];
+            inter[tn] = mat[tm * marix_ld + col_idx];
           }
         }
 
@@ -154,7 +160,13 @@ struct GEMVKernel {
 
       #pragma clang loop unroll(full)
       for(int tm = 0; tm < TM; tm++) {
-        out_vec[out_row + tm] = result[tm];
+        if(kDoAxpby) {
+          out_vec[out_row + tm] = 
+              static_cast<T>(alpha) * result[tm] + 
+              static_cast<T>(beta) * bias[(out_row + tm) * bias_stride];
+        } else {
+          out_vec[out_row + tm] = result[tm];
+        }
       }
 
     }
@@ -172,7 +184,8 @@ template <
   const int BM, /* Threadgroup rows (in threads) */
   const int BN, /* Threadgroup cols (in threads) */
   const int TM, /* Thread rows (in elements) */
-  const int TN > /* Thread cols (in elements) */ 
+  const int TN, /* Thread cols (in elements) */ 
+  const bool kDoAxpby> /* Do out = alpha * out + beta * bias */
 struct GEMVTKernel {
 
   // - The matrix of size (M = in_vec_size, N = out_vec_size) is divided up 
@@ -197,11 +210,16 @@ struct GEMVTKernel {
   MLX_MTL_CONST short tgp_mem_size = BN * BM * TN;
 
   static METAL_FUNC void run(
-      const device T* mat,
-      const device T* in_vec,
-      device T* out_vec, 
-      const constant int& in_vec_size [[buffer(3)]],
-      const constant int& out_vec_size [[buffer(4)]],
+      const device T* mat [[buffer(0)]],
+      const device T* in_vec [[buffer(1)]],
+      const device T* bias [[buffer(2)]],
+      device T* out_vec [[buffer(3)]], 
+      const constant int& in_vec_size [[buffer(4)]],
+      const constant int& out_vec_size [[buffer(5)]],
+      const constant int& marix_ld [[buffer(6)]],
+      const constant float& alpha [[buffer(7)]],
+      const constant float& beta [[buffer(8)]],
+      const constant int& bias_stride [[buffer(14)]],
       threadgroup T* tgp_memory [[threadgroup(0)]],
       uint3 tid [[threadgroup_position_in_grid]],
       uint3 lid [[thread_position_in_threadgroup]],
@@ -245,7 +263,7 @@ struct GEMVTKernel {
           #pragma clang loop unroll(full)
           for(int tm = 0; tm < TM; tm++) {
             for(int tn = 0; tn < TN; tn++) {
-              inter[tn] = mat[(bm + tm) * out_vec_size + out_col + tn];
+              inter[tn] = mat[(bm + tm) * marix_ld + out_col + tn];
             }
             for(int tn = 0; tn < TN; tn++) {
               result[tn] += v_coeff[tm] * inter[tn];
@@ -257,7 +275,7 @@ struct GEMVTKernel {
             v_coeff[tm] = in_vec[bm + tm];
 
             for(int tn = 0; tn < TN; tn++) {
-              inter[tn] = mat[(bm + tm) * out_vec_size + out_col + tn];
+              inter[tn] = mat[(bm + tm) * marix_ld + out_col + tn];
             }
             for(int tn = 0; tn < TN; tn++) {
               result[tn] += v_coeff[tm] * inter[tn];
@@ -292,13 +310,17 @@ struct GEMVTKernel {
 
       #pragma clang loop unroll(full)
       for(int j = 0; j < TN; j++) {
-        out_vec[out_col + j] = result[j];
+
+        if(kDoAxpby) {
+          out_vec[out_col + j] = 
+              static_cast<T>(alpha) * result[j] + 
+              static_cast<T>(beta) * bias[(out_col + j) * bias_stride];
+        } else {
+          out_vec[out_col + j] = result[j];
+        }
       }
     }
-
   }
-
-
 };
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -310,78 +332,64 @@ template <
     const int BM, /* Threadgroup rows (in threads) */
     const int BN, /* Threadgroup cols (in threads) */
     const int TM, /* Thread rows (in elements) */
-    const int TN> /* Thread cols (in elements) */
+    const int TN, /* Thread cols (in elements) */
+    const bool kDoNCBatch, /* Batch ndim > 1 */
+    const bool kDoAxpby> /* Do out = alpha * out + beta * bias */
 [[kernel, max_total_threads_per_threadgroup(BM * BN)]] void gemv(
     const device T* mat [[buffer(0)]],
     const device T* in_vec [[buffer(1)]],
-    device T* out_vec [[buffer(2)]], 
-    const constant int& in_vec_size [[buffer(3)]],
-    const constant int& out_vec_size [[buffer(4)]],
-    const constant int& vector_batch_stride [[buffer(5)]],
-    const constant int& matrix_batch_stride [[buffer(6)]],
+    const device T* bias [[buffer(2)]],
+    device T* out_vec [[buffer(3)]], 
+    const constant int& in_vec_size [[buffer(4)]],
+    const constant int& out_vec_size [[buffer(5)]],
+    const constant int& marix_ld [[buffer(6)]],
+    const constant float& alpha [[buffer(7)]],
+    const constant float& beta [[buffer(8)]],
+    const constant int& batch_ndim [[buffer(9)]],
+    const constant int* batch_shape [[buffer(10)]],
+    const constant size_t* vector_batch_stride [[buffer(11)]],
+    const constant size_t* matrix_batch_stride [[buffer(12)]],
+    const constant size_t* bias_batch_stride [[buffer(13)]],
+    const constant int& bias_stride [[buffer(14)]],
     uint3 tid [[threadgroup_position_in_grid]],
     uint3 lid [[thread_position_in_threadgroup]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
 
-  using gemv_kernel = GEMVKernel<T, BM, BN, TM, TN>;
+  using gemv_kernel = GEMVKernel<T, BM, BN, TM, TN, kDoAxpby>;
   threadgroup T tgp_memory[gemv_kernel::tgp_mem_size];
 
   // Update batch offsets
-  in_vec += tid.z * vector_batch_stride;
-  mat += tid.z * matrix_batch_stride;
-  out_vec += tid.z * out_vec_size;
-
-  gemv_kernel::run( 
-    mat, 
-    in_vec, 
-    out_vec,
-    in_vec_size,
-    out_vec_size,
-    tgp_memory,
-    tid,
-    lid,
-    simd_gid,
-    simd_lid
-  );
-
-}
-
-template <
-    typename T, 
-    const int BM, /* Threadgroup rows (in threads) */
-    const int BN, /* Threadgroup cols (in threads) */
-    const int TM, /* Thread rows (in elements) */
-    const int TN> /* Thread cols (in elements) */
-[[kernel, max_total_threads_per_threadgroup(BM * BN)]] void gemv_nc(
-    const device T* mat [[buffer(0)]],
-    const device T* in_vec [[buffer(1)]],
-    device T* out_vec [[buffer(2)]], 
-    const constant int& in_vec_size [[buffer(3)]],
-    const constant int& out_vec_size [[buffer(4)]],
-    const constant int& nc_dim [[buffer(5)]],
-    const device int* nc_shape [[buffer(6)]],
-    const device size_t* nc_strides_vec [[buffer(7)]],
-    const device size_t* nc_strides_mat [[buffer(8)]],
-    uint3 tid [[threadgroup_position_in_grid]],
-    uint3 lid [[thread_position_in_threadgroup]],
-    uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]]) {
-
-  using gemv_kernel = GEMVKernel<T, BM, BN, TM, TN>;
-  threadgroup T tgp_memory[gemv_kernel::tgp_mem_size];
-
-  // Update batch offsets
-  in_vec += elem_to_loc(tid.z, nc_shape, nc_strides_vec, nc_dim);
-  mat += elem_to_loc(tid.z, nc_shape, nc_strides_mat, nc_dim);
+  if(kDoNCBatch) {
+    in_vec += elem_to_loc(tid.z, batch_shape, vector_batch_stride, batch_ndim);
+    mat += elem_to_loc(tid.z, batch_shape, matrix_batch_stride, batch_ndim);
+
+    if(kDoAxpby) {
+      bias += elem_to_loc(tid.z, batch_shape, bias_batch_stride, batch_ndim);
+    }
+
+  } else {
+    in_vec += tid.z * vector_batch_stride[0];
+    mat += tid.z * matrix_batch_stride[0];
+
+    if(kDoAxpby) {
+      bias += tid.z * bias_batch_stride[0];
+    }
+  }
+
   out_vec += tid.z * out_vec_size;
 
   gemv_kernel::run( 
     mat, 
     in_vec, 
+    bias,
     out_vec,
     in_vec_size,
     out_vec_size,
+    marix_ld,
+    alpha,
+    beta,
+    bias_stride,
     tgp_memory,
     tid,
     lid,
@@ -392,41 +400,34 @@ template <
 }
 
 
-#define instantiate_gemv_c(name, itype, bm, bn, tm, tn) \
-  template [[host_name("gemv_" #name "_bm" #bm "_bn" #bn "_tm" #tm "_tn" #tn)]] \
-  [[kernel]] void gemv<itype, bm, bn, tm, tn>( \
+#define instantiate_gemv_helper(name, itype, bm, bn, tm, tn, nc, axpby) \
+  template [[host_name("gemv_" #name "_bm" #bm "_bn" #bn "_tm" #tm "_tn" #tn "_nc" #nc "_axpby" #axpby)]] \
+  [[kernel]] void gemv<itype, bm, bn, tm, tn, nc, axpby>( \
     const device itype* mat [[buffer(0)]], \
-    const device itype* vec [[buffer(1)]], \
-    device itype* out [[buffer(2)]], \
-    const constant int& in_vec_size [[buffer(3)]], \
-    const constant int& out_vec_size [[buffer(4)]], \
-    const constant int& vector_batch_stride [[buffer(5)]], \
-    const constant int& matrix_batch_stride [[buffer(6)]], \
-    uint3 tid [[threadgroup_position_in_grid]], \
-    uint3 lid [[thread_position_in_threadgroup]], \
-    uint simd_gid [[simdgroup_index_in_threadgroup]], \
-    uint simd_lid [[thread_index_in_simdgroup]]);
-
-#define instantiate_gemv_nc(name, itype, bm, bn, tm, tn) \
-  template [[host_name("gemv_" #name "_bm" #bm "_bn" #bn "_tm" #tm "_tn" #tn "_nc")]] \
-  [[kernel]] void gemv_nc<itype, bm, bn, tm, tn>( \
-    const device itype* mat [[buffer(0)]], \
-    const device itype* vec [[buffer(1)]], \
-    device itype* out [[buffer(2)]], \
-    const constant int& in_vec_size [[buffer(3)]], \
-    const constant int& out_vec_size [[buffer(4)]], \
-    const constant int& nc_dim [[buffer(5)]], \
-    const device int* nc_shape [[buffer(6)]], \
-    const device size_t* nc_strides_vec [[buffer(7)]], \
-    const device size_t* nc_strides_mat [[buffer(8)]], \
+    const device itype* in_vec [[buffer(1)]], \
+    const device itype* bias [[buffer(2)]], \
+    device itype* out_vec [[buffer(3)]], \
+    const constant int& in_vec_size [[buffer(4)]], \
+    const constant int& out_vec_size [[buffer(5)]], \
+    const constant int& marix_ld [[buffer(6)]], \
+    const constant float& alpha [[buffer(7)]], \
+    const constant float& beta [[buffer(8)]], \
+    const constant int& batch_ndim [[buffer(9)]], \
+    const constant int* batch_shape [[buffer(10)]], \
+    const constant size_t* vector_batch_stride [[buffer(11)]], \
+    const constant size_t* matrix_batch_stride [[buffer(12)]], \
+    const constant size_t* bias_batch_stride [[buffer(13)]], \
+    const constant int& bias_stride [[buffer(14)]], \
     uint3 tid [[threadgroup_position_in_grid]], \
     uint3 lid [[thread_position_in_threadgroup]], \
     uint simd_gid [[simdgroup_index_in_threadgroup]], \
     uint simd_lid [[thread_index_in_simdgroup]]);
 
 #define instantiate_gemv(name, itype, bm, bn, tm, tn) \
-  instantiate_gemv_c(name, itype, bm, bn, tm, tn) \
-  instantiate_gemv_nc(name, itype, bm, bn, tm, tn)
+  instantiate_gemv_helper(name, itype, bm, bn, tm, tn, 0, 0) \
+  instantiate_gemv_helper(name, itype, bm, bn, tm, tn, 0, 1) \
+  instantiate_gemv_helper(name, itype, bm, bn, tm, tn, 1, 0) \
+  instantiate_gemv_helper(name, itype, bm, bn, tm, tn, 1, 1)
 
 #define instantiate_gemv_blocks(name, itype) \
   instantiate_gemv(name, itype, 4, 32, 1, 4) \
@@ -446,77 +447,64 @@ template <
     const int BM, /* Threadgroup rows (in threads) */
     const int BN, /* Threadgroup cols (in threads) */
     const int TM, /* Thread rows (in elements) */
-    const int TN> /* Thread cols (in elements) */
+    const int TN, /* Thread cols (in elements) */
+    const bool kDoNCBatch, /* Batch ndim > 1 */
+    const bool kDoAxpby> /* Do out = alpha * out + beta * bias */
 [[kernel, max_total_threads_per_threadgroup(BM * BN)]] void gemv_t(
     const device T* mat [[buffer(0)]],
     const device T* in_vec [[buffer(1)]],
-    device T* out_vec [[buffer(2)]], 
-    const constant int& in_vec_size [[buffer(3)]],
-    const constant int& out_vec_size [[buffer(4)]],
-    const constant int& vector_batch_stride [[buffer(5)]],
-    const constant int& matrix_batch_stride [[buffer(6)]],
+    const device T* bias [[buffer(2)]],
+    device T* out_vec [[buffer(3)]], 
+    const constant int& in_vec_size [[buffer(4)]],
+    const constant int& out_vec_size [[buffer(5)]],
+    const constant int& marix_ld [[buffer(6)]],
+    const constant float& alpha [[buffer(7)]],
+    const constant float& beta [[buffer(8)]],
+    const constant int& batch_ndim [[buffer(9)]],
+    const constant int* batch_shape [[buffer(10)]],
+    const constant size_t* vector_batch_stride [[buffer(11)]],
+    const constant size_t* matrix_batch_stride [[buffer(12)]],
+    const constant size_t* bias_batch_stride [[buffer(13)]],
+    const constant int& bias_stride [[buffer(14)]],
     uint3 tid [[threadgroup_position_in_grid]],
     uint3 lid [[thread_position_in_threadgroup]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
 
-  using gemv_kernel = GEMVTKernel<T, BM, BN, TM, TN>;
+  using gemv_kernel = GEMVTKernel<T, BM, BN, TM, TN, kDoAxpby>;
   threadgroup T tgp_memory[gemv_kernel::tgp_mem_size];
 
   // Update batch offsets
-  in_vec += tid.z * vector_batch_stride;
-  mat += tid.z * matrix_batch_stride;
-  out_vec += tid.z * out_vec_size;
-
-  gemv_kernel::run( 
-    mat, 
-    in_vec, 
-    out_vec,
-    in_vec_size,
-    out_vec_size,
-    tgp_memory,
-    tid,
-    lid,
-    simd_gid,
-    simd_lid
-  );
-}
-
-template <
-    typename T, 
-    const int BM, /* Threadgroup rows (in threads) */
-    const int BN, /* Threadgroup cols (in threads) */
-    const int TM, /* Thread rows (in elements) */
-    const int TN> /* Thread cols (in elements) */
-[[kernel, max_total_threads_per_threadgroup(BM * BN)]] void gemv_t_nc(
-    const device T* mat [[buffer(0)]],
-    const device T* in_vec [[buffer(1)]],
-    device T* out_vec [[buffer(2)]], 
-    const constant int& in_vec_size [[buffer(3)]],
-    const constant int& out_vec_size [[buffer(4)]],
-    const constant int& nc_dim [[buffer(5)]],
-    const device int* nc_shape [[buffer(6)]],
-    const device size_t* nc_strides_vec [[buffer(7)]],
-    const device size_t* nc_strides_mat [[buffer(8)]],
-    uint3 tid [[threadgroup_position_in_grid]],
-    uint3 lid [[thread_position_in_threadgroup]],
-    uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]]) {
-
-  using gemv_kernel = GEMVTKernel<T, BM, BN, TM, TN>;
-  threadgroup T tgp_memory[gemv_kernel::tgp_mem_size];
-
-  // Update batch offsets
-  in_vec += elem_to_loc(tid.z, nc_shape, nc_strides_vec, nc_dim);
-  mat += elem_to_loc(tid.z, nc_shape, nc_strides_mat, nc_dim);
+  if(kDoNCBatch) {
+    in_vec += elem_to_loc(tid.z, batch_shape, vector_batch_stride, batch_ndim);
+    mat += elem_to_loc(tid.z, batch_shape, matrix_batch_stride, batch_ndim);
+
+    if(kDoAxpby) {
+      bias += elem_to_loc(tid.z, batch_shape, bias_batch_stride, batch_ndim);
+    }
+
+  } else {
+    in_vec += tid.z * vector_batch_stride[0];
+    mat += tid.z * matrix_batch_stride[0];
+
+    if(kDoAxpby) {
+      bias += tid.z * bias_batch_stride[0];
+    }
+  }
+
   out_vec += tid.z * out_vec_size;
 
   gemv_kernel::run( 
     mat, 
     in_vec, 
+    bias,
     out_vec,
     in_vec_size,
     out_vec_size,
+    marix_ld,
+    alpha,
+    beta,
+    bias_stride,
     tgp_memory,
     tid,
     lid,
@@ -526,41 +514,34 @@ template <
 
 }
 
-#define instantiate_gemv_t_c(name, itype, bm, bn, tm, tn) \
-  template [[host_name("gemv_t_" #name "_bm" #bm "_bn" #bn "_tm" #tm "_tn" #tn)]] \
-  [[kernel]] void gemv_t<itype, bm, bn, tm, tn>( \
+#define instantiate_gemv_t_helper(name, itype, bm, bn, tm, tn, nc, axpby) \
+  template [[host_name("gemv_t_" #name "_bm" #bm "_bn" #bn "_tm" #tm "_tn" #tn "_nc" #nc "_axpby" #axpby)]] \
+  [[kernel]] void gemv_t<itype, bm, bn, tm, tn, nc, axpby>( \
     const device itype* mat [[buffer(0)]], \
-    const device itype* vec [[buffer(1)]], \
-    device itype* out [[buffer(2)]], \
-    const constant int& in_vec_size [[buffer(3)]], \
-    const constant int& out_vec_size [[buffer(4)]], \
-    const constant int& vector_batch_stride [[buffer(5)]], \
-    const constant int& matrix_batch_stride [[buffer(6)]], \
-    uint3 tid [[threadgroup_position_in_grid]], \
-    uint3 lid [[thread_position_in_threadgroup]], \
-    uint simd_gid [[simdgroup_index_in_threadgroup]], \
-    uint simd_lid [[thread_index_in_simdgroup]]);
-
-#define instantiate_gemv_t_nc(name, itype, bm, bn, tm, tn) \
-  template [[host_name("gemv_t_" #name "_bm" #bm "_bn" #bn "_tm" #tm "_tn" #tn "_nc")]] \
-  [[kernel]] void gemv_t_nc<itype, bm, bn, tm, tn>( \
-    const device itype* mat [[buffer(0)]], \
-    const device itype* vec [[buffer(1)]], \
-    device itype* out [[buffer(2)]], \
-    const constant int& in_vec_size [[buffer(3)]], \
-    const constant int& out_vec_size [[buffer(4)]], \
-    const constant int& nc_dim [[buffer(5)]], \
-    const device int* nc_shape [[buffer(6)]], \
-    const device size_t* nc_strides_vec [[buffer(7)]], \
-    const device size_t* nc_strides_mat [[buffer(8)]], \
+    const device itype* in_vec [[buffer(1)]], \
+    const device itype* bias [[buffer(2)]], \
+    device itype* out_vec [[buffer(3)]], \
+    const constant int& in_vec_size [[buffer(4)]], \
+    const constant int& out_vec_size [[buffer(5)]], \
+    const constant int& marix_ld [[buffer(6)]], \
+    const constant float& alpha [[buffer(7)]], \
+    const constant float& beta [[buffer(8)]], \
+    const constant int& batch_ndim [[buffer(9)]], \
+    const constant int* batch_shape [[buffer(10)]], \
+    const constant size_t* vector_batch_stride [[buffer(11)]], \
+    const constant size_t* matrix_batch_stride [[buffer(12)]], \
+    const constant size_t* bias_batch_stride [[buffer(13)]], \
+    const constant int& bias_stride [[buffer(14)]], \
     uint3 tid [[threadgroup_position_in_grid]], \
     uint3 lid [[thread_position_in_threadgroup]], \
     uint simd_gid [[simdgroup_index_in_threadgroup]], \
     uint simd_lid [[thread_index_in_simdgroup]]);
 
 #define instantiate_gemv_t(name, itype, bm, bn, tm, tn) \
-  instantiate_gemv_t_c(name, itype, bm, bn, tm, tn) \
-  instantiate_gemv_t_nc(name, itype, bm, bn, tm, tn)
+  instantiate_gemv_t_helper(name, itype, bm, bn, tm, tn, 0, 0) \
+  instantiate_gemv_t_helper(name, itype, bm, bn, tm, tn, 0, 1) \
+  instantiate_gemv_t_helper(name, itype, bm, bn, tm, tn, 1, 0) \
+  instantiate_gemv_t_helper(name, itype, bm, bn, tm, tn, 1, 1)
 
 #define instantiate_gemv_t_blocks(name, itype) \
   instantiate_gemv_t(name, itype, 8, 8, 4, 1) \

mlx/backend/metal/kernels/layer_norm.metal

@@ -0,0 +1,553 @@
+// Copyright © 2024 Apple Inc.
+
+#include <metal_common>
+#include <metal_simdgroup>
+
+#include "mlx/backend/metal/kernels/bf16.h"
+#include "mlx/backend/metal/kernels/defines.h"
+#include "mlx/backend/metal/kernels/utils.h"
+
+using namespace metal;
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void layer_norm_single_row(
+    const device T* x,
+    const device T* w,
+    const device T* b,
+    device T* out,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    constant uint& b_stride,
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  float sumx = 0;
+  float sumx2 = 0;
+  float thread_x[N_READS];
+
+  constexpr int SIMD_SIZE = 32;
+
+  threadgroup float local_sumx[SIMD_SIZE];
+  threadgroup float local_sumx2[SIMD_SIZE];
+  threadgroup float local_mean[1];
+  threadgroup float local_normalizer[1];
+
+  x += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+  b += b_stride * lid * N_READS;
+
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      thread_x[i] = x[i];
+      sumx2 += thread_x[i] * thread_x[i];
+      sumx += thread_x[i];
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        thread_x[i] = x[i];
+        sumx2 += thread_x[i] * thread_x[i];
+        sumx += thread_x[i];
+      }
+    }
+  }
+
+  sumx = simd_sum(sumx);
+  sumx2 = simd_sum(sumx2);
+
+  //  Initialize shared memory
+  if (simd_group_id == 0) {
+    local_sumx[simd_lane_id] = 0;
+    local_sumx2[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write simd accumulations into shared memory
+  if (simd_lane_id == 0) {
+    local_sumx[simd_group_id] = sumx;
+    local_sumx2[simd_group_id] = sumx2;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Accumulate over simd groups
+  if (simd_group_id == 0) {
+    sumx = simd_sum(local_sumx[simd_lane_id]);
+    sumx2 = simd_sum(local_sumx2[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      float mean = sumx / axis_size;
+      float variance = sumx2 / axis_size - mean * mean;
+
+      local_mean[0] = mean;
+      local_normalizer[0] = metal::precise::rsqrt(variance + eps);
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  float mean = local_mean[0];
+  float normalizer = local_normalizer[0];
+
+  // Write the outputs
+  out += gid * axis_size + lid * N_READS;
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      thread_x[i] = (thread_x[i] - mean) * normalizer;
+      out[i] = w[w_stride * i] * static_cast<T>(thread_x[i]) + b[b_stride * i];
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        thread_x[i] = (thread_x[i] - mean) * normalizer;
+        out[i] = w[w_stride * i] * static_cast<T>(thread_x[i]) + b[b_stride * i];
+      }
+    }
+  }
+}
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void layer_norm_looped(
+    const device T* x,
+    const device T* w,
+    const device T* b,
+    device T* out,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    constant uint& b_stride,
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint lsize [[threads_per_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  float sumx = 0;
+  float sumx2 = 0;
+
+  constexpr int SIMD_SIZE = 32;
+
+  threadgroup float local_sumx[SIMD_SIZE];
+  threadgroup float local_sumx2[SIMD_SIZE];
+  threadgroup float local_mean[1];
+  threadgroup float local_normalizer[1];
+
+  x += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+  b += b_stride * lid * N_READS;
+
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        float xi = x[i + r];
+        sumx2 += xi * xi;
+        sumx += xi;
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          float xi = x[i + r];
+          sumx2 += xi * xi;
+          sumx += xi;
+        }
+      }
+    }
+  }
+
+  sumx = simd_sum(sumx);
+  sumx2 = simd_sum(sumx2);
+
+  //  Initialize shared memory
+  if (simd_group_id == 0) {
+    local_sumx[simd_lane_id] = 0;
+    local_sumx2[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write simd accumulations into shared memory
+  if (simd_lane_id == 0) {
+    local_sumx[simd_group_id] = sumx;
+    local_sumx2[simd_group_id] = sumx2;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Accumulate over simd groups
+  if (simd_group_id == 0) {
+    sumx = simd_sum(local_sumx[simd_lane_id]);
+    sumx2 = simd_sum(local_sumx2[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      float mean = sumx / axis_size;
+      float variance = sumx2 / axis_size - mean * mean;
+
+      local_mean[0] = mean;
+      local_normalizer[0] = metal::precise::rsqrt(variance + eps);
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  float mean = local_mean[0];
+  float normalizer = local_normalizer[0];
+
+  // Write the outputs
+  out += gid * axis_size + lid * N_READS;
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        float xi = (x[r + i] - mean) * normalizer;
+        out[r + i] = w[w_stride * (i + r)] * static_cast<T>(xi) + b[b_stride * (i + r)];
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          float xi = (x[r + i] - mean) * normalizer;
+          out[r + i] = w[w_stride * (i + r)] * static_cast<T>(xi) + b[b_stride * (i + r)];
+        }
+      }
+    }
+  }
+}
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void vjp_layer_norm_single_row(
+    const device T* x,
+    const device T* w,
+    const device T* g,
+    device T* gx,
+    device T* gw,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  // Advance the input pointers
+  x += gid * axis_size + lid * N_READS;
+  g += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+
+  // Allocate registers for the computation and accumulators
+  float thread_x[N_READS];
+  float thread_w[N_READS];
+  float thread_g[N_READS];
+  float sumx = 0;
+  float sumx2 = 0;
+  float sumwg = 0;
+  float sumwgx = 0;
+
+  constexpr int SIMD_SIZE = 32;
+
+  threadgroup float local_sumx[SIMD_SIZE];
+  threadgroup float local_sumx2[SIMD_SIZE];
+  threadgroup float local_sumwg[SIMD_SIZE];
+  threadgroup float local_sumwgx[SIMD_SIZE];
+  threadgroup float local_mean[1];
+  threadgroup float local_normalizer[1];
+  threadgroup float local_meanwg[1];
+  threadgroup float local_meanwgx[1];
+
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      thread_x[i] = x[i];
+      thread_w[i] = w[i * w_stride];
+      thread_g[i] = g[i];
+      float wg = thread_w[i] * thread_g[i];
+      sumx += thread_x[i];
+      sumx2 += thread_x[i] * thread_x[i];
+      sumwg += wg;
+      sumwgx += wg * thread_x[i];
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        thread_x[i] = x[i];
+        thread_w[i] = w[i * w_stride];
+        thread_g[i] = g[i];
+        float wg = thread_w[i] * thread_g[i];
+        sumx += thread_x[i];
+        sumx2 += thread_x[i] * thread_x[i];
+        sumwg += wg;
+        sumwgx += wg * thread_x[i];
+      }
+    }
+  }
+
+  sumx = simd_sum(sumx);
+  sumx2 = simd_sum(sumx2);
+  sumwg = simd_sum(sumwg);
+  sumwgx = simd_sum(sumwgx);
+
+  //  Initialize shared memory
+  if (simd_group_id == 0) {
+    local_sumx[simd_lane_id] = 0;
+    local_sumx2[simd_lane_id] = 0;
+    local_sumwg[simd_lane_id] = 0;
+    local_sumwgx[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write simd accumulations into shared memory
+  if (simd_lane_id == 0) {
+    local_sumx[simd_group_id] = sumx;
+    local_sumx2[simd_group_id] = sumx2;
+    local_sumwg[simd_group_id] = sumwg;
+    local_sumwgx[simd_group_id] = sumwgx;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Accumulate over simd groups
+  if (simd_group_id == 0) {
+    sumx = simd_sum(local_sumx[simd_lane_id]);
+    sumx2 = simd_sum(local_sumx2[simd_lane_id]);
+    sumwg = simd_sum(local_sumwg[simd_lane_id]);
+    sumwgx = simd_sum(local_sumwgx[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      float mean = sumx / axis_size;
+      float variance = sumx2 / axis_size - mean * mean;
+
+      local_mean[0] = mean;
+      local_normalizer[0] = metal::precise::rsqrt(variance + eps);
+      local_meanwg[0] = sumwg / axis_size;
+      local_meanwgx[0] = sumwgx / axis_size;
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  float mean = local_mean[0];
+  float normalizer = local_normalizer[0];
+  float meanwg = local_meanwg[0];
+  float meanwgxc = local_meanwgx[0] - meanwg * mean;
+  float normalizer2 = normalizer * normalizer;
+
+  // Write the outputs
+  gx += gid * axis_size + lid * N_READS;
+  gw += gid * axis_size + lid * N_READS;
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      thread_x[i] = (thread_x[i] - mean) * normalizer;
+      gx[i] = static_cast<T>(normalizer * (thread_w[i] * thread_g[i] - meanwg) -
+                             thread_x[i] * meanwgxc * normalizer2);
+      gw[i] = static_cast<T>(thread_g[i] * thread_x[i]);
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        thread_x[i] = (thread_x[i] - mean) * normalizer;
+        gx[i] = static_cast<T>(normalizer * (thread_w[i] * thread_g[i] - meanwg) -
+                               thread_x[i] * meanwgxc * normalizer2);
+        gw[i] = static_cast<T>(thread_g[i] * thread_x[i]);
+      }
+    }
+  }
+}
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void vjp_layer_norm_looped(
+    const device T* x,
+    const device T* w,
+    const device T* g,
+    device T* gx,
+    device T* gw,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint lsize [[threads_per_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  // Advance the input pointers
+  x += gid * axis_size + lid * N_READS;
+  g += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+
+  // Allocate registers for the accumulators
+  float sumx = 0;
+  float sumx2 = 0;
+  float sumwg = 0;
+  float sumwgx = 0;
+
+  constexpr int SIMD_SIZE = 32;
+
+  threadgroup float local_sumx[SIMD_SIZE];
+  threadgroup float local_sumx2[SIMD_SIZE];
+  threadgroup float local_sumwg[SIMD_SIZE];
+  threadgroup float local_sumwgx[SIMD_SIZE];
+  threadgroup float local_mean[1];
+  threadgroup float local_normalizer[1];
+  threadgroup float local_meanwg[1];
+  threadgroup float local_meanwgx[1];
+
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        float xi = x[i + r];
+        float wi = w[(i + r) * w_stride];
+        float gi = g[i + r];
+        float wg = wi * gi;
+        sumx += xi;
+        sumx2 += xi * xi;
+        sumwg += wg;
+        sumwgx += wg * xi;
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          float xi = x[i + r];
+          float wi = w[(i + r) * w_stride];
+          float gi = g[i + r];
+          float wg = wi * gi;
+          sumx += xi;
+          sumx2 += xi * xi;
+          sumwg += wg;
+          sumwgx += wg * xi;
+        }
+      }
+    }
+  }
+
+  sumx = simd_sum(sumx);
+  sumx2 = simd_sum(sumx2);
+  sumwg = simd_sum(sumwg);
+  sumwgx = simd_sum(sumwgx);
+
+  //  Initialize shared memory
+  if (simd_group_id == 0) {
+    local_sumx[simd_lane_id] = 0;
+    local_sumx2[simd_lane_id] = 0;
+    local_sumwg[simd_lane_id] = 0;
+    local_sumwgx[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write simd accumulations into shared memory
+  if (simd_lane_id == 0) {
+    local_sumx[simd_group_id] = sumx;
+    local_sumx2[simd_group_id] = sumx2;
+    local_sumwg[simd_group_id] = sumwg;
+    local_sumwgx[simd_group_id] = sumwgx;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Accumulate over simd groups
+  if (simd_group_id == 0) {
+    sumx = simd_sum(local_sumx[simd_lane_id]);
+    sumx2 = simd_sum(local_sumx2[simd_lane_id]);
+    sumwg = simd_sum(local_sumwg[simd_lane_id]);
+    sumwgx = simd_sum(local_sumwgx[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      float mean = sumx / axis_size;
+      float variance = sumx2 / axis_size - mean * mean;
+
+      local_mean[0] = mean;
+      local_normalizer[0] = metal::precise::rsqrt(variance + eps);
+      local_meanwg[0] = sumwg / axis_size;
+      local_meanwgx[0] = sumwgx / axis_size;
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  float mean = local_mean[0];
+  float normalizer = local_normalizer[0];
+  float meanwg = local_meanwg[0];
+  float meanwgxc = local_meanwgx[0] - meanwg * mean;
+  float normalizer2 = normalizer * normalizer;
+
+  // Write the outputs
+  gx += gid * axis_size + lid * N_READS;
+  gw += gid * axis_size + lid * N_READS;
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        float xi = (x[i + r] - mean) * normalizer;
+        float wi = w[(i + r) * w_stride];
+        float gi = g[i + r];
+        gx[i + r] = static_cast<T>(normalizer * (wi * gi - meanwg) -
+                                   xi * meanwgxc * normalizer2);
+        gw[i + r] = static_cast<T>(gi * xi);
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          float xi = (x[i + r] - mean) * normalizer;
+          float wi = w[(i + r) * w_stride];
+          float gi = g[i + r];
+          gx[i + r] = static_cast<T>(normalizer * (wi * gi - meanwg) -
+                                     xi * meanwgxc * normalizer2);
+          gw[i + r] = static_cast<T>(gi * xi);
+        }
+      }
+    }
+  }
+}
+
+// clang-format off
+#define instantiate_layer_norm_single_row(name, itype)            \
+  template [[host_name("layer_norm" #name)]] [[kernel]] void      \
+  layer_norm_single_row<itype>(                                   \
+      const device itype* x,                                      \
+      const device itype* w,                                      \
+      const device itype* b,                                      \
+      device itype* out,                                          \
+      constant float& eps,                                        \
+      constant uint& axis_size,                                   \
+      constant uint& w_stride,                                    \
+      constant uint& b_stride,                                    \
+      uint gid [[thread_position_in_grid]],                       \
+      uint lid [[thread_position_in_threadgroup]],                \
+      uint simd_lane_id [[thread_index_in_simdgroup]],            \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);     \
+  template [[host_name("vjp_layer_norm" #name)]] [[kernel]] void  \
+  vjp_layer_norm_single_row<itype>(                               \
+      const device itype* x,                                      \
+      const device itype* w,                                      \
+      const device itype* g,                                      \
+      device itype* gx,                                           \
+      device itype* gw,                                           \
+      constant float& eps,                                        \
+      constant uint& axis_size,                                   \
+      constant uint& w_stride,                                    \
+      uint gid [[thread_position_in_grid]],                       \
+      uint lid [[thread_position_in_threadgroup]],                \
+      uint simd_lane_id [[thread_index_in_simdgroup]],            \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+
+#define instantiate_layer_norm_looped(name, itype)                       \
+  template [[host_name("layer_norm_looped" #name)]] [[kernel]] void      \
+  layer_norm_looped<itype>(                                              \
+      const device itype* x,                                             \
+      const device itype* w,                                             \
+      const device itype* b,                                             \
+      device itype* out,                                                 \
+      constant float& eps,                                               \
+      constant uint& axis_size,                                          \
+      constant uint& w_stride,                                           \
+      constant uint& b_stride,                                           \
+      uint gid [[thread_position_in_grid]],                              \
+      uint lid [[thread_position_in_threadgroup]],                       \
+      uint lsize [[threads_per_threadgroup]],                            \
+      uint simd_lane_id [[thread_index_in_simdgroup]],                   \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);            \
+  template [[host_name("vjp_layer_norm_looped" #name)]] [[kernel]] void  \
+  vjp_layer_norm_looped<itype>(                                          \
+      const device itype* x,                                             \
+      const device itype* w,                                             \
+      const device itype* g,                                             \
+      device itype* gx,                                                  \
+      device itype* gb,                                                  \
+      constant float& eps,                                               \
+      constant uint& axis_size,                                          \
+      constant uint& w_stride,                                           \
+      uint gid [[thread_position_in_grid]],                              \
+      uint lid [[thread_position_in_threadgroup]],                       \
+      uint lsize [[threads_per_threadgroup]],                            \
+      uint simd_lane_id [[thread_index_in_simdgroup]],                   \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+
+#define instantiate_layer_norm(name, itype)      \
+  instantiate_layer_norm_single_row(name, itype) \
+  instantiate_layer_norm_looped(name, itype)
+
+instantiate_layer_norm(float32, float)
+instantiate_layer_norm(float16, half)
+instantiate_layer_norm(bfloat16, bfloat16_t)
+    // clang-format on
+

mlx/backend/metal/kernels/quantized.metal

@@ -15,14 +15,6 @@ using namespace metal;
 
 MLX_MTL_CONST int SIMD_SIZE = 32;
 
-template <typename T> struct AccT {
-  typedef T acc_t;
-};
-
-template <> struct AccT<bfloat16_t> {
-  typedef float acc_t;
-};
-
 
 template <typename T, typename U, int values_per_thread, int bits>
 inline U load_vector(const device T *x, thread U *x_thread) {
@@ -60,6 +52,51 @@ inline U load_vector(const device T *x, thread U *x_thread) {
   return sum;
 }
 
+template <typename T, typename U, int values_per_thread, int bits>
+inline U load_vector_safe(const device T *x, thread U *x_thread, int N) {
+  static_assert(bits == 2 || bits == 4 || bits == 8, "Template undefined for bits not in {2, 4, 8}");
+
+  U sum = 0;
+
+  if (bits == 2) {
+    for (int i = 0; i < N; i += 4) {
+      sum += x[i] + x[i+1] + x[i+2] + x[i+3];
+      x_thread[i] = x[i];
+      x_thread[i+1] = x[i+1] / 4.0f;
+      x_thread[i+2] = x[i+2] / 16.0f;
+      x_thread[i+3] = x[i+3] / 64.0f;
+    }
+    for (int i=N; i<values_per_thread; i++) {
+      x_thread[i] = 0;
+    }
+  }
+
+  else if (bits == 4) {
+    for (int i = 0; i < N; i += 4) {
+      sum += x[i] + x[i+1] + x[i+2] + x[i+3];
+      x_thread[i] = x[i];
+      x_thread[i+1] = x[i+1] / 16.0f;
+      x_thread[i+2] = x[i+2] / 256.0f;
+      x_thread[i+3] = x[i+3] / 4096.0f;
+    }
+    for (int i=N; i<values_per_thread; i++) {
+      x_thread[i] = 0;
+    }
+  }
+
+  else if (bits == 8) {
+    for (int i = 0; i < N; i++) {
+      sum += x[i];
+      x_thread[i] = x[i];
+    }
+    for (int i=N; i<values_per_thread; i++) {
+      x_thread[i] = 0;
+    }
+  }
+
+  return sum;
+}
+
 template <typename U, int values_per_thread, int bits>
 inline U qdot(const device uint8_t* w, const thread U *x_thread, U scale, U bias, U sum) {
   static_assert(bits == 2 || bits == 4 || bits == 8, "Template undefined for bits not in {2, 4, 8}");
@@ -96,6 +133,74 @@ inline U qdot(const device uint8_t* w, const thread U *x_thread, U scale, U bias
   return scale * accum + sum * bias;
 }
 
+template <typename U, int values_per_thread, int bits>
+inline U qdot_safe(const device uint8_t* w, const thread U *x_thread, U scale, U bias, U sum, int N) {
+  static_assert(bits == 2 || bits == 4 || bits == 8, "Template undefined for bits not in {2, 4, 8}");
+
+  U accum = 0;
+
+  if (bits == 2) {
+    for (int i = 0; i < (N / 4); i++) {
+      accum += (
+          x_thread[4*i] * (w[i] & 0x03)
+          + x_thread[4*i+1] * (w[i] & 0x0c)
+          + x_thread[4*i+2] * (w[i] & 0x30)
+          + x_thread[4*i+3] * (w[i] & 0xc0));
+    }
+  }
+
+  else if (bits == 4) {
+    const device uint16_t* ws = (const device uint16_t*)w;
+    for (int i = 0; i < (N / 4); i++) {
+      accum += (
+          x_thread[4*i] * (ws[i] & 0x000f)
+          + x_thread[4*i+1] * (ws[i] & 0x00f0)
+          + x_thread[4*i+2] * (ws[i] & 0x0f00)
+          + x_thread[4*i+3] * (ws[i] & 0xf000));
+    }
+  }
+
+  else if (bits == 8) {
+    for (int i = 0; i < N; i++) {
+      accum += x_thread[i] * w[i];
+    }
+  }
+
+  return scale * accum + sum * bias;
+}
+
+template <typename U, int values_per_thread, int bits>
+inline void qouter(const thread uint8_t* w, U x, U scale, U bias, thread U* result) {
+  static_assert(bits == 2 || bits == 4 || bits == 8, "Template undefined for bits not in {2, 4, 8}");
+
+  if (bits == 2) {
+    U s[4] = {scale, scale / 4.0f, scale / 16.0f, scale / 64.0f};
+    for (int i = 0; i < (values_per_thread / 4); i++) {
+      result[4*i] += x * (s[0] * (w[i] & 0x03) + bias);
+      result[4*i+1] += x * (s[1] * (w[i] & 0x0c) + bias);
+      result[4*i+2] += x * (s[2] * (w[i] & 0x30) + bias);
+      result[4*i+3] += x * (s[3] * (w[i] & 0xc0) + bias);
+    }
+  }
+
+  else if (bits == 4) {
+    const thread uint16_t* ws = (const thread uint16_t*)w;
+    U s[4] = {scale, scale / 16.0f, scale / 256.0f, scale / 4096.0f};
+    for (int i = 0; i < (values_per_thread / 4); i++) {
+      result[4*i] += x * (s[0] * (ws[i] & 0x000f) + bias);
+      result[4*i+1] += x * (s[1] * (ws[i] & 0x00f0) + bias);
+      result[4*i+2] += x * (s[2] * (ws[i] & 0x0f00) + bias);
+      result[4*i+3] += x * (s[3] * (ws[i] & 0xf000) + bias);
+    }
+  }
+
+  else if (bits == 8) {
+    for (int i = 0; i < values_per_thread; i++) {
+      result[i] += x * (scale * w[i] + bias);
+    }
+  }
+}
+
 template <typename T, int group_size, int bits, int packs_per_thread>
 [[kernel]] void qmv_fast(
     const device uint32_t* w [[buffer(0)]],
@@ -204,7 +309,8 @@ template <typename T, const int group_size, const int bits>
     x += tid.z * in_vec_size + simd_lid * values_per_thread;
     y += tid.z * out_vec_size + out_row;
 
-    for (int k = 0; k < in_vec_size; k += block_size) {
+    int k = 0;
+    for (; k < in_vec_size-block_size; k += block_size) {
       U sum = load_vector<T, U, values_per_thread, bits>(x, x_thread);
 
       for (int row = 0; out_row + row < out_vec_size; row++) {
@@ -222,6 +328,18 @@ template <typename T, const int group_size, const int bits>
       biases += block_size / group_size;
       x += block_size;
     }
+    const int remaining = clamp(static_cast<int>(in_vec_size - k - simd_lid * values_per_thread), 0, values_per_thread);
+    U sum = load_vector_safe<T, U, values_per_thread, bits>(x, x_thread, remaining);
+
+    for (int row = 0; out_row + row < out_vec_size; row++) {
+      const device uint8_t* wl = (const device uint8_t *)(w + row * in_vec_size_w);
+      const device T* sl = scales + row * in_vec_size_g;
+      const device T* bl = biases + row * in_vec_size_g;
+
+      U s = sl[0];
+      U b = bl[0];
+      result[row] += qdot<U, values_per_thread, bits>(wl, x_thread, s, b, sum);
+    }
 
     for (int row = 0; out_row + row < out_vec_size; row++) {
       result[row] = simd_sum(result[row]);
@@ -239,7 +357,8 @@ template <typename T, const int group_size, const int bits>
     x += tid.z * in_vec_size + simd_lid * values_per_thread;
     y += tid.z * out_vec_size + used_out_row;
 
-    for (int k = 0; k < in_vec_size; k += block_size) {
+    int k = 0;
+    for (; k < in_vec_size-block_size; k += block_size) {
       U sum = load_vector<T, U, values_per_thread, bits>(x, x_thread);
 
       for (int row = 0; row < results_per_simdgroup; row++) {
@@ -257,6 +376,18 @@ template <typename T, const int group_size, const int bits>
       biases += block_size / group_size;
       x += block_size;
     }
+    const int remaining = clamp(static_cast<int>(in_vec_size - k - simd_lid * values_per_thread), 0, values_per_thread);
+    U sum = load_vector_safe<T, U, values_per_thread, bits>(x, x_thread, remaining);
+
+    for (int row = 0; row < results_per_simdgroup; row++) {
+      const device uint8_t* wl = (const device uint8_t *)(w + row * in_vec_size_w);
+      const device T* sl = scales + row * in_vec_size_g;
+      const device T* bl = biases + row * in_vec_size_g;
+
+      U s = sl[0];
+      U b = bl[0];
+      result[row] += qdot_safe<U, values_per_thread, bits>(wl, x_thread, s, b, sum, remaining);
+    }
 
     for (int row = 0; row < results_per_simdgroup; row++) {
       result[row] = simd_sum(result[row]);
@@ -268,7 +399,7 @@ template <typename T, const int group_size, const int bits>
 }
 
 
-template <typename T, const int BM, const int BN, const int group_size, const int bits>
+template <typename T, const int group_size, const int bits>
 [[kernel]] void qvm(
     const device T* x [[buffer(0)]],
     const device uint32_t* w [[buffer(1)]],
@@ -278,39 +409,28 @@ template <typename T, const int BM, const int BN, const int group_size, const in
     const constant int& in_vec_size [[buffer(5)]],
     const constant int& out_vec_size [[buffer(6)]],
     uint3 tid [[threadgroup_position_in_grid]],
-    uint lid [[thread_index_in_threadgroup]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
 
-  static_assert(BM == SIMD_SIZE, "qvm expects BM to be equal to SIMD_SIZE");
-  static_assert(BN == BM, "qvm expects a block size of 32x32");
+  constexpr int num_simdgroups = 8;
+  constexpr int pack_factor = 32 / bits;
+  constexpr int blocksize = SIMD_SIZE;
 
-  (void)lid;
-
-  constexpr int bitmask = (1 << bits) - 1;
-  constexpr int el_per_int = 32 / bits;
-  constexpr int colgroup = BN * el_per_int;
-  constexpr int groups_per_block = colgroup / group_size;
-
-  typedef typename AccT<T>::acc_t U;
-  threadgroup U scales_block[BM * groups_per_block];
-  threadgroup U biases_block[BM * groups_per_block];
-  threadgroup U x_block[BM];
+  typedef float U;
 
   thread uint32_t w_local;
-  thread U result[el_per_int] = {0};
+  thread U result[pack_factor] = {0};
   thread U scale = 1;
   thread U bias = 0;
   thread U x_local = 0;
 
   // Adjust positions
-  const int out_vec_size_w = out_vec_size / el_per_int;
+  const int out_vec_size_w = out_vec_size / pack_factor;
   const int out_vec_size_g = out_vec_size / group_size;
-  int out_col_start = tid.y * (BN * el_per_int);
-  int out_col = out_col_start + simd_gid * el_per_int;
-  w += out_col / el_per_int;
-  scales += out_col_start / group_size;
-  biases += out_col_start / group_size;
+  int out_col = tid.y * (num_simdgroups * pack_factor) + simd_gid * pack_factor;
+  w += out_col / pack_factor;
+  scales += out_col / group_size;
+  biases += out_col / group_size;
   x += tid.z * in_vec_size;
   y += tid.z * out_vec_size + out_col;
 
@@ -318,53 +438,39 @@ template <typename T, const int BM, const int BN, const int group_size, const in
     return;
   }
 
-  // Loop over in_vec in blocks of colgroup
-  for (int i=0; i<in_vec_size; i+=BM) {
-    int offset_lid = simd_lid + i;
-    int offset_gid = simd_gid + i;
-    bool thread_in_bounds = offset_lid < in_vec_size;
-    bool group_in_bounds = offset_gid < in_vec_size;
+  // Loop over in_vec in blocks of blocksize
+  int i = 0;
+  for (; i + blocksize <= in_vec_size; i += blocksize) {
+    x_local = x[i + simd_lid];
+    scale = scales[(i + simd_lid) * out_vec_size_g];
+    bias = biases[(i + simd_lid) * out_vec_size_g];
+    w_local = w[(i + simd_lid) * out_vec_size_w];
 
-    // Load the vec to shared memory
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (simd_gid == 0) {
-      x_block[simd_lid] = (thread_in_bounds) ? x[offset_lid] : 0;
-    }
-
-    // Load the scales and biases to shared memory
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    if (simd_lid < groups_per_block && group_in_bounds) {
-      scales_block[simd_gid * groups_per_block + simd_lid] = scales[offset_gid * out_vec_size_g + simd_lid];
-      biases_block[simd_gid * groups_per_block + simd_lid] = biases[offset_gid * out_vec_size_g + simd_lid];
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Load in_vec, scale, bias to registers
-    x_local = x_block[simd_lid];
-    scale = scales_block[simd_lid * groups_per_block + (simd_gid * el_per_int) / group_size];
-    bias = biases_block[simd_lid * groups_per_block + (simd_gid * el_per_int) / group_size];
-
-    // Load the matrix elements
-    w_local = (thread_in_bounds) ? w[offset_lid * out_vec_size_w] : 0;
-
-    // Do all the work.
-    #pragma clang loop unroll(full)
-    for (int k=0; k<el_per_int; k++) {
-      result[k] += (scale * static_cast<U>(w_local & bitmask) + bias) * x_local;
-      w_local >>= bits;
-    }
+    qouter<U, pack_factor, bits>((thread uint8_t *)&w_local, x_local, scale, bias, result);
   }
+  if (static_cast<int>(i + simd_lid) < in_vec_size) {
+    x_local = x[i + simd_lid];
+    scale = scales[(i + simd_lid) * out_vec_size_g];
+    bias = biases[(i + simd_lid) * out_vec_size_g];
+    w_local = w[(i + simd_lid) * out_vec_size_w];
+  } else {
+    x_local = 0;
+    scale = 0;
+    bias = 0;
+    w_local = 0;
+  }
+  qouter<U, pack_factor, bits>((thread uint8_t *)&w_local, x_local, scale, bias, result);
 
   // Accumulate in the simdgroup
   #pragma clang loop unroll(full)
-  for (int k=0; k<el_per_int; k++) {
+  for (int k=0; k<pack_factor; k++) {
     result[k] = simd_sum(result[k]);
   }
 
   // Store the result
   if (simd_lid == 0) {
     #pragma clang loop unroll(full)
-    for (int k=0; k<el_per_int; k++) {
+    for (int k=0; k<pack_factor; k++) {
       y[k] = static_cast<T>(result[k]);
     }
   }
@@ -414,6 +520,7 @@ template <typename T, const int BM, const int BK, const int BN, const int group_
   const int K_g = K / group_size;
   const int y_row = tid.y * BM;
   const int y_col = tid.x * BN;
+
   x += y_row * K;
   w += y_col * K_w;
   scales += y_col * K_g;
@@ -466,7 +573,10 @@ template <typename T, const int BM, const int BK, const int BN, const int group_
           const device uint32_t * w_local = w + offset_row * K_w + offset_col;
           threadgroup T * Ws_local = Ws + offset_row * BK + offset_col * el_per_int;
 
-          if (y_row + offset_row < N) {
+          // y_col corresponds to the row of the weight matrix and added to
+          // offset_row it should be less than the total number of rows
+          // otherwise skip.
+          if (y_col + offset_row < N) {
             uint32_t wi = *w_local;
             T scale = scales_block[offset_row * groups_per_block + offset_col / (group_size / el_per_int)];
             T bias = biases_block[offset_row * groups_per_block + offset_col / (group_size / el_per_int)];
@@ -619,7 +729,7 @@ template <typename T, const int BM, const int BK, const int BN, const int group_
           const device uint32_t * w_local = w + offset_row * N_w + offset_col;
           threadgroup T * Ws_local = Ws + offset_row * BN + offset_col * el_per_int;
 
-          if (y_row + offset_row < K) {
+          if (k + offset_row < K) {
             uint32_t wi = *w_local;
             T scale = scales_block[offset_row * groups_per_block + offset_col / (group_size / el_per_int)];
             T bias = biases_block[offset_row * groups_per_block + offset_col / (group_size / el_per_int)];
@@ -738,7 +848,7 @@ instantiate_qmv_types( 32, 8)
 
 #define instantiate_qvm(name, itype, group_size, bits) \
   template [[host_name("qvm_" #name "_gs_" #group_size "_b_" #bits)]] \
-  [[kernel]] void qvm<itype, 32, 32, group_size, bits>( \
+  [[kernel]] void qvm<itype, group_size, bits>( \
     const device itype* x [[buffer(0)]], \
     const device uint32_t* w [[buffer(1)]], \
     const device itype* scales [[buffer(2)]], \
@@ -747,7 +857,6 @@ instantiate_qmv_types( 32, 8)
     const constant int& in_vec_size [[buffer(5)]], \
     const constant int& out_vec_size [[buffer(6)]], \
     uint3 tid [[threadgroup_position_in_grid]], \
-    uint lid [[thread_index_in_threadgroup]], \
     uint simd_gid [[simdgroup_index_in_threadgroup]], \
     uint simd_lid [[thread_index_in_simdgroup]]);
 

mlx/backend/metal/kernels/reduction/kernels/reduce_col.metal

@@ -6,6 +6,69 @@
 
 using namespace metal;
 
+///////////////////////////////////////////////////////////////////////////////
+// Small column reduce kernel
+///////////////////////////////////////////////////////////////////////////////
+
+template <typename T, typename U, typename Op>
+[[kernel]] void col_reduce_small(
+    const device T *in [[buffer(0)]],
+    device U *out [[buffer(1)]],
+    const constant size_t& reduction_size [[buffer(2)]],
+    const constant size_t& reduction_stride [[buffer(3)]],
+    const constant size_t& out_size [[buffer(4)]],
+    const constant int* shape [[buffer(5)]],
+    const constant size_t* strides [[buffer(6)]],
+    const constant int& ndim [[buffer(7)]],
+    const constant size_t& non_col_reductions [[buffer(8)]],
+    const constant int* non_col_shapes [[buffer(9)]],
+    const constant size_t* non_col_strides [[buffer(10)]],
+    const constant int& non_col_ndim [[buffer(11)]],
+    uint tid [[thread_position_in_grid]]) {
+
+  // Appease the compiler
+  (void)out_size;
+
+  Op op;
+  U total_val = Op::init;
+
+  auto out_idx = tid;
+
+  in += elem_to_loc(
+        out_idx,
+        shape + non_col_ndim,
+        strides + non_col_ndim,
+        ndim - non_col_ndim);
+
+  for(uint i = 0; i < non_col_reductions; i++) {
+    size_t in_idx = elem_to_loc(i, non_col_shapes, non_col_strides, non_col_ndim);
+
+    for(uint j = 0; j < reduction_size; j++, in_idx += reduction_stride) {
+      U val = static_cast<U>(in[in_idx]);
+      total_val = op(total_val, val);
+    }
+  }
+
+  out[out_idx] = total_val;
+}
+
+#define instantiate_col_reduce_small(name, itype, otype, op) \
+  template [[host_name("col_reduce_small_" #name)]] \
+  [[kernel]] void col_reduce_small<itype, otype, op>( \
+      const device itype *in [[buffer(0)]], \
+      device otype *out [[buffer(1)]], \
+      const constant size_t& reduction_size [[buffer(2)]], \
+      const constant size_t& reduction_stride [[buffer(3)]], \
+      const constant size_t& out_size [[buffer(4)]], \
+      const constant int* shape [[buffer(5)]],  \
+      const constant size_t* strides [[buffer(6)]],  \
+      const constant int& ndim [[buffer(7)]],  \
+      const constant size_t& non_col_reductions [[buffer(8)]], \
+      const constant int* non_col_shapes [[buffer(9)]], \
+      const constant size_t* non_col_strides [[buffer(10)]], \
+      const constant int& non_col_ndim [[buffer(11)]], \
+      uint tid [[thread_position_in_grid]]);
+
 ///////////////////////////////////////////////////////////////////////////////
 // Column reduce helper
 ///////////////////////////////////////////////////////////////////////////////
@@ -171,9 +234,11 @@ template <typename T, typename U, typename Op, int N_READS = REDUCE_N_READS>
 ///////////////////////////////////////////////////////////////////////////////
 
 #define instantiate_same_col_reduce_helper(name, tname, type, op) \
+  instantiate_col_reduce_small(name ##tname, type, type, op<type>) \
   instantiate_col_reduce_general(name ##tname, type, type, op<type>)
 
 #define instantiate_same_col_reduce_na_helper(name, tname, type, op) \
+  instantiate_col_reduce_small(name ##tname, type, type, op<type>) \
   instantiate_col_reduce_general_no_atomics(name ##tname, type, type, op<type>)
 
 instantiate_reduce_ops(instantiate_same_col_reduce_helper, instantiate_reduce_helper_types)
@@ -181,4 +246,8 @@ instantiate_reduce_ops(instantiate_same_col_reduce_na_helper, instantiate_reduce
 
 instantiate_col_reduce_general(sumbool_, bool, uint32_t, Sum<uint32_t>)
 instantiate_reduce_from_types(instantiate_col_reduce_general, and, bool, And)
-instantiate_reduce_from_types(instantiate_col_reduce_general, or, bool, Or)
+instantiate_reduce_from_types(instantiate_col_reduce_general, or, bool, Or)
+
+instantiate_col_reduce_small(sumbool_, bool, uint32_t, Sum<uint32_t>)
+instantiate_reduce_from_types(instantiate_col_reduce_small, and, bool, And)
+instantiate_reduce_from_types(instantiate_col_reduce_small, or, bool, Or)

mlx/backend/metal/kernels/reduction/kernels/reduce_row.metal

@@ -108,15 +108,17 @@ template <typename T, typename U, typename Op>
     const short i_ed = short(reduction_size);
     const short i_jump = reductions_per_thread;
 
-    for(short r = r_st; r < r_ed; r += r_jump) {
+    if(r_st < r_jump) {
+      for(short r = r_st; r < r_ed; r += r_jump) {
 
-      uint in_idx = elem_to_loc(out_idx + r * out_size, shape, strides, ndim);
-      const device T * in_row = in + in_idx;
+        uint in_idx = elem_to_loc(out_idx + r * out_size, shape, strides, ndim);
+        const device T * in_row = in + in_idx;
+
+        for(short i = i_st; i < i_ed; i += i_jump) {
+          total_val = op(static_cast<U>(in_row[i]), total_val);
+        }
 
-      for(short i = i_st; i < i_ed; i += i_jump) {
-        total_val = op(static_cast<U>(in_row[i]), total_val);
       }
-
     }
 
   }

mlx/backend/metal/kernels/rms_norm.metal

@@ -0,0 +1,435 @@
+// Copyright © 2024 Apple Inc.
+
+#include <metal_common>
+#include <metal_simdgroup>
+
+#include "mlx/backend/metal/kernels/bf16.h"
+#include "mlx/backend/metal/kernels/defines.h"
+#include "mlx/backend/metal/kernels/utils.h"
+
+using namespace metal;
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void rms_single_row(
+    const device T* x,
+    const device T* w,
+    device T* out,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    threadgroup float* local_inv_mean [[threadgroup(0)]],
+    threadgroup float* local_sums [[threadgroup(1)]],
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  float acc = 0;
+  x += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      float xi = x[i];
+      acc += xi * xi;
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        float xi = x[i];
+        acc += xi * xi;
+      }
+    }
+  }
+  acc = simd_sum(acc);
+  //  Initialize shared memory
+  if (simd_group_id == 0) {
+    local_sums[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write simd accumulations into shared memory
+  if (simd_lane_id == 0) {
+    local_sums[simd_group_id] = acc;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Accumulate over simd groups
+  if (simd_group_id == 0) {
+    acc = simd_sum(local_sums[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      local_inv_mean[0] = metal::precise::rsqrt(acc / axis_size + eps);
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write the outputs
+  out += gid * axis_size + lid * N_READS;
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      out[i] = w[w_stride * i] * static_cast<T>(x[i] * local_inv_mean[0]);
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        out[i] = w[w_stride * i] * static_cast<T>(x[i] * local_inv_mean[0]);
+      }
+    }
+  }
+}
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void rms_looped(
+    const device T* x,
+    const device T* w,
+    device T* out,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    threadgroup float* local_inv_mean [[threadgroup(0)]],
+    threadgroup float* local_sums [[threadgroup(1)]],
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint lsize [[threads_per_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  float acc = 0;
+  x += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        float xi = x[i + r];
+        acc += xi * xi;
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          float xi = x[i + r];
+          acc += xi * xi;
+        }
+      }
+    }
+  }
+  acc = simd_sum(acc);
+  //  Initialize shared memory
+  if (simd_group_id == 0) {
+    local_sums[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write simd accumulations into shared memory
+  if (simd_lane_id == 0) {
+    local_sums[simd_group_id] = acc;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Accumulate over simd groups
+  if (simd_group_id == 0) {
+    acc = simd_sum(local_sums[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      local_inv_mean[0] = metal::precise::rsqrt(acc / axis_size + eps);
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+
+  // Write the outputs
+  out += gid * axis_size + lid * N_READS;
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        out[r + i] = w[w_stride * (i + r)] *
+            static_cast<T>(x[r + i] * local_inv_mean[0]);
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          out[r + i] = w[w_stride * (i + r)] *
+              static_cast<T>(x[r + i] * local_inv_mean[0]);
+        }
+      }
+    }
+  }
+}
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void vjp_rms_single_row(
+    const device T* x,
+    const device T* w,
+    const device T* g,
+    device T* gx,
+    device T* gw,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  // Advance the input pointers
+  x += gid * axis_size + lid * N_READS;
+  g += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+
+  // Allocate registers for the computation and accumulators
+  float thread_x[N_READS];
+  float thread_w[N_READS];
+  float thread_g[N_READS];
+  float sumx2 = 0;
+  float sumgwx = 0;
+
+  // Allocate shared memory to implement the reduction
+  constexpr int SIMD_SIZE = 32;
+  threadgroup float local_sumx2[SIMD_SIZE];
+  threadgroup float local_sumgwx[SIMD_SIZE];
+  threadgroup float local_normalizer[1];
+  threadgroup float local_meangwx[1];
+
+  // Read and accumulate locally
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      thread_x[i] = x[i];
+      thread_w[i] = w[w_stride * i];
+      thread_g[i] = g[i];
+
+      sumx2 += thread_x[i] * thread_x[i];
+      sumgwx += thread_x[i] * thread_w[i] * thread_g[i];
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        thread_x[i] = x[i];
+        thread_w[i] = w[w_stride * i];
+        thread_g[i] = g[i];
+
+        sumx2 += thread_x[i] * thread_x[i];
+        sumgwx += thread_x[i] * thread_w[i] * thread_g[i];
+      }
+    }
+  }
+
+  // Accumulate across threads
+  sumx2 = simd_sum(sumx2);
+  sumgwx = simd_sum(sumgwx);
+  if (simd_group_id == 0) {
+    local_sumx2[simd_lane_id] = 0;
+    local_sumgwx[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  if (simd_lane_id == 0) {
+    local_sumx2[simd_group_id] = sumx2;
+    local_sumgwx[simd_group_id] = sumgwx;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  if (simd_group_id == 0) {
+    sumx2 = simd_sum(local_sumx2[simd_lane_id]);
+    sumgwx = simd_sum(local_sumgwx[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      local_meangwx[0] = sumgwx / axis_size;
+      local_normalizer[0] = metal::precise::rsqrt(sumx2 / axis_size + eps);
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  float meangwx = local_meangwx[0];
+  float normalizer = local_normalizer[0];
+  float normalizer3 = normalizer * normalizer * normalizer;
+
+  // Write the outputs
+  gx += gid * axis_size + lid * N_READS;
+  gw += gid * axis_size + lid * N_READS;
+  if (lid * N_READS + N_READS <= axis_size) {
+    for (int i = 0; i < N_READS; i++) {
+      gx[i] = static_cast<T>(thread_g[i] * thread_w[i] * normalizer - thread_x[i] * meangwx * normalizer3);
+      gw[i] = static_cast<T>(thread_g[i] * thread_x[i] * normalizer);
+    }
+  } else {
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        gx[i] = static_cast<T>(thread_g[i] * thread_w[i] * normalizer - thread_x[i] * meangwx * normalizer3);
+        gw[i] = static_cast<T>(thread_g[i] * thread_x[i] * normalizer);
+      }
+    }
+  }
+}
+
+template <typename T, int N_READS = RMS_N_READS>
+[[kernel]] void vjp_rms_looped(
+    const device T* x,
+    const device T* w,
+    const device T* g,
+    device T* gx,
+    device T* gw,
+    constant float& eps,
+    constant uint& axis_size,
+    constant uint& w_stride,
+    uint gid [[threadgroup_position_in_grid]],
+    uint lid [[thread_position_in_threadgroup]],
+    uint lsize [[threads_per_threadgroup]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
+  // Advance the input pointers
+  x += gid * axis_size + lid * N_READS;
+  g += gid * axis_size + lid * N_READS;
+  w += w_stride * lid * N_READS;
+
+  // Allocate registers for the accumulators
+  float sumx2 = 0;
+  float sumgwx = 0;
+
+  // Allocate shared memory to implement the reduction
+  constexpr int SIMD_SIZE = 32;
+  threadgroup float local_sumx2[SIMD_SIZE];
+  threadgroup float local_sumgwx[SIMD_SIZE];
+  threadgroup float local_normalizer[1];
+  threadgroup float local_meangwx[1];
+
+  // Read and accumulate locally
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        float xi = x[i + r];
+        float wi = w[w_stride * (i + r)];
+        float gi = g[i + r];
+
+        sumx2 += xi * xi;
+        sumgwx += xi * wi * gi;
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          float xi = x[i + r];
+          float wi = w[w_stride * (i + r)];
+          float gi = g[i + r];
+
+          sumx2 += xi * xi;
+          sumgwx += xi * wi * gi;
+        }
+      }
+    }
+  }
+
+  // Accumulate across threads
+  sumx2 = simd_sum(sumx2);
+  sumgwx = simd_sum(sumgwx);
+  if (simd_group_id == 0) {
+    local_sumx2[simd_lane_id] = 0;
+    local_sumgwx[simd_lane_id] = 0;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  if (simd_lane_id == 0) {
+    local_sumx2[simd_group_id] = sumx2;
+    local_sumgwx[simd_group_id] = sumgwx;
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  if (simd_group_id == 0) {
+    sumx2 = simd_sum(local_sumx2[simd_lane_id]);
+    sumgwx = simd_sum(local_sumgwx[simd_lane_id]);
+    if (simd_lane_id == 0) {
+      local_meangwx[0] = sumgwx / axis_size;
+      local_normalizer[0] = metal::precise::rsqrt(sumx2 / axis_size + eps);
+    }
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  float meangwx = local_meangwx[0];
+  float normalizer = local_normalizer[0];
+  float normalizer3 = normalizer * normalizer * normalizer;
+
+  // Write the outputs
+  gx += gid * axis_size + lid * N_READS;
+  gw += gid * axis_size + lid * N_READS;
+  for (uint r = 0; r < axis_size; r += lsize * N_READS) {
+    if (r + lid * N_READS + N_READS <= axis_size) {
+      for (int i = 0; i < N_READS; i++) {
+        float xi = x[i + r];
+        float wi = w[w_stride * (i + r)];
+        float gi = g[i + r];
+
+        gx[i + r] = static_cast<T>(gi * wi * normalizer - xi * meangwx * normalizer3);
+        gw[i + r] = static_cast<T>(gi * xi * normalizer);
+      }
+    } else {
+      for (int i = 0; i < N_READS; i++) {
+        if ((r + lid * N_READS + i) < axis_size) {
+          float xi = x[i + r];
+          float wi = w[w_stride * (i + r)];
+          float gi = g[i + r];
+
+          gx[i + r] = static_cast<T>(gi * wi * normalizer - xi * meangwx * normalizer3);
+          gw[i + r] = static_cast<T>(gi * xi * normalizer);
+        }
+      }
+    }
+  }
+}
+
+// clang-format off
+#define instantiate_rms_single_row(name, itype)               \
+  template [[host_name("rms" #name)]] [[kernel]] void         \
+  rms_single_row<itype>(                                      \
+      const device itype* x,                                  \
+      const device itype* w,                                  \
+      device itype* out,                                      \
+      constant float& eps,                                    \
+      constant uint& axis_size,                               \
+      constant uint& w_stride,                                \
+      threadgroup float* local_inv_mean [[threadgroup(0)]],   \
+      threadgroup float* local_sums [[threadgroup(1)]],       \
+      uint gid [[thread_position_in_grid]],                   \
+      uint lid [[thread_position_in_threadgroup]],            \
+      uint simd_lane_id [[thread_index_in_simdgroup]],        \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]); \
+                                                              \
+  template [[host_name("vjp_rms" #name)]] [[kernel]] void     \
+  vjp_rms_single_row<itype>(                                  \
+      const device itype* x,                                  \
+      const device itype* w,                                  \
+      const device itype* g,                                  \
+      device itype* gx,                                       \
+      device itype* gw,                                       \
+      constant float& eps,                                    \
+      constant uint& axis_size,                               \
+      constant uint& w_stride,                                \
+      uint gid [[thread_position_in_grid]],                   \
+      uint lid [[thread_position_in_threadgroup]],            \
+      uint simd_lane_id [[thread_index_in_simdgroup]],        \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+
+#define instantiate_rms_looped(name, itype)                      \
+  template [[host_name("rms_looped" #name)]] [[kernel]] void     \
+  rms_looped<itype>(                                             \
+      const device itype* x,                                     \
+      const device itype* w,                                     \
+      device itype* out,                                         \
+      constant float& eps,                                       \
+      constant uint& axis_size,                                  \
+      constant uint& w_stride,                                   \
+      threadgroup float* local_inv_mean [[threadgroup(0)]],      \
+      threadgroup float* local_sums [[threadgroup(1)]],          \
+      uint gid [[thread_position_in_grid]],                      \
+      uint lid [[thread_position_in_threadgroup]],               \
+      uint lsize [[threads_per_threadgroup]],                    \
+      uint simd_lane_id [[thread_index_in_simdgroup]],           \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);    \
+                                                                 \
+  template [[host_name("vjp_rms_looped" #name)]] [[kernel]] void \
+  vjp_rms_looped<itype>(                                         \
+      const device itype* x,                                     \
+      const device itype* w,                                     \
+      const device itype* g,                                     \
+      device itype* gx,                                          \
+      device itype* gw,                                          \
+      constant float& eps,                                       \
+      constant uint& axis_size,                                  \
+      constant uint& w_stride,                                   \
+      uint gid [[thread_position_in_grid]],                      \
+      uint lid [[thread_position_in_threadgroup]],               \
+      uint lsize [[threads_per_threadgroup]],                    \
+      uint simd_lane_id [[thread_index_in_simdgroup]],           \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);
+
+#define instantiate_rms(name, itype)      \
+  instantiate_rms_single_row(name, itype) \
+  instantiate_rms_looped(name, itype)
+
+instantiate_rms(float32, float)
+instantiate_rms(float16, half)
+instantiate_rms(bfloat16, bfloat16_t)
+    // clang-format on

mlx/backend/metal/kernels/rope.metal

@@ -5,11 +5,12 @@
 #include "mlx/backend/metal/kernels/bf16.h"
 #include "mlx/backend/metal/kernels/utils.h"
 
-template <typename T, bool traditional>
+template <typename T, bool traditional, bool forward>
 [[kernel]] void rope(
     const device T *in [[buffer(0)]],
     device T * out [[buffer(1)]],
     constant const size_t strides[3],
+    constant const size_t out_strides[3],
     constant const int& offset,
     constant const float& base,
     constant const float& scale,
@@ -19,13 +20,13 @@ template <typename T, bool traditional>
   uint in_index_1, in_index_2;
   uint out_index_1, out_index_2;
   if (traditional) {
-    out_index_1 = 2 * (pos.x + grid.x * (pos.y + grid.y * pos.z));
+    out_index_1 = 2 * pos.x * out_strides[2] + pos.y * out_strides[1] + pos.z * out_strides[0];
     out_index_2 = out_index_1 + 1;
     in_index_1 = 2 * pos.x * strides[2] + pos.y * strides[1] + pos.z * strides[0];
     in_index_2 = in_index_1 + strides[2];
   } else {
-    out_index_1 = pos.x + 2*(grid.x * (pos.y + grid.y * pos.z));
-    out_index_2 = out_index_1 + grid.x;
+    out_index_1 = pos.x * out_strides[2] + pos.y * out_strides[1] + pos.z * out_strides[0];
+    out_index_2 = out_index_1 + grid.x * out_strides[2];
     in_index_1 = pos.x * strides[2] + pos.y * strides[1] + pos.z * strides[0];
     in_index_2 = in_index_1 + grid.x * strides[2];
   }
@@ -42,27 +43,41 @@ template <typename T, bool traditional>
   // Read and write the output
   float x1 = static_cast<float>(in[in_index_1]);
   float x2 = static_cast<float>(in[in_index_2]);
-  float rx1 = x1 * costheta - x2 * sintheta;
-  float rx2 = x1 * sintheta + x2 * costheta;
+  float rx1;
+  float rx2;
+  if (forward) {
+    rx1 = x1 * costheta - x2 * sintheta;
+    rx2 = x1 * sintheta + x2 * costheta;
+  } else {
+    rx1 = x2 * sintheta + x1 * costheta;
+    rx2 = x2 * costheta - x1 * sintheta;
+  }
   out[out_index_1] = static_cast<T>(rx1);
   out[out_index_2] = static_cast<T>(rx2);
 }
 
-#define instantiate_rope(name, type, traditional) \
+#define instantiate_rope(name, type, traditional, forward) \
   template [[host_name("rope_" #name)]] \
-  [[kernel]] void rope<type, traditional>( \
+  [[kernel]] void rope<type, traditional, forward>( \
       const device type* in [[buffer(0)]], \
       device type* out [[buffer(1)]], \
     constant const size_t strides[3], \
+    constant const size_t out_strides[3], \
     constant const int& offset, \
     constant const float& base, \
     constant const float& scale, \
     uint3 pos [[thread_position_in_grid]], \
     uint3 grid [[threads_per_grid]]);
 
-instantiate_rope(traditional_float16, half, true)
-instantiate_rope(traditional_bfloat16, bfloat16_t, true)
-instantiate_rope(traditional_float32, float, true)
-instantiate_rope(float16, half, false)
-instantiate_rope(bfloat16, bfloat16_t, false)
-instantiate_rope(float32, float, false)
+instantiate_rope(traditional_float16, half, true, true)
+instantiate_rope(traditional_bfloat16, bfloat16_t, true, true)
+instantiate_rope(traditional_float32, float, true, true)
+instantiate_rope(float16, half, false, true)
+instantiate_rope(bfloat16, bfloat16_t, false, true)
+instantiate_rope(float32, float, false, true)
+instantiate_rope(vjp_traditional_float16, half, true, false)
+instantiate_rope(vjp_traditional_bfloat16, bfloat16_t, true, false)
+instantiate_rope(vjp_traditional_float32, float, true, false)
+instantiate_rope(vjp_float16, half, false, false)
+instantiate_rope(vjp_bfloat16, bfloat16_t, false, false)
+instantiate_rope(vjp_float32, float, false, false)

mlx/backend/metal/kernels/scan.metal

@@ -451,7 +451,7 @@ instantiate_scan_helper(sum_int32_int32,         int32_t,     int32_t,     CumSu
 //instantiate_scan_helper(sum_int64_int64,         int64_t,     int64_t,     CumSum, 2)
 instantiate_scan_helper(sum_float16_float16,     half,        half,        CumSum, 4)
 instantiate_scan_helper(sum_float32_float32,     float,       float,       CumSum, 4)
-//instantiate_scan_helper(sum_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumSum, 4)
+instantiate_scan_helper(sum_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumSum, 4)
 //instantiate_scan_helper(sum_complex64_complex64, complex64_t, complex64_t, CumSum)
 //instantiate_scan_helper(prod_bool__bool_,         bool,        bool,        CumProd, 4)
 instantiate_scan_helper(prod_uint8_uint8,         uint8_t,     uint8_t,     CumProd, 4)
@@ -464,7 +464,7 @@ instantiate_scan_helper(prod_int32_int32,         int32_t,     int32_t,     CumP
 //instantiate_scan_helper(prod_int64_int64,         int64_t,     int64_t,     CumProd, 2)
 instantiate_scan_helper(prod_float16_float16,     half,        half,        CumProd, 4)
 instantiate_scan_helper(prod_float32_float32,     float,       float,       CumProd, 4)
-//instantiate_scan_helper(prod_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumProd, 4)
+instantiate_scan_helper(prod_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumProd, 4)
 //instantiate_scan_helper(prod_complex64_complex64, complex64_t, complex64_t, CumProd)
 //instantiate_scan_helper(max_bool__bool_,         bool,        bool,        CumMax, 4)
 instantiate_scan_helper(max_uint8_uint8,         uint8_t,     uint8_t,     CumMax, 4)
@@ -477,7 +477,7 @@ instantiate_scan_helper(max_int32_int32,         int32_t,     int32_t,     CumMa
 //instantiate_scan_helper(max_int64_int64,         int64_t,     int64_t,     CumMax, 2)
 instantiate_scan_helper(max_float16_float16,     half,        half,        CumMax, 4)
 instantiate_scan_helper(max_float32_float32,     float,       float,       CumMax, 4)
-//instantiate_scan_helper(max_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumMax, 4)
+instantiate_scan_helper(max_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumMax, 4)
 //instantiate_scan_helper(max_complex64_complex64, complex64_t, complex64_t, CumMax)
 //instantiate_scan_helper(min_bool__bool_,         bool,        bool,        CumMin, 4)
 instantiate_scan_helper(min_uint8_uint8,         uint8_t,     uint8_t,     CumMin, 4)
@@ -490,5 +490,5 @@ instantiate_scan_helper(min_int32_int32,         int32_t,     int32_t,     CumMi
 //instantiate_scan_helper(min_int64_int64,         int64_t,     int64_t,     CumMin, 2)
 instantiate_scan_helper(min_float16_float16,     half,        half,        CumMin, 4)
 instantiate_scan_helper(min_float32_float32,     float,       float,       CumMin, 4)
-//instantiate_scan_helper(min_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumMin, 4)
+instantiate_scan_helper(min_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumMin, 4)
 //instantiate_scan_helper(min_complex64_complex64, complex64_t, complex64_t, CumMin)

mlx/backend/metal/kernels/scatter.metal

@@ -20,7 +20,6 @@ METAL_FUNC void scatter_1d_index_impl(
   const constant int* out_shape [[buffer(3)]],
   const constant size_t* out_strides [[buffer(4)]],
   const constant size_t& upd_size [[buffer(5)]],
-  const constant bool& upd_col_contiguous [[buffer(6)]],
   const thread array<const device IdxT*, NIDX>& idx_buffers,
   uint2 gid [[thread_position_in_grid]]) {
 
@@ -33,11 +32,7 @@ METAL_FUNC void scatter_1d_index_impl(
     out_idx += idx_val * out_strides[i];
   }
 
-  if (!upd_col_contiguous) {
-    op.atomic_update(out, updates[gid.y * upd_size + gid.x], out_idx + gid.x);
-  } else {
-    op.atomic_update(out, updates[gid.x * upd_size + gid.y], out_idx + gid.x);
-  }
+  op.atomic_update(out, updates[gid.y * upd_size + gid.x], out_idx + gid.x);
 }
 
 #define make_scatter_1d_index(IDX_ARG, IDX_ARR) \
@@ -48,7 +43,6 @@ template <typename T, typename IdxT, typename Op, int NIDX> \
   const constant int* out_shape [[buffer(3)]], \
   const constant size_t* out_strides [[buffer(4)]], \
   const constant size_t& upd_size [[buffer(5)]], \
-  const constant bool& upd_col_contiguous [[buffer(6)]], \
   IDX_ARG(IdxT) \
   uint2 gid [[thread_position_in_grid]]) { \
   \
@@ -60,7 +54,6 @@ template <typename T, typename IdxT, typename Op, int NIDX> \
     out_shape, \
     out_strides, \
     upd_size, \
-    upd_col_contiguous, \
     idx_buffers, \
     gid); \
   \
@@ -195,7 +188,6 @@ template [[host_name("scatter_1d_index" name "_" #nidx)]] \
   const constant int* out_shape [[buffer(3)]], \
   const constant size_t* out_strides [[buffer(4)]], \
   const constant size_t& upd_size [[buffer(5)]], \
-  const constant bool& upd_col_contiguous [[buffer(6)]], \
   IDX_ARG(idx_t) \
   uint2 gid [[thread_position_in_grid]]);
 

mlx/backend/metal/kernels/softmax.metal

@@ -1,6 +1,5 @@
 // Copyright © 2023 Apple Inc.
 
-#include <metal_atomic>
 #include <metal_common>
 #include <metal_simdgroup>
 
@@ -12,46 +11,48 @@ using namespace metal;
 
 template <typename T>
 inline T softmax_exp(T x) {
-  // Softmax doesn't need high precision exponential cause it is gonna be x
-  // will be in (-oo, 0] anyway and subsequently it will be divided by
-  // sum(exp(x_i)).
+  // Softmax doesn't need high precision exponential cause x is gonna be in
+  // (-oo, 0] anyway and subsequently it will be divided by sum(exp(x_i)).
   return fast::exp(x);
 }
 
-template <typename T, int N_READS = SOFTMAX_N_READS>
+template <typename T, typename AccT = T, int N_READS = SOFTMAX_N_READS>
 [[kernel]] void softmax_single_row(
     const device T* in,
     device T* out,
     constant int& axis_size,
-    threadgroup T* local_max [[threadgroup(0)]],
-    threadgroup T* local_normalizer [[threadgroup(1)]],
     uint gid [[threadgroup_position_in_grid]],
     uint _lid [[thread_position_in_threadgroup]],
     uint simd_lane_id [[thread_index_in_simdgroup]],
     uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
   int lid = _lid;
 
-  T ld[N_READS];
+  constexpr int SIMD_SIZE = 32;
+
+  threadgroup AccT local_max[SIMD_SIZE];
+  threadgroup AccT local_normalizer[SIMD_SIZE];
+
+  AccT ld[N_READS];
 
   in += gid * axis_size + lid * N_READS;
   if (lid * N_READS + N_READS <= axis_size) {
-    for (int i=0; i<N_READS; i++) {
-        ld[i] = in[i];
+    for (int i = 0; i < N_READS; i++) {
+      ld[i] = AccT(in[i]);
     }
   } else {
-      for (int i = 0; i < N_READS; i++) {
-        ld[i] =
-            ((lid * N_READS + i) < axis_size) ? in[i] : T(Limits<T>::finite_min);
-      }
+    for (int i = 0; i < N_READS; i++) {
+      ld[i] = ((lid * N_READS + i) < axis_size) ? AccT(in[i])
+                                                : Limits<AccT>::finite_min;
+    }
   }
   if (simd_group_id == 0) {
-    local_max[simd_lane_id] = Limits<T>::finite_min;
+    local_max[simd_lane_id] = Limits<AccT>::finite_min;
     local_normalizer[simd_lane_id] = 0;
   }
   threadgroup_barrier(mem_flags::mem_threadgroup);
 
   // Get the max
-  T maxval = Limits<T>::finite_min;
+  AccT maxval = Limits<AccT>::finite_min;
   for (int i = 0; i < N_READS; i++) {
     maxval = (maxval < ld[i]) ? ld[i] : maxval;
   }
@@ -70,9 +71,9 @@ template <typename T, int N_READS = SOFTMAX_N_READS>
   maxval = local_max[0];
 
   // Compute exp(x_i - maxval) and store the partial sums in local_normalizer
-  T normalizer = 0;
+  AccT normalizer = 0;
   for (int i = 0; i < N_READS; i++) {
-    T exp_x = softmax_exp(ld[i] - maxval);
+    AccT exp_x = softmax_exp(ld[i] - maxval);
     ld[i] = exp_x;
     normalizer += exp_x;
   }
@@ -93,25 +94,23 @@ template <typename T, int N_READS = SOFTMAX_N_READS>
   // Normalize and write to the output
   out += gid * axis_size + lid * N_READS;
   if (lid * N_READS + N_READS <= axis_size) {
-    for (int i=0; i<N_READS; i++) {
-        out[i] = ld[i] * normalizer;
+    for (int i = 0; i < N_READS; i++) {
+      out[i] = T(ld[i] * normalizer);
     }
   } else {
-      for (int i = 0; i < N_READS; i++) {
-        if ((lid * N_READS + i) < axis_size) {
-          out[i] = ld[i] * normalizer;
-        }
+    for (int i = 0; i < N_READS; i++) {
+      if ((lid * N_READS + i) < axis_size) {
+        out[i] = T(ld[i] * normalizer);
       }
+    }
   }
 }
 
-template <typename T, int N_READS = SOFTMAX_N_READS>
+template <typename T, typename AccT = T, int N_READS = SOFTMAX_N_READS>
 [[kernel]] void softmax_looped(
     const device T* in,
     device T* out,
     constant int& axis_size,
-    threadgroup T* local_max [[threadgroup(0)]],
-    threadgroup T* local_normalizer [[threadgroup(1)]],
     uint gid [[threadgroup_position_in_grid]],
     uint lid [[thread_position_in_threadgroup]],
     uint lsize [[threads_per_threadgroup]],
@@ -119,22 +118,27 @@ template <typename T, int N_READS = SOFTMAX_N_READS>
     uint simd_group_id [[simdgroup_index_in_threadgroup]]) {
   in += gid * axis_size;
 
+  constexpr int SIMD_SIZE = 32;
+
+  threadgroup AccT local_max[SIMD_SIZE];
+  threadgroup AccT local_normalizer[SIMD_SIZE];
+
   // Get the max and the normalizer in one go
-  T prevmax;
-  T maxval = Limits<T>::finite_min;
-  T normalizer = 0;
+  AccT prevmax;
+  AccT maxval = Limits<AccT>::finite_min;
+  AccT normalizer = 0;
   for (int r = 0; r < static_cast<int>(ceildiv(axis_size, N_READS * lsize));
        r++) {
     int offset = r * lsize * N_READS + lid * N_READS;
-    T vals[N_READS];
+    AccT vals[N_READS];
     if (offset + N_READS <= axis_size) {
       for (int i = 0; i < N_READS; i++) {
-        vals[i] = in[offset + i];
+        vals[i] = AccT(in[offset + i]);
       }
     } else {
       for (int i = 0; i < N_READS; i++) {
-        vals[i] =
-            (offset + i < axis_size) ? in[offset + i] : T(Limits<T>::finite_min);
+        vals[i] = (offset + i < axis_size) ? AccT(in[offset + i])
+                                           : Limits<AccT>::finite_min;
       }
     }
     prevmax = maxval;
@@ -180,49 +184,66 @@ template <typename T, int N_READS = SOFTMAX_N_READS>
        r++) {
     int offset = r * lsize * N_READS + lid * N_READS;
     if (offset + N_READS <= axis_size) {
-      for (int i=0; i<N_READS; i++) {
-        out[offset + i] = softmax_exp(in[offset + i] - maxval) * normalizer;
+      for (int i = 0; i < N_READS; i++) {
+        out[offset + i] = T(softmax_exp(in[offset + i] - maxval) * normalizer);
       }
     } else {
       for (int i = 0; i < N_READS; i++) {
         if (offset + i < axis_size) {
-          out[offset + i] = softmax_exp(in[offset + i] - maxval) * normalizer;
+          out[offset + i] =
+              T(softmax_exp(in[offset + i] - maxval) * normalizer);
         }
       }
     }
   }
 }
 
-#define instantiate_softmax_single_row(name, itype)           \
+// clang-format off
+#define instantiate_softmax(name, itype)  \
   template [[host_name("softmax_" #name)]] [[kernel]] void    \
   softmax_single_row<itype>(                                  \
       const device itype* in,                                 \
       device itype* out,                                      \
       constant int& axis_size,                                \
-      threadgroup itype* local_max [[threadgroup(0)]],        \
-      threadgroup itype* local_normalizer [[threadgroup(1)]], \
       uint gid [[thread_position_in_grid]],                   \
       uint _lid [[thread_position_in_threadgroup]],           \
       uint simd_lane_id [[thread_index_in_simdgroup]],        \
-      uint simd_group_id [[simdgroup_index_in_threadgroup]]);
-
-#define instantiate_softmax_looped(name, itype)                   \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]); \
   template [[host_name("softmax_looped_" #name)]] [[kernel]] void \
   softmax_looped<itype>(                                          \
       const device itype* in,                                     \
       device itype* out,                                          \
       constant int& axis_size,                                    \
-      threadgroup itype* local_max [[threadgroup(0)]],            \
-      threadgroup itype* local_normalizer [[threadgroup(1)]],     \
       uint gid [[threadgroup_position_in_grid]],                  \
       uint lid [[thread_position_in_threadgroup]],                \
       uint lsize [[threads_per_threadgroup]],                     \
       uint simd_lane_id [[thread_index_in_simdgroup]],            \
       uint simd_group_id [[simdgroup_index_in_threadgroup]]);
 
-#define instantiate_softmax(name, itype)      \
-  instantiate_softmax_single_row(name, itype) \
-      instantiate_softmax_looped(name, itype)
+#define instantiate_softmax_precise(name, itype)                   \
+  template [[host_name("softmax_precise_" #name)]] [[kernel]] void \
+  softmax_single_row<itype, float>(                                \
+      const device itype* in,                                      \
+      device itype* out,                                           \
+      constant int& axis_size,                                     \
+      uint gid [[thread_position_in_grid]],                        \
+      uint _lid [[thread_position_in_threadgroup]],                \
+      uint simd_lane_id [[thread_index_in_simdgroup]],             \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);      \
+  template [[host_name("softmax_looped_precise_" #name)]] [[kernel]] void \
+  softmax_looped<itype, float>(                                           \
+      const device itype* in,                                             \
+      device itype* out,                                                  \
+      constant int& axis_size,                                            \
+      uint gid [[threadgroup_position_in_grid]],                          \
+      uint lid [[thread_position_in_threadgroup]],                        \
+      uint lsize [[threads_per_threadgroup]],                             \
+      uint simd_lane_id [[thread_index_in_simdgroup]],                    \
+      uint simd_group_id [[simdgroup_index_in_threadgroup]]);
 
-instantiate_softmax(float32, float) instantiate_softmax(float16, half)
-    instantiate_softmax(bfloat16, bfloat16_t)
+instantiate_softmax(float32, float)
+instantiate_softmax(float16, half)
+instantiate_softmax(bfloat16, bfloat16_t)
+instantiate_softmax_precise(float16, half)
+instantiate_softmax_precise(bfloat16, bfloat16_t)
+// clang-format on

mlx/backend/metal/kernels/steel/conv/loaders/loader_channel_l.h

@@ -394,7 +394,7 @@ struct Conv2DWeightBlockLoader {
       const constant ImplicitGemmConv2DParams* gemm_params_,
       uint simd_group_id [[simdgroup_index_in_threadgroup]],
       uint simd_lane_id [[thread_index_in_simdgroup]])
-      : src_ld(params_->wt_strides[0]),
+      : src_ld(params_ -> wt_strides[0]),
         thread_idx(simd_group_id * 32 + simd_lane_id),
         bi(thread_idx / TCOLS),
         bj(vec_size * (thread_idx % TCOLS)),

mlx/backend/metal/kernels/steel/conv/loaders/loader_channel_n.h

@@ -244,7 +244,7 @@ struct Conv2DWeightBlockLoaderSmallChannels {
       const constant ImplicitGemmConv2DParams* gemm_params_,
       uint simd_group_id [[simdgroup_index_in_threadgroup]],
       uint simd_lane_id [[thread_index_in_simdgroup]])
-      : src_ld(params_->wt_strides[0]),
+      : src_ld(params_ -> wt_strides[0]),
         thread_idx(simd_group_id * 32 + simd_lane_id),
         bi(thread_idx / TCOLS),
         bj(vec_size * (thread_idx % TCOLS)),

mlx/backend/metal/kernels/steel/conv/loaders/loader_general.h

@@ -220,7 +220,7 @@ struct Conv2DWeightBlockLoaderGeneral {
       const short base_ww_,
       uint simd_group_id [[simdgroup_index_in_threadgroup]],
       uint simd_lane_id [[thread_index_in_simdgroup]])
-      : src_ld(params_->wt_strides[0]),
+      : src_ld(params_ -> wt_strides[0]),
         thread_idx(simd_group_id * 32 + simd_lane_id),
         bi(thread_idx / TCOLS),
         bj(vec_size * (thread_idx % TCOLS)),

mlx/backend/metal/kernels/steel/gemm/gemm.h

@@ -140,7 +140,7 @@ struct GEMMKernel {
   static METAL_FUNC void run(
       const device T* A [[buffer(0)]],
       const device T* B [[buffer(1)]],
-      device U* C [[buffer(2)]],
+      device U* D [[buffer(2)]],
       const constant GEMMParams* params [[buffer(3)]],
       threadgroup T* As [[threadgroup(0)]],
       threadgroup T* Bs [[threadgroup(1)]],
@@ -167,7 +167,7 @@ struct GEMMKernel {
 
     A += transpose_a ? c_row : c_row * params->lda;
     B += transpose_b ? c_col * params->ldb : c_col;
-    C += c_row * params->ldc + c_col;
+    D += c_row * params->ldd + c_col;
 
     // Prepare threadgroup loading operations
     thread loader_a_t loader_a(A, params->lda, As, simd_group_id, simd_lane_id);
@@ -214,7 +214,7 @@ struct GEMMKernel {
       }
 
       // Store results to device memory
-      mma_op.store_result(C, params->ldc);
+      mma_op.store_result(D, params->ldd);
       return;
 
     }
@@ -237,7 +237,7 @@ struct GEMMKernel {
             tgp_bn,
             leftover_bk);
 
-        mma_op.store_result(C, params->ldc);
+        mma_op.store_result(D, params->ldd);
         return;
 
       } else if (tgp_bn == BN) {
@@ -252,7 +252,7 @@ struct GEMMKernel {
             tgp_bn,
             leftover_bk);
 
-        mma_op.store_result_safe(C, params->ldc, short2(tgp_bn, tgp_bm));
+        mma_op.store_result_safe(D, params->ldd, short2(tgp_bn, tgp_bm));
         return;
 
       } else if (tgp_bm == BM) {
@@ -267,7 +267,7 @@ struct GEMMKernel {
             tgp_bn,
             leftover_bk);
 
-        mma_op.store_result_safe(C, params->ldc, short2(tgp_bn, tgp_bm));
+        mma_op.store_result_safe(D, params->ldd, short2(tgp_bn, tgp_bm));
         return;
 
       } else {
@@ -282,7 +282,7 @@ struct GEMMKernel {
             tgp_bn,
             leftover_bk);
 
-        mma_op.store_result_safe(C, params->ldc, short2(tgp_bn, tgp_bm));
+        mma_op.store_result_safe(D, params->ldd, short2(tgp_bn, tgp_bm));
         return;
       }
     }

mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm.metal

@@ -1,6 +1,7 @@
 // Copyright © 2024 Apple Inc.
 
 #include "mlx/backend/metal/kernels/bf16.h"
+#include "mlx/backend/metal/kernels/utils.h"
 #include "mlx/backend/metal/kernels/steel/gemm/gemm.h"
 
 using namespace metal;
@@ -23,8 +24,10 @@ template <typename T,
 [[kernel, max_total_threads_per_threadgroup(WM * WN * 32)]] void gemm(
     const device T *A [[buffer(0)]],
     const device T *B [[buffer(1)]],
-    device T *C [[buffer(2)]],
-    const constant GEMMParams* params [[buffer(3)]],
+    device T *D [[buffer(3)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    const constant int* batch_shape [[buffer(6)]],
+    const constant size_t* batch_strides [[buffer(7)]],
     uint simd_lane_id [[thread_index_in_simdgroup]],
     uint simd_group_id [[simdgroup_index_in_threadgroup]],
     uint3 tid [[threadgroup_position_in_grid]],
@@ -36,12 +39,25 @@ template <typename T,
     threadgroup T Bs[gemm_kernel::tgp_mem_size_b];
 
     // Adjust for batch
-    A += params->batch_stride_a * tid.z;
-    B += params->batch_stride_b * tid.z;
-    C += params->batch_stride_c * tid.z;
+    if(params->batch_ndim > 1) {
+      const constant size_t* A_bstrides = batch_strides;
+      const constant size_t* B_bstrides = batch_strides + params->batch_ndim;
+
+      ulong2 batch_offsets = elem_to_loc_broadcast(
+          tid.z, batch_shape, A_bstrides, B_bstrides, params->batch_ndim);
+
+      A += batch_offsets.x;
+      B += batch_offsets.y;
+      
+    } else {
+      A += params->batch_stride_a * tid.z;
+      B += params->batch_stride_b * tid.z;
+    }
+    
+    D += params->batch_stride_d * tid.z;
 
     gemm_kernel::run( 
-      A, B, C, 
+      A, B, D, 
       params,
       As, Bs,
       simd_lane_id, simd_group_id, tid, lid
@@ -57,8 +73,10 @@ template <typename T,
   [[kernel]] void gemm<itype, bm, bn, bk, wm, wn, trans_a, trans_b, mn_aligned, k_aligned>( \
       const device itype *A [[buffer(0)]], \
       const device itype *B [[buffer(1)]], \
-      device itype *C [[buffer(2)]], \
-      const constant GEMMParams* params [[buffer(3)]], \
+      device itype *D [[buffer(3)]], \
+      const constant GEMMParams* params [[buffer(4)]], \
+      const constant int* batch_shape [[buffer(6)]], \
+      const constant size_t* batch_strides [[buffer(7)]], \
       uint simd_lane_id [[thread_index_in_simdgroup]], \
       uint simd_group_id [[simdgroup_index_in_threadgroup]], \
       uint3 tid [[threadgroup_position_in_grid]], \

mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_addmm.metal

@@ -27,7 +27,10 @@ template <typename T,
     const device T *B [[buffer(1)]],
     const device T *C [[buffer(2)]],
     device T *D [[buffer(3)]],
-    const constant GEMMAddMMParams* params [[buffer(4)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    const constant GEMMAddMMParams* addmm_params [[buffer(5)]],
+    const constant int* batch_shape [[buffer(6)]],
+    const constant size_t* batch_strides [[buffer(7)]],
     uint simd_lane_id [[thread_index_in_simdgroup]],
     uint simd_group_id [[simdgroup_index_in_threadgroup]],
     uint3 tid [[threadgroup_position_in_grid]],
@@ -50,9 +53,24 @@ template <typename T,
     threadgroup T Bs[gemm_kernel::tgp_mem_size_b];
 
     // Adjust for batch
-    A += params->batch_stride_a * tid.z;
-    B += params->batch_stride_b * tid.z;
-    C += params->batch_stride_c * tid.z;
+    if(params->batch_ndim > 1) {
+      const constant size_t* A_bstrides = batch_strides;
+      const constant size_t* B_bstrides = batch_strides + params->batch_ndim;
+      const constant size_t* C_bstrides = B_bstrides + params->batch_ndim;
+
+      ulong3 batch_offsets = elem_to_loc_broadcast(
+          tid.z, batch_shape, A_bstrides, B_bstrides, C_bstrides, params->batch_ndim);
+
+      A += batch_offsets.x;
+      B += batch_offsets.y;
+      C += batch_offsets.z;
+      
+    } else {
+      A += params->batch_stride_a * tid.z;
+      B += params->batch_stride_b * tid.z;
+      C += addmm_params->batch_stride_c * tid.z;
+    }
+
     D += params->batch_stride_d * tid.z;
 
     const int tid_y = ((tid.y) << params->swizzle_log) +
@@ -71,9 +89,10 @@ template <typename T,
 
     A += transpose_a ? c_row : c_row * params->lda;
     B += transpose_b ? c_col * params->ldb : c_col;
-    C += c_row * params->ldc + c_col * params->fdc;
     D += c_row * params->ldd + c_col;
 
+    C += c_row * addmm_params->ldc + c_col * addmm_params->fdc;
+
     // Prepare threadgroup loading operations
     thread loader_a_t loader_a(A, params->lda, As, simd_group_id, simd_lane_id);
     thread loader_b_t loader_b(B, params->ldb, Bs, simd_group_id, simd_lane_id);
@@ -83,7 +102,7 @@ template <typename T,
 
     int gemm_k_iterations = params->gemm_k_iterations_aligned;
 
-    const Epilogue epilogue_op(params->alpha, params->beta);
+    const Epilogue epilogue_op(addmm_params->alpha, addmm_params->beta);
 
     ///////////////////////////////////////////////////////////////////////////////
     // MNK aligned loop
@@ -121,7 +140,7 @@ template <typename T,
       }
 
       // Store results to device memory
-      mma_op.store_result(D, params->ldd, C, params->ldc, params->fdc, epilogue_op);
+      mma_op.store_result(D, params->ldd, C, addmm_params->ldc, addmm_params->fdc, epilogue_op);
       return;
 
     }
@@ -145,7 +164,7 @@ template <typename T,
             leftover_bk,
             LoopAlignment<true, true, K_aligned>{});
 
-        mma_op.store_result(D, params->ldd, C, params->ldc, params->fdc, epilogue_op);
+        mma_op.store_result(D, params->ldd, C, addmm_params->ldc, addmm_params->fdc, epilogue_op);
         return;
 
       } else if (tgp_bn == BN) {
@@ -163,7 +182,7 @@ template <typename T,
 
         return mma_op.store_result_safe(
             D, params->ldd, 
-            C, params->ldc, params->fdc,
+            C, addmm_params->ldc, addmm_params->fdc,
             short2(tgp_bn, tgp_bm), 
             epilogue_op);
 
@@ -182,7 +201,7 @@ template <typename T,
 
         return mma_op.store_result_safe(
             D, params->ldd, 
-            C, params->ldc, params->fdc,
+            C, addmm_params->ldc, addmm_params->fdc,
             short2(tgp_bn, tgp_bm), 
             epilogue_op);
 
@@ -201,7 +220,7 @@ template <typename T,
 
         return mma_op.store_result_safe(
             D, params->ldd, 
-            C, params->ldc, params->fdc,
+            C, addmm_params->ldc, addmm_params->fdc,
             short2(tgp_bn, tgp_bm), 
             epilogue_op);
       }
@@ -219,7 +238,10 @@ template <typename T,
       const device itype *B [[buffer(1)]], \
       const device itype *C [[buffer(2)]], \
       device itype *D [[buffer(3)]], \
-      const constant GEMMAddMMParams* params [[buffer(4)]], \
+      const constant GEMMParams* gemm_params [[buffer(4)]], \
+      const constant GEMMAddMMParams* params [[buffer(5)]], \
+      const constant int* batch_shape [[buffer(6)]], \
+      const constant size_t* batch_strides [[buffer(7)]], \
       uint simd_lane_id [[thread_index_in_simdgroup]], \
       uint simd_group_id [[simdgroup_index_in_threadgroup]], \
       uint3 tid [[threadgroup_position_in_grid]], \

mlx/backend/metal/kernels/steel/gemm/mma.h

@@ -144,9 +144,9 @@ struct BlockMMA {
   }
 
   /* Store results from simdgroup_matrix results into device memory */
-  METAL_FUNC void store_result(device U* C, const int ldc) const {
+  METAL_FUNC void store_result(device U* D, const int ldd) const {
     // Adjust for simdgroup and thread location
-    C += (sm + tm) * ldc + tn + sn;
+    D += (sm + tm) * ldd + tn + sn;
 
     // Loop over all simdgroup tiles
     STEEL_PRAGMA_UNROLL
@@ -155,22 +155,22 @@ struct BlockMMA {
       for (short j = 0; j < TN; j++) {
         // Get accumulated result and associated offset in C
         thread const auto& accum = results[i * TN + j].thread_elements();
-        int offset = (i * TM_stride) * ldc + (j * TN_stride);
+        int offset = (i * TM_stride) * ldd + (j * TN_stride);
 
         // Apply epilogue
         U outs[2] = {Epilogue::apply(accum[0]), Epilogue::apply(accum[1])};
 
-        // Write out C
-        C[offset] = outs[0];
-        C[offset + 1] = outs[1];
+        // Write out D
+        D[offset] = outs[0];
+        D[offset + 1] = outs[1];
       }
     }
   }
 
   METAL_FUNC void
-  store_result_safe(device U* C, const int ldc, short2 dst_tile_dims) const {
+  store_result_safe(device U* D, const int ldd, short2 dst_tile_dims) const {
     // Adjust for simdgroup and thread location
-    C += (sm + tm) * ldc + (tn + sn);
+    D += (sm + tm) * ldd + (tn + sn);
     dst_tile_dims -= short2(tn + sn, sm + tm);
 
     if (dst_tile_dims.x <= 0 || dst_tile_dims.y <= 0)
@@ -183,15 +183,15 @@ struct BlockMMA {
         for (int j = 0; j < TN; j++) {
           // Get accumulated result and associated offset in C
           thread const auto& accum = results[i * TN + j].thread_elements();
-          int offset = (i * TM_stride) * ldc + (j * TN_stride);
+          int offset = (i * TM_stride) * ldd + (j * TN_stride);
 
           // Apply epilogue and output C
           if (j * TN_stride < dst_tile_dims.x) {
-            C[offset] = Epilogue::apply(accum[0]);
+            D[offset] = Epilogue::apply(accum[0]);
           }
 
           if (j * TN_stride + 1 < dst_tile_dims.x) {
-            C[offset + 1] = Epilogue::apply(accum[1]);
+            D[offset + 1] = Epilogue::apply(accum[1]);
           }
         }
       }

mlx/backend/metal/kernels/steel/gemm/params.h

@@ -16,17 +16,19 @@ struct GEMMParams {
 
   const int lda;
   const int ldb;
-  const int ldc;
+  const int ldd;
 
   const int tiles_n;
   const int tiles_m;
 
   const int batch_stride_a;
   const int batch_stride_b;
-  const int batch_stride_c;
+  const int batch_stride_d;
 
   const int swizzle_log;
   const int gemm_k_iterations_aligned;
+
+  const int batch_ndim;
 };
 
 struct GEMMSpiltKParams {
@@ -49,30 +51,13 @@ struct GEMMSpiltKParams {
 };
 
 struct GEMMAddMMParams {
-  const int M;
-  const int N;
-  const int K;
-
-  const int lda;
-  const int ldb;
   const int ldc;
-  const int ldd;
+  const int fdc;
 
-  const int tiles_n;
-  const int tiles_m;
-
-  const int batch_stride_a;
-  const int batch_stride_b;
   const int batch_stride_c;
-  const int batch_stride_d;
-
-  const int swizzle_log;
-  const int gemm_k_iterations_aligned;
 
   const float alpha;
   const float beta;
-
-  const int fdc;
 };
 
 } // namespace steel

mlx/backend/metal/kernels/steel/utils.h

@@ -5,4 +5,41 @@
 #include <metal_stdlib>
 
 #define STEEL_CONST static constant constexpr const
-#define STEEL_PRAGMA_UNROLL _Pragma("clang loop unroll(full)")
+#define STEEL_PRAGMA_UNROLL _Pragma("clang loop unroll(full)")
+
+METAL_FUNC ulong2 elem_to_loc_broadcast(
+    uint elem,
+    constant const int* shape,
+    constant const size_t* a_strides,
+    constant const size_t* b_strides,
+    int ndim) {
+  ulong loc_a{0};
+  ulong loc_b{0};
+  for (int i = ndim - 1; i >= 0 && elem > 0; --i) {
+    int pos_in_dim = (elem % shape[i]);
+    elem /= shape[i];
+    loc_a += pos_in_dim * a_strides[i];
+    loc_b += pos_in_dim * b_strides[i];
+  }
+  return ulong2(loc_a, loc_b);
+}
+
+METAL_FUNC ulong3 elem_to_loc_broadcast(
+    uint elem,
+    constant const int* shape,
+    constant const size_t* a_strides,
+    constant const size_t* b_strides,
+    constant const size_t* c_strides,
+    int ndim) {
+  ulong loc_a{0};
+  ulong loc_b{0};
+  ulong loc_c{0};
+  for (int i = ndim - 1; i >= 0 && elem > 0; --i) {
+    int pos_in_dim = (elem % shape[i]);
+    elem /= shape[i];
+    loc_a += pos_in_dim * a_strides[i];
+    loc_b += pos_in_dim * b_strides[i];
+    loc_c += pos_in_dim * c_strides[i];
+  }
+  return ulong3(loc_a, loc_b, loc_c);
+}

mlx/backend/metal/kernels/unary.h

@@ -7,6 +7,7 @@
 
 #include "mlx/backend/metal/kernels/bf16.h"
 #include "mlx/backend/metal/kernels/erf.h"
+#include "mlx/backend/metal/kernels/expm1f.h"
 #include "mlx/backend/metal/kernels/utils.h"
 
 namespace {
@@ -183,6 +184,13 @@ struct Exp {
   }
 };
 
+struct Expm1 {
+  template <typename T>
+  T operator()(T x) {
+    return static_cast<T>(expm1f(static_cast<float>(x)));
+  };
+};
+
 struct Floor {
   template <typename T>
   T operator()(T x) {

mlx/backend/metal/kernels/unary.metal

@@ -71,6 +71,7 @@ instantiate_unary_types(ceil, Ceil)
 instantiate_unary_float(cos, Cos)
 instantiate_unary_float(cosh, Cosh)
 instantiate_unary_float(exp, Exp)
+instantiate_unary_float(expm1, Expm1)
 instantiate_unary_types(floor, Floor)
 instantiate_unary_float(log, Log)
 instantiate_unary_float(log2, Log2)

mlx/backend/metal/kernels/utils.h

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2024 Apple Inc.
 
 #pragma once
 
@@ -65,12 +65,18 @@ struct Limits<bool> {
 // Indexing utils
 ///////////////////////////////////////////////////////////////////////////////
 
-inline size_t elem_to_loc(
+#define MLX_MTL_PRAGMA_UNROLL _Pragma("clang loop unroll(full)")
+
+///////////////////////////////////////////////////////////////////////////////
+// Single Array with generic dims
+
+template <typename stride_t>
+METAL_FUNC stride_t elem_to_loc(
     uint elem,
     device const int* shape,
-    device const size_t* strides,
+    device const stride_t* strides,
     int ndim) {
-  size_t loc = 0;
+  stride_t loc = 0;
   for (int i = ndim - 1; i >= 0 && elem > 0; --i) {
     loc += (elem % shape[i]) * strides[i];
     elem /= shape[i];
@@ -78,12 +84,13 @@ inline size_t elem_to_loc(
   return loc;
 }
 
-inline size_t elem_to_loc(
+template <typename stride_t>
+METAL_FUNC stride_t elem_to_loc(
     uint elem,
     constant const int* shape,
-    constant const size_t* strides,
+    constant const stride_t* strides,
     int ndim) {
-  size_t loc = 0;
+  stride_t loc = 0;
   for (int i = ndim - 1; i >= 0 && elem > 0; --i) {
     loc += (elem % shape[i]) * strides[i];
     elem /= shape[i];
@@ -91,52 +98,59 @@ inline size_t elem_to_loc(
   return loc;
 }
 
-template <int NDIM>
-inline uint3 elem_to_loc_3_nd(
+// Non templated version to handle arbitrary dims
+template <typename stride_t>
+METAL_FUNC stride_t elem_to_loc(
     uint3 elem,
-    constant const int shape[NDIM],
-    constant const size_t a_strides[NDIM],
-    constant const size_t b_strides[NDIM],
-    constant const size_t c_strides[NDIM]) {
-  uint3 loc = {
-      static_cast<uint>(
-          elem.x * a_strides[NDIM - 1] + elem.y * a_strides[NDIM - 2]),
-      static_cast<uint>(
-          elem.x * b_strides[NDIM - 1] + elem.y * b_strides[NDIM - 2]),
-      static_cast<uint>(
-          elem.x * c_strides[NDIM - 1] + elem.y * c_strides[NDIM - 2])};
-  for (int d = NDIM - 3; d >= 0; --d) {
-    uint l = elem.z % shape[d];
-    loc.x += l * a_strides[d];
-    loc.y += l * b_strides[d];
-    loc.z += l * c_strides[d];
+    constant const int* shape,
+    constant const stride_t* strides,
+    int ndim) {
+  stride_t loc = elem.x * strides[ndim - 1] + elem.y * strides[ndim - 2];
+  for (int d = ndim - 3; d >= 0; --d) {
+    loc += (elem.z % shape[d]) * strides[d];
     elem.z /= shape[d];
   }
   return loc;
 }
 
+///////////////////////////////////////////////////////////////////////////////
+// Single Array with fixed N dims
+
+template <typename stride_t>
+METAL_FUNC stride_t elem_to_loc_1(uint elem, constant const stride_t& stride) {
+  return elem * stride;
+}
+
+template <typename stride_t>
+METAL_FUNC stride_t
+elem_to_loc_2(uint2 elem, constant const stride_t strides[2]) {
+  return elem.x * strides[1] + elem.y * strides[0];
+}
+
+template <typename stride_t>
+METAL_FUNC stride_t
+elem_to_loc_3(uint3 elem, constant const stride_t strides[3]) {
+  return elem.x * strides[2] + elem.y * strides[1] + elem.z * strides[0];
+}
+
 template <int NDIM>
-inline uint2 elem_to_loc_2_nd(
-    uint3 elem,
-    constant const int shape[NDIM],
-    constant const size_t a_strides[NDIM],
-    constant const size_t b_strides[NDIM]) {
-  uint2 loc = {
-      static_cast<uint>(
-          elem.x * a_strides[NDIM - 1] + elem.y * a_strides[NDIM - 2]),
-      static_cast<uint>(
-          elem.x * b_strides[NDIM - 1] + elem.y * b_strides[NDIM - 2])};
-  for (int d = NDIM - 3; d >= 0; --d) {
-    uint l = elem.z % shape[d];
-    loc.x += l * a_strides[d];
-    loc.y += l * b_strides[d];
-    elem.z /= shape[d];
+METAL_FUNC size_t elem_to_loc_nd(
+    uint elem,
+    device const int* shape,
+    device const size_t* strides) {
+  size_t loc = (elem % shape[NDIM - 1]) * strides[NDIM - 1];
+
+  MLX_MTL_PRAGMA_UNROLL
+  for (int d = NDIM - 2; d >= 0; --d) {
+    elem /= shape[d + 1];
+    loc += (elem % shape[d]) * strides[d];
   }
+
   return loc;
 }
 
 template <int NDIM>
-inline size_t elem_to_loc_nd(
+METAL_FUNC size_t elem_to_loc_nd(
     uint3 elem,
     constant const int shape[NDIM],
     constant const size_t strides[NDIM]) {
@@ -148,33 +162,59 @@ inline size_t elem_to_loc_nd(
   return loc;
 }
 
-inline size_t elem_to_loc_1(uint elem, constant const size_t& stride) {
-  return elem * stride;
+template <int NDIM>
+METAL_FUNC int64_t elem_to_loc_nd(
+    uint elem,
+    constant const int shape[NDIM],
+    constant const int64_t strides[NDIM]) {
+  int64_t loc = (elem % shape[NDIM - 1]) * strides[NDIM - 1];
+
+  MLX_MTL_PRAGMA_UNROLL
+  for (int d = NDIM - 2; d >= 0; --d) {
+    elem /= shape[d + 1];
+    loc += (elem % shape[d]) * strides[d];
+  }
+
+  return loc;
 }
 
-inline size_t elem_to_loc_2(uint2 elem, constant const size_t strides[2]) {
-  return elem.x * strides[1] + elem.y * strides[0];
-}
-
-inline size_t elem_to_loc_3(uint3 elem, constant const size_t strides[3]) {
-  return elem.x * strides[2] + elem.y * strides[1] + elem.z * strides[0];
-}
-
-// Non templated version to handle arbitrary dims
-inline size_t elem_to_loc(
+template <int NDIM>
+METAL_FUNC int64_t elem_to_loc_nd(
     uint3 elem,
-    constant const int* shape,
-    constant const size_t* strides,
-    int ndim) {
-  size_t loc = elem.x * strides[ndim - 1] + elem.y * strides[ndim - 2];
-  for (int d = ndim - 3; d >= 0; --d) {
+    constant const int shape[NDIM],
+    constant const int64_t strides[NDIM]) {
+  int64_t loc = elem.x * strides[NDIM - 1] + elem.y * strides[NDIM - 2];
+  for (int d = NDIM - 3; d >= 0; --d) {
     loc += (elem.z % shape[d]) * strides[d];
     elem.z /= shape[d];
   }
   return loc;
 }
 
-inline uint3 elem_to_loc_3_nd(
+///////////////////////////////////////////////////////////////////////////////
+// Multiple Arrays with generic dims
+
+METAL_FUNC uint2 elem_to_loc_2_nd(
+    uint3 elem,
+    constant const int* shape,
+    constant const size_t* a_strides,
+    constant const size_t* b_strides,
+    int ndim) {
+  uint2 loc = {
+      static_cast<uint>(
+          elem.x * a_strides[ndim - 1] + elem.y * a_strides[ndim - 2]),
+      static_cast<uint>(
+          elem.x * b_strides[ndim - 1] + elem.y * b_strides[ndim - 2])};
+  for (int d = ndim - 3; d >= 0; --d) {
+    uint l = elem.z % shape[d];
+    loc.x += l * a_strides[d];
+    loc.y += l * b_strides[d];
+    elem.z /= shape[d];
+  }
+  return loc;
+}
+
+METAL_FUNC uint3 elem_to_loc_3_nd(
     uint3 elem,
     constant const int* shape,
     constant const size_t* a_strides,
@@ -198,18 +238,21 @@ inline uint3 elem_to_loc_3_nd(
   return loc;
 }
 
-inline uint2 elem_to_loc_2_nd(
+///////////////////////////////////////////////////////////////////////////////
+// Multiple Arrays with fixed N dims
+
+template <int NDIM>
+METAL_FUNC uint2 elem_to_loc_2_nd(
     uint3 elem,
-    constant const int* shape,
-    constant const size_t* a_strides,
-    constant const size_t* b_strides,
-    int ndim) {
+    constant const int shape[NDIM],
+    constant const size_t a_strides[NDIM],
+    constant const size_t b_strides[NDIM]) {
   uint2 loc = {
       static_cast<uint>(
-          elem.x * a_strides[ndim - 1] + elem.y * a_strides[ndim - 2]),
+          elem.x * a_strides[NDIM - 1] + elem.y * a_strides[NDIM - 2]),
       static_cast<uint>(
-          elem.x * b_strides[ndim - 1] + elem.y * b_strides[ndim - 2])};
-  for (int d = ndim - 3; d >= 0; --d) {
+          elem.x * b_strides[NDIM - 1] + elem.y * b_strides[NDIM - 2])};
+  for (int d = NDIM - 3; d >= 0; --d) {
     uint l = elem.z % shape[d];
     loc.x += l * a_strides[d];
     loc.y += l * b_strides[d];
@@ -219,55 +262,26 @@ inline uint2 elem_to_loc_2_nd(
 }
 
 template <int NDIM>
-inline uint elem_to_loc_nd(
-    uint elem,
-    device const int* shape,
-    device const size_t* strides);
-
-template <>
-inline uint elem_to_loc_nd<1>(
-    uint elem,
-    device const int* shape,
-    device const size_t* strides) {
-  return (elem % shape[0]) * strides[0];
-}
-
-template <>
-inline uint elem_to_loc_nd<2>(
-    uint elem,
-    device const int* shape,
-    device const size_t* strides) {
-  uint loc = (elem % shape[1]) * strides[1];
-  elem /= shape[1];
-  loc += (elem % shape[0]) * strides[0];
-  return loc;
-}
-
-template <>
-inline uint elem_to_loc_nd<3>(
-    uint elem,
-    device const int* shape,
-    device const size_t* strides) {
-  uint loc = (elem % shape[2]) * strides[2];
-  elem /= shape[2];
-  loc += (elem % shape[1]) * strides[1];
-  elem /= shape[1];
-  loc += (elem % shape[0]) * strides[0];
-  return loc;
-}
-
-template <>
-inline uint elem_to_loc_nd<4>(
-    uint elem,
-    device const int* shape,
-    device const size_t* strides) {
-  uint loc = (elem % shape[3]) * strides[3];
-  elem /= shape[3];
-  loc += (elem % shape[2]) * strides[2];
-  elem /= shape[2];
-  loc += (elem % shape[1]) * strides[1];
-  elem /= shape[1];
-  loc += (elem % shape[0]) * strides[0];
+METAL_FUNC uint3 elem_to_loc_3_nd(
+    uint3 elem,
+    constant const int shape[NDIM],
+    constant const size_t a_strides[NDIM],
+    constant const size_t b_strides[NDIM],
+    constant const size_t c_strides[NDIM]) {
+  uint3 loc = {
+      static_cast<uint>(
+          elem.x * a_strides[NDIM - 1] + elem.y * a_strides[NDIM - 2]),
+      static_cast<uint>(
+          elem.x * b_strides[NDIM - 1] + elem.y * b_strides[NDIM - 2]),
+      static_cast<uint>(
+          elem.x * c_strides[NDIM - 1] + elem.y * c_strides[NDIM - 2])};
+  for (int d = NDIM - 3; d >= 0; --d) {
+    uint l = elem.z % shape[d];
+    loc.x += l * a_strides[d];
+    loc.y += l * b_strides[d];
+    loc.z += l * c_strides[d];
+    elem.z /= shape[d];
+  }
   return loc;
 }
 

mlx/backend/metal/matmul.cpp

@@ -191,6 +191,69 @@ inline void mps_matmul(
       });
 }
 
+inline auto collapse_batches(const array& a, const array& b) {
+  // Get and check the shape for the batched dims
+  std::vector<int> A_bshape{a.shape().begin(), a.shape().end() - 2};
+  std::vector<int> B_bshape{b.shape().begin(), b.shape().end() - 2};
+  if (A_bshape != B_bshape) {
+    std::ostringstream msg;
+    msg << "[matmul] Got matrices with incorrectly broadcasted shapes: " << "A "
+        << a.shape() << ", B " << b.shape() << ".";
+    throw std::runtime_error(msg.str());
+  }
+
+  std::vector<size_t> A_bstride{a.strides().begin(), a.strides().end() - 2};
+  std::vector<size_t> B_bstride{b.strides().begin(), b.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] =
+      collapse_contiguous_dims(A_bshape, std::vector{A_bstride, B_bstride});
+
+  auto A_batch_stride = batch_strides[0];
+  auto B_batch_stride = batch_strides[1];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    A_batch_stride.push_back(0);
+    B_batch_stride.push_back(0);
+  }
+
+  return std::make_tuple(batch_shape, A_batch_stride, B_batch_stride);
+}
+
+inline auto collapse_batches(const array& a, const array& b, const array& c) {
+  // Get and check the shape for the batched dims
+  std::vector<int> A_bshape{a.shape().begin(), a.shape().end() - 2};
+  std::vector<int> B_bshape{b.shape().begin(), b.shape().end() - 2};
+  std::vector<int> C_bshape{c.shape().begin(), c.shape().end() - 2};
+  if (A_bshape != B_bshape || A_bshape != C_bshape) {
+    std::ostringstream msg;
+    msg << "[addmm] Got matrices with incorrectly broadcasted shapes: " << "A "
+        << a.shape() << ", B " << b.shape() << ", B " << c.shape() << ".";
+    throw std::runtime_error(msg.str());
+  }
+
+  std::vector<size_t> A_bstride{a.strides().begin(), a.strides().end() - 2};
+  std::vector<size_t> B_bstride{b.strides().begin(), b.strides().end() - 2};
+  std::vector<size_t> C_bstride{c.strides().begin(), c.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] = collapse_contiguous_dims(
+      A_bshape, std::vector{A_bstride, B_bstride, C_bstride});
+
+  auto A_batch_stride = batch_strides[0];
+  auto B_batch_stride = batch_strides[1];
+  auto C_batch_stride = batch_strides[2];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    A_batch_stride.push_back(0);
+    B_batch_stride.push_back(0);
+    C_batch_stride.push_back(0);
+  }
+
+  return std::make_tuple(
+      batch_shape, A_batch_stride, B_batch_stride, C_batch_stride);
+}
+
 } // namespace
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -211,22 +274,33 @@ void steel_matmul(
     int ldb,
     bool transpose_a,
     bool transpose_b,
-    std::vector<array>& copies) {
+    std::vector<array>& copies,
+    std::vector<int> batch_shape /* = {} */,
+    std::vector<size_t> A_batch_stride /* = {} */,
+    std::vector<size_t> B_batch_stride /* = {} */) {
   using namespace mlx::steel;
 
-  // Coalesce (B, M, K) X (K, N) to (B*M, K) X (K, N)
-  if (batch_size_out > 1 && !transpose_a &&
-      a.data_size() == batch_size_out * M * K && b.size() == K * N) {
-    M = M * batch_size_out;
-    batch_size_out = 1;
+  if (batch_shape.empty()) {
+    /////////////////////////////////////////////////////////////////////////////
+    // Check and collapse batch dimensions
+    auto [batch_shape_, A_bstride_, B_bstride_] = collapse_batches(a, b);
+
+    batch_shape = batch_shape_;
+    A_batch_stride = A_bstride_;
+    B_batch_stride = B_bstride_;
+    // Collapse batches into M if needed
+    if (batch_size_out > 1 && !transpose_a && batch_shape.size() == 1 &&
+        a.strides()[a.ndim() - 2] == K && A_batch_stride.back() == M * K &&
+        B_batch_stride.back() == 0) {
+      M *= batch_shape.back();
+      batch_size_out = 1;
+
+      A_batch_stride = {0};
+      B_batch_stride = {0};
+      batch_shape = {1};
+    }
   }
 
-  // Account for batch sizes and basic broadcasting
-  int batch_size_a = a.data_size() / (M * K);
-  int batch_size_b = b.data_size() / (K * N);
-
-  int matrix_stride_a = (batch_size_a == 1) ? 0 : M * K;
-  int matrix_stride_b = (batch_size_b == 1) ? 0 : K * N;
   int matrix_stride_out = M * N;
 
   /////////////////////////////////////////////////////////////////////////////
@@ -257,11 +331,11 @@ void steel_matmul(
           << (transpose_b ? 't' : 'n') << "_" << type_to_name(a) << "_"
           << type_to_name(C_split) << "_bm" << bm << "_bn" << bn << "_bk" << bk
           << "_wm" << wm << "_wn" << wn << "_MN_"
-          << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned"
-          << "_K_" << ((K % bk == 0) ? "t" : "n") << "aligned";
+          << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned" << "_K_"
+          << ((K % bk == 0) ? "t" : "n") << "aligned";
 
     // Encode and dispatch gemm kernel
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
@@ -269,25 +343,25 @@ void steel_matmul(
     int tm = (M + bm - 1) / bm;
 
     GEMMSpiltKParams params{
-        M,
-        N,
-        K,
-        lda,
-        ldb,
-        N,
-        tn,
-        tm,
-        split_k_partitions,
-        split_k_partition_stride,
-        split_k_partition_size,
-        gemm_k_iterations};
+        /* const int M = */ M,
+        /* const int N = */ N,
+        /* const int K = */ K,
+        /* const int lda = */ lda,
+        /* const int ldb = */ ldb,
+        /* const int ldc = */ N,
+        /* const int tiles_n = */ tn,
+        /* const int tiles_m = */ tm,
+        /* const int split_k_partitions = */ split_k_partitions,
+        /* const int split_k_partition_stride = */ split_k_partition_stride,
+        /* const int split_k_partition_size = */ split_k_partition_size,
+        /* const int gemm_k_iterations_aligned = */ gemm_k_iterations};
 
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(tn, tm, split_k_partitions);
 
-    set_array_buffer(compute_encoder, a, 0);
-    set_array_buffer(compute_encoder, b, 1);
-    set_array_buffer(compute_encoder, C_split, 2);
+    compute_encoder.set_input_array(a, 0);
+    compute_encoder.set_input_array(b, 1);
+    compute_encoder.set_output_array(C_split, 2);
 
     compute_encoder->setBytes(&params, sizeof(GEMMSpiltKParams), 3);
     compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
@@ -305,8 +379,8 @@ void steel_matmul(
       compute_encoder->setComputePipelineState(kernel);
 
       // Set the arguments for the kernel
-      set_array_buffer(compute_encoder, C_split, 0);
-      set_array_buffer(compute_encoder, out, 1);
+      compute_encoder.set_input_array(C_split, 0);
+      compute_encoder.set_output_array(out, 1);
       compute_encoder->setBytes(&split_k_partitions, sizeof(int), 2);
       compute_encoder->setBytes(&split_k_partition_stride, sizeof(int), 3);
       compute_encoder->setBytes(&N, sizeof(int), 4);
@@ -347,11 +421,11 @@ void steel_matmul(
         << (transpose_b ? 't' : 'n') << "_" << type_to_name(a) << "_"
         << type_to_name(out) << "_bm" << bm << "_bn" << bn << "_bk" << bk
         << "_wm" << wm << "_wn" << wn << "_MN_"
-        << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned"
-        << "_K_" << ((K % bk == 0) ? "t" : "n") << "aligned";
+        << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned" << "_K_"
+        << ((K % bk == 0) ? "t" : "n") << "aligned";
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
@@ -364,19 +438,20 @@ void steel_matmul(
 
   // Prepare steel matmul params
   GEMMParams params{
-      M,
-      N,
-      K,
-      lda,
-      ldb,
-      N,
-      tn,
-      tm,
-      matrix_stride_a,
-      matrix_stride_b,
-      matrix_stride_out,
-      swizzle_log,
-      (K / bk)};
+      /* const int M = */ M,
+      /* const int N = */ N,
+      /* const int K = */ K,
+      /* const int lda = */ lda,
+      /* const int ldb = */ ldb,
+      /* const int ldd = */ N,
+      /* const int tiles_n = */ tn,
+      /* const int tiles_m = */ tm,
+      /* const int batch_stride_a = */ int(A_batch_stride.back()),
+      /* const int batch_stride_b = */ int(B_batch_stride.back()),
+      /* const int batch_stride_d = */ matrix_stride_out,
+      /* const int swizzle_log = */ swizzle_log,
+      /* const int gemm_k_iterations_aligned = */ (K / bk),
+      /* const int batch_ndim = */ int(batch_shape.size())};
 
   // Prepare launch grid params
   int tile = 1 << swizzle_log;
@@ -386,37 +461,25 @@ void steel_matmul(
   MTL::Size group_dims = MTL::Size(32, wn, wm);
   MTL::Size grid_dims = MTL::Size(tn, tm, batch_size_out);
 
-  // Launch only 1 kernel in the case of simple batching / broadcasting
-  if (batch_size_out == std::max(batch_size_a, batch_size_b) &&
-      (batch_size_a == batch_size_b ||
-       std::min(batch_size_a, batch_size_b) == 1)) {
-    set_array_buffer(compute_encoder, a, 0);
-    set_array_buffer(compute_encoder, b, 1);
-    set_array_buffer(compute_encoder, out, 2);
+  std::vector<size_t> batch_strides = A_batch_stride;
+  batch_strides.insert(
+      batch_strides.end(), B_batch_stride.begin(), B_batch_stride.end());
 
-    compute_encoder->setBytes(&params, sizeof(GEMMParams), 3);
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
-  } else { // Otherwise launch kernels with set offsets
+  // Launch kernel
+  compute_encoder.set_input_array(a, 0);
+  compute_encoder.set_input_array(b, 1);
+  compute_encoder.set_output_array(out, 3);
 
-    MTL::Size grid_dims_single = MTL::Size(tn, tm, 1);
+  compute_encoder->setBytes(&params, sizeof(GEMMParams), 4);
 
-    for (int i = 0; i < batch_size_out; ++i) {
-      auto a_off = elem_to_loc(M * K * i, a.shape(), a.strides());
-      auto b_off = elem_to_loc(K * N * i, b.shape(), b.strides());
+  compute_encoder->setBytes(
+      batch_shape.data(), sizeof(int) * batch_shape.size(), 6);
+  compute_encoder->setBytes(
+      batch_strides.data(), sizeof(size_t) * batch_strides.size(), 7);
 
-      auto a_buf = static_cast<const MTL::Buffer*>(a.buffer().ptr());
-      auto b_buf = static_cast<const MTL::Buffer*>(b.buffer().ptr());
-      auto out_buf = static_cast<const MTL::Buffer*>(out.buffer().ptr());
-
-      compute_encoder->setBuffer(a_buf, a_off * a.itemsize(), 0);
-      compute_encoder->setBuffer(b_buf, b_off * b.itemsize(), 1);
-      compute_encoder->setBuffer(out_buf, i * M * N * out.itemsize(), 2);
-
-      compute_encoder->setBytes(&params, sizeof(GEMMParams), 3);
-      compute_encoder->dispatchThreadgroups(grid_dims_single, group_dims);
-    }
-  }
+  compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
 
+  // Clear copies
   d.get_command_buffer(s.index)->addCompletedHandler(
       [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
   return;
@@ -424,16 +487,25 @@ void steel_matmul(
 
 void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
-  if (!is_floating_point(out.dtype())) {
+  if (!issubdtype(out.dtype(), floating)) {
     throw std::runtime_error(
         "[matmul] Does not yet support non-floating point types.");
   }
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
   auto& s = stream();
   auto& d = metal::device(s.device);
 
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
+  // Return 0s if either input is empty
+  if (a_pre.size() == 0 || b_pre.size() == 0) {
+    array zero = array(0, a_pre.dtype());
+    copy_gpu(zero, out, CopyType::Scalar, s);
+    auto command_buffer = d.get_command_buffer(s.index);
+    command_buffer->addCompletedHandler([zero](MTL::CommandBuffer*) {});
+    return;
+  }
+
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
   /////////////////////////////////////////////////////////////////////////////
   // Init checks and prep
@@ -444,9 +516,9 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto check_transpose = [&copies, &s](const array& arr) {
     auto stx = arr.strides()[arr.ndim() - 2];
     auto sty = arr.strides()[arr.ndim() - 1];
-    if (stx == arr.shape(-1) && sty == 1) {
+    if (sty == 1) {
       return std::make_tuple(false, stx, arr);
-    } else if (stx == 1 && sty == arr.shape(-2)) {
+    } else if (stx == 1) {
       return std::make_tuple(true, sty, arr);
     } else {
       array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
@@ -464,8 +536,25 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   int N = b.shape(-1);
   int K = a.shape(-1);
 
+  /////////////////////////////////////////////////////////////////////////////
+  // Check and collapse batch dimensions
+
+  auto [batch_shape, A_batch_stride, B_batch_stride] = collapse_batches(a, b);
+
   auto batch_size_out = out.size() / (M * N);
 
+  // Collapse batches into M if needed
+  if (batch_size_out > 1 && !a_transposed && batch_shape.size() == 1 &&
+      a.strides()[a.ndim() - 2] == K && A_batch_stride.back() == M * K &&
+      B_batch_stride.back() == 0) {
+    M *= batch_shape.back();
+    batch_size_out = 1;
+
+    A_batch_stride = {0};
+    B_batch_stride = {0};
+    batch_shape = {1};
+  }
+
   /////////////////////////////////////////////////////////////////////////////
   // Gemv specialization
 
@@ -482,20 +571,18 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
 
     int mat_cols = transpose_mat ? out_vector_len : in_vector_len;
     int mat_rows = transpose_mat ? in_vector_len : out_vector_len;
+    int mat_ld = is_b_matrix ? b_cols : a_cols;
 
-    int batch_size_mat = mat.data_size() / (mat_cols * mat_rows);
-    int stride_mat = batch_size_mat == 1 ? 0 : mat_cols * mat_rows;
+    auto batch_strides_mat = is_b_matrix ? B_batch_stride : A_batch_stride;
+    auto batch_strides_vec = is_b_matrix ? A_batch_stride : B_batch_stride;
 
-    int batch_size_vec = vec.data_size() / in_vector_len;
-    int stride_vec = batch_size_vec == 1 ? 0 : in_vector_len;
+    int stride_mat = batch_strides_mat.back();
+    int stride_vec = batch_strides_vec.back();
 
     // Determine if inputs have simple batching / broadcasting
-    bool contiguous_kernel =
-        (batch_size_out == std::max(batch_size_mat, batch_size_vec) &&
-         (batch_size_mat == batch_size_vec ||
-          std::min(batch_size_mat, batch_size_vec) == 1));
+    bool contiguous_kernel = (batch_shape.size() == 1);
 
-    int nc_dim = out.ndim() - 2;
+    int batch_ndim = batch_shape.size();
 
     // Determine dispatch kernel
     int tm = 4, tn = 4;
@@ -531,13 +618,10 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     }
 
     kname << "_bm" << bm << "_bn" << bn << "_tm" << tm << "_tn" << tn;
-
-    if (!contiguous_kernel) {
-      kname << "_nc";
-    }
+    kname << "_nc" << !contiguous_kernel << "_axpby0";
 
     // Encode and dispatch kernel
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
@@ -545,27 +629,20 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     MTL::Size group_dims = MTL::Size(bn, bm, 1);
     MTL::Size grid_dims = MTL::Size(n_tgp, 1, batch_size_out);
 
-    set_array_buffer(compute_encoder, mat, 0);
-    set_array_buffer(compute_encoder, vec, 1);
-    set_array_buffer(compute_encoder, out, 2);
+    compute_encoder.set_input_array(mat, 0);
+    compute_encoder.set_input_array(vec, 1);
+    compute_encoder.set_output_array(out, 3);
 
-    compute_encoder->setBytes(&in_vector_len, sizeof(int), 3);
-    compute_encoder->setBytes(&out_vector_len, sizeof(int), 4);
+    compute_encoder->setBytes(&in_vector_len, sizeof(int), 4);
+    compute_encoder->setBytes(&out_vector_len, sizeof(int), 5);
+    compute_encoder->setBytes(&mat_ld, sizeof(int), 6);
 
-    if (contiguous_kernel) {
-      compute_encoder->setBytes(&stride_vec, sizeof(int), 5);
-      compute_encoder->setBytes(&stride_mat, sizeof(int), 6);
-    } else {
-      // In case of complex broadcasting, we consider the shape[:-2] and
-      // strides [:-2] to determine the location of a batch
-      // nc_dim = out.ndim() - 2
-      compute_encoder->setBytes(&nc_dim, sizeof(int), 5);
-      compute_encoder->setBytes(out.shape().data(), nc_dim * sizeof(int), 6);
-      compute_encoder->setBytes(
-          vec.strides().data(), nc_dim * sizeof(size_t), 7);
-      compute_encoder->setBytes(
-          mat.strides().data(), nc_dim * sizeof(size_t), 8);
-    }
+    compute_encoder->setBytes(&batch_ndim, sizeof(int), 9);
+    compute_encoder->setBytes(batch_shape.data(), batch_ndim * sizeof(int), 10);
+    compute_encoder->setBytes(
+        batch_strides_vec.data(), batch_ndim * sizeof(size_t), 11);
+    compute_encoder->setBytes(
+        batch_strides_mat.data(), batch_ndim * sizeof(size_t), 12);
 
     compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
 
@@ -573,7 +650,6 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
         [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
     return;
   }
-
   /////////////////////////////////////////////////////////////////////////////
   // Gemm specialization
 
@@ -598,25 +674,28 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
 
   return steel_matmul(
-      s,
-      d,
-      a,
-      b,
-      out,
-      M,
-      N,
-      K,
-      batch_size_out,
-      a_cols,
-      b_cols,
-      a_transposed,
-      b_transposed,
-      copies);
+      /* const Stream& s = */ s,
+      /* metal::Device& d = */ d,
+      /* const array& a = */ a,
+      /* const array& b = */ b,
+      /* array& out = */ out,
+      /* int M = */ M,
+      /* int N = */ N,
+      /* int K = */ K,
+      /* int batch_size_out = */ batch_size_out,
+      /* int lda = */ a_cols,
+      /* int ldb = */ b_cols,
+      /* bool transpose_a = */ a_transposed,
+      /* bool transpose_b = */ b_transposed,
+      /* std::vector<array>& = */ copies,
+      /* std::vector<int> batch_shape = */ batch_shape,
+      /* std::vector<size_t> A_batch_stride = */ A_batch_stride,
+      /* std::vector<size_t> B_batch_stride = */ B_batch_stride);
 }
 
 void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 3);
-  if (!is_floating_point(out.dtype())) {
+  if (!issubdtype(out.dtype(), floating)) {
     throw std::runtime_error(
         "[matmul] Does not yet support non-floating point types.");
   }
@@ -637,9 +716,9 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto check_transpose = [&copies, &s](const array& arr) {
     auto stx = arr.strides()[arr.ndim() - 2];
     auto sty = arr.strides()[arr.ndim() - 1];
-    if (stx == arr.shape(-1) && sty == 1) {
+    if (sty == 1) {
       return std::make_tuple(false, stx, arr);
-    } else if (stx == 1 && sty == arr.shape(-2)) {
+    } else if (stx == 1) {
       return std::make_tuple(true, sty, arr);
     } else {
       array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
@@ -657,33 +736,151 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   int N = b.shape(-1);
   int K = a.shape(-1);
 
-  auto batch_size_out = out.size() / (M * N);
-
   array c = c_pre;
   int ldc = c.strides()[c.ndim() - 2];
   int fdc = c.strides()[c.ndim() - 1];
-  int matrix_stride_c = c.ndim() <= 2 ? 0 : c.strides()[c.ndim() - 3];
 
   int lda = a_cols;
   int ldb = b_cols;
+  int ldd = N;
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Check and collapse batch dimensions
+  auto [batch_shape, A_batch_stride, B_batch_stride, C_batch_stride] =
+      collapse_batches(a, b, c);
+
+  auto batch_size_out = out.size() / (M * N);
+
+  // Collapse batches into M if needed
+  if (batch_size_out > 1 && !transpose_a && batch_shape.size() == 1 &&
+      a.strides()[a.ndim() - 2] == K && A_batch_stride.back() == M * K &&
+      C_batch_stride.back() == M * c.strides()[c.ndim() - 2] &&
+      B_batch_stride.back() == 0) {
+    M *= batch_shape.back();
+    batch_size_out = 1;
+
+    A_batch_stride = {0};
+    B_batch_stride = {0};
+    C_batch_stride = {0};
+    batch_shape = {1};
+  }
+
+  int matrix_stride_out = M * N;
+
+  /////////////////////////////////////////////////////////////////////////////
+  // Gemv specialization
+
+  // Route to gemv if needed
+  if (std::min(M, N) == 1) {
+    // Collect problem info
+    bool is_b_matrix = N != 1;
+
+    auto& mat = is_b_matrix ? b : a;
+    auto& vec = is_b_matrix ? a : b;
+    bool transpose_mat = is_b_matrix ? !transpose_b : transpose_a;
+    int in_vector_len = K;
+    int out_vector_len = is_b_matrix ? N : M;
+
+    int mat_cols = transpose_mat ? out_vector_len : in_vector_len;
+    int mat_rows = transpose_mat ? in_vector_len : out_vector_len;
+    int mat_ld = is_b_matrix ? b_cols : a_cols;
+
+    auto batch_strides_mat = is_b_matrix ? B_batch_stride : A_batch_stride;
+    auto batch_strides_vec = is_b_matrix ? A_batch_stride : B_batch_stride;
+
+    int stride_mat = batch_strides_mat.back();
+    int stride_vec = batch_strides_vec.back();
+
+    // Determine if inputs have simple batching / broadcasting
+    bool contiguous_kernel = (batch_shape.size() == 1);
+
+    int batch_ndim = batch_shape.size();
+
+    // Determine dispatch kernel
+    int tm = 4, tn = 4;
+    int bm, bn, n_out_per_tgp;
+    std::ostringstream kname;
+
+    if (transpose_mat) {
+      bm = 8;
+      bn = 8;
+      if (out_vector_len >= 24576) {
+        bn = 128;
+      } else if (out_vector_len >= 16384) {
+        bn = 64;
+      } else if (out_vector_len >= 8192) {
+        bn = 16;
+      }
+
+      // Specialized kernel for very small outputs
+      tn = out_vector_len < tn ? 1 : tn;
+
+      n_out_per_tgp = bn * tn;
+      kname << "gemv_t_" << type_to_name(out);
+
+    } else {
+      bm = out_vector_len >= 4096 ? 8 : 4;
+      bn = 32;
+
+      // Specialized kernel for very small outputs
+      tm = out_vector_len < tm ? 1 : tm;
+
+      n_out_per_tgp = bm * tm;
+      kname << "gemv_" << type_to_name(out);
+    }
+
+    kname << "_bm" << bm << "_bn" << bn << "_tm" << tm << "_tn" << tn;
+    kname << "_nc" << !contiguous_kernel << "_axpby1";
+
+    // Encode and dispatch kernel
+    auto& compute_encoder = d.get_command_encoder(s.index);
+    auto kernel = d.get_kernel(kname.str());
+    compute_encoder->setComputePipelineState(kernel);
+
+    int n_tgp = (out_vector_len + n_out_per_tgp - 1) / n_out_per_tgp;
+    MTL::Size group_dims = MTL::Size(bn, bm, 1);
+    MTL::Size grid_dims = MTL::Size(n_tgp, 1, batch_size_out);
+
+    compute_encoder.set_input_array(mat, 0);
+    compute_encoder.set_input_array(vec, 1);
+    compute_encoder.set_input_array(c, 2);
+    compute_encoder.set_output_array(out, 3);
+
+    compute_encoder->setBytes(&in_vector_len, sizeof(int), 4);
+    compute_encoder->setBytes(&out_vector_len, sizeof(int), 5);
+    compute_encoder->setBytes(&mat_ld, sizeof(int), 6);
+
+    compute_encoder->setBytes(&alpha_, sizeof(float), 7);
+    compute_encoder->setBytes(&beta_, sizeof(float), 8);
+
+    compute_encoder->setBytes(&batch_ndim, sizeof(int), 9);
+    compute_encoder->setBytes(batch_shape.data(), batch_ndim * sizeof(int), 10);
+    compute_encoder->setBytes(
+        batch_strides_vec.data(), batch_ndim * sizeof(size_t), 11);
+    compute_encoder->setBytes(
+        batch_strides_mat.data(), batch_ndim * sizeof(size_t), 12);
+    compute_encoder->setBytes(
+        C_batch_stride.data(), batch_ndim * sizeof(size_t), 13);
+
+    int bias_stride = c.strides()[c.ndim() - 1];
+    compute_encoder->setBytes(&bias_stride, sizeof(int), 14);
+
+    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
+
+    d.get_command_buffer(s.index)->addCompletedHandler(
+        [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
+    return;
+  }
 
   using namespace mlx::steel;
 
-  // Account for batch sizes and basic broadcasting
-  int batch_size_a = a.data_size() / (M * K);
-  int batch_size_b = b.data_size() / (K * N);
-
-  int matrix_stride_a = (batch_size_a == 1) ? 0 : M * K;
-  int matrix_stride_b = (batch_size_b == 1) ? 0 : K * N;
-  int matrix_stride_out = M * N;
+  /////////////////////////////////////////////////////////////////////////////
+  // Split K specialization
 
   int _tm = M / 16;
   int _tn = N / 16;
   int _tk = K / 16;
 
-  /////////////////////////////////////////////////////////////////////////////
-  // Split K specialization
-
   if (batch_size_out == 1 && (_tm * _tn) <= 32 && _tk >= 8) {
     int bm = M < 40 ? 16 : 32;
     int bn = N < 40 ? 16 : 32;
@@ -705,11 +902,11 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
           << (transpose_b ? 't' : 'n') << "_" << type_to_name(a) << "_"
           << type_to_name(C_split) << "_bm" << bm << "_bn" << bn << "_bk" << bk
           << "_wm" << wm << "_wn" << wn << "_MN_"
-          << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned"
-          << "_K_" << ((K % bk == 0) ? "t" : "n") << "aligned";
+          << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned" << "_K_"
+          << ((K % bk == 0) ? "t" : "n") << "aligned";
 
     // Encode and dispatch gemm kernel
-    auto compute_encoder = d.get_command_encoder(s.index);
+    auto& compute_encoder = d.get_command_encoder(s.index);
     auto kernel = d.get_kernel(kname.str());
     compute_encoder->setComputePipelineState(kernel);
 
@@ -733,9 +930,9 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
     MTL::Size group_dims = MTL::Size(32, wn, wm);
     MTL::Size grid_dims = MTL::Size(tn, tm, split_k_partitions);
 
-    set_array_buffer(compute_encoder, a, 0);
-    set_array_buffer(compute_encoder, b, 1);
-    set_array_buffer(compute_encoder, C_split, 2);
+    compute_encoder.set_input_array(a, 0);
+    compute_encoder.set_input_array(b, 1);
+    compute_encoder.set_output_array(C_split, 2);
 
     compute_encoder->setBytes(&params, sizeof(GEMMSpiltKParams), 3);
     compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
@@ -748,12 +945,12 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       compute_encoder->setComputePipelineState(kernel);
 
       // Set the arguments for the kernel
-      set_array_buffer(compute_encoder, C_split, 0);
-      set_array_buffer(compute_encoder, out, 1);
+      compute_encoder.set_input_array(C_split, 0);
+      compute_encoder.set_output_array(out, 1);
       compute_encoder->setBytes(&split_k_partitions, sizeof(int), 2);
       compute_encoder->setBytes(&split_k_partition_stride, sizeof(int), 3);
       compute_encoder->setBytes(&N, sizeof(int), 4);
-      set_array_buffer(compute_encoder, c, 5);
+      compute_encoder.set_input_array(c, 5);
       compute_encoder->setBytes(&ldc, sizeof(int), 6);
       compute_encoder->setBytes(&fdc, sizeof(int), 7);
       compute_encoder->setBytes(&alpha_, sizeof(float), 8);
@@ -794,12 +991,12 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
         << (transpose_b ? 't' : 'n') << "_" << type_to_name(a) << "_"
         << type_to_name(out) << "_bm" << bm << "_bn" << bn << "_bk" << bk
         << "_wm" << wm << "_wn" << wn << "_MN_"
-        << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned"
-        << "_K_" << ((K % bk == 0) ? "t" : "n") << "aligned"
+        << ((M % bm == 0 && N % bn == 0) ? "t" : "n") << "aligned" << "_K_"
+        << ((K % bk == 0) ? "t" : "n") << "aligned"
         << ((alpha_ == 1. && beta_ == 1.) ? "_add" : "_axpby");
 
   // Encode and dispatch kernel
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   auto kernel = d.get_kernel(kname.str());
   compute_encoder->setComputePipelineState(kernel);
 
@@ -809,25 +1006,29 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   // TODO: Explore device-based tuning for swizzle
   int swizzle_log = 0; // tm >= 6 ? 3 : (tm <= 3 ? 0 : 2);
 
+  // Prepare steel matmul params
+  GEMMParams gemm_params{
+      /* const int M = */ M,
+      /* const int N = */ N,
+      /* const int K = */ K,
+      /* const int lda = */ lda,
+      /* const int ldb = */ ldb,
+      /* const int ldd = */ N,
+      /* const int tiles_n = */ tn,
+      /* const int tiles_m = */ tm,
+      /* const int batch_stride_a = */ int(A_batch_stride.back()),
+      /* const int batch_stride_b = */ int(B_batch_stride.back()),
+      /* const int batch_stride_d = */ matrix_stride_out,
+      /* const int swizzle_log = */ swizzle_log,
+      /* const int gemm_k_iterations_aligned = */ (K / bk),
+      /* const int batch_ndim = */ int(batch_shape.size())};
+
   GEMMAddMMParams params{
-      M,
-      N,
-      K,
-      lda,
-      ldb,
-      ldc,
-      N,
-      tn,
-      tm,
-      matrix_stride_a,
-      matrix_stride_b,
-      matrix_stride_c,
-      matrix_stride_out,
-      swizzle_log,
-      (K / bk),
-      alpha_,
-      beta_,
-      fdc};
+      /* const int ldc = */ ldc,
+      /* const int fdc = */ fdc,
+      /* const int batch_stride_c = */ int(C_batch_stride.back()),
+      /* const float alpha = */ alpha_,
+      /* const float beta = */ beta_};
 
   int tile = 1 << swizzle_log;
   tm = (tm + tile - 1) / tile;
@@ -836,40 +1037,27 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   MTL::Size group_dims = MTL::Size(32, wn, wm);
   MTL::Size grid_dims = MTL::Size(tn, tm, batch_size_out);
 
-  // Launch only 1 kernel in the case of simple batching / broadcasting
-  if (batch_size_out == std::max(batch_size_a, batch_size_b) &&
-      (batch_size_a == batch_size_b ||
-       std::min(batch_size_a, batch_size_b) == 1)) {
-    set_array_buffer(compute_encoder, a, 0);
-    set_array_buffer(compute_encoder, b, 1);
-    set_array_buffer(compute_encoder, c, 2);
-    set_array_buffer(compute_encoder, out, 3);
+  std::vector<size_t> batch_strides = A_batch_stride;
+  batch_strides.insert(
+      batch_strides.end(), B_batch_stride.begin(), B_batch_stride.end());
+  batch_strides.insert(
+      batch_strides.end(), C_batch_stride.begin(), C_batch_stride.end());
 
-    compute_encoder->setBytes(&params, sizeof(GEMMAddMMParams), 4);
-    compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
-  } else { // Otherwise launch kernels with set offsets
+  // Launch kernel
+  compute_encoder.set_input_array(a, 0);
+  compute_encoder.set_input_array(b, 1);
+  compute_encoder.set_input_array(c, 2);
+  compute_encoder.set_output_array(out, 3);
 
-    MTL::Size grid_dims_single = MTL::Size(tn, tm, 1);
+  compute_encoder->setBytes(&gemm_params, sizeof(GEMMParams), 4);
+  compute_encoder->setBytes(&params, sizeof(GEMMAddMMParams), 5);
 
-    for (int i = 0; i < batch_size_out; ++i) {
-      auto a_off = elem_to_loc(M * K * i, a.shape(), a.strides());
-      auto b_off = elem_to_loc(K * N * i, b.shape(), b.strides());
-      auto c_off = elem_to_loc(M * N * i, c.shape(), c.strides());
+  compute_encoder->setBytes(
+      batch_shape.data(), sizeof(int) * batch_shape.size(), 6);
+  compute_encoder->setBytes(
+      batch_strides.data(), sizeof(size_t) * batch_strides.size(), 7);
 
-      auto a_buf = static_cast<const MTL::Buffer*>(a.buffer().ptr());
-      auto b_buf = static_cast<const MTL::Buffer*>(b.buffer().ptr());
-      auto c_buf = static_cast<const MTL::Buffer*>(c.buffer().ptr());
-      auto out_buf = static_cast<const MTL::Buffer*>(out.buffer().ptr());
-
-      compute_encoder->setBuffer(a_buf, a_off * a.itemsize(), 0);
-      compute_encoder->setBuffer(b_buf, b_off * b.itemsize(), 1);
-      compute_encoder->setBuffer(c_buf, c_off * c.itemsize(), 2);
-      compute_encoder->setBuffer(out_buf, i * M * N * out.itemsize(), 3);
-
-      compute_encoder->setBytes(&params, sizeof(GEMMAddMMParams), 4);
-      compute_encoder->dispatchThreadgroups(grid_dims_single, group_dims);
-    }
-  }
+  compute_encoder->dispatchThreadgroups(grid_dims, group_dims);
 
   d.get_command_buffer(s.index)->addCompletedHandler(
       [copies](MTL::CommandBuffer*) mutable { copies.clear(); });

mlx/backend/metal/matmul.h

@@ -26,6 +26,9 @@ void steel_matmul(
     int ldb,
     bool transpose_a,
     bool transpose_b,
-    std::vector<array>& copies);
+    std::vector<array>& copies,
+    std::vector<int> batch_shape = {},
+    std::vector<size_t> A_batch_stride = {},
+    std::vector<size_t> B_batch_stride = {});
 
 } // namespace mlx::core

mlx/backend/metal/metal.cpp

@@ -1,10 +1,10 @@
 // Copyright © 2023-2024 Apple Inc.
-
 #include <cstdlib>
 #include <future>
 #include <memory>
 
 #include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/utils.h"
 #include "mlx/primitives.h"
 #include "mlx/scheduler.h"
 
@@ -74,6 +74,8 @@ std::function<void()> make_task(
       if (arr.is_tracer()) {
         inputs = arr.inputs();
       }
+
+      debug_set_primitive_buffer_label(command_buffer, arr.primitive());
       arr.primitive().eval_gpu(arr.inputs(), outputs);
     }
     std::vector<std::shared_ptr<array::Data>> buffers;
@@ -86,7 +88,6 @@ std::function<void()> make_task(
     if (!arr.is_tracer()) {
       arr.detach();
     }
-
     if (p) {
       metal::device(s.device).end_encoding(s.index);
       scheduler::notify_new_task(s);
@@ -108,4 +109,31 @@ std::function<void()> make_task(
   return task;
 }
 
+bool start_capture(std::string path, id object) {
+  auto pool = new_scoped_memory_pool();
+
+  auto descriptor = MTL::CaptureDescriptor::alloc()->init();
+  descriptor->setCaptureObject(object);
+
+  if (path.length() > 0) {
+    auto string = NS::String::string(path.c_str(), NS::UTF8StringEncoding);
+    auto url = NS::URL::fileURLWithPath(string);
+    descriptor->setDestination(MTL::CaptureDestinationGPUTraceDocument);
+    descriptor->setOutputURL(url);
+  }
+
+  auto manager = MTL::CaptureManager::sharedCaptureManager();
+  return manager->startCapture(descriptor, nullptr);
+}
+
+bool start_capture(std::string path) {
+  auto& device = metal::device(mlx::core::Device::gpu);
+  return start_capture(path, device.mtl_device());
+}
+
+void stop_capture() {
+  auto manager = MTL::CaptureManager::sharedCaptureManager();
+  manager->stopCapture();
+}
+
 } // namespace mlx::core::metal

mlx/backend/metal/metal.h

@@ -2,15 +2,11 @@
 
 #pragma once
 
-#include <future>
-#include <memory>
-#include <vector>
-
 #include "mlx/array.h"
-#include "mlx/stream.h"
 
 namespace mlx::core::metal {
 
+/* Check if the Metal backend is available. */
 bool is_available();
 
 /* Get the actively used memory in bytes.
@@ -58,12 +54,8 @@ size_t set_memory_limit(size_t limit, bool relaxed = true);
  * */
 size_t set_cache_limit(size_t limit);
 
-void new_stream(Stream stream);
-std::shared_ptr<void> new_scoped_memory_pool();
-
-std::function<void()> make_task(
-    array& arr,
-    std::vector<std::shared_future<void>> deps,
-    std::shared_ptr<std::promise<void>> p);
+/** Capture a GPU trace, saving it to an absolute file `path` */
+bool start_capture(std::string path = "");
+void stop_capture();
 
 } // namespace mlx::core::metal

mlx/backend/metal/metal_impl.h

@@ -0,0 +1,22 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#pragma once
+
+#include <future>
+#include <memory>
+#include <vector>
+
+#include "mlx/array.h"
+#include "mlx/stream.h"
+
+namespace mlx::core::metal {
+
+void new_stream(Stream stream);
+std::shared_ptr<void> new_scoped_memory_pool();
+
+std::function<void()> make_task(
+    array& arr,
+    std::vector<std::shared_future<void>> deps,
+    std::shared_ptr<std::promise<void>> p);
+
+} // namespace mlx::core::metal

mlx/backend/metal/normalization.cpp

@@ -0,0 +1,418 @@
+// Copyright © 2024 Apple Inc.
+#include <algorithm>
+
+#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/kernels/defines.h"
+#include "mlx/backend/metal/reduce.h"
+#include "mlx/backend/metal/utils.h"
+#include "mlx/fast_primitives.h"
+
+namespace mlx::core::fast {
+
+void RMSNorm::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& s = stream();
+  auto& d = metal::device(s.device);
+  auto& out = outputs[0];
+
+  // Make sure that the last dimension is contiguous
+  std::vector<array> copies;
+  auto check_input = [&copies, &s](const array& x) -> const array& {
+    bool no_copy = x.strides()[x.ndim() - 1] == 1;
+    if (x.ndim() > 1) {
+      auto s = x.strides()[x.ndim() - 2];
+      no_copy &= (s == 0 || s == x.shape().back());
+    }
+    if (no_copy) {
+      return x;
+    } else {
+      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
+      copy_gpu(x, copies.back(), CopyType::General, s);
+      return copies.back();
+    }
+  };
+  const array& x = check_input(inputs[0]);
+  const array& w = inputs[1];
+
+  if (x.is_donatable()) {
+    out.move_shared_buffer(x);
+  } else {
+    out.set_data(
+        allocator::malloc_or_wait(x.data_size() * x.itemsize()),
+        x.data_size(),
+        x.strides(),
+        x.flags());
+  }
+
+  auto axis_size = static_cast<uint32_t>(x.shape().back());
+  int n_rows = x.data_size() / axis_size;
+
+  const int simd_size = 32;
+  const int n_reads = RMS_N_READS;
+  const int looped_limit = RMS_LOOPED_LIMIT;
+  std::string op_name = "rms";
+  if (axis_size > looped_limit) {
+    op_name += "_looped";
+  }
+  op_name += type_to_name(out);
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  {
+    auto kernel = d.get_kernel(op_name);
+
+    MTL::Size grid_dims, group_dims;
+    if (axis_size <= looped_limit) {
+      size_t threadgroup_needed = (axis_size + n_reads - 1) / n_reads;
+      size_t simds_needed = (threadgroup_needed + simd_size - 1) / simd_size;
+      size_t threadgroup_size = simd_size * simds_needed;
+      assert(threadgroup_size <= kernel->maxTotalThreadsPerThreadgroup());
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    } else {
+      size_t threadgroup_size = kernel->maxTotalThreadsPerThreadgroup();
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    }
+
+    uint32_t w_stride = w.strides()[0];
+    compute_encoder->setComputePipelineState(kernel);
+    compute_encoder.set_input_array(
+        x.data_shared_ptr() == nullptr ? out : x, 0);
+    compute_encoder.set_input_array(w, 1);
+    compute_encoder.set_output_array(out, 2);
+    compute_encoder->setBytes(&eps_, sizeof(float), 3);
+    compute_encoder->setBytes(&axis_size, sizeof(int), 4);
+    compute_encoder->setBytes(&w_stride, sizeof(uint32_t), 5);
+    compute_encoder->setThreadgroupMemoryLength(
+        16 * 8, 0); // minimum of 16 bytes
+    compute_encoder->setThreadgroupMemoryLength(simd_size * sizeof(float), 1);
+    compute_encoder->dispatchThreads(grid_dims, group_dims);
+  }
+  d.get_command_buffer(s.index)->addCompletedHandler(
+      [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
+}
+
+void RMSNormVJP::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& s = stream();
+  auto& d = metal::device(s.device);
+
+  // Ensure row contiguity. We could relax this step by checking that the array
+  // is contiguous (no broadcasts or holes) and that the input strides are the
+  // same as the cotangent strides but for now this is simpler.
+  std::vector<array> copies;
+  auto check_input = [&copies, &s](const array& x) -> const array& {
+    if (x.flags().row_contiguous) {
+      return x;
+    }
+    copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
+    copy_gpu(x, copies.back(), CopyType::General, s);
+    return copies.back();
+  };
+  const array& x = check_input(inputs[0]);
+  const array& w = inputs[1];
+  const array& g = check_input(inputs[2]);
+  array& gx = outputs[0];
+  array& gw = outputs[1];
+
+  // Allocate space for the outputs
+  bool x_in_gx = false;
+  bool g_in_gx = false;
+  if (x.is_donatable()) {
+    gx.move_shared_buffer(x);
+    x_in_gx = true;
+  } else if (g.is_donatable()) {
+    gx.move_shared_buffer(g);
+    g_in_gx = true;
+  } else {
+    gx.set_data(allocator::malloc_or_wait(gx.nbytes()));
+  }
+
+  auto axis_size = static_cast<uint32_t>(x.shape().back());
+  int n_rows = x.data_size() / axis_size;
+
+  // Allocate a temporary to store the gradients for w and initialize the
+  // gradient accumulator to 0.
+  array gw_temp({n_rows, x.shape().back()}, gw.dtype(), nullptr, {});
+  bool g_in_gw = false;
+  if (!g_in_gx && g.is_donatable()) {
+    gw_temp.move_shared_buffer(g);
+    g_in_gw = true;
+  } else {
+    gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes()));
+  }
+  copies.push_back(gw_temp);
+  {
+    array zero(0, gw.dtype());
+    copy_gpu(zero, gw, CopyType::Scalar, s);
+    copies.push_back(std::move(zero));
+  }
+
+  const int simd_size = 32;
+  const int n_reads = RMS_N_READS;
+  const int looped_limit = RMS_LOOPED_LIMIT;
+  std::string op_name = "vjp_rms";
+  if (axis_size > looped_limit) {
+    op_name += "_looped";
+  }
+  op_name += type_to_name(gx);
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  {
+    auto kernel = d.get_kernel(op_name);
+
+    MTL::Size grid_dims, group_dims;
+    if (axis_size <= looped_limit) {
+      size_t threadgroup_needed = (axis_size + n_reads - 1) / n_reads;
+      size_t simds_needed = (threadgroup_needed + simd_size - 1) / simd_size;
+      size_t threadgroup_size = simd_size * simds_needed;
+      assert(threadgroup_size <= kernel->maxTotalThreadsPerThreadgroup());
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    } else {
+      size_t threadgroup_size = kernel->maxTotalThreadsPerThreadgroup();
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    }
+
+    uint32_t w_stride = w.strides()[0];
+    compute_encoder->setComputePipelineState(kernel);
+    compute_encoder.set_input_array(x_in_gx ? gx : x, 0);
+    compute_encoder.set_input_array(w, 1);
+    compute_encoder.set_input_array(g_in_gx ? gx : (g_in_gw ? gw_temp : g), 2);
+    compute_encoder.set_output_array(gx, 3);
+    compute_encoder.set_output_array(gw_temp, 4);
+    compute_encoder->setBytes(&eps_, sizeof(float), 5);
+    compute_encoder->setBytes(&axis_size, sizeof(int), 6);
+    compute_encoder->setBytes(&w_stride, sizeof(uint32_t), 7);
+    compute_encoder->dispatchThreads(grid_dims, group_dims);
+  }
+
+  ReductionPlan plan(
+      ReductionOpType::ContiguousStridedReduce, {n_rows}, {axis_size});
+  strided_reduce_general_dispatch(
+      gw_temp, gw, "sum", plan, {0}, compute_encoder, d, s);
+
+  d.get_command_buffer(s.index)->addCompletedHandler(
+      [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
+}
+
+void LayerNorm::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& s = stream();
+  auto& d = metal::device(s.device);
+  auto& out = outputs[0];
+
+  // Make sure that the last dimension is contiguous
+  std::vector<array> copies;
+  auto check_input = [&copies, &s](const array& x) -> const array& {
+    bool no_copy = x.strides()[x.ndim() - 1] == 1;
+    if (x.ndim() > 1) {
+      auto s = x.strides()[x.ndim() - 2];
+      no_copy &= (s == 0 || s == x.shape().back());
+    }
+    if (no_copy) {
+      return x;
+    } else {
+      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
+      copy_gpu(x, copies.back(), CopyType::General, s);
+      return copies.back();
+    }
+  };
+  const array& x = check_input(inputs[0]);
+  const array& w = inputs[1];
+  const array& b = inputs[2];
+
+  if (x.is_donatable()) {
+    out.move_shared_buffer(x);
+  } else {
+    out.set_data(
+        allocator::malloc_or_wait(x.data_size() * x.itemsize()),
+        x.data_size(),
+        x.strides(),
+        x.flags());
+  }
+
+  auto axis_size = static_cast<uint32_t>(x.shape().back());
+  int n_rows = x.data_size() / axis_size;
+
+  const int simd_size = 32;
+  const int n_reads = RMS_N_READS;
+  const int looped_limit = RMS_LOOPED_LIMIT;
+  std::string op_name = "layer_norm";
+  if (axis_size > looped_limit) {
+    op_name += "_looped";
+  }
+  op_name += type_to_name(out);
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  {
+    auto kernel = d.get_kernel(op_name);
+
+    MTL::Size grid_dims, group_dims;
+    if (axis_size <= looped_limit) {
+      size_t threadgroup_needed = (axis_size + n_reads - 1) / n_reads;
+      size_t simds_needed = (threadgroup_needed + simd_size - 1) / simd_size;
+      size_t threadgroup_size = simd_size * simds_needed;
+      assert(threadgroup_size <= kernel->maxTotalThreadsPerThreadgroup());
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    } else {
+      size_t threadgroup_size = kernel->maxTotalThreadsPerThreadgroup();
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    }
+
+    uint32_t w_stride = (w.ndim() == 1) ? w.strides()[0] : 0;
+    uint32_t b_stride = (b.ndim() == 1) ? b.strides()[0] : 0;
+    compute_encoder->setComputePipelineState(kernel);
+    compute_encoder.set_input_array(
+        x.data_shared_ptr() == nullptr ? out : x, 0);
+    compute_encoder.set_input_array(w, 1);
+    compute_encoder.set_input_array(b, 2);
+    compute_encoder.set_output_array(out, 3);
+    compute_encoder->setBytes(&eps_, sizeof(float), 4);
+    compute_encoder->setBytes(&axis_size, sizeof(int), 5);
+    compute_encoder->setBytes(&w_stride, sizeof(uint32_t), 6);
+    compute_encoder->setBytes(&b_stride, sizeof(uint32_t), 7);
+    compute_encoder->dispatchThreads(grid_dims, group_dims);
+  }
+  d.get_command_buffer(s.index)->addCompletedHandler(
+      [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
+}
+
+void LayerNormVJP::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& s = stream();
+  auto& d = metal::device(s.device);
+
+  // Ensure row contiguity. We could relax this step by checking that the array
+  // is contiguous (no broadcasts or holes) and that the input strides are the
+  // same as the cotangent strides but for now this is simpler.
+  std::vector<array> copies;
+  auto check_input = [&copies, &s](const array& x) -> const array& {
+    if (x.flags().row_contiguous) {
+      return x;
+    }
+    copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
+    copy_gpu(x, copies.back(), CopyType::General, s);
+    return copies.back();
+  };
+  const array& x = check_input(inputs[0]);
+  const array& w = inputs[1];
+  const array& b = inputs[2];
+  const array& g = check_input(inputs[3]);
+  array& gx = outputs[0];
+  array& gw = outputs[1];
+  array& gb = outputs[2];
+
+  // Allocate space for the outputs
+  bool x_in_gx = false;
+  bool g_in_gx = false;
+  if (x.is_donatable()) {
+    gx.move_shared_buffer(x);
+    x_in_gx = true;
+  } else if (g.is_donatable()) {
+    gx.move_shared_buffer(g);
+    g_in_gx = true;
+  } else {
+    gx.set_data(allocator::malloc_or_wait(gx.nbytes()));
+  }
+
+  auto axis_size = static_cast<uint32_t>(x.shape().back());
+  int n_rows = x.data_size() / axis_size;
+
+  // Allocate a temporary to store the gradients for w and initialize the
+  // gradient accumulator to 0.
+  array gw_temp({n_rows, x.shape().back()}, gw.dtype(), nullptr, {});
+  bool g_in_gw = false;
+  if (!g_in_gx && g.is_donatable()) {
+    gw_temp.move_shared_buffer(g);
+    g_in_gw = true;
+  } else {
+    gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes()));
+  }
+  copies.push_back(gw_temp);
+  {
+    array zero(0, gw.dtype());
+    copy_gpu(zero, gw, CopyType::Scalar, s);
+    copy_gpu(zero, gb, CopyType::Scalar, s);
+    copies.push_back(std::move(zero));
+  }
+
+  // Finish with the gradient for b in case we had a b
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  if (gb.ndim() == 1 && gb.size() == axis_size) {
+    ReductionPlan plan(
+        ReductionOpType::ContiguousStridedReduce, {n_rows}, {axis_size});
+    strided_reduce_general_dispatch(
+        g_in_gx ? gx : (g_in_gw ? gw_temp : g),
+        gb,
+        "sum",
+        plan,
+        {0},
+        compute_encoder,
+        d,
+        s);
+  }
+
+  const int simd_size = 32;
+  const int n_reads = RMS_N_READS;
+  const int looped_limit = RMS_LOOPED_LIMIT;
+  std::string op_name = "vjp_layer_norm";
+  if (axis_size > looped_limit) {
+    op_name += "_looped";
+  }
+  op_name += type_to_name(gx);
+  {
+    auto kernel = d.get_kernel(op_name);
+
+    MTL::Size grid_dims, group_dims;
+    if (axis_size <= looped_limit) {
+      size_t threadgroup_needed = (axis_size + n_reads - 1) / n_reads;
+      size_t simds_needed = (threadgroup_needed + simd_size - 1) / simd_size;
+      size_t threadgroup_size = simd_size * simds_needed;
+      assert(threadgroup_size <= kernel->maxTotalThreadsPerThreadgroup());
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    } else {
+      size_t threadgroup_size = kernel->maxTotalThreadsPerThreadgroup();
+      size_t n_threads = n_rows * threadgroup_size;
+      grid_dims = MTL::Size(n_threads, 1, 1);
+      group_dims = MTL::Size(threadgroup_size, 1, 1);
+    }
+
+    uint32_t w_stride = (w.ndim() == 1) ? w.strides()[0] : 0;
+    compute_encoder->setComputePipelineState(kernel);
+    compute_encoder.set_input_array(x_in_gx ? gx : x, 0);
+    compute_encoder.set_input_array(w, 1);
+    compute_encoder.set_input_array(g_in_gx ? gx : (g_in_gw ? gw_temp : g), 2);
+    compute_encoder.set_output_array(gx, 3);
+    compute_encoder.set_output_array(gw_temp, 4);
+    compute_encoder->setBytes(&eps_, sizeof(float), 5);
+    compute_encoder->setBytes(&axis_size, sizeof(int), 6);
+    compute_encoder->setBytes(&w_stride, sizeof(uint32_t), 7);
+    compute_encoder->dispatchThreads(grid_dims, group_dims);
+  }
+
+  if (gw.ndim() == 1 && gw.size() == axis_size) {
+    ReductionPlan plan(
+        ReductionOpType::ContiguousStridedReduce, {n_rows}, {axis_size});
+    strided_reduce_general_dispatch(
+        gw_temp, gw, "sum", plan, {0}, compute_encoder, d, s);
+  }
+
+  d.get_command_buffer(s.index)->addCompletedHandler(
+      [copies](MTL::CommandBuffer*) mutable { copies.clear(); });
+}
+
+} // namespace mlx::core::fast

mlx/backend/metal/primitives.cpp

@@ -17,7 +17,7 @@ namespace mlx::core {
 
 namespace {
 
-static constexpr int METAL_MAX_INDEX_ARRAYS = 10;
+constexpr int METAL_MAX_INDEX_ARRAYS = 10;
 
 void binary_op(
     const std::vector<array>& inputs,
@@ -68,18 +68,18 @@ void binary_op(
   auto& s = out.primitive().stream();
   auto& d = metal::device(s.device);
   auto kernel = d.get_kernel(kname.str());
-  auto compute_encoder = d.get_command_encoder(s.index);
+  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;
-  set_array_buffer(compute_encoder, donate_a ? outputs[0] : a, 0);
-  set_array_buffer(
-      compute_encoder, donate_b ? (donate_a ? outputs[1] : outputs[0]) : b, 1);
-  set_array_buffer(compute_encoder, outputs[0], 2);
-  set_array_buffer(compute_encoder, outputs[1], 3);
+  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();
@@ -167,13 +167,13 @@ void binary_op(
   auto& s = out.primitive().stream();
   auto& d = metal::device(s.device);
   auto kernel = d.get_kernel(kname.str());
-  auto compute_encoder = d.get_command_encoder(s.index);
+  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;
-  set_array_buffer(compute_encoder, donate_a ? out : a, 0);
-  set_array_buffer(compute_encoder, donate_b ? out : b, 1);
-  set_array_buffer(compute_encoder, out, 2);
+  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();
@@ -253,12 +253,12 @@ void ternary_op(
   auto& s = out.primitive().stream();
   auto& d = metal::device(s.device);
   auto kernel = d.get_kernel(kname.str());
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
-  set_array_buffer(compute_encoder, a, 0);
-  set_array_buffer(compute_encoder, b, 1);
-  set_array_buffer(compute_encoder, c, 2);
-  set_array_buffer(compute_encoder, out, 3);
+  compute_encoder.set_input_array(a, 0);
+  compute_encoder.set_input_array(b, 1);
+  compute_encoder.set_input_array(c, 2);
+  compute_encoder.set_output_array(out, 3);
 
   if (topt == TernaryOpType::General) {
     auto ndim = shape.size();
@@ -339,11 +339,11 @@ void unary_op(
   }
   MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
 
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
-  set_array_buffer(
-      compute_encoder, in.data_shared_ptr() == nullptr ? out : in, 0);
-  set_array_buffer(compute_encoder, out, 1);
+  compute_encoder.set_input_array(
+      in.data_shared_ptr() == nullptr ? out : in, 0);
+  compute_encoder.set_output_array(out, 1);
   if (!contig) {
     compute_encoder->setBytes(in.shape().data(), in.ndim() * sizeof(int), 2);
     compute_encoder->setBytes(
@@ -365,7 +365,7 @@ void Add::eval_gpu(const std::vector<array>& inputs, array& out) {
 }
 
 template <typename T>
-void arange_set_scalars(T start, T next, MTL::ComputeCommandEncoder* enc) {
+void arange_set_scalars(T start, T next, CommandEncoder& enc) {
   enc->setBytes(&start, sizeof(T), 0);
   T step = next - start;
   enc->setBytes(&step, sizeof(T), 1);
@@ -384,7 +384,7 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   MTL::Size grid_dims = MTL::Size(nthreads, 1, 1);
   MTL::Size group_dims = MTL::Size(
       std::min(nthreads, kernel->maxTotalThreadsPerThreadgroup()), 1, 1);
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
 
   switch (out.dtype()) {
@@ -427,7 +427,7 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
       throw std::runtime_error("[Arange::eval_gpu] Does not support complex64");
   }
 
-  set_array_buffer(compute_encoder, out, 2);
+  compute_encoder.set_output_array(out, 2);
   compute_encoder->dispatchThreads(grid_dims, group_dims);
 }
 
@@ -487,7 +487,7 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   // ArgReduce
   int simd_size = 32;
   int n_reads = 4;
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   {
     auto kernel = d.get_kernel(op_name + type_to_name(in));
     NS::UInteger thread_group_size = std::min(
@@ -502,8 +502,8 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
     MTL::Size grid_dims = MTL::Size(n_threads, 1, 1);
     MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
     compute_encoder->setComputePipelineState(kernel);
-    set_array_buffer(compute_encoder, in, 0);
-    set_array_buffer(compute_encoder, out, 1);
+    compute_encoder.set_input_array(in, 0);
+    compute_encoder.set_output_array(out, 1);
     if (ndim == 0) {
       // Pass place holders so metal doesn't complain
       int shape_ = 0;
@@ -552,6 +552,9 @@ void Concatenate::eval_gpu(const std::vector<array>& inputs, array& out) {
   flags.row_contiguous = false;
   flags.col_contiguous = false;
   flags.contiguous = false;
+  auto& d = metal::device(stream().device);
+  auto& compute_encoder = d.get_command_encoder(stream().index);
+  auto concurrent_ctx = compute_encoder.start_concurrent();
   for (int i = 0; i < inputs.size(); i++) {
     array out_slice(inputs[i].shape(), out.dtype(), nullptr, {});
     size_t data_offset = strides[axis_] * sizes[i];
@@ -615,6 +618,10 @@ void Exp::eval_gpu(const std::vector<array>& inputs, array& out) {
   unary_op(inputs, out, "exp");
 }
 
+void Expm1::eval_gpu(const std::vector<array>& inputs, array& out) {
+  unary_op(inputs, out, "expm1");
+}
+
 void Full::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto in = inputs[0];
   CopyType ctype;
@@ -696,6 +703,10 @@ void Minimum::eval_gpu(const std::vector<array>& inputs, array& out) {
   binary_op(inputs, out, "min");
 }
 
+void NumberOfElements::eval_gpu(const std::vector<array>& inputs, array& out) {
+  eval(inputs, out);
+}
+
 void Floor::eval_gpu(const std::vector<array>& inputs, array& out) {
   unary_op(inputs, out, "floor");
 }
@@ -783,10 +794,10 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   MTL::Size grid_dims = MTL::Size(num_keys, half_size + odd, 1);
   NS::UInteger thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
   MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
-  auto compute_encoder = d.get_command_encoder(s.index);
+  auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder->setComputePipelineState(kernel);
-  set_array_buffer(compute_encoder, keys, 0);
-  set_array_buffer(compute_encoder, out, 1);
+  compute_encoder.set_input_array(keys, 0);
+  compute_encoder.set_output_array(out, 1);
   compute_encoder->setBytes(&odd, sizeof(bool), 2);
   compute_encoder->setBytes(&bytes_per_key, sizeof(size_t), 3);
 
@@ -805,20 +816,20 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
 void Reshape::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (in.flags().row_contiguous) {
-    auto flags = in.flags();
-    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());
-  } else {
+
+  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
+
+  if (copy_necessary) {
     copy_gpu(in, out, CopyType::General);
+  } else {
+    shared_buffer_reshape(in, out_strides, out);
   }
 }
 
 void Round::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  if (not is_integral(in.dtype())) {
+  if (issubdtype(in.dtype(), inexact)) {
     unary_op(inputs, out, "round");
   } else {
     // No-op integer types
@@ -861,7 +872,73 @@ void Sqrt::eval_gpu(const std::vector<array>& inputs, array& out) {
 }
 
 void Slice::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
+  assert(inputs.size() == 1);
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  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);
+
+  // Do copy if needed
+  if (copy_needed) {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    std::vector<int64_t> ostrides{out.strides().begin(), out.strides().end()};
+    copy_gpu_inplace(
+        /* const array& in = */ in,
+        /* array& out = */ out,
+        /* const std::vector<int>& data_shape = */ out.shape(),
+        /* const std::vector<stride_t>& i_strides = */ inp_strides,
+        /* const std::vector<stride_t>& o_strides = */ ostrides,
+        /* int64_t i_offset = */ data_offset,
+        /* int64_t o_offset = */ 0,
+        /* CopyType ctype = */ CopyType::General,
+        /* const Stream& s = */ stream());
+  } else {
+    std::vector<size_t> ostrides{inp_strides.begin(), inp_strides.end()};
+    shared_buffer_slice(in, ostrides, data_offset, out);
+  }
+}
+
+void SliceUpdate::eval_gpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 2);
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  auto& in = inputs[0];
+  auto& upd = inputs[1];
+
+  if (upd.size() == 0) {
+    out.copy_shared_buffer(in);
+    return;
+  }
+
+  // Check if materialization is needed
+  auto ctype = in.flags().contiguous && in.size() == in.data_size()
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_gpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+
+  // Calculate out strides, initial offset and if copy needs to be made
+  auto [data_offset, out_strides] = prepare_slice(out);
+
+  // Do copy
+  std::vector<int64_t> upd_strides{upd.strides().begin(), upd.strides().end()};
+  copy_gpu_inplace<int64_t>(
+      /* const array& src = */ upd,
+      /* array& dst = */ out,
+      /* const std::vector<int>& data_shape = */ upd.shape(),
+      /* const std::vector<stride_t>& i_strides = */ upd_strides,
+      /* const std::vector<stride_t>& o_strides = */ out_strides,
+      /* int64_t i_offset = */ 0,
+      /* int64_t o_offset = */ data_offset,
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      /* const Stream& s = */ stream());
 }
 
 void StopGradient::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -890,4 +967,14 @@ void QRF::eval_gpu(
   throw std::runtime_error("[QRF::eval_gpu] Metal QR factorization NYI.");
 }
 
+void SVD::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  throw std::runtime_error("[SVD::eval_gpu] Metal SVD NYI.");
+}
+
+void Inverse::eval_gpu(const std::vector<array>& inputs, array& output) {
+  throw std::runtime_error("[Inverse::eval_gpu] Metal inversion NYI.");
+}
+
 } // namespace mlx::core