only interrupt during an eval

* Fix typo; Fix lint warning when reuse the same name

* Add missing period

.circleci/config.yml

@@ -92,10 +92,12 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF
+              CMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF \
+              CMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               python3 setup.py develop
       - run:
@@ -146,7 +148,9 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install -e . -v
+            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
+            CMAKE_ARGS="CMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+              pip install -e . -v
       - run:
           name: Generate package stubs
           command: |

CMakeLists.txt

@@ -9,6 +9,7 @@ if(NOT MLX_VERSION)
   string(REGEX MATCH "#define MLX_VERSION_PATCH ([0-9]+)" _ "${_mlx_h_version}")
   set(_patch ${CMAKE_MATCH_1})
   set(MLX_PROJECT_VERSION "${_major}.${_minor}.${_patch}")
+  set(MLX_VERSION ${MLX_PROJECT_VERSION})
 else()
   string(REGEX REPLACE "^([0-9]+\.[0-9]+\.[0-9]+).*" "\\1" MLX_PROJECT_VERSION
                        ${MLX_VERSION})
@@ -41,8 +42,6 @@ option(MLX_BUILD_BLAS_FROM_SOURCE "Build OpenBLAS from source code" OFF)
 option(MLX_METAL_JIT "Use JIT compilation for Metal kernels" OFF)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
 
-add_compile_definitions("MLX_VERSION=${MLX_VERSION}")
-
 # --------------------- Processor tests -------------------------
 message(
   STATUS
@@ -77,7 +76,6 @@ include(FetchContent)
 cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
-set_target_properties(mlx PROPERTIES COMPILE_WARNING_AS_ERROR ON)
 
 if(MLX_BUILD_METAL)
   set(METAL_LIB "-framework Metal")
@@ -232,6 +230,14 @@ if(MPI_FOUND)
   endif()
 endif()
 
+message(STATUS "Downloading json")
+FetchContent_Declare(
+  json
+  URL https://github.com/nlohmann/json/releases/download/v3.11.3/json.tar.xz)
+FetchContent_MakeAvailable(json)
+target_include_directories(
+  mlx PRIVATE $<BUILD_INTERFACE:${json_SOURCE_DIR}/single_include/nlohmann>)
+
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/mlx)
 
 target_include_directories(

benchmarks/python/gather_bench.py

@@ -1,7 +1,6 @@
 # Copyright © 2023-2024 Apple Inc.
 
 import argparse
-from time import time
 
 import mlx.core as mx
 import torch

docs/src/dev/extensions.rst

@@ -22,12 +22,12 @@ You can do that in MLX directly:
 This function performs that operation while leaving the implementation and
 function transformations to MLX.
 
-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:
+However, you may want to customize the underlying implementation, perhaps to
+make it faster. In this tutorial we will go through adding custom extensions.
+It will cover:
 
 * The structure of the MLX library.
-* Implementing a CPU operation that redirects to Accelerate_ when appropriate.
+* Implementing a CPU operation.
 * Implementing a GPU operation using metal.
 * Adding the ``vjp`` and ``jvp`` function transformation.
 * Building a custom extension and binding it to python.
@@ -45,7 +45,7 @@ Operations
 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
+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++:
 
@@ -55,7 +55,7 @@ C++:
     *  Scale and sum two vectors element-wise
     *  z = alpha * x + beta * y
     *
-    *  Follow numpy style broadcasting between x and y
+    *  Use NumPy-style broadcasting between x and y
     *  Inputs are upcasted to floats if needed
     **/
     array axpby(
@@ -66,7 +66,7 @@ C++:
         StreamOrDevice s = {} // Stream on which to schedule the operation
     );
 
-The simplest way to this operation is in terms of existing operations:
+The simplest way to implement this is with existing operations:
 
 .. code-block:: C++
 
@@ -153,9 +153,6 @@ more concrete:
       private:
         float alpha_;
         float beta_;
-
-        /** Fall back implementation for evaluation on CPU */
-        void eval(const std::vector<array>& inputs, array& out);
     };
 
 The :class:`Axpby` class derives from the base :class:`Primitive` class. The
@@ -188,7 +185,7 @@ Let's reimplement our operation now in terms of our :class:`Axpby` primitive.
         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);
 
@@ -234,49 +231,59 @@ the execution of the computation graph, and calls :meth:`Axpby::eval_cpu` or
 Implementing the CPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-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`.
+Let's start by implementing :meth:`Axpby::eval_cpu`.
 
-Our naive method will go over each element of the output array, find the
+The 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++
 
-    template <typename T>
-    void axpby_impl(
-            const array& x,
-            const array& y,
-            array& out,
-            float alpha_,
-            float beta_) {
-        // We only allocate memory when we are ready to fill the output
-        // malloc_or_wait synchronously allocates available memory
-        // There may be a wait executed here if the allocation is requested
-        // under memory-pressured conditions
-        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  template <typename T>
+  void axpby_impl(
+      const mx::array& x,
+      const mx::array& y,
+      mx::array& out,
+      float alpha_,
+      float beta_,
+      mx::Stream stream) {
+    // Allocate the output with `malloc_or_wait` which synchronously allocates
+    // memory, potentially waiting if the system is under memory pressure
+    out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
 
-        // Collect input and output data pointers
-        const T* x_ptr = x.data<T>();
-        const T* y_ptr = y.data<T>();
-        T* out_ptr = out.data<T>();
+    // Get the CPU command encoder and register input and output arrays
+    auto& encoder = mx::cpu::get_command_encoder(stream);
+    encoder.set_input_array(x);
+    encoder.set_input_array(y);
+    encoder.set_output_array(out);
 
-        // Cast alpha and beta to the relevant types
-        T alpha = static_cast<T>(alpha_);
-        T beta = static_cast<T>(beta_);
+    // Launch the CPU kernel
+    encoder.dispatch([x_ptr = x.data<T>(),
+                      y_ptr = y.data<T>(),
+                      out_ptr = out.data<T>(),
+                      size = out.size(),
+                      shape = out.shape(),
+                      x_strides = x.strides(),
+                      y_strides = y.strides(),
+                      alpha_,
+                      beta_]() {
 
-        // Do the element-wise operation for each output
-        for (size_t out_idx = 0; out_idx < out.size(); out_idx++) {
-            // Map linear indices to offsets in x and y
-            auto x_offset = elem_to_loc(out_idx, x.shape(), x.strides());
-            auto y_offset = elem_to_loc(out_idx, y.shape(), y.strides());
+      // Cast alpha and beta to the relevant types
+      T alpha = static_cast<T>(alpha_);
+      T beta = static_cast<T>(beta_);
 
-            // We allocate the output to be contiguous and regularly strided
-            // (defaults to row major) and hence it doesn't need additional mapping
-            out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
-        }
-    }
+      // Do the element-wise operation for each output
+      for (size_t out_idx = 0; out_idx < size; out_idx++) {
+        // Map linear indices to offsets in x and y
+        auto x_offset = mx::elem_to_loc(out_idx, shape, x_strides);
+        auto y_offset = mx::elem_to_loc(out_idx, shape, y_strides);
+
+        // We allocate the output to be contiguous and regularly strided
+        // (defaults to row major) and hence it doesn't need additional mapping
+        out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
+      }
+    });
+  }
 
 Our implementation should work for all incoming floating point arrays.
 Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
@@ -284,112 +291,32 @@ Accordingly, we add dispatches for ``float32``, ``float16``, ``bfloat16`` and
 
 .. code-block:: C++
 
-    /** Fall back implementation for evaluation on CPU */
-    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) {
-            return axpby_impl<float>(x, y, out, alpha_, beta_);
-        } else if (out.dtype() == float16) {
-            return axpby_impl<float16_t>(x, y, out, alpha_, beta_);
-        } else if (out.dtype() == bfloat16) {
-            return axpby_impl<bfloat16_t>(x, y, out, alpha_, beta_);
-        } else if (out.dtype() == complex64) {
-            return axpby_impl<complex64_t>(x, y, out, alpha_, beta_);
-        } else {
-            throw std::runtime_error(
-                "[Axpby] Only supports floating point types.");
-        }
-    }
-
-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 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 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++
-
-    template <typename T>
-    void axpby_impl_accelerate(
-            const array& x,
-            const array& y,
-            array& out,
-            float alpha_,
-            float beta_) {
-        // 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
-        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);
-
-        // Get x and y pointers for catlas_saxpby
-        const T* x_ptr = x.data<T>();
-        T* y_ptr = out.data<T>();
-
-        T alpha = static_cast<T>(alpha_);
-        T beta = static_cast<T>(beta_);
-
-        // Call the inplace accelerate operator
-        catlas_saxpby(
-            /* N = */ out.size(),
-            /* ALPHA = */ alpha,
-            /* X = */ x_ptr,
-            /* INCX = */ 1,
-            /* BETA = */ beta,
-            /* Y = */ y_ptr,
-            /* INCY = */ 1);
-    }
-
-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,
-      const std::vector<array>& outputs) {
-        assert(inputs.size() == 2);
-        auto& x = inputs[0];
-        auto& y = inputs[1];
-        auto& out = outputs[0];
+        const std::vector<mx::array>& inputs,
+        std::vector<mx::array>& outputs) {
+      auto& x = inputs[0];
+      auto& y = inputs[1];
+      auto& out = outputs[0];
 
-        // Accelerate specialization for contiguous single precision float arrays
-        if (out.dtype() == float32 &&
-            ((x.flags().row_contiguous && y.flags().row_contiguous) ||
-            (x.flags().col_contiguous && y.flags().col_contiguous))) {
-            axpby_impl_accelerate<float>(x, y, out, alpha_, beta_);
-            return;
-        }
-
-        // Fall back to common back-end if specializations are not available
-        eval(inputs, outputs);
+      // Dispatch to the correct dtype
+      if (out.dtype() == mx::float32) {
+        return axpby_impl<float>(x, y, out, alpha_, beta_, stream());
+      } else if (out.dtype() == mx::float16) {
+        return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_, stream());
+      } else if (out.dtype() == mx::bfloat16) {
+        return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_, stream());
+      } else if (out.dtype() == mx::complex64) {
+        return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_, stream());
+      } else {
+        throw std::runtime_error(
+            "Axpby is only supported for floating point types.");
+      }
     }
 
 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.
+primitive here.
 
 Implementing the GPU Back-end
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -824,7 +751,7 @@ Results
 ^^^^^^^
 
 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.
+with the naive :meth:`simple_axpby` we first defined.
 
 .. code-block:: python
 
@@ -832,13 +759,11 @@ with the naive :meth:`simple_axpby` we first defined on the CPU.
     from mlx_sample_extensions import axpby
     import time
 
-    mx.set_default_device(mx.cpu)
-
     def simple_axpby(x: mx.array, y: mx.array, alpha: float, beta: float) -> mx.array:
         return alpha * x + beta * y
 
-    M = 256
-    N = 512
+    M = 4096
+    N = 4096
 
     x = mx.random.normal((M, N))
     y = mx.random.normal((M, N))
@@ -849,24 +774,24 @@ with the naive :meth:`simple_axpby` we first defined on the CPU.
 
     def bench(f):
         # Warm up
-        for i in range(100):
+        for i in range(5):
             z = f(x, y, alpha, beta)
             mx.eval(z)
 
         # Timed run
         s = time.time()
-        for i in range(5000):
+        for i in range(100):
             z = f(x, y, alpha, beta)
             mx.eval(z)
         e = time.time()
-        return e - s
+        return 1000 * (e - s) / 100
 
     simple_time = bench(simple_axpby)
     custom_time = bench(axpby)
 
-    print(f"Simple axpby: {simple_time:.3f} s | Custom axpby: {custom_time:.3f} s")
+    print(f"Simple axpby: {simple_time:.3f} ms | Custom axpby: {custom_time:.3f} ms")
 
-The results are ``Simple axpby: 0.114 s | Custom axpby: 0.109 s``. We see
+The results are ``Simple axpby: 1.559 ms | Custom axpby: 0.774 ms``. We see
 modest improvements right away!
 
 This operation is now good to be used to build other operations, in

examples/extensions/CMakeLists.txt

@@ -10,7 +10,6 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 option(BUILD_SHARED_LIBS "Build extensions as a shared library" ON)
 
 # ----------------------------- Dependencies -----------------------------
-find_package(MLX CONFIG REQUIRED)
 find_package(
   Python 3.8
   COMPONENTS Interpreter Development.Module
@@ -21,6 +20,12 @@ execute_process(
   OUTPUT_VARIABLE nanobind_ROOT)
 find_package(nanobind CONFIG REQUIRED)
 
+execute_process(
+  COMMAND "${Python_EXECUTABLE}" -m mlx --cmake-dir
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+  OUTPUT_VARIABLE MLX_ROOT)
+find_package(MLX CONFIG REQUIRED)
+
 # ----------------------------- Extensions -----------------------------
 
 # Add library

examples/extensions/axpby/axpby.cpp

@@ -1,20 +1,14 @@
-// Copyright © 2023-2024 Apple Inc.
+// Copyright © 2023-2025 Apple Inc.
 
-#include <cassert>
 #include <iostream>
 #include <sstream>
 
-#include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/utils.h"
 
 #include "axpby/axpby.h"
 
-#ifdef ACCELERATE_NEW_LAPACK
-#include <vecLib/cblas_new.h>
-#endif
-
 #ifdef _METAL_
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/utils.h"
@@ -76,136 +70,67 @@ void axpby_impl(
     const mx::array& y,
     mx::array& out,
     float alpha_,
-    float beta_) {
-  // We only allocate memory when we are ready to fill the output
-  // malloc_or_wait synchronously allocates available memory
-  // There may be a wait executed here if the allocation is requested
-  // under memory-pressured conditions
+    float beta_,
+    mx::Stream stream) {
+  // Allocate the output with `malloc_or_wait` which synchronously allocates
+  // memory, potentially waiting if the system is under memory pressure
   out.set_data(mx::allocator::malloc_or_wait(out.nbytes()));
 
-  // Collect input and output data pointers
-  const T* x_ptr = x.data<T>();
-  const T* y_ptr = y.data<T>();
-  T* out_ptr = out.data<T>();
+  // Get the CPU command encoder and register input and output arrays
+  auto& encoder = mx::cpu::get_command_encoder(stream);
+  encoder.set_input_array(x);
+  encoder.set_input_array(y);
+  encoder.set_output_array(out);
 
-  // Cast alpha and beta to the relevant types
-  T alpha = static_cast<T>(alpha_);
-  T beta = static_cast<T>(beta_);
+  // Launch the CPU kernel
+  encoder.dispatch([x_ptr = x.data<T>(),
+                    y_ptr = y.data<T>(),
+                    out_ptr = out.data<T>(),
+                    size = out.size(),
+                    shape = out.shape(),
+                    x_strides = x.strides(),
+                    y_strides = y.strides(),
+                    alpha_,
+                    beta_]() {
+    // Cast alpha and beta to the relevant types
+    T alpha = static_cast<T>(alpha_);
+    T beta = static_cast<T>(beta_);
 
-  // Do the element-wise operation for each output
-  for (size_t out_idx = 0; out_idx < out.size(); out_idx++) {
-    // Map linear indices to offsets in x and y
-    auto x_offset = mx::elem_to_loc(out_idx, x.shape(), x.strides());
-    auto y_offset = mx::elem_to_loc(out_idx, y.shape(), y.strides());
+    // Do the element-wise operation for each output
+    for (size_t out_idx = 0; out_idx < size; out_idx++) {
+      // Map linear indices to offsets in x and y
+      auto x_offset = mx::elem_to_loc(out_idx, shape, x_strides);
+      auto y_offset = mx::elem_to_loc(out_idx, shape, y_strides);
 
-    // We allocate the output to be contiguous and regularly strided
-    // (defaults to row major) and hence it doesn't need additional mapping
-    out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
-  }
+      // We allocate the output to be contiguous and regularly strided
+      // (defaults to row major) and hence it doesn't need additional mapping
+      out_ptr[out_idx] = alpha * x_ptr[x_offset] + beta * y_ptr[y_offset];
+    }
+  });
 }
 
-/** Fall back implementation for evaluation on CPU */
-void Axpby::eval(
+void Axpby::eval_cpu(
     const std::vector<mx::array>& inputs,
     std::vector<mx::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() == mx::float32) {
-    return axpby_impl<float>(x, y, out, alpha_, beta_);
+    return axpby_impl<float>(x, y, out, alpha_, beta_, stream());
   } else if (out.dtype() == mx::float16) {
-    return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_);
+    return axpby_impl<mx::float16_t>(x, y, out, alpha_, beta_, stream());
   } else if (out.dtype() == mx::bfloat16) {
-    return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_);
+    return axpby_impl<mx::bfloat16_t>(x, y, out, alpha_, beta_, stream());
   } else if (out.dtype() == mx::complex64) {
-    return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_);
+    return axpby_impl<mx::complex64_t>(x, y, out, alpha_, beta_, stream());
   } else {
     throw std::runtime_error(
         "Axpby is only supported for floating point types.");
   }
 }
 
-///////////////////////////////////////////////////////////////////////////////
-// Primitive Accelerate Backend Implementation
-///////////////////////////////////////////////////////////////////////////////
-
-#ifdef ACCELERATE_NEW_LAPACK
-
-template <typename T>
-void axpby_impl_accelerate(
-    const mx::array& x,
-    const mx::array& y,
-    mx::array& out,
-    float alpha_,
-    float beta_) {
-  // 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(mx::allocator::malloc_or_wait(out.nbytes()));
-
-  // We then copy over the elements using the contiguous vector specialization
-  copy_inplace(y, out, mx::CopyType::Vector);
-
-  // Get x and y pointers for catlas_saxpby
-  const T* x_ptr = x.data<T>();
-  T* y_ptr = out.data<T>();
-
-  T alpha = static_cast<T>(alpha_);
-  T beta = static_cast<T>(beta_);
-
-  // Call the inplace accelerate operator
-  catlas_saxpby(
-      /* N = */ out.size(),
-      /* ALPHA = */ alpha,
-      /* X = */ x_ptr,
-      /* INCX = */ 1,
-      /* BETA = */ beta,
-      /* Y = */ y_ptr,
-      /* INCY = */ 1);
-}
-
-/** Evaluate primitive on CPU using accelerate specializations */
-void Axpby::eval_cpu(
-    const std::vector<mx::array>& inputs,
-    std::vector<mx::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() == mx::float32 &&
-      ((x.flags().row_contiguous && y.flags().row_contiguous) ||
-       (x.flags().col_contiguous && y.flags().col_contiguous))) {
-    axpby_impl_accelerate<float>(x, y, out, alpha_, beta_);
-    return;
-  }
-
-  // Fall back to common backend if specializations are not available
-  eval(inputs, outputs);
-}
-
-#else // Accelerate not available
-
-/** Evaluate primitive on CPU falling back to common backend */
-void Axpby::eval_cpu(
-    const std::vector<mx::array>& inputs,
-    std::vector<mx::array>& outputs) {
-  eval(inputs, outputs);
-}
-
-#endif
-
 ///////////////////////////////////////////////////////////////////////////////
 // Primitive Metal Backend Implementation
 ///////////////////////////////////////////////////////////////////////////////
@@ -217,7 +142,6 @@ void Axpby::eval_gpu(
     const std::vector<mx::array>& inputs,
     std::vector<mx::array>& outputs) {
   // Prepare inputs
-  assert(inputs.size() == 2);
   auto& x = inputs[0];
   auto& y = inputs[1];
   auto& out = outputs[0];

examples/extensions/axpby/axpby.h

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2025 Apple Inc.
 
 #pragma once
 
@@ -85,11 +85,6 @@ class Axpby : public mx::Primitive {
  private:
   float alpha_;
   float beta_;
-
-  /** Fall back implementation for evaluation on CPU */
-  void eval(
-      const std::vector<mx::array>& inputs,
-      std::vector<mx::array>& outputs);
 };
 
 } // namespace my_ext

examples/extensions/axpby/axpby.metal

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2025 Apple Inc.
 
 #include <metal_stdlib>
 

mlx/CMakeLists.txt

@@ -17,9 +17,13 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/transforms.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/linalg.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/metal.h)
 
+# Define MLX_VERSION only in the version.cpp file.
+add_library(mlx_version STATIC ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp)
+target_compile_definitions(mlx_version PRIVATE MLX_VERSION="${MLX_VERSION}")
+target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:mlx_version>)
+
 if(MSVC)
   # Disable some MSVC warnings to speed up compilation.
   target_compile_options(mlx PUBLIC /wd4068 /wd4244 /wd4267 /wd4804)

mlx/array.cpp

@@ -56,6 +56,18 @@ std::vector<array> array::make_arrays(
   return outputs;
 }
 
+array array::unsafe_weak_copy(const array& other) {
+  auto cpy = array(other.shape(), other.dtype(), nullptr, {});
+  cpy.set_data(
+      other.buffer(),
+      other.data_size(),
+      other.strides(),
+      other.flags(),
+      [](auto) {});
+  cpy.array_desc_->data_ptr = other.array_desc_->data_ptr;
+  return cpy;
+}
+
 array::array(std::initializer_list<float> data)
     : array_desc_(std::make_shared<ArrayDesc>(
           Shape{static_cast<ShapeElem>(data.size())},
@@ -76,35 +88,27 @@ array::array(allocator::Buffer data, Shape shape, Dtype dtype, Deleter deleter)
   set_data(data, deleter);
 }
 
-array::array(
-    allocator::Buffer data,
-    Shape shape,
-    Dtype dtype,
-    Strides strides,
-    size_t data_size,
-    Flags flags,
-    Deleter deleter)
-    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
-  set_data(data, data_size, std::move(strides), flags, deleter);
-}
-
 void array::detach() {
+  array_desc_->primitive = nullptr;
+  for (auto& s : array_desc_->siblings) {
+    s.array_desc_->primitive = nullptr;
+  }
   for (auto& s : array_desc_->siblings) {
     s.array_desc_->inputs.clear();
     s.array_desc_->siblings.clear();
     s.array_desc_->position = 0;
-    s.array_desc_->primitive = nullptr;
   }
   array_desc_->inputs.clear();
   array_desc_->siblings.clear();
   array_desc_->position = 0;
-  array_desc_->primitive = nullptr;
 }
 
 bool array::is_available() const {
   if (status() == Status::available) {
     return true;
-  } else if (status() == Status::evaluated && event().is_signaled()) {
+  } else if (
+      status() == Status::evaluated &&
+      (!event().valid() || event().is_signaled())) {
     set_status(Status::available);
     return true;
   }
@@ -113,7 +117,10 @@ bool array::is_available() const {
 
 void array::wait() {
   if (!is_available()) {
-    event().wait();
+    if (event().valid()) {
+      event().wait();
+      detach_event();
+    }
     set_status(Status::available);
   }
 }
@@ -174,34 +181,13 @@ 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,
-    const Strides& strides,
-    Flags flags,
-    size_t data_size,
-    size_t offset /* = 0 */) {
-  array_desc_->data = std::move(other.array_desc_->data);
-  array_desc_->strides = strides;
-  array_desc_->flags = flags;
-  array_desc_->data_size = data_size;
-  auto char_offset = sizeof(char) * itemsize() * offset;
-  auto data_ptr = other.array_desc_->data_ptr;
-  other.array_desc_->data_ptr = nullptr;
-  array_desc_->data_ptr =
-      static_cast<void*>(static_cast<char*>(data_ptr) + char_offset);
-}
-
-void array::move_shared_buffer(array other) {
-  move_shared_buffer(other, other.strides(), other.flags(), other.data_size());
-}
-
 array::~array() {
   if (array_desc_ == nullptr) {
     return;
   }
 
-  // Ignore arrays that might be detached during eval
-  if (status() == array::Status::scheduled) {
+  // Detached/detaching
+  if (array_desc_->primitive == nullptr) {
     return;
   }
 

mlx/array.h

@@ -199,6 +199,13 @@ class array {
       const std::shared_ptr<Primitive>& primitive,
       const std::vector<array>& inputs);
 
+  /**
+   * Get a new array that refers to the same data as the input but with a
+   * non-owning pointer to it. Note the array is detached from the graph and has
+   * no inputs, siblings or primitive.
+   */
+  static array unsafe_weak_copy(const array& other);
+
   /** A unique identifier for an array. */
   std::uintptr_t id() const {
     return reinterpret_cast<std::uintptr_t>(array_desc_.get());
@@ -243,18 +250,6 @@ class array {
     bool col_contiguous : 1;
   };
 
-  /** Build an array from all the info held by the array description. Including
-   * the buffer, strides, flags.
-   */
-  explicit array(
-      allocator::Buffer data,
-      Shape shape,
-      Dtype dtype,
-      Strides strides,
-      size_t data_size,
-      Flags flags,
-      Deleter deleter = allocator::free);
-
   /** The array's primitive. */
   Primitive& primitive() const {
     return *(array_desc_->primitive);
@@ -365,11 +360,6 @@ class array {
     // For example, the status of `x` in `auto x = a + b`.
     unscheduled,
 
-    // The ouptut of a computation which has been scheduled but `eval_*` has
-    // not yet been called on the array's primitive. A possible
-    // status of `x` in `auto x = a + b; eval(x);`
-    scheduled,
-
     // The array's `eval_*` function has been run, but the computation is not
     // necessarily complete. The array will have memory allocated and if it is
     // not a tracer then it will be detached from the graph.
@@ -406,6 +396,10 @@ class array {
     array_desc_->event = std::move(e);
   }
 
+  void detach_event() const {
+    array_desc_->event = Event{};
+  }
+
   // Mark the array as a tracer array (true) or not.
   void set_tracer(bool is_tracer) {
     array_desc_->is_tracer = is_tracer;
@@ -431,15 +425,6 @@ class array {
 
   void copy_shared_buffer(const array& other);
 
-  void move_shared_buffer(
-      array other,
-      const Strides& strides,
-      Flags flags,
-      size_t data_size,
-      size_t offset = 0);
-
-  void move_shared_buffer(array other);
-
   void overwrite_descriptor(const array& other) {
     array_desc_ = other.array_desc_;
   }

mlx/backend/common/binary.h

@@ -38,8 +38,7 @@ inline void set_binary_op_output_data(
     const array& a,
     const array& b,
     array& out,
-    BinaryOpType bopt,
-    bool donate_with_move = false) {
+    BinaryOpType bopt) {
   bool b_donatable = is_donatable(b, out);
   bool a_donatable = is_donatable(a, out);
   switch (bopt) {
@@ -49,11 +48,7 @@ inline void set_binary_op_output_data(
       break;
     case BinaryOpType::ScalarVector:
       if (b_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(b);
-        } else {
-          out.copy_shared_buffer(b);
-        }
+        out.copy_shared_buffer(b);
       } else {
         out.set_data(
             allocator::malloc_or_wait(b.data_size() * out.itemsize()),
@@ -64,11 +59,7 @@ inline void set_binary_op_output_data(
       break;
     case BinaryOpType::VectorScalar:
       if (a_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(a);
-        } else {
-          out.copy_shared_buffer(a);
-        }
+        out.copy_shared_buffer(a);
       } else {
         out.set_data(
             allocator::malloc_or_wait(a.data_size() * out.itemsize()),
@@ -79,17 +70,9 @@ inline void set_binary_op_output_data(
       break;
     case BinaryOpType::VectorVector:
       if (a_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(a);
-        } else {
-          out.copy_shared_buffer(a);
-        }
+        out.copy_shared_buffer(a);
       } else if (b_donatable) {
-        if (donate_with_move) {
-          out.move_shared_buffer(b);
-        } else {
-          out.copy_shared_buffer(b);
-        }
+        out.copy_shared_buffer(b);
       } else {
         out.set_data(
             allocator::malloc_or_wait(a.data_size() * out.itemsize()),
@@ -100,18 +83,10 @@ inline void set_binary_op_output_data(
       break;
     case BinaryOpType::General:
       if (a_donatable && a.flags().row_contiguous && a.size() == out.size()) {
-        if (donate_with_move) {
-          out.move_shared_buffer(a);
-        } else {
-          out.copy_shared_buffer(a);
-        }
+        out.copy_shared_buffer(a);
       } else if (
           b_donatable && b.flags().row_contiguous && b.size() == out.size()) {
-        if (donate_with_move) {
-          out.move_shared_buffer(b);
-        } else {
-          out.copy_shared_buffer(b);
-        }
+        out.copy_shared_buffer(b);
       } else {
         out.set_data(allocator::malloc_or_wait(out.nbytes()));
       }

mlx/backend/common/common.cpp

@@ -39,7 +39,7 @@ void AsStrided::eval(const std::vector<array>& inputs, array& out) {
   // rely on data_size anyway.
   size_t data_size = out.size();
 
-  return move_or_copy(in, out, strides_, flags, data_size, offset_);
+  return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
 }
 
 void broadcast(const array& in, array& out) {
@@ -56,7 +56,7 @@ void broadcast(const array& in, array& out) {
   if (out.size() > in.size()) {
     flags.row_contiguous = flags.col_contiguous = false;
   }
-  move_or_copy(in, out, strides, flags, in.data_size());
+  out.copy_shared_buffer(in, strides, flags, in.data_size());
 }
 
 void Broadcast::eval(const std::vector<array>& inputs, array& out) {
@@ -69,7 +69,7 @@ void BroadcastAxes::eval(const std::vector<array>& inputs, array& out) {
 
 void Copy::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  move_or_copy(inputs[0], out);
+  out.copy_shared_buffer(inputs[0]);
 }
 
 void CustomTransforms::eval(
@@ -78,7 +78,7 @@ void CustomTransforms::eval(
   assert(inputs.size() > outputs.size());
   for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
        i++, j++) {
-    move_or_copy(inputs[j], outputs[i]);
+    outputs[i].copy_shared_buffer(inputs[j]);
   }
 }
 
@@ -87,7 +87,7 @@ void Depends::eval(
     std::vector<array>& outputs) {
   assert(inputs.size() > outputs.size());
   for (int i = 0; i < outputs.size(); i++) {
-    move_or_copy(inputs[i], outputs[i]);
+    outputs[i].copy_shared_buffer(inputs[i]);
   }
 }
 
@@ -98,7 +98,7 @@ void ExpandDims::eval(const std::vector<array>& inputs, array& out) {
   for (auto ax : axes_) {
     strides.insert(strides.begin() + ax, 1);
   }
-  move_or_copy(in, out, strides, in.flags(), in.data_size());
+  out.copy_shared_buffer(in, strides, in.flags(), in.data_size());
 }
 
 void NumberOfElements::eval(const std::vector<array>& inputs, array& out) {
@@ -210,7 +210,7 @@ void shared_buffer_reshape(
     auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
     flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
   }
-  move_or_copy(in, out, out_strides, flags, in.data_size());
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
 }
 
 void Split::eval(
@@ -276,12 +276,12 @@ void Squeeze::eval(const std::vector<array>& inputs, array& out) {
       strides.push_back(in.strides(i));
     }
   }
-  move_or_copy(in, out, strides, in.flags(), in.data_size());
+  out.copy_shared_buffer(in, strides, in.flags(), in.data_size());
 }
 
 void StopGradient::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  move_or_copy(inputs[0], out);
+  out.copy_shared_buffer(inputs[0]);
 }
 
 void Transpose::eval(const std::vector<array>& inputs, array& out) {
@@ -315,7 +315,7 @@ void Transpose::eval(const std::vector<array>& inputs, array& out) {
       b_stride *= out.shape(ri);
     }
   }
-  move_or_copy(in, out, out_strides, flags, in.data_size());
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
 }
 
 } // namespace mlx::core

mlx/backend/common/compiled.cpp

@@ -161,8 +161,7 @@ void compiled_allocate_outputs(
     std::vector<array>& outputs,
     const std::vector<array>& inputs_,
     const std::unordered_set<uintptr_t>& constant_ids_,
-    bool contiguous,
-    bool move_buffers /* = false */) {
+    bool contiguous) {
   if (contiguous) {
     int o = 0;
     Strides strides;
@@ -178,11 +177,7 @@ void compiled_allocate_outputs(
       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);
-        }
+        outputs[o++].copy_shared_buffer(in);
       }
       // Get representative input flags to properly set non-donated outputs
       if (strides.empty() && in.size() == outputs[0].size()) {
@@ -210,13 +205,8 @@ void compiled_allocate_outputs(
       if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
           in.itemsize() == outputs[o].itemsize() && in.is_donatable() &&
           constant_ids_.find(inputs_[i].id()) == constant_ids_.end()) {
-        if (move_buffers) {
-          outputs[o].move_shared_buffer(
-              in, outputs[o].strides(), in.flags(), in.data_size());
-        } else {
-          outputs[o].copy_shared_buffer(
-              in, outputs[o].strides(), in.flags(), in.data_size());
-        }
+        outputs[o].copy_shared_buffer(
+            in, outputs[o].strides(), in.flags(), in.data_size());
         o++;
       }
     }

mlx/backend/common/compiled.h

@@ -62,7 +62,6 @@ void compiled_allocate_outputs(
     std::vector<array>& outputs,
     const std::vector<array>& inputs_,
     const std::unordered_set<uintptr_t>& constant_ids_,
-    bool contiguous,
-    bool move_buffers = false);
+    bool contiguous);
 
 } // namespace mlx::core

mlx/backend/common/copy.h

@@ -22,4 +22,25 @@ enum class CopyType {
   GeneralGeneral
 };
 
+inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
+  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.copy_shared_buffer(in);
+      return true;
+    } else {
+      out.set_data(
+          allocator::malloc_or_wait(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+      return false;
+    }
+  } else {
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    return false;
+  }
+}
+
 } // namespace mlx::core

mlx/backend/common/load.cpp

@@ -3,7 +3,8 @@
 #include <algorithm>
 #include <utility>
 
-#include "mlx/backend/common/load.h"
+#include "mlx/primitives.h"
+#include "mlx/scheduler.h"
 
 namespace {
 
@@ -26,26 +27,31 @@ void swap_endianness(uint8_t* data_bytes, size_t N) {
 
 namespace mlx::core {
 
-void load(
-    array& out,
-    size_t offset,
-    const std::shared_ptr<io::Reader>& reader,
-    bool swap_endianness_) {
-  reader->read(out.data<char>(), out.nbytes(), offset);
-
-  if (swap_endianness_) {
-    switch (out.itemsize()) {
-      case 2:
-        swap_endianness<2>(out.data<uint8_t>(), out.data_size());
-        break;
-      case 4:
-        swap_endianness<4>(out.data<uint8_t>(), out.data_size());
-        break;
-      case 8:
-        swap_endianness<8>(out.data<uint8_t>(), out.data_size());
-        break;
+void Load::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  auto read_task = [out_ptr = out.data<char>(),
+                    size = out.size(),
+                    itemsize = out.itemsize(),
+                    offset = offset_,
+                    reader = reader_,
+                    swap_endianness_ = swap_endianness_]() mutable {
+    reader->read(out_ptr, size * itemsize, offset);
+    if (swap_endianness_) {
+      switch (itemsize) {
+        case 2:
+          swap_endianness<2>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+        case 4:
+          swap_endianness<4>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+        case 8:
+          swap_endianness<8>(reinterpret_cast<uint8_t*>(out_ptr), size);
+          break;
+      }
     }
-  }
+  };
+  auto fut = io::thread_pool().enqueue(std::move(read_task)).share();
+  scheduler::enqueue(stream(), [fut = std::move(fut)]() { fut.wait(); });
 }
 
 } // namespace mlx::core

mlx/backend/common/load.h

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

mlx/backend/common/slicing.cpp

@@ -36,7 +36,7 @@ void shared_buffer_slice(
   flags.col_contiguous = is_col_contiguous;
   flags.contiguous = (no_bsx_size == data_size);
 
-  move_or_copy(in, out, out_strides, flags, data_size, data_offset);
+  out.copy_shared_buffer(in, out_strides, flags, data_size, data_offset);
 }
 
 void slice(

mlx/backend/common/ternary.h

@@ -36,15 +36,10 @@ inline void set_ternary_op_output_data(
     const array& b,
     const array& c,
     array& out,
-    TernaryOpType topt,
-    bool donate_with_move = false) {
-  auto maybe_donate = [&out, donate_with_move](const array& x) {
+    TernaryOpType topt) {
+  auto maybe_donate = [&out](const array& x) {
     if (is_donatable(x, out)) {
-      if (donate_with_move) {
-        out.move_shared_buffer(x);
-      } else {
-        out.copy_shared_buffer(x);
-      }
+      out.copy_shared_buffer(x);
       return true;
     }
     return false;

mlx/backend/common/utils.cpp

@@ -4,28 +4,6 @@
 
 namespace mlx::core {
 
-void move_or_copy(const array& in, array& out) {
-  if (in.is_donatable()) {
-    out.move_shared_buffer(in);
-  } else {
-    out.copy_shared_buffer(in);
-  }
-}
-
-void move_or_copy(
-    const array& in,
-    array& out,
-    const Strides& strides,
-    array::Flags flags,
-    size_t data_size,
-    size_t offset /* = 0 */) {
-  if (in.is_donatable()) {
-    out.move_shared_buffer(in, strides, flags, data_size, offset);
-  } else {
-    out.copy_shared_buffer(in, strides, flags, data_size, offset);
-  }
-}
-
 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
     const Shape& shape,
     const std::vector<Strides>& strides,

mlx/backend/common/utils.h

@@ -159,15 +159,6 @@ inline bool is_donatable(const array& in, const array& out) {
       in.buffer_size() <= out.nbytes() + donation_extra;
 }
 
-void move_or_copy(const array& in, array& out);
-void move_or_copy(
-    const array& in,
-    array& out,
-    const Strides& strides,
-    array::Flags flags,
-    size_t data_size,
-    size_t offset = 0);
-
 std::pair<bool, Strides> prepare_reshape(const array& in, const array& out);
 
 void shared_buffer_reshape(

mlx/backend/cpu/CMakeLists.txt

@@ -44,7 +44,9 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/eigh.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/encoder.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
@@ -65,6 +67,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/inverse.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/cholesky.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/unary.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_BINARY_DIR}/compiled_preamble.cpp)
 
 if(MLX_BUILD_ACCELERATE)

mlx/backend/cpu/arange.h

@@ -2,76 +2,27 @@
 
 #pragma once
 
-#include "mlx/allocator.h"
 #include "mlx/array.h"
+#include "mlx/backend/cpu/encoder.h"
 
 namespace mlx::core {
 
 namespace {
 
 template <typename T>
-void arange(T start, T next, array& out, size_t size) {
+void arange(T start, T next, array& out, size_t size, Stream stream) {
   auto ptr = out.data<T>();
   auto step_size = next - start;
-  for (int i = 0; i < size; ++i) {
-    ptr[i] = start;
-    start += step_size;
-  }
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(out);
+  encoder.dispatch([ptr, start, step_size, size]() mutable {
+    for (int i = 0; i < size; ++i) {
+      ptr[i] = start;
+      start += step_size;
+    }
+  });
 }
 
 } // namespace
 
-void arange(
-    const std::vector<array>& inputs,
-    array& out,
-    double start,
-    double step) {
-  assert(inputs.size() == 0);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  switch (out.dtype()) {
-    case bool_:
-      throw std::runtime_error("Bool type unsupported for arange.");
-      break;
-    case uint8:
-      arange<uint8_t>(start, start + step, out, out.size());
-      break;
-    case uint16:
-      arange<uint16_t>(start, start + step, out, out.size());
-      break;
-    case uint32:
-      arange<uint32_t>(start, start + step, out, out.size());
-      break;
-    case uint64:
-      arange<uint64_t>(start, start + step, out, out.size());
-      break;
-    case int8:
-      arange<int8_t>(start, start + step, out, out.size());
-      break;
-    case int16:
-      arange<int16_t>(start, start + step, out, out.size());
-      break;
-    case int32:
-      arange<int32_t>(start, start + step, out, out.size());
-      break;
-    case int64:
-      arange<int64_t>(start, start + step, out, out.size());
-      break;
-    case float16:
-      arange<float16_t>(start, start + step, out, out.size());
-      break;
-    case float32:
-      arange<float>(start, start + step, out, out.size());
-      break;
-    case float64:
-      arange<double>(start, start + step, out, out.size());
-      break;
-    case bfloat16:
-      arange<bfloat16_t>(start, start + step, out, out.size());
-      break;
-    case complex64:
-      arange<complex64_t>(start, start + step, out, out.size());
-      break;
-  }
-}
-
 } // namespace mlx::core

mlx/backend/cpu/arg_reduce.cpp

@@ -3,6 +3,7 @@
 #include <cassert>
 
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
@@ -17,15 +18,18 @@ void arg_reduce(const array& in, array& out, const OpT& op, int axis) {
   Shape shape = in.shape();
   strides.erase(strides.begin() + axis);
   shape.erase(shape.begin() + axis);
+  auto in_ptr = in.data<InT>();
+  auto out_ptr = out.data<uint32_t>();
+
   for (uint32_t i = 0; i < out.size(); ++i) {
     auto loc = elem_to_loc(i, shape, strides);
-    auto in_ptr = in.data<InT>() + loc;
+    auto local_in_ptr = in_ptr + loc;
     uint32_t ind_v = 0;
-    InT v = (*in_ptr);
-    for (uint32_t j = 0; j < axis_size; ++j, in_ptr += axis_stride) {
-      op(j, (*in_ptr), &ind_v, &v);
+    InT v = (*local_in_ptr);
+    for (uint32_t j = 0; j < axis_size; ++j, local_in_ptr += axis_stride) {
+      op(j, (*local_in_ptr), &ind_v, &v);
     }
-    out.data<uint32_t>()[i] = ind_v;
+    out_ptr[i] = ind_v;
   }
 }
 
@@ -65,51 +69,58 @@ void ArgReduce::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  switch (in.dtype()) {
-    case bool_:
-      arg_reduce_dispatch<bool>(in, out, reduce_type_, axis_);
-      break;
-    case uint8:
-      arg_reduce_dispatch<uint8_t>(in, out, reduce_type_, axis_);
-      break;
-    case uint16:
-      arg_reduce_dispatch<uint16_t>(in, out, reduce_type_, axis_);
-      break;
-    case uint32:
-      arg_reduce_dispatch<uint32_t>(in, out, reduce_type_, axis_);
-      break;
-    case uint64:
-      arg_reduce_dispatch<uint64_t>(in, out, reduce_type_, axis_);
-      break;
-    case int8:
-      arg_reduce_dispatch<int8_t>(in, out, reduce_type_, axis_);
-      break;
-    case int16:
-      arg_reduce_dispatch<int16_t>(in, out, reduce_type_, axis_);
-      break;
-    case int32:
-      arg_reduce_dispatch<int32_t>(in, out, reduce_type_, axis_);
-      break;
-    case int64:
-      arg_reduce_dispatch<int64_t>(in, out, reduce_type_, axis_);
-      break;
-    case float16:
-      arg_reduce_dispatch<float16_t>(in, out, reduce_type_, axis_);
-      break;
-    case float32:
-      arg_reduce_dispatch<float>(in, out, reduce_type_, axis_);
-      break;
-    case bfloat16:
-      arg_reduce_dispatch<bfloat16_t>(in, out, reduce_type_, axis_);
-      break;
-    case float64:
-      arg_reduce_dispatch<double>(in, out, reduce_type_, axis_);
-      break;
-    case complex64:
-      arg_reduce_dispatch<complex64_t>(in, out, reduce_type_, axis_);
-      break;
-  }
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    reduce_type_ = reduce_type_,
+                    axis_ = axis_]() mutable {
+    switch (in.dtype()) {
+      case bool_:
+        arg_reduce_dispatch<bool>(in, out, reduce_type_, axis_);
+        break;
+      case uint8:
+        arg_reduce_dispatch<uint8_t>(in, out, reduce_type_, axis_);
+        break;
+      case uint16:
+        arg_reduce_dispatch<uint16_t>(in, out, reduce_type_, axis_);
+        break;
+      case uint32:
+        arg_reduce_dispatch<uint32_t>(in, out, reduce_type_, axis_);
+        break;
+      case uint64:
+        arg_reduce_dispatch<uint64_t>(in, out, reduce_type_, axis_);
+        break;
+      case int8:
+        arg_reduce_dispatch<int8_t>(in, out, reduce_type_, axis_);
+        break;
+      case int16:
+        arg_reduce_dispatch<int16_t>(in, out, reduce_type_, axis_);
+        break;
+      case int32:
+        arg_reduce_dispatch<int32_t>(in, out, reduce_type_, axis_);
+        break;
+      case int64:
+        arg_reduce_dispatch<int64_t>(in, out, reduce_type_, axis_);
+        break;
+      case float16:
+        arg_reduce_dispatch<float16_t>(in, out, reduce_type_, axis_);
+        break;
+      case float32:
+        arg_reduce_dispatch<float>(in, out, reduce_type_, axis_);
+        break;
+      case bfloat16:
+        arg_reduce_dispatch<bfloat16_t>(in, out, reduce_type_, axis_);
+        break;
+      case float64:
+        arg_reduce_dispatch<double>(in, out, reduce_type_, axis_);
+        break;
+      case complex64:
+        arg_reduce_dispatch<complex64_t>(in, out, reduce_type_, axis_);
+        break;
+    }
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/binary.cpp

@@ -8,6 +8,7 @@
 #include "mlx/backend/cpu/binary.h"
 #include "mlx/backend/cpu/binary_ops.h"
 #include "mlx/backend/cpu/binary_two.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
 
@@ -16,51 +17,218 @@ namespace mlx::core {
 namespace {
 
 template <typename Op>
-void comparison_op(const array& a, const array& b, array& out, Op op) {
-  switch (a.dtype()) {
-    case bool_:
-      binary_op<bool, bool>(a, b, out, op);
-      break;
-    case uint8:
-      binary_op<uint8_t, bool>(a, b, out, op);
-      break;
-    case uint16:
-      binary_op<uint16_t, bool>(a, b, out, op);
-      break;
-    case uint32:
-      binary_op<uint32_t, bool>(a, b, out, op);
-      break;
-    case uint64:
-      binary_op<uint64_t, bool>(a, b, out, op);
-      break;
-    case int8:
-      binary_op<int8_t, bool>(a, b, out, op);
-      break;
-    case int16:
-      binary_op<int16_t, bool>(a, b, out, op);
-      break;
-    case int32:
-      binary_op<int32_t, bool>(a, b, out, op);
-      break;
-    case int64:
-      binary_op<int64_t, bool>(a, b, out, op);
-      break;
-    case float16:
-      binary_op<float16_t, bool>(a, b, out, op);
-      break;
-    case float32:
-      binary_op<float, bool>(a, b, out, op);
-      break;
-    case float64:
-      binary_op<double, bool>(a, b, out, op);
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t, bool>(a, b, out, op);
-      break;
-    case complex64:
-      binary_op<complex64_t, bool>(a, b, out, op);
-      break;
-  }
+void binary(const array& a, const array& b, array& out, Op op, Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool, Op>(a, b, out, bopt);
+        break;
+      case uint8:
+        binary_op<uint8_t, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, Op>(a, b, out, bopt);
+        break;
+      case float16:
+        binary_op<float16_t, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, Op>(a, b, out, bopt);
+        break;
+      case complex64:
+        binary_op<complex64_t, Op>(a, b, out, bopt);
+        break;
+    }
+  });
+}
+
+template <typename Op>
+void comparison_op(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (a.dtype()) {
+      case bool_:
+        binary_op<bool, bool, Op>(a, b, out, bopt);
+        break;
+      case uint8:
+        binary_op<uint8_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, bool, Op>(a, b, out, bopt);
+        break;
+      case float16:
+        binary_op<float16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, bool, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, bool, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case complex64:
+        binary_op<complex64_t, bool, Op>(a, b, out, bopt);
+        break;
+    }
+  });
+}
+
+template <typename Op>
+void binary_float(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case float16:
+        binary_op<float16_t, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, Op>(a, b, out, bopt);
+        break;
+      default:
+        throw std::runtime_error(
+            "[binary_float] Only supports non-complex floating point types.");
+    }
+  });
+}
+
+template <typename Op>
+void binary_int(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool, Op>(a, b, out, bopt);
+      case uint8:
+        binary_op<uint8_t, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, Op>(a, b, out, bopt);
+        break;
+      default:
+        throw std::runtime_error("[binary_int] Type not supported");
+        break;
+    }
+  });
 }
 
 } // namespace
@@ -69,7 +237,7 @@ void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Add());
+  binary(a, b, out, detail::Add(), stream());
 }
 
 void DivMod::eval_cpu(
@@ -78,70 +246,89 @@ void DivMod::eval_cpu(
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  auto integral_op = [](auto x, auto y) {
-    return std::make_pair(x / y, x % y);
-  };
-  auto float_op = [](auto x, auto y) {
-    return std::make_pair(std::trunc(x / y), std::fmod(x, y));
-  };
-  switch (outputs[0].dtype()) {
-    case bool_:
-      binary_op<bool>(a, b, outputs, integral_op);
-    case uint8:
-      binary_op<uint8_t>(a, b, outputs, integral_op);
-      break;
-    case uint16:
-      binary_op<uint16_t>(a, b, outputs, integral_op);
-      break;
-    case uint32:
-      binary_op<uint32_t>(a, b, outputs, integral_op);
-      break;
-    case uint64:
-      binary_op<uint64_t>(a, b, outputs, integral_op);
-      break;
-    case int8:
-      binary_op<int8_t>(a, b, outputs, integral_op);
-      break;
-    case int16:
-      binary_op<int16_t>(a, b, outputs, integral_op);
-      break;
-    case int32:
-      binary_op<int32_t>(a, b, outputs, integral_op);
-      break;
-    case int64:
-      binary_op<int64_t>(a, b, outputs, integral_op);
-      break;
-    case float16:
-      binary_op<float16_t>(a, b, outputs, float_op);
-      break;
-    case float32:
-      binary_op<float>(a, b, outputs, float_op);
-      break;
-    case float64:
-      binary_op<double>(a, b, outputs, float_op);
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, outputs, float_op);
-      break;
-    case complex64:
-      // Should never get here
-      throw std::runtime_error("[DivMod] Complex type not supported");
-      break;
-  }
+  auto bopt = get_binary_op_type(a, b);
+  auto& out_a = outputs[0];
+  auto& out_b = outputs[1];
+  set_binary_op_output_data(a, b, out_a, bopt);
+  set_binary_op_output_data(a, b, out_b, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out_a);
+  encoder.set_output_array(out_b);
+
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out_a = array::unsafe_weak_copy(out_a),
+                    out_b = array::unsafe_weak_copy(out_b),
+                    bopt]() mutable {
+    auto integral_op = [](auto x, auto y) {
+      return std::make_pair(x / y, x % y);
+    };
+    auto float_op = [](auto x, auto y) {
+      return std::make_pair(std::trunc(x / y), std::fmod(x, y));
+    };
+
+    switch (out_a.dtype()) {
+      case bool_:
+        binary_op<bool>(a, b, out_a, out_b, integral_op, bopt);
+      case uint8:
+        binary_op<uint8_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int8:
+        binary_op<int8_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int16:
+        binary_op<int16_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int32:
+        binary_op<int32_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case int64:
+        binary_op<int64_t>(a, b, out_a, out_b, integral_op, bopt);
+        break;
+      case float16:
+        binary_op<float16_t>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case float32:
+        binary_op<float>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case float64:
+        binary_op<double>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t>(a, b, out_a, out_b, float_op, bopt);
+        break;
+      case complex64:
+        // Should never get here
+        throw std::runtime_error("[DivMod] Complex type not supported");
+        break;
+    }
+  });
 }
 
 void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Divide());
+  binary(a, b, out, detail::Divide(), stream());
 }
 
 void Remainder::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Remainder());
+  binary(a, b, out, detail::Remainder(), stream());
 }
 
 void Equal::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -149,181 +336,143 @@ void Equal::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& a = inputs[0];
   auto& b = inputs[1];
   if (equal_nan_) {
-    switch (a.dtype()) {
-      case float16:
-        binary_op<float16_t, bool>(a, b, out, detail::NaNEqual());
-        break;
-      case float32:
-        binary_op<float, bool>(a, b, out, detail::NaNEqual());
-        break;
-      case float64:
-        binary_op<double, bool>(a, b, out, detail::NaNEqual());
-        break;
-      case bfloat16:
-        binary_op<bfloat16_t, bool>(a, b, out, detail::NaNEqual());
-        break;
-      case complex64:
-        binary_op<complex64_t, bool>(a, b, out, detail::NaNEqual());
-        break;
-      default:
-        throw std::runtime_error(
-            "[NanEqual::eval_cpu] Only for floating point types.");
-    }
+    auto bopt = get_binary_op_type(a, b);
+    set_binary_op_output_data(a, b, out, bopt);
+
+    auto& encoder = cpu::get_command_encoder(stream());
+    encoder.set_input_array(a);
+    encoder.set_input_array(b);
+    encoder.set_output_array(out);
+    encoder.dispatch([a = array::unsafe_weak_copy(a),
+                      b = array::unsafe_weak_copy(b),
+                      out = array::unsafe_weak_copy(out),
+                      bopt]() mutable {
+      switch (a.dtype()) {
+        case float16:
+          binary_op<float16_t, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case float32:
+          binary_op<float, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case float64:
+          binary_op<double, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case bfloat16:
+          binary_op<bfloat16_t, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        case complex64:
+          binary_op<complex64_t, bool, detail::NaNEqual>(a, b, out, bopt);
+          break;
+        default:
+          throw std::runtime_error(
+              "[NanEqual::eval_cpu] Only for floating point types.");
+      }
+    });
   } else {
-    comparison_op(a, b, out, detail::Equal());
+    comparison_op(a, b, out, detail::Equal(), stream());
   }
 }
 
 void Greater::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::Greater());
+  comparison_op(inputs[0], inputs[1], out, detail::Greater(), stream());
 }
 
 void GreaterEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::GreaterEqual());
+  comparison_op(inputs[0], inputs[1], out, detail::GreaterEqual(), stream());
 }
 
 void Less::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::Less());
+  comparison_op(inputs[0], inputs[1], out, detail::Less(), stream());
 }
 
 void LessEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::LessEqual());
+  comparison_op(inputs[0], inputs[1], out, detail::LessEqual(), stream());
 }
 
 void LogAddExp::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  switch (out.dtype()) {
-    case float16:
-      binary_op<float16_t>(a, b, out, detail::LogAddExp());
-      break;
-    case float32:
-      binary_op<float>(a, b, out, detail::LogAddExp());
-      break;
-    case float64:
-      binary_op<double>(a, b, out, detail::LogAddExp());
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, out, detail::LogAddExp());
-      break;
-    default:
-      throw std::runtime_error(
-          "[LogAddExp::eval_cpu] Only supports non-complex floating point types.");
-  }
+  binary_float(a, b, out, detail::LogAddExp(), stream());
 }
 
 void LogicalAnd::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2); // LogicalAnd requires two input arrays
   auto& in1 = inputs[0];
   auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalAnd());
+  binary(in1, in2, out, detail::LogicalAnd(), stream());
 }
 
 void LogicalOr::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2); // LogicalOr requires two input arrays
   auto& in1 = inputs[0];
   auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalOr());
+  binary(in1, in2, out, detail::LogicalOr(), stream());
 }
 
 void Maximum::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Maximum());
+  binary(a, b, out, detail::Maximum(), stream());
 }
 
 void Minimum::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Minimum());
+  binary(a, b, out, detail::Minimum(), stream());
 }
 
 void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Multiply());
+  binary(a, b, out, detail::Multiply(), stream());
 }
 
 void NotEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::NotEqual());
+  comparison_op(inputs[0], inputs[1], out, detail::NotEqual(), stream());
 }
 
 void Power::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Power());
+  binary(a, b, out, detail::Power(), stream());
 }
 
 void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  binary(a, b, out, detail::Subtract());
+  binary(a, b, out, detail::Subtract(), stream());
 }
 
 void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
   auto& b = inputs[1];
-  auto dispatch_type = [&a, &b, &out](auto op) {
-    switch (out.dtype()) {
-      case bool_:
-        binary_op<bool>(a, b, out, op);
-      case uint8:
-        binary_op<uint8_t>(a, b, out, op);
-        break;
-      case uint16:
-        binary_op<uint16_t>(a, b, out, op);
-        break;
-      case uint32:
-        binary_op<uint32_t>(a, b, out, op);
-        break;
-      case uint64:
-        binary_op<uint64_t>(a, b, out, op);
-        break;
-      case int8:
-        binary_op<int8_t>(a, b, out, op);
-        break;
-      case int16:
-        binary_op<int16_t>(a, b, out, op);
-        break;
-      case int32:
-        binary_op<int32_t>(a, b, out, op);
-        break;
-      case int64:
-        binary_op<int64_t>(a, b, out, op);
-        break;
-      default:
-        throw std::runtime_error(
-            "[BitwiseBinary::eval_cpu] Type not supported");
-        break;
-    }
-  };
   switch (op_) {
     case BitwiseBinary::And:
-      dispatch_type(detail::BitwiseAnd());
+      binary_int(a, b, out, detail::BitwiseAnd(), stream());
       break;
     case BitwiseBinary::Or:
-      dispatch_type(detail::BitwiseOr());
+      binary_int(a, b, out, detail::BitwiseOr(), stream());
       break;
     case BitwiseBinary::Xor:
-      dispatch_type(detail::BitwiseXor());
+      binary_int(a, b, out, detail::BitwiseXor(), stream());
       break;
     case BitwiseBinary::LeftShift:
-      dispatch_type(detail::LeftShift());
+      binary_int(a, b, out, detail::LeftShift(), stream());
       break;
     case BitwiseBinary::RightShift:
-      dispatch_type(detail::RightShift());
+      binary_int(a, b, out, detail::RightShift(), stream());
       break;
   }
 }
@@ -332,23 +481,7 @@ void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 2);
   const auto& a = inputs[0];
   const auto& b = inputs[1];
-  switch (out.dtype()) {
-    case float16:
-      binary_op<float16_t>(a, b, out, detail::ArcTan2());
-      break;
-    case float32:
-      binary_op<float>(a, b, out, detail::ArcTan2());
-      break;
-    case float64:
-      binary_op<double>(a, b, out, detail::ArcTan2());
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, out, detail::ArcTan2());
-      break;
-    default:
-      throw std::runtime_error(
-          "[ArcTan2::eval_cpu] Only supports non-complex floating point types.");
-  }
+  binary_float(a, b, out, detail::ArcTan2(), stream());
 }
 
 } // namespace mlx::core

mlx/backend/cpu/binary.h

@@ -3,7 +3,6 @@
 #pragma once
 #include <cassert>
 
-#include "mlx/allocator.h"
 #include "mlx/array.h"
 #include "mlx/backend/common/binary.h"
 #include "mlx/backend/common/utils.h"
@@ -14,22 +13,18 @@ namespace mlx::core {
 
 template <typename Op>
 struct VectorScalar {
-  Op op;
-
-  VectorScalar(Op op_) : op(op_) {}
-
   template <typename T, typename U>
   void operator()(const T* a, const T* b, U* dst, int size) {
     T scalar = *b;
     constexpr int N = simd::max_size<T>;
     while (size >= N) {
-      simd::store(dst, op(simd::load<T, N>(a), simd::Simd<T, N>(scalar)));
+      simd::store(dst, Op{}(simd::load<T, N>(a), simd::Simd<T, N>(scalar)));
       dst += N;
       a += N;
       size -= N;
     }
     while (size-- > 0) {
-      *dst = op(*a, scalar);
+      *dst = Op{}(*a, scalar);
       dst++;
       a++;
     }
@@ -38,22 +33,18 @@ struct VectorScalar {
 
 template <typename Op>
 struct ScalarVector {
-  Op op;
-
-  ScalarVector(Op op_) : op(op_) {}
-
   template <typename T, typename U>
   void operator()(const T* a, const T* b, U* dst, int size) {
     T scalar = *a;
     constexpr int N = simd::max_size<T>;
     while (size >= N) {
-      simd::store(dst, op(simd::Simd<T, N>(scalar), simd::load<T, N>(b)));
+      simd::store(dst, Op{}(simd::Simd<T, N>(scalar), simd::load<T, N>(b)));
       dst += N;
       b += N;
       size -= N;
     }
     while (size-- > 0) {
-      *dst = op(scalar, *b);
+      *dst = Op{}(scalar, *b);
       dst++;
       b++;
     }
@@ -62,22 +53,18 @@ struct ScalarVector {
 
 template <typename Op>
 struct VectorVector {
-  Op op;
-
-  VectorVector(Op op_) : op(op_) {}
-
   template <typename T, typename U>
   void operator()(const T* a, const T* b, U* dst, int size) {
     constexpr int N = simd::max_size<T>;
     while (size >= N) {
-      simd::store(dst, op(simd::load<T, N>(a), simd::load<T, N>(b)));
+      simd::store(dst, Op{}(simd::load<T, N>(a), simd::load<T, N>(b)));
       dst += N;
       a += N;
       b += N;
       size -= N;
     }
     while (size-- > 0) {
-      *dst = op(*a, *b);
+      *dst = Op{}(*a, *b);
       dst++;
       a++;
       b++;
@@ -90,7 +77,6 @@ void binary_op_dims(
     const T* a,
     const T* b,
     U* out,
-    Op op,
     const Shape& shape,
     const Strides& a_strides,
     const Strides& b_strides,
@@ -104,12 +90,12 @@ void binary_op_dims(
   for (int i = 0; i < N; i++) {
     if constexpr (D > 1) {
       binary_op_dims<T, U, Op, D - 1, Strided>(
-          a, b, out, op, shape, a_strides, b_strides, out_strides, axis + 1);
+          a, b, out, shape, a_strides, b_strides, out_strides, axis + 1);
     } else {
       if constexpr (Strided) {
-        op(a, b, out, stride_out);
+        Op{}(a, b, out, stride_out);
       } else {
-        *out = op(*a, *b);
+        *out = Op{}(*a, *b);
       }
     }
     out += stride_out;
@@ -120,66 +106,38 @@ void binary_op_dims(
 
 template <typename T, typename U, bool Strided, typename Op>
 void binary_op_dispatch_dims(
-    const array& a,
-    const array& b,
-    array& out,
-    Op op,
+    const T* a,
+    const T* b,
+    U* out,
     int dim,
+    int size,
     const Shape& shape,
     const Strides& a_strides,
     const Strides& b_strides,
     const Strides& out_strides) {
-  const T* a_ptr = a.data<T>();
-  const T* b_ptr = b.data<T>();
-  U* out_ptr = out.data<U>();
   switch (dim) {
     case 1:
       binary_op_dims<T, U, Op, 1, Strided>(
-          a_ptr,
-          b_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          out_strides,
-          0);
+          a, b, out, shape, a_strides, b_strides, out_strides, 0);
       return;
     case 2:
       binary_op_dims<T, U, Op, 2, Strided>(
-          a_ptr,
-          b_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          out_strides,
-          0);
+          a, b, out, shape, a_strides, b_strides, out_strides, 0);
       return;
     case 3:
       binary_op_dims<T, U, Op, 3, Strided>(
-          a_ptr,
-          b_ptr,
-          out_ptr,
-          op,
-          shape,
-          a_strides,
-          b_strides,
-          out_strides,
-          0);
+          a, b, out, shape, a_strides, b_strides, out_strides, 0);
       return;
   }
 
   ContiguousIterator a_it(shape, a_strides, dim - 3);
   ContiguousIterator b_it(shape, b_strides, dim - 3);
   auto stride = out_strides[dim - 4];
-  for (int64_t elem = 0; elem < a.size(); elem += stride) {
+  for (int64_t elem = 0; elem < size; elem += stride) {
     binary_op_dims<T, U, Op, 3, Strided>(
-        a_ptr + a_it.loc,
-        b_ptr + b_it.loc,
-        out_ptr + elem,
-        op,
+        a + a_it.loc,
+        b + b_it.loc,
+        out + elem,
         shape,
         a_strides,
         b_strides,
@@ -191,40 +149,41 @@ void binary_op_dispatch_dims(
 }
 
 template <typename T, typename U, typename Op>
-void binary_op(const array& a, const array& b, array& out, Op op) {
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
-
+void binary_op(const array& a, const array& b, array& out, BinaryOpType bopt) {
   // The full computation is scalar scalar so call the base op once
+  auto a_ptr = a.data<T>();
+  auto b_ptr = b.data<T>();
+
+  auto out_ptr = out.data<U>();
   if (bopt == BinaryOpType::ScalarScalar) {
-    *(out.data<U>()) = op(*a.data<T>(), *b.data<T>());
+    *out_ptr = Op{}(*a_ptr, *b_ptr);
     return;
   }
 
   // The full computation is scalar vector so delegate to the op
   if (bopt == BinaryOpType::ScalarVector) {
-    ScalarVector{op}(a.data<T>(), b.data<T>(), out.data<U>(), b.data_size());
+    ScalarVector<Op>{}(a_ptr, b_ptr, out_ptr, b.data_size());
     return;
   }
 
   // The full computation is vector scalar so delegate to the op
   if (bopt == BinaryOpType::VectorScalar) {
-    VectorScalar{op}(a.data<T>(), b.data<T>(), out.data<U>(), a.data_size());
+    VectorScalar<Op>{}(a_ptr, b_ptr, out_ptr, a.data_size());
     return;
   }
 
   // The full computation is vector vector so delegate to the op
   if (bopt == BinaryOpType::VectorVector) {
-    VectorVector{op}(a.data<T>(), b.data<T>(), out.data<U>(), out.size());
+    VectorVector<Op>{}(a_ptr, b_ptr, out_ptr, a.size());
     return;
   }
 
   // General computation so let's try to optimize
   auto [new_shape, new_strides] = collapse_contiguous_dims(
       a.shape(), {a.strides(), b.strides(), out.strides()});
-  const auto& a_strides = new_strides[0];
-  const auto& b_strides = new_strides[1];
-  const auto& strides = new_strides[2];
+  auto& a_strides = new_strides[0];
+  auto& b_strides = new_strides[1];
+  auto& strides = new_strides[2];
 
   // Get the left-most dim such that the array is row contiguous after
   auto leftmost_rc_dim = [&strides](const auto& arr_strides) {
@@ -248,7 +207,8 @@ void binary_op(const array& a, const array& b, array& out, Op op) {
 
   auto ndim = new_shape.size();
 
-  // Case 1: LxM and FxM where L and F are broadcastable and M is row contiguous
+  // Case 1: LxM and FxM where L and F are broadcastable and M is row
+  // contiguous
   int dim = ndim;
   if (int d = std::max(a_rc_dim, b_rc_dim); d < ndim) {
     bopt = BinaryOpType::VectorVector;
@@ -275,99 +235,59 @@ void binary_op(const array& a, const array& b, array& out, Op op) {
 
   switch (bopt) {
     case BinaryOpType::VectorVector:
-      binary_op_dispatch_dims<T, U, true>(
-          a,
-          b,
-          out,
-          VectorVector{op},
+      binary_op_dispatch_dims<T, U, true, VectorVector<Op>>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
           dim,
+          a.size(),
           new_shape,
           a_strides,
           b_strides,
           strides);
       break;
     case BinaryOpType::VectorScalar:
-      binary_op_dispatch_dims<T, U, true>(
-          a,
-          b,
-          out,
-          VectorScalar{op},
+      binary_op_dispatch_dims<T, U, true, VectorScalar<Op>>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
           dim,
+          a.size(),
           new_shape,
           a_strides,
           b_strides,
           strides);
       break;
     case BinaryOpType::ScalarVector:
-      binary_op_dispatch_dims<T, U, true>(
-          a,
-          b,
-          out,
-          ScalarVector{op},
+      binary_op_dispatch_dims<T, U, true, ScalarVector<Op>>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
           dim,
+          a.size(),
           new_shape,
           a_strides,
           b_strides,
           strides);
       break;
     default:
-      binary_op_dispatch_dims<T, U, false>(
-          a, b, out, op, dim, new_shape, a_strides, b_strides, strides);
+      binary_op_dispatch_dims<T, U, false, Op>(
+          a_ptr,
+          b_ptr,
+          out_ptr,
+          dim,
+          a.size(),
+          new_shape,
+          a_strides,
+          b_strides,
+          strides);
       break;
   }
 }
 
 template <typename T, typename Op>
-void binary_op(const array& a, const array& b, array& out, Op op) {
-  binary_op<T, T>(a, b, out, op);
-}
-
-template <typename Op>
-void binary(const array& a, const array& b, array& out, Op op) {
-  switch (out.dtype()) {
-    case bool_:
-      binary_op<bool>(a, b, out, op);
-      break;
-    case uint8:
-      binary_op<uint8_t>(a, b, out, op);
-      break;
-    case uint16:
-      binary_op<uint16_t>(a, b, out, op);
-      break;
-    case uint32:
-      binary_op<uint32_t>(a, b, out, op);
-      break;
-    case uint64:
-      binary_op<uint64_t>(a, b, out, op);
-      break;
-    case int8:
-      binary_op<int8_t>(a, b, out, op);
-      break;
-    case int16:
-      binary_op<int16_t>(a, b, out, op);
-      break;
-    case int32:
-      binary_op<int32_t>(a, b, out, op);
-      break;
-    case int64:
-      binary_op<int64_t>(a, b, out, op);
-      break;
-    case float16:
-      binary_op<float16_t>(a, b, out, op);
-      break;
-    case float32:
-      binary_op<float>(a, b, out, op);
-      break;
-    case float64:
-      binary_op<double>(a, b, out, op);
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, out, op);
-      break;
-    case complex64:
-      binary_op<complex64_t>(a, b, out, op);
-      break;
-  }
+void binary_op(const array& a, const array& b, array& out, BinaryOpType bopt) {
+  binary_op<T, T, Op>(a, b, out, bopt);
 }
 
 } // namespace mlx::core

mlx/backend/cpu/binary_two.h

@@ -58,14 +58,14 @@ void binary_op_dispatch_dims(
     Op op) {
   auto [shape, strides] = collapse_contiguous_dims(
       a.shape(), {a.strides(), b.strides(), out_a.strides()});
-  const auto& a_strides = strides[0];
-  const auto& b_strides = strides[1];
-  const auto& out_strides = strides[2];
   const T* a_ptr = a.data<T>();
   const T* b_ptr = b.data<T>();
   U* out_a_ptr = out_a.data<U>();
   U* out_b_ptr = out_b.data<U>();
 
+  const auto& a_strides = strides[0];
+  const auto& b_strides = strides[1];
+  const auto& out_strides = strides[2];
   int ndim = shape.size();
   switch (ndim) {
     case 1:
@@ -120,14 +120,10 @@ template <typename T, typename U = T, typename Op>
 void binary_op(
     const array& a,
     const array& b,
-    std::vector<array>& outputs,
-    Op op) {
-  auto bopt = get_binary_op_type(a, b);
-  auto& out_a = outputs[0];
-  auto& out_b = outputs[1];
-  set_binary_op_output_data(a, b, out_a, bopt);
-  set_binary_op_output_data(a, b, out_b, bopt);
-
+    array& out_a,
+    array& out_b,
+    Op op,
+    BinaryOpType bopt) {
   // The full computation is scalar scalar so call the base op once
   if (bopt == BinaryOpType::General) {
     binary_op_dispatch_dims<T, U, Op>(a, b, out_a, out_b, op);
@@ -141,14 +137,14 @@ void binary_op(
   if (bopt == BinaryOpType::ScalarScalar) {
     std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
   } else if (bopt == BinaryOpType::ScalarVector) {
-    for (size_t i = 0; i < b.size(); ++i) {
+    for (size_t i = 0; i < b.data_size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
       b_ptr++;
     }
   } else if (bopt == BinaryOpType::VectorScalar) {
-    for (size_t i = 0; i < a.size(); ++i) {
+    for (size_t i = 0; i < a.data_size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
@@ -165,58 +161,6 @@ void binary_op(
   }
 }
 
-template <typename Op>
-void binary(
-    const array& a,
-    const array& b,
-    std::vector<array>& outputs,
-    Op op) {
-  switch (outputs[0].dtype()) {
-    case bool_:
-      binary_op<bool>(a, b, outputs, op);
-      break;
-    case uint8:
-      binary_op<uint8_t>(a, b, outputs, op);
-      break;
-    case uint16:
-      binary_op<uint16_t>(a, b, outputs, op);
-      break;
-    case uint32:
-      binary_op<uint32_t>(a, b, outputs, op);
-      break;
-    case uint64:
-      binary_op<uint64_t>(a, b, outputs, op);
-      break;
-    case int8:
-      binary_op<int8_t>(a, b, outputs, op);
-      break;
-    case int16:
-      binary_op<int16_t>(a, b, outputs, op);
-      break;
-    case int32:
-      binary_op<int32_t>(a, b, outputs, op);
-      break;
-    case int64:
-      binary_op<int64_t>(a, b, outputs, op);
-      break;
-    case float16:
-      binary_op<float16_t>(a, b, outputs, op);
-      break;
-    case float32:
-      binary_op<float>(a, b, outputs, op);
-      break;
-    case float64:
-      binary_op<double>(a, b, outputs, op);
-      break;
-    case bfloat16:
-      binary_op<bfloat16_t>(a, b, outputs, op);
-      break;
-    case complex64:
-      binary_op<complex64_t>(a, b, outputs, op);
-      break;
-  }
-}
-
 } // namespace
 
 } // namespace mlx::core

mlx/backend/cpu/cholesky.cpp

@@ -2,6 +2,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/linalg.h"
 #include "mlx/primitives.h"
@@ -9,7 +10,7 @@
 namespace mlx::core {
 
 template <typename T>
-void cholesky_impl(const array& a, array& factor, bool upper) {
+void cholesky_impl(const array& a, array& factor, bool upper, Stream stream) {
   // Lapack uses the column-major convention. We take advantage of the fact that
   // the matrix should be symmetric:
   //   (A)ᵀ = A
@@ -17,60 +18,63 @@ void cholesky_impl(const array& a, array& factor, bool upper) {
   // triangular matrix, so uplo is the opposite of what we would expect from
   // upper
 
-  char uplo = (upper) ? 'L' : 'U';
-
   // The decomposition is computed in place, so just copy the input to the
   // output.
   copy(
       a,
       factor,
-      a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream);
 
-  const int N = a.shape(-1);
-  const size_t num_matrices = a.size() / (N * N);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(factor);
+  encoder.dispatch([matrix = factor.data<T>(),
+                    upper,
+                    N = a.shape(-1),
+                    size = a.size()]() mutable {
+    char uplo = (upper) ? 'L' : 'U';
+    size_t num_matrices = size / (N * N);
+    for (int i = 0; i < num_matrices; i++) {
+      // Compute Cholesky factorization.
+      int info;
+      potrf<T>(
+          /* uplo = */ &uplo,
+          /* n = */ &N,
+          /* a = */ matrix,
+          /* lda = */ &N,
+          /* info = */ &info);
 
-  T* matrix = factor.data<T>();
-
-  for (int i = 0; i < num_matrices; i++) {
-    // Compute Cholesky factorization.
-    int info;
-    potrf<T>(
-        /* uplo = */ &uplo,
-        /* n = */ &N,
-        /* a = */ matrix,
-        /* lda = */ &N,
-        /* info = */ &info);
-
-    // TODO: We do nothing when the matrix is not positive semi-definite
-    // because throwing an error would result in a crash. If we figure out how
-    // to catch errors from the implementation we should throw.
-    if (info < 0) {
-      std::stringstream msg;
-      msg << "[cholesky] Cholesky decomposition failed with error code "
-          << info;
-      throw std::runtime_error(msg.str());
-    }
-
-    // Zero out the upper/lower triangle while advancing the pointer to the
-    // next matrix at the same time.
-    for (int row = 0; row < N; row++) {
-      if (upper) {
-        std::fill(matrix, matrix + row, 0);
-      } else {
-        std::fill(matrix + row + 1, matrix + N, 0);
+      // TODO: We do nothing when the matrix is not positive semi-definite
+      // because throwing an error would result in a crash. If we figure out how
+      // to catch errors from the implementation we should throw.
+      if (info < 0) {
+        std::stringstream msg;
+        msg << "[Cholesky::eval_cpu] Cholesky decomposition failed with error code "
+            << info;
+        throw std::runtime_error(msg.str());
+      }
+
+      // Zero out the upper/lower triangle while advancing the pointer to the
+      // next matrix at the same time.
+      for (int row = 0; row < N; row++) {
+        if (upper) {
+          std::fill(matrix, matrix + row, 0);
+        } else {
+          std::fill(matrix + row + 1, matrix + N, 0);
+        }
+        matrix += N;
       }
-      matrix += N;
     }
-  }
+  });
 }
 
 void Cholesky::eval_cpu(const std::vector<array>& inputs, array& output) {
   switch (inputs[0].dtype()) {
     case float32:
-      cholesky_impl<float>(inputs[0], output, upper_);
+      cholesky_impl<float>(inputs[0], output, upper_, stream());
       break;
     case float64:
-      cholesky_impl<double>(inputs[0], output, upper_);
+      cholesky_impl<double>(inputs[0], output, upper_, stream());
       break;
     default:
       throw std::runtime_error(

mlx/backend/cpu/compiled.cpp

@@ -11,6 +11,7 @@
 
 #include "mlx/backend/common/compiled.h"
 #include "mlx/backend/cpu/compiled_preamble.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
@@ -288,6 +289,7 @@ void Compiled::eval_cpu(
   // Figure out which kernel we are using
   auto& shape = outputs[0].shape();
   auto contiguous = compiled_check_contiguity(inputs, shape);
+  auto& encoder = cpu::get_command_encoder(stream());
 
   // Handle all broadcasting and collect function input arguments
   std::vector<void*> args;
@@ -298,6 +300,7 @@ void Compiled::eval_cpu(
       continue;
     }
     auto& x = inputs[i];
+    encoder.set_input_array(x);
     args.push_back((void*)x.data<void>());
 
     if (contiguous || is_scalar(x)) {
@@ -356,18 +359,25 @@ void Compiled::eval_cpu(
   });
 
   compiled_allocate_outputs(
-      inputs, outputs, inputs_, constant_ids_, contiguous, false);
+      inputs, outputs, inputs_, constant_ids_, contiguous);
 
   for (auto& x : outputs) {
     args.push_back(x.data<void>());
+    encoder.set_output_array(x);
   }
+  Shape out_shape;
   if (!contiguous) {
-    args.push_back((void*)outputs[0].shape().data());
+    out_shape = outputs[0].shape();
+    args.push_back((void*)out_shape.data());
   } else {
     args.push_back((void*)outputs[0].data_size());
   }
   auto fun = (void (*)(void**))fn_ptr;
-  fun(args.data());
+  encoder.dispatch(
+      [fun,
+       args = std::move(args),
+       strides = std::move(strides),
+       out_shape = std::move(out_shape)]() mutable { fun(args.data()); });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/copy.cpp

@@ -5,6 +5,7 @@
 #include "mlx/allocator.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
 
 namespace mlx::core {
@@ -13,19 +14,19 @@ namespace {
 
 template <typename SrcT, typename DstT>
 void copy_single(const array& src, array& dst) {
-  auto val = static_cast<DstT>(src.data<SrcT>()[0]);
+  auto src_ptr = src.data<SrcT>();
   auto dst_ptr = dst.data<DstT>();
-  for (int i = 0; i < dst.size(); ++i) {
-    dst_ptr[i] = val;
-  }
+  auto size = dst.size();
+  auto val = static_cast<DstT>(src_ptr[0]);
+  std::fill_n(dst_ptr, size, val);
 }
 
 template <typename SrcT, typename DstT>
 void copy_vector(const array& src, array& dst) {
   auto src_ptr = src.data<SrcT>();
   auto dst_ptr = dst.data<DstT>();
-  size_t size = src.data_size();
-  std::copy(src_ptr, src_ptr + src.data_size(), dst_ptr);
+  auto size = src.data_size();
+  std::copy(src_ptr, src_ptr + size, dst_ptr);
 }
 
 template <typename SrcT, typename DstT, int D>
@@ -60,36 +61,57 @@ void copy_general_general(
     const Strides& i_strides,
     const Strides& o_strides,
     int64_t i_offset,
-    int64_t o_offset) {
+    int64_t o_offset,
+    const std::optional<array>& dynamic_i_offset,
+    const std::optional<array>& dynamic_o_offset) {
+  auto src_ptr = src.data<SrcT>() + i_offset;
+  auto dst_ptr = dst.data<DstT>() + o_offset;
+  auto i_offset_ptr =
+      dynamic_i_offset ? dynamic_i_offset->data<int64_t>() : nullptr;
+  auto o_offset_ptr =
+      dynamic_o_offset ? dynamic_o_offset->data<int64_t>() : nullptr;
+  auto size = src.size();
   if (data_shape.empty()) {
-    auto val = static_cast<DstT>(*(src.data<SrcT>() + i_offset));
-    auto dst_ptr = dst.data<DstT>() + o_offset;
+    auto val = static_cast<DstT>(*src_ptr);
     *dst_ptr = val;
     return;
   }
   auto [shape, strides] =
       collapse_contiguous_dims(data_shape, {i_strides, o_strides});
-  auto src_ptr = src.data<SrcT>() + i_offset;
-  auto dst_ptr = dst.data<DstT>() + o_offset;
+
   int ndim = shape.size();
-  if (ndim == 1) {
-    copy_dims<SrcT, DstT, 1>(
-        src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
-    return;
-  } else if (ndim == 2) {
-    copy_dims<SrcT, DstT, 2>(
-        src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
-    return;
-  } else if (ndim == 3) {
-    copy_dims<SrcT, DstT, 3>(
-        src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+  if (ndim < 3) {
+    if (i_offset_ptr) {
+      src_ptr += i_offset_ptr[0];
+    }
+    if (o_offset_ptr) {
+      dst_ptr += o_offset_ptr[0];
+    }
+
+    if (ndim == 1) {
+      copy_dims<SrcT, DstT, 1>(
+          src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+    } else if (ndim == 2) {
+      copy_dims<SrcT, DstT, 2>(
+          src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+    } else if (ndim == 3) {
+      copy_dims<SrcT, DstT, 3>(
+          src_ptr, dst_ptr, shape, strides[0], strides[1], 0);
+    }
     return;
   }
+  if (i_offset_ptr) {
+    src_ptr += i_offset_ptr[0];
+  }
+  if (o_offset_ptr) {
+    dst_ptr += o_offset_ptr[0];
+  }
+
   ContiguousIterator in(shape, strides[0], ndim - 3);
   ContiguousIterator out(shape, strides[1], ndim - 3);
   auto stride = std::accumulate(
       shape.end() - 3, shape.end(), 1, std::multiplies<int64_t>());
-  for (int64_t elem = 0; elem < src.size(); elem += stride) {
+  for (int64_t elem = 0; elem < size; elem += stride) {
     copy_dims<SrcT, DstT, 3>(
         src_ptr + in.loc,
         dst_ptr + out.loc,
@@ -105,7 +127,15 @@ void copy_general_general(
 template <typename SrcT, typename DstT>
 inline void copy_general_general(const array& src, array& dst) {
   copy_general_general<SrcT, DstT>(
-      src, dst, src.shape(), src.strides(), dst.strides(), 0, 0);
+      src,
+      dst,
+      src.shape(),
+      src.strides(),
+      dst.strides(),
+      0,
+      0,
+      std::nullopt,
+      std::nullopt);
 }
 
 template <typename SrcT, typename DstT>
@@ -116,7 +146,9 @@ void copy_general(
     const Strides& i_strides,
     const Strides&,
     int64_t i_offset,
-    int64_t o_offset) {
+    int64_t o_offset,
+    const std::optional<array>& dynamic_i_offset,
+    const std::optional<array>& dynamic_o_offset) {
   copy_general_general<SrcT, DstT>(
       src,
       dst,
@@ -124,7 +156,9 @@ void copy_general(
       i_strides,
       make_contiguous_strides(data_shape),
       i_offset,
-      o_offset);
+      o_offset,
+      dynamic_i_offset,
+      dynamic_o_offset);
 }
 
 template <typename SrcT, typename DstT>
@@ -136,7 +170,9 @@ inline void copy_general(const array& src, array& dst) {
       src.strides(),
       make_contiguous_strides(src.shape()),
       0,
-      0);
+      0,
+      std::nullopt,
+      std::nullopt);
 }
 
 template <typename SrcT, typename DstT, typename... Args>
@@ -259,35 +295,27 @@ inline void copy_inplace_dispatch(
 
 } // namespace
 
-void copy_inplace(const array& src, array& dst, CopyType ctype) {
-  copy_inplace_dispatch(src, dst, ctype);
+void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(src);
+  encoder.set_output_array(dst);
+  encoder.dispatch(
+      [src = array::unsafe_weak_copy(src),
+       dst = array::unsafe_weak_copy(dst),
+       ctype]() mutable { copy_inplace_dispatch(src, dst, ctype); });
 }
 
-void copy(const array& src, array& dst, CopyType ctype) {
-  // Allocate the output
-  switch (ctype) {
-    case CopyType::Vector:
-      if (src.is_donatable() && src.itemsize() == dst.itemsize()) {
-        dst.copy_shared_buffer(src);
-      } else {
-        auto size = src.data_size();
-        dst.set_data(
-            allocator::malloc_or_wait(size * dst.itemsize()),
-            size,
-            src.strides(),
-            src.flags());
-      }
-      break;
-    case CopyType::Scalar:
-    case CopyType::General:
-    case CopyType::GeneralGeneral:
-      dst.set_data(allocator::malloc_or_wait(dst.nbytes()));
-      break;
+void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
+  bool donated = set_copy_output_data(src, dst, ctype);
+  if (donated && src.dtype() == dst.dtype()) {
+    // If the output has the same type as the input then there is nothing to
+    // copy, just use the buffer.
+    return;
   }
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;
   }
-  copy_inplace(src, dst, ctype);
+  copy_inplace(src, dst, ctype, stream);
 }
 
 void copy_inplace(
@@ -298,24 +326,51 @@ void copy_inplace(
     const Strides& o_strides,
     int64_t i_offset,
     int64_t o_offset,
-    CopyType ctype) {
-  switch (ctype) {
-    case CopyType::General:
-    case CopyType::GeneralGeneral:
-      copy_inplace_dispatch(
-          src,
-          dst,
-          ctype,
-          data_shape,
-          i_strides,
-          o_strides,
-          i_offset,
-          o_offset);
-      break;
-    case CopyType::Scalar:
-    case CopyType::Vector:
-      copy_inplace_dispatch(src, dst, ctype);
-  }
+    CopyType ctype,
+    Stream stream,
+    const std::optional<array>& dynamic_i_offset, /* = std::nullopt */
+    const std::optional<array>& dynamic_o_offset /* = std::nullopt */) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(src);
+  encoder.set_output_array(dst);
+  auto weak_copy_if_set = [](auto x) -> std::optional<array> {
+    if (x) {
+      return array::unsafe_weak_copy(*x);
+    } else {
+      return std::nullopt;
+    }
+  };
+  encoder.dispatch(
+      [src = array::unsafe_weak_copy(src),
+       dst = array::unsafe_weak_copy(dst),
+       data_shape,
+       i_strides,
+       o_strides,
+       i_offset,
+       o_offset,
+       ctype,
+       dynamic_i_offset = weak_copy_if_set(dynamic_i_offset),
+       dynamic_o_offset = weak_copy_if_set(dynamic_o_offset)]() mutable {
+        switch (ctype) {
+          case CopyType::General:
+          case CopyType::GeneralGeneral:
+            copy_inplace_dispatch(
+                src,
+                dst,
+                ctype,
+                data_shape,
+                i_strides,
+                o_strides,
+                i_offset,
+                o_offset,
+                dynamic_i_offset,
+                dynamic_o_offset);
+            break;
+          case CopyType::Scalar:
+          case CopyType::Vector:
+            copy_inplace_dispatch(src, dst, ctype);
+        }
+      });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/copy.h

@@ -2,14 +2,16 @@
 
 #pragma once
 
+#include <optional>
+
 #include "mlx/array.h"
 #include "mlx/backend/common/copy.h"
 #include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
-void copy(const array& src, array& dst, CopyType ctype);
-void copy_inplace(const array& src, array& dst, CopyType ctype);
+void copy(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream);
 
 void copy_inplace(
     const array& src,
@@ -19,6 +21,9 @@ void copy_inplace(
     const Strides& o_strides,
     int64_t i_offset,
     int64_t o_offset,
-    CopyType ctype);
+    CopyType ctype,
+    Stream stream,
+    const std::optional<array>& dynamic_i_offset = std::nullopt,
+    const std::optional<array>& dynamic_o_offset = std::nullopt);
 
 } // namespace mlx::core

mlx/backend/cpu/distributed.cpp

@@ -0,0 +1,94 @@
+// Copyright © 2024 Apple Inc.
+
+#include <cassert>
+
+#include "mlx/allocator.h"
+#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/distributed/primitives.h"
+
+namespace mlx::core::distributed {
+
+std::pair<array, bool> ensure_row_contiguous(const array& arr, Stream stream) {
+  if (arr.flags().row_contiguous) {
+    return {arr, false};
+  } else {
+    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+    copy(arr, arr_copy, CopyType::General, stream);
+    return {arr_copy, true};
+  }
+};
+
+void AllReduce::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto donate_or_copy = [s = stream()](const array& in, array& out) {
+    if (in.flags().row_contiguous) {
+      if (in.is_donatable()) {
+        out.copy_shared_buffer(in);
+      } else {
+        out.set_data(allocator::malloc_or_wait(out.nbytes()));
+      }
+      return in;
+    } else {
+      array arr_copy(in.shape(), in.dtype(), nullptr, {});
+      copy(in, arr_copy, CopyType::General, s);
+      out.copy_shared_buffer(arr_copy);
+      return arr_copy;
+    }
+  };
+
+  auto in = donate_or_copy(inputs[0], outputs[0]);
+  switch (reduce_type_) {
+    case Sum:
+      distributed::detail::all_sum(group(), in, outputs[0], stream());
+      break;
+    default:
+      throw std::runtime_error("Only all reduce sum is supported for now");
+  }
+}
+
+void AllGather::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto [in, copied] = ensure_row_contiguous(inputs[0], stream());
+  outputs[0].set_data(allocator::malloc_or_wait(outputs[0].nbytes()));
+  distributed::detail::all_gather(group(), in, outputs[0], stream());
+  if (copied) {
+    auto& enc = cpu::get_command_encoder(stream());
+    enc.add_temporary(in);
+  }
+}
+
+void Send::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto [in, copied] = ensure_row_contiguous(inputs[0], stream());
+  distributed::detail::send(group(), in, dst_, stream());
+  outputs[0].copy_shared_buffer(inputs[0]);
+  if (copied) {
+    auto& enc = cpu::get_command_encoder(stream());
+    enc.add_temporary(in);
+  }
+}
+
+void Recv::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 0);
+  assert(outputs.size() == 1);
+
+  outputs[0].set_data(allocator::malloc_or_wait(outputs[0].nbytes()));
+  distributed::detail::recv(group(), outputs[0], src_, stream());
+}
+
+} // namespace mlx::core::distributed

mlx/backend/cpu/eigh.cpp

@@ -3,6 +3,7 @@
 #include "mlx/allocator.h"
 #include "mlx/array.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/linalg.h"
 #include "mlx/primitives.h"
@@ -16,59 +17,72 @@ void eigh_impl(
     array& vectors,
     array& values,
     const std::string& uplo,
-    bool compute_eigenvectors) {
+    bool compute_eigenvectors,
+    Stream stream) {
   auto vec_ptr = vectors.data<T>();
   auto eig_ptr = values.data<T>();
-
   char jobz = compute_eigenvectors ? 'V' : 'N';
-  auto N = vectors.shape(-1);
 
-  // Work query
-  int lwork = -1;
-  int liwork = -1;
-  int info;
-  {
-    T work;
-    int iwork;
-    syevd<T>(
-        &jobz,
-        uplo.c_str(),
-        &N,
-        nullptr,
-        &N,
-        nullptr,
-        &work,
-        &lwork,
-        &iwork,
-        &liwork,
-        &info);
-    lwork = static_cast<int>(work);
-    liwork = iwork;
-  }
-
-  auto work_buf = array::Data{allocator::malloc_or_wait(sizeof(T) * lwork)};
-  auto iwork_buf = array::Data{allocator::malloc_or_wait(sizeof(int) * liwork)};
-  for (size_t i = 0; i < vectors.size() / (N * N); ++i) {
-    syevd<T>(
-        &jobz,
-        uplo.c_str(),
-        &N,
-        vec_ptr,
-        &N,
-        eig_ptr,
-        static_cast<T*>(work_buf.buffer.raw_ptr()),
-        &lwork,
-        static_cast<int*>(iwork_buf.buffer.raw_ptr()),
-        &liwork,
-        &info);
-    vec_ptr += N * N;
-    eig_ptr += N;
-    if (info != 0) {
-      std::stringstream msg;
-      msg << "[Eigh::eval_cpu] Eigenvalue decomposition failed with error code "
-          << info;
-      throw std::runtime_error(msg.str());
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(vectors);
+  encoder.set_output_array(values);
+  encoder.dispatch([vec_ptr,
+                    eig_ptr,
+                    jobz,
+                    uplo = uplo[0],
+                    N = vectors.shape(-1),
+                    size = vectors.size()]() mutable {
+    // Work query
+    int lwork = -1;
+    int liwork = -1;
+    int info;
+    {
+      T work;
+      int iwork;
+      syevd<T>(
+          &jobz,
+          &uplo,
+          &N,
+          nullptr,
+          &N,
+          nullptr,
+          &work,
+          &lwork,
+          &iwork,
+          &liwork,
+          &info);
+      lwork = static_cast<int>(work);
+      liwork = iwork;
     }
+
+    auto work_buf = array::Data{allocator::malloc_or_wait(sizeof(T) * lwork)};
+    auto iwork_buf =
+        array::Data{allocator::malloc_or_wait(sizeof(int) * liwork)};
+    for (size_t i = 0; i < size / (N * N); ++i) {
+      syevd<T>(
+          &jobz,
+          &uplo,
+          &N,
+          vec_ptr,
+          &N,
+          eig_ptr,
+          static_cast<T*>(work_buf.buffer.raw_ptr()),
+          &lwork,
+          static_cast<int*>(iwork_buf.buffer.raw_ptr()),
+          &liwork,
+          &info);
+      vec_ptr += N * N;
+      eig_ptr += N;
+      if (info != 0) {
+        std::stringstream msg;
+        msg << "[Eigh::eval_cpu] Eigenvalue decomposition failed with error code "
+            << info;
+        throw std::runtime_error(msg.str());
+      }
+    }
+  });
+  if (!compute_eigenvectors) {
+    encoder.add_temporary(vectors);
   }
 }
 
@@ -89,7 +103,8 @@ void Eigh::eval_cpu(
   copy(
       a,
       vectors,
-      a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream());
 
   if (compute_eigenvectors_) {
     // Set the strides and flags so the eigenvectors
@@ -107,14 +122,15 @@ void Eigh::eval_cpu(
         flags.col_contiguous = true;
       }
     }
-    vectors.move_shared_buffer(vectors, strides, flags, vectors.data_size());
+    vectors.copy_shared_buffer(vectors, strides, flags, vectors.data_size());
   }
   switch (a.dtype()) {
     case float32:
-      eigh_impl<float>(vectors, values, uplo_, compute_eigenvectors_);
+      eigh_impl<float>(vectors, values, uplo_, compute_eigenvectors_, stream());
       break;
     case float64:
-      eigh_impl<double>(vectors, values, uplo_, compute_eigenvectors_);
+      eigh_impl<double>(
+          vectors, values, uplo_, compute_eigenvectors_, stream());
       break;
     default:
       throw std::runtime_error(

mlx/backend/cpu/encoder.cpp

@@ -0,0 +1,16 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cpu/encoder.h"
+
+namespace mlx::core::cpu {
+
+CommandEncoder& get_command_encoder(Stream stream) {
+  static std::unordered_map<int, CommandEncoder> encoder_map;
+  auto it = encoder_map.find(stream.index);
+  if (it == encoder_map.end()) {
+    it = encoder_map.emplace(stream.index, stream).first;
+  }
+  return it->second;
+}
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/encoder.h

@@ -0,0 +1,53 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <unordered_map>
+
+#include "mlx/array.h"
+#include "mlx/scheduler.h"
+
+namespace mlx::core::cpu {
+
+struct CommandEncoder {
+  CommandEncoder(Stream stream) : stream_(stream) {}
+
+  CommandEncoder(const CommandEncoder&) = delete;
+  CommandEncoder& operator=(const CommandEncoder&) = delete;
+  CommandEncoder(CommandEncoder&&) = delete;
+  CommandEncoder& operator=(CommandEncoder&&) = delete;
+
+  void set_input_array(const array& a) {}
+  void set_output_array(array& a) {}
+
+  // Hold onto a temporary until any already scheduled tasks which use it as
+  // an input are complete.
+  void add_temporary(array arr) {
+    temporaries_.push_back(std::move(arr));
+  }
+
+  void add_temporaries(std::vector<array> arrays) {
+    temporaries_.insert(
+        temporaries_.end(),
+        std::make_move_iterator(arrays.begin()),
+        std::make_move_iterator(arrays.end()));
+  }
+
+  std::vector<array>& temporaries() {
+    return temporaries_;
+  }
+
+  template <class F, class... Args>
+  void dispatch(F&& f, Args&&... args) {
+    auto task = std::bind(std::forward<F>(f), std::forward<Args>(args)...);
+    scheduler::enqueue(stream_, std::move(task));
+  }
+
+ private:
+  Stream stream_;
+  std::vector<array> temporaries_;
+};
+
+CommandEncoder& get_command_encoder(Stream stream);
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/eval.cpp

@@ -0,0 +1,44 @@
+// Copyright © 2025 Apple Inc.
+#include "mlx/backend/cpu/eval.h"
+#include "mlx/backend/cpu/encoder.h"
+#include "mlx/primitives.h"
+#include "mlx/scheduler.h"
+#include "mlx/utils.h"
+
+namespace mlx::core::cpu {
+
+void eval(array& arr) {
+  auto s = arr.primitive().stream();
+
+  auto outputs = arr.outputs();
+  {
+    // If the array is a tracer hold a reference
+    // to its inputs so they don't get donated
+    std::vector<array> inputs;
+    if (arr.is_tracer()) {
+      inputs = arr.inputs();
+    }
+    arr.primitive().eval_cpu(arr.inputs(), outputs);
+  }
+
+  std::unordered_set<std::shared_ptr<array::Data>> buffers;
+  for (auto& in : arr.inputs()) {
+    buffers.insert(in.data_shared_ptr());
+  }
+  for (auto& s : arr.siblings()) {
+    buffers.insert(s.data_shared_ptr());
+  }
+  // Remove the output if it was donated to by an input
+  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+    buffers.erase(it);
+  }
+  auto& encoder = cpu::get_command_encoder(s);
+  scheduler::notify_new_task(s);
+  encoder.dispatch([s,
+                    buffers = std::move(buffers),
+                    temps = std::move(encoder.temporaries())]() {
+    scheduler::notify_task_completion(s);
+  });
+}
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/eval.h

@@ -0,0 +1,12 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/stream.h"
+
+namespace mlx::core::cpu {
+
+void eval(array& arr);
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/fft.cpp

@@ -4,6 +4,7 @@
 
 #include "mlx/3rdparty/pocketfft.h"
 #include "mlx/allocator.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
@@ -38,46 +39,78 @@ void FFT::eval_cpu(const std::vector<array>& inputs, array& out) {
         });
     scale /= nelem;
   }
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+
   if (in.dtype() == complex64 && out.dtype() == complex64) {
     auto in_ptr =
         reinterpret_cast<const std::complex<float>*>(in.data<complex64_t>());
     auto out_ptr =
         reinterpret_cast<std::complex<float>*>(out.data<complex64_t>());
-    pocketfft::c2c(
-        shape,
-        strides_in,
-        strides_out,
-        axes_,
-        !inverse_,
-        in_ptr,
-        out_ptr,
-        scale);
+    encoder.dispatch([shape = std::move(shape),
+                      strides_in = std::move(strides_in),
+                      strides_out = std::move(strides_out),
+                      axes = axes_,
+                      inverse = inverse_,
+                      in_ptr,
+                      out_ptr,
+                      scale]() {
+      pocketfft::c2c(
+          shape,
+          strides_in,
+          strides_out,
+          axes,
+          !inverse,
+          in_ptr,
+          out_ptr,
+          scale);
+    });
   } else if (in.dtype() == float32 && out.dtype() == complex64) {
     auto in_ptr = in.data<float>();
     auto out_ptr =
         reinterpret_cast<std::complex<float>*>(out.data<complex64_t>());
-    pocketfft::r2c(
-        shape,
-        strides_in,
-        strides_out,
-        axes_,
-        !inverse_,
-        in_ptr,
-        out_ptr,
-        scale);
+    encoder.dispatch([shape = std::move(shape),
+                      strides_in = std::move(strides_in),
+                      strides_out = std::move(strides_out),
+                      axes = axes_,
+                      inverse = inverse_,
+                      in_ptr,
+                      out_ptr,
+                      scale]() {
+      pocketfft::r2c(
+          shape,
+          strides_in,
+          strides_out,
+          axes,
+          !inverse,
+          in_ptr,
+          out_ptr,
+          scale);
+    });
   } else if (in.dtype() == complex64 && out.dtype() == float32) {
     auto in_ptr =
         reinterpret_cast<const std::complex<float>*>(in.data<complex64_t>());
     auto out_ptr = out.data<float>();
-    pocketfft::c2r(
-        shape,
-        strides_in,
-        strides_out,
-        axes_,
-        !inverse_,
-        in_ptr,
-        out_ptr,
-        scale);
+    encoder.dispatch([shape = std::move(shape),
+                      strides_in = std::move(strides_in),
+                      strides_out = std::move(strides_out),
+                      axes = axes_,
+                      inverse = inverse_,
+                      in_ptr,
+                      out_ptr,
+                      scale]() {
+      pocketfft::c2r(
+          shape,
+          strides_in,
+          strides_out,
+          axes,
+          !inverse,
+          in_ptr,
+          out_ptr,
+          scale);
+    });
   } else {
     throw std::runtime_error(
         "[FFT] Received unexpected input and output type combination.");

mlx/backend/cpu/gemm.h

@@ -7,14 +7,20 @@ namespace mlx::core {
 
 template <typename T>
 void matmul(
-    const array& a,
-    const array& b,
-    array& out,
+    const T* a,
+    const T* b,
+    T* out,
     bool a_transposed,
     bool b_transposed,
     size_t lda,
     size_t ldb,
+    size_t ldc,
     float alpha,
-    float beta);
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides);
 
 } // namespace mlx::core

mlx/backend/cpu/gemms/bnns.cpp

@@ -9,39 +9,46 @@
 
 namespace mlx::core {
 
-BNNSDataType to_bnns_dtype(Dtype mlx_dtype) {
-  uint32_t size_bits = size_of(mlx_dtype) * 8;
-  switch (kindof(mlx_dtype)) {
-    case Dtype::Kind::b:
-      return BNNSDataTypeBoolean;
-    case Dtype::Kind::u:
-      return BNNSDataType(BNNSDataTypeUIntBit | size_bits);
-    case Dtype::Kind::i:
-      return BNNSDataType(BNNSDataTypeIntBit | size_bits);
-    case Dtype::Kind::f:
-      return BNNSDataType(BNNSDataTypeFloatBit | size_bits);
-    case Dtype::Kind::V:
-      return BNNSDataTypeBFloat16;
-    case Dtype::Kind::c:
-      throw std::invalid_argument("BNNS does not support complex types");
-  }
+template <typename T>
+constexpr BNNSDataType to_bnns_dtype();
+
+template <>
+constexpr BNNSDataType to_bnns_dtype<float>() {
+  return BNNSDataType(BNNSDataTypeFloatBit | 32);
+}
+template <>
+constexpr BNNSDataType to_bnns_dtype<float16_t>() {
+  return BNNSDataType(BNNSDataTypeFloatBit | 16);
 }
 
+template <>
+constexpr BNNSDataType to_bnns_dtype<bfloat16_t>() {
+  return BNNSDataTypeBFloat16;
+}
+
+template <typename T>
 void matmul_bnns(
-    const array& a,
-    const array& b,
-    array& out,
+    const T* a,
+    const T* b,
+    T* out,
     bool a_transposed,
     bool b_transposed,
     size_t lda,
     size_t ldb,
+    size_t ldc,
     float alpha,
-    float beta) {
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
 
-  BNNSDataType bnns_dtype = to_bnns_dtype(out.dtype());
+  BNNSDataType bnns_dtype = to_bnns_dtype<T>();
 
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
@@ -115,14 +122,14 @@ void matmul_bnns(
   auto bnns_filter =
       BNNSFilterCreateLayerBroadcastMatMul(&gemm_params, nullptr);
 
-  for (int i = 0; i < (a.size() / (M * K)); ++i) {
+  for (int i = 0; i < batch_size; ++i) {
     BNNSFilterApplyTwoInput(
         bnns_filter,
-        a.data<uint8_t>() +
-            elem_to_loc(M * K * i, a.shape(), a.strides()) * a.itemsize(),
-        b.data<uint8_t>() +
-            elem_to_loc(K * N * i, b.shape(), b.strides()) * b.itemsize(),
-        out.data<uint8_t>() + M * N * i * out.itemsize());
+        reinterpret_cast<const uint8_t*>(
+            a + elem_to_loc(M * K * i, a_shape, a_strides)),
+        reinterpret_cast<const uint8_t*>(
+            b + elem_to_loc(K * N * i, b_shape, b_strides)),
+        reinterpret_cast<uint8_t*>(out + M * N * i));
   }
 
   BNNSFilterDestroy(bnns_filter);
@@ -131,30 +138,72 @@ void matmul_bnns(
 
 template <>
 void matmul<float16_t>(
-    const array& a,
-    const array& b,
-    array& out,
+    const float16_t* a,
+    const float16_t* b,
+    float16_t* out,
     bool a_transposed,
     bool b_transposed,
     size_t lda,
     size_t ldb,
+    size_t ldc,
     float alpha,
-    float beta) {
-  matmul_bnns(a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta);
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  matmul_bnns(
+      a,
+      b,
+      out,
+      a_transposed,
+      b_transposed,
+      lda,
+      ldb,
+      ldc,
+      alpha,
+      beta,
+      batch_size,
+      a_shape,
+      a_strides,
+      b_shape,
+      b_strides);
 }
 
 template <>
 void matmul<bfloat16_t>(
-    const array& a,
-    const array& b,
-    array& out,
+    const bfloat16_t* a,
+    const bfloat16_t* b,
+    bfloat16_t* out,
     bool a_transposed,
     bool b_transposed,
     size_t lda,
     size_t ldb,
+    size_t ldc,
     float alpha,
-    float beta) {
-  matmul_bnns(a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta);
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  matmul_bnns(
+      a,
+      b,
+      out,
+      a_transposed,
+      b_transposed,
+      lda,
+      ldb,
+      ldc,
+      alpha,
+      beta,
+      batch_size,
+      a_shape,
+      a_strides,
+      b_shape,
+      b_strides);
 }
 
 } // namespace mlx::core

mlx/backend/cpu/gemms/cblas.cpp

@@ -8,20 +8,27 @@ namespace mlx::core {
 
 template <>
 void matmul<float>(
-    const array& a,
-    const array& b,
-    array& out,
+    const float* a,
+    const float* b,
+    float* out,
     bool a_transposed,
     bool b_transposed,
     size_t lda,
     size_t ldb,
+    size_t ldc,
     float alpha,
-    float beta) {
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
 
-  for (int i = 0; i < (a.size() / (M * K)); ++i) {
+  for (int i = 0; i < batch_size; ++i) {
     cblas_sgemm(
         CblasRowMajor,
         a_transposed ? CblasTrans : CblasNoTrans, // transA
@@ -29,34 +36,40 @@ void matmul<float>(
         M,
         N,
         K,
-        alpha, // alpha
-        a.data<float>() + elem_to_loc(M * K * i, a.shape(), a.strides()),
+        alpha,
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
         lda,
-        b.data<float>() + elem_to_loc(K * N * i, b.shape(), b.strides()),
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
         ldb,
-        beta, // beta
-        out.data<float>() + M * N * i,
-        out.shape(-1) // ldc
-    );
+        beta,
+        out + M * N * i,
+        ldc);
   }
 }
 
 template <>
 void matmul<double>(
-    const array& a,
-    const array& b,
-    array& out,
+    const double* a,
+    const double* b,
+    double* out,
     bool a_transposed,
     bool b_transposed,
     size_t lda,
     size_t ldb,
+    size_t ldc,
     float alpha,
-    float beta) {
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
 
-  for (int i = 0; i < (a.size() / (M * K)); ++i) {
+  for (int i = 0; i < batch_size; ++i) {
     cblas_dgemm(
         CblasRowMajor,
         a_transposed ? CblasTrans : CblasNoTrans, // transA
@@ -64,15 +77,14 @@ void matmul<double>(
         M,
         N,
         K,
-        alpha, // alpha
-        a.data<double>() + elem_to_loc(M * K * i, a.shape(), a.strides()),
+        alpha,
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
         lda,
-        b.data<double>() + elem_to_loc(K * N * i, b.shape(), b.strides()),
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
         ldb,
-        beta, // beta
-        out.data<double>() + M * N * i,
-        out.shape(-1) // ldc
-    );
+        beta,
+        out + M * N * i,
+        ldc);
   }
 }
 

mlx/backend/cpu/gemms/no_bf16.cpp

@@ -6,15 +6,21 @@ namespace mlx::core {
 
 template <>
 void matmul<bfloat16_t>(
-    const array&,
-    const array&,
-    array&,
+    const bfloat16_t*,
+    const bfloat16_t*,
+    bfloat16_t*,
     bool,
     bool,
     size_t,
     size_t,
+    size_t,
     float,
-    float) {
+    float,
+    size_t,
+    const Shape&,
+    const Strides&,
+    const Shape&,
+    const Strides&) {
   throw std::runtime_error("[Matmul::eval_cpu] bfloat16 not supported.");
 }
 

mlx/backend/cpu/gemms/no_fp16.cpp

@@ -6,15 +6,21 @@ namespace mlx::core {
 
 template <>
 void matmul<float16_t>(
-    const array&,
-    const array&,
-    array&,
+    const float16_t*,
+    const float16_t*,
+    float16_t*,
     bool,
     bool,
     size_t,
     size_t,
+    size_t,
     float,
-    float) {
+    float,
+    size_t,
+    const Shape&,
+    const Strides&,
+    const Shape&,
+    const Strides&) {
   throw std::runtime_error("[Matmul::eval_cpu] float16 not supported.");
 }
 

mlx/backend/cpu/hadamard.cpp

@@ -4,16 +4,17 @@
 
 #include "mlx/backend/common/hadamard.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 // n = 2^k component
 template <typename T>
-void hadamard_n(array& out, int n, int m, float scale) {
-  for (int b = 0; b < out.size() / n; b++) {
+void hadamard_n(T* out, int n, int m, float scale, size_t size) {
+  for (int b = 0; b < size / n; b++) {
     size_t loc = b * n;
-    T* data_ptr = out.data<T>() + loc;
+    T* data_ptr = out + loc;
     int h = 1;
     int n_over_2 = n / 2;
     while (h < n) {
@@ -36,7 +37,7 @@ void hadamard_n(array& out, int n, int m, float scale) {
 
 // m component
 template <typename T>
-void hadamard_m(array& out, int n, int m, float scale) {
+void hadamard_m(T* out, int n, int m, float scale, size_t size) {
   auto h_matrices = hadamard_matrices();
   auto& matrix = h_matrices[m];
   auto start = 1;
@@ -51,9 +52,9 @@ void hadamard_m(array& out, int n, int m, float scale) {
     end = matrix.find('\n', start);
   }
 
-  for (int b = 0; b < out.size() / m / n; b++) {
+  for (int b = 0; b < size / m / n; b++) {
     size_t loc = b * n * m;
-    T* data_ptr = out.data<T>() + loc;
+    T* data_ptr = out + loc;
     for (int i = 0; i < n; i++) {
       std::vector<float> out(m);
       for (int j = 0; j < m; j++) {
@@ -74,12 +75,17 @@ void hadamard_m(array& out, int n, int m, float scale) {
 }
 
 template <typename T>
-void hadamard(array& out, int n, int m, float scale) {
-  float n_scale = m > 1 ? 1.0 : scale;
-  hadamard_n<T>(out, n, m, n_scale);
-  if (m > 1) {
-    hadamard_m<T>(out, n, m, scale);
-  }
+void hadamard(array& out, int n, int m, float scale, Stream stream) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(out);
+  auto out_ptr = out.data<T>();
+  encoder.dispatch([out_ptr, size = out.size(), n, m, scale]() {
+    float n_scale = m > 1 ? 1.0 : scale;
+    hadamard_n<T>(out_ptr, n, m, n_scale, size);
+    if (m > 1) {
+      hadamard_m<T>(out_ptr, n, m, scale, size);
+    }
+  });
 }
 
 void Hadamard::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -87,18 +93,26 @@ void Hadamard::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Copy input to output
-  copy(in, out, CopyType::General);
+  if (in.flags().row_contiguous && in.is_donatable()) {
+    out.copy_shared_buffer(in);
+  } else {
+    copy(
+        in,
+        out,
+        in.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+        stream());
+  }
 
   int axis = out.ndim() - 1;
   auto [n, m] = decompose_hadamard(out.shape(axis));
 
   switch (in.dtype()) {
     case float32:
-      return hadamard<float>(out, n, m, scale_);
+      return hadamard<float>(out, n, m, scale_, stream());
     case float16:
-      return hadamard<float16_t>(out, n, m, scale_);
+      return hadamard<float16_t>(out, n, m, scale_, stream());
     case bfloat16:
-      return hadamard<bfloat16_t>(out, n, m, scale_);
+      return hadamard<bfloat16_t>(out, n, m, scale_, stream());
     default:
       throw std::invalid_argument("[hadamard] Unsupported type.");
   }

mlx/backend/cpu/indexing.cpp

@@ -8,6 +8,7 @@
 
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 
 namespace mlx::core {
 
@@ -21,6 +22,40 @@ inline size_t offset_neg_idx(uint32_t idx, size_t) {
   return idx;
 }
 
+struct None {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) = x;
+  }
+};
+struct Sum {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) += x;
+  }
+};
+
+struct Prod {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) *= x;
+  }
+};
+
+struct Max {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) = (*y > x) ? *y : x;
+  }
+};
+
+struct Min {
+  template <typename T>
+  void operator()(T x, T* y) {
+    (*y) = (*y < x) ? *y : x;
+  }
+};
+
 template <typename T, typename IdxT>
 void gather(
     const array& src,
@@ -73,13 +108,14 @@ void gather(
   size_t ind_size = slice_size == 0 ? 0 : out.size() / slice_size;
   const T* src_ptr = src.data<T>();
   T* dst_ptr = out.data<T>();
-  size_t out_idx = 0;
 
   std::vector<ContiguousIterator> its(inds.begin(), inds.end());
   ContiguousIterator src_it;
   if (!can_copy && src.ndim() > 0) {
     src_it = ContiguousIterator(slice_sizes, src.strides(), src.ndim());
   }
+
+  size_t out_idx = 0;
   for (int idx = 0; idx < ind_size; idx++) {
     size_t src_idx = 0;
     for (int ii = 0; ii < inds.size(); ++ii) {
@@ -164,43 +200,56 @@ void Gather::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
   auto& src = inputs[0];
-  std::vector<array> inds(inputs.begin() + 1, inputs.end());
-
-  if (inds.empty()) {
-    dispatch_gather<uint8_t>(src, inds, out, axes_, slice_sizes_);
-    return;
+  std::vector<array> inds;
+  for (auto it = inputs.begin() + 1; it < inputs.end(); ++it) {
+    inds.push_back(array::unsafe_weak_copy(*it));
   }
-
-  switch (inds[0].dtype()) {
-    case uint8:
+  auto& encoder = cpu::get_command_encoder(stream());
+  for (auto& in : inputs) {
+    encoder.set_input_array(in);
+  }
+  encoder.set_output_array(out);
+  encoder.dispatch([axes_ = axes_,
+                    slice_sizes_ = slice_sizes_,
+                    src = array::unsafe_weak_copy(src),
+                    inds = std::move(inds),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    if (inds.empty()) {
       dispatch_gather<uint8_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case uint16:
-      dispatch_gather<uint16_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case uint32:
-      dispatch_gather<uint32_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case uint64:
-      dispatch_gather<uint64_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int8:
-      dispatch_gather<int8_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int16:
-      dispatch_gather<int16_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int32:
-      dispatch_gather<int32_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    case int64:
-      dispatch_gather<int64_t>(src, inds, out, axes_, slice_sizes_);
-      break;
-    default:
-      throw std::runtime_error(
-          "[Gather::eval_cpu] Cannot gather with indices type.");
-      break;
-  }
+      return;
+    }
+
+    switch (inds[0].dtype()) {
+      case uint8:
+        dispatch_gather<uint8_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case uint16:
+        dispatch_gather<uint16_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case uint32:
+        dispatch_gather<uint32_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case uint64:
+        dispatch_gather<uint64_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int8:
+        dispatch_gather<int8_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int16:
+        dispatch_gather<int16_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int32:
+        dispatch_gather<int32_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      case int64:
+        dispatch_gather<int64_t>(src, inds, out, axes_, slice_sizes_);
+        break;
+      default:
+        throw std::runtime_error(
+            "[Gather::eval_cpu] Cannot gather with indices type.");
+        break;
+    }
+  });
 }
 template <typename T, typename IdxT>
 void gather_axis(
@@ -235,6 +284,7 @@ void gather_axis(
   for (int i = axis + 1; i < ind.ndim(); ++i) {
     size_post *= ind.shape(i);
   }
+
   size_t stride_pre = size_post * ind_ax_size;
   for (size_t i = 0; i < size_pre; i++) {
     for (size_t k = 0; k < size_post; k++) {
@@ -305,38 +355,48 @@ void dispatch_gather_axis(
 
 void GatherAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
   auto& src = inputs[0];
   auto& inds = inputs[1];
-  switch (inds.dtype()) {
-    case uint8:
-      dispatch_gather_axis<uint8_t>(src, inds, out, axis_);
-      break;
-    case uint16:
-      dispatch_gather_axis<uint16_t>(src, inds, out, axis_);
-      break;
-    case uint32:
-      dispatch_gather_axis<uint32_t>(src, inds, out, axis_);
-      break;
-    case uint64:
-      dispatch_gather_axis<uint64_t>(src, inds, out, axis_);
-      break;
-    case int8:
-      dispatch_gather_axis<int8_t>(src, inds, out, axis_);
-      break;
-    case int16:
-      dispatch_gather_axis<int16_t>(src, inds, out, axis_);
-      break;
-    case int32:
-      dispatch_gather_axis<int32_t>(src, inds, out, axis_);
-      break;
-    case int64:
-      dispatch_gather_axis<int64_t>(src, inds, out, axis_);
-      break;
-    default:
-      throw std::runtime_error(
-          "[GatherAxis::eval_cpu] Cannot gather with indices type.");
-      break;
-  }
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(src);
+  encoder.set_input_array(inds);
+  encoder.set_output_array(out);
+  encoder.dispatch([axis_ = axis_,
+                    src = array::unsafe_weak_copy(src),
+                    inds = array::unsafe_weak_copy(inds),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (inds.dtype()) {
+      case uint8:
+        dispatch_gather_axis<uint8_t>(src, inds, out, axis_);
+        break;
+      case uint16:
+        dispatch_gather_axis<uint16_t>(src, inds, out, axis_);
+        break;
+      case uint32:
+        dispatch_gather_axis<uint32_t>(src, inds, out, axis_);
+        break;
+      case uint64:
+        dispatch_gather_axis<uint64_t>(src, inds, out, axis_);
+        break;
+      case int8:
+        dispatch_gather_axis<int8_t>(src, inds, out, axis_);
+        break;
+      case int16:
+        dispatch_gather_axis<int16_t>(src, inds, out, axis_);
+        break;
+      case int32:
+        dispatch_gather_axis<int32_t>(src, inds, out, axis_);
+        break;
+      case int64:
+        dispatch_gather_axis<int64_t>(src, inds, out, axis_);
+        break;
+      default:
+        throw std::runtime_error(
+            "[GatherAxis::eval_cpu] Cannot gather with indices type.");
+        break;
+    }
+  });
 }
 
 template <typename InT, typename IdxT, typename OpT>
@@ -344,8 +404,7 @@ void scatter(
     const array& updates,
     array& out,
     const std::vector<array>& inds,
-    const std::vector<int>& axes,
-    const OpT& op) {
+    const std::vector<int>& axes) {
   int nind = inds.size();
   auto inds_ndim = updates.ndim() - out.ndim();
   size_t n_updates = nind ? inds[0].size() : 1;
@@ -361,9 +420,11 @@ void scatter(
   ContiguousIterator update_it(updates);
   ContiguousIterator out_it(update_shape, out.strides(), out.ndim());
 
+  auto out_ptr = out.data<InT>();
+  auto upd_ptr = updates.data<InT>();
   for (int i = 0; i < n_updates; ++i) {
     size_t out_offset = 0;
-    for (int j = 0; j < nind; ++j) {
+    for (int j = 0; j < inds.size(); ++j) {
       auto ax = axes[j];
       auto idx_loc = its[j].loc;
       its[j].step();
@@ -373,8 +434,7 @@ void scatter(
     }
     update_it.seek(i * update_size);
     for (int j = 0; j < update_size; ++j) {
-      op(updates.data<InT>()[update_it.loc],
-         out.data<InT>() + out_offset + out_it.loc);
+      OpT{}(upd_ptr[update_it.loc], out_ptr + out_offset + out_it.loc);
       update_it.step();
       out_it.step();
     }
@@ -392,26 +452,19 @@ void dispatch_scatter_inds(
     Scatter::ReduceType rtype) {
   switch (rtype) {
     case Scatter::None:
-      scatter<InT, IdxT>(
-          updates, out, indices, axes, [](auto x, auto* y) { (*y) = x; });
+      scatter<InT, IdxT, None>(updates, out, indices, axes);
       break;
     case Scatter::Sum:
-      scatter<InT, IdxT>(
-          updates, out, indices, axes, [](auto x, auto* y) { (*y) += x; });
+      scatter<InT, IdxT, Sum>(updates, out, indices, axes);
       break;
     case Scatter::Prod:
-      scatter<InT, IdxT>(
-          updates, out, indices, axes, [](auto x, auto* y) { (*y) *= x; });
+      scatter<InT, IdxT, Prod>(updates, out, indices, axes);
       break;
     case Scatter::Max:
-      scatter<InT, IdxT>(updates, out, indices, axes, [](auto x, auto* y) {
-        (*y) = (*y > x) ? *y : x;
-      });
+      scatter<InT, IdxT, Max>(updates, out, indices, axes);
       break;
     case Scatter::Min:
-      scatter<InT, IdxT>(updates, out, indices, axes, [](auto x, auto* y) {
-        (*y) = (*y < x) ? *y : x;
-      });
+      scatter<InT, IdxT, Min>(updates, out, indices, axes);
       break;
   }
 }
@@ -463,67 +516,75 @@ void Scatter::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() >= 2);
 
   auto& src = inputs[0];
-  std::vector<array> inds(inputs.begin() + 1, inputs.end() - 1);
   auto& updates = inputs.back();
 
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype);
+  copy(src, out, ctype, stream());
 
-  switch (src.dtype()) {
-    case bool_:
-      dispatch_scatter<bool>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint8:
-      dispatch_scatter<uint8_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint16:
-      dispatch_scatter<uint16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint32:
-      dispatch_scatter<uint32_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case uint64:
-      dispatch_scatter<uint64_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int8:
-      dispatch_scatter<int8_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int16:
-      dispatch_scatter<int16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int32:
-      dispatch_scatter<int32_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case int64:
-      dispatch_scatter<int64_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case float16:
-      dispatch_scatter<float16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case float32:
-      dispatch_scatter<float>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case float64:
-      dispatch_scatter<double>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case bfloat16:
-      dispatch_scatter<bfloat16_t>(out, inds, updates, axes_, reduce_type_);
-      break;
-    case complex64:
-      dispatch_scatter<complex64_t>(out, inds, updates, axes_, reduce_type_);
-      break;
+  auto& encoder = cpu::get_command_encoder(stream());
+  std::vector<array> inds;
+  for (auto it = inputs.begin() + 1; it < inputs.end() - 1; ++it) {
+    encoder.set_input_array(*it);
+    inds.push_back(array::unsafe_weak_copy(*it));
   }
+  encoder.set_input_array(updates);
+  encoder.set_output_array(out);
+  encoder.dispatch([axes_ = axes_,
+                    reduce_type_ = reduce_type_,
+                    updates = array::unsafe_weak_copy(updates),
+                    inds = std::move(inds),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        dispatch_scatter<bool>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint8:
+        dispatch_scatter<uint8_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint16:
+        dispatch_scatter<uint16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint32:
+        dispatch_scatter<uint32_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case uint64:
+        dispatch_scatter<uint64_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int8:
+        dispatch_scatter<int8_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int16:
+        dispatch_scatter<int16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int32:
+        dispatch_scatter<int32_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case int64:
+        dispatch_scatter<int64_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case float16:
+        dispatch_scatter<float16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case float32:
+        dispatch_scatter<float>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case float64:
+        dispatch_scatter<double>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case bfloat16:
+        dispatch_scatter<bfloat16_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+      case complex64:
+        dispatch_scatter<complex64_t>(out, inds, updates, axes_, reduce_type_);
+        break;
+    }
+  });
 }
 
 template <typename T, typename IdxT, typename OpT>
-void scatter_axis(
-    array& out,
-    const array idx,
-    const array& upd,
-    int axis,
-    const OpT& op) {
+void scatter_axis(array& out, const array idx, const array& upd, int axis) {
   auto strides = idx.strides();
   strides.erase(strides.begin() + axis);
   auto shape = idx.shape();
@@ -557,8 +618,9 @@ void scatter_axis(
       for (int j = 0; j < idx_ax_size; ++j) {
         auto ind_val = offset_neg_idx(
             idx_ptr[idx_it.loc + j * idx_ax_stride], dst_ax_size);
-        op(upd_ptr[upd_it.loc + j * upd_ax_stride],
-           dst_ptr + k + ind_val * dst_ax_stride);
+        OpT{}(
+            upd_ptr[upd_it.loc + j * upd_ax_stride],
+            dst_ptr + k + ind_val * dst_ax_stride);
       }
       idx_it.step();
       upd_it.step();
@@ -576,12 +638,10 @@ void dispatch_scatter_axis_op(
     ScatterAxis::ReduceType rtype) {
   switch (rtype) {
     case ScatterAxis::None:
-      scatter_axis<InT, IdxT>(
-          out, idx, updates, axis, [](auto x, auto* y) { (*y) = x; });
+      scatter_axis<InT, IdxT, None>(out, idx, updates, axis);
       break;
     case ScatterAxis::Sum:
-      scatter_axis<InT, IdxT>(
-          out, idx, updates, axis, [](auto x, auto* y) { (*y) += x; });
+      scatter_axis<InT, IdxT, Sum>(out, idx, updates, axis);
       break;
   }
 }
@@ -634,53 +694,65 @@ void ScatterAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype);
+  copy(src, out, ctype, stream());
 
-  switch (src.dtype()) {
-    case bool_:
-      dispatch_scatter_axis<bool>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case uint8:
-      dispatch_scatter_axis<uint8_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case uint16:
-      dispatch_scatter_axis<uint16_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case uint32:
-      dispatch_scatter_axis<uint32_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case uint64:
-      dispatch_scatter_axis<uint64_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case int8:
-      dispatch_scatter_axis<int8_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case int16:
-      dispatch_scatter_axis<int16_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case int32:
-      dispatch_scatter_axis<int32_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case int64:
-      dispatch_scatter_axis<int64_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case float16:
-      dispatch_scatter_axis<float16_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case float32:
-      dispatch_scatter_axis<float>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case float64:
-      dispatch_scatter_axis<double>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case bfloat16:
-      dispatch_scatter_axis<bfloat16_t>(out, idx, updates, axis_, reduce_type_);
-      break;
-    case complex64:
-      dispatch_scatter_axis<complex64_t>(
-          out, idx, updates, axis_, reduce_type_);
-      break;
-  }
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(idx);
+  encoder.set_input_array(updates);
+  encoder.set_output_array(out);
+  encoder.dispatch([axis_ = axis_,
+                    reduce_type_ = reduce_type_,
+                    idx = array::unsafe_weak_copy(idx),
+                    updates = array::unsafe_weak_copy(updates),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        dispatch_scatter_axis<bool>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint8:
+        dispatch_scatter_axis<uint8_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint16:
+        dispatch_scatter_axis<uint16_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint32:
+        dispatch_scatter_axis<uint32_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case uint64:
+        dispatch_scatter_axis<uint64_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int8:
+        dispatch_scatter_axis<int8_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int16:
+        dispatch_scatter_axis<int16_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int32:
+        dispatch_scatter_axis<int32_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case int64:
+        dispatch_scatter_axis<int64_t>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case float16:
+        dispatch_scatter_axis<float16_t>(
+            out, idx, updates, axis_, reduce_type_);
+        break;
+      case float32:
+        dispatch_scatter_axis<float>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case float64:
+        dispatch_scatter_axis<double>(out, idx, updates, axis_, reduce_type_);
+        break;
+      case bfloat16:
+        dispatch_scatter_axis<bfloat16_t>(
+            out, idx, updates, axis_, reduce_type_);
+        break;
+      case complex64:
+        dispatch_scatter_axis<complex64_t>(
+            out, idx, updates, axis_, reduce_type_);
+        break;
+    }
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/inverse.cpp

@@ -2,20 +2,21 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 template <typename T>
-void general_inv(array& inv, int N, int i) {
+void general_inv(T* inv, int N) {
   int info;
   auto ipiv = array::Data{allocator::malloc_or_wait(sizeof(int) * N)};
   // Compute LU factorization.
   getrf<T>(
       /* m = */ &N,
       /* n = */ &N,
-      /* a = */ inv.data<T>() + N * N * i,
+      /* a = */ inv,
       /* lda = */ &N,
       /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
       /* info = */ &info);
@@ -53,7 +54,7 @@ void general_inv(array& inv, int N, int i) {
   // Compute inverse.
   getri<T>(
       /* m = */ &N,
-      /* a = */ inv.data<T>() + N * N * i,
+      /* a = */ inv,
       /* lda = */ &N,
       /* ipiv = */ static_cast<int*>(ipiv.buffer.raw_ptr()),
       /* work = */ static_cast<T*>(scratch.buffer.raw_ptr()),
@@ -68,29 +69,28 @@ void general_inv(array& inv, int N, int i) {
 }
 
 template <typename T>
-void tri_inv(array& inv, int N, int i, bool upper) {
+void tri_inv(T* inv, int N, bool upper) {
   const char uplo = upper ? 'L' : 'U';
   const char diag = 'N';
-  T* data = inv.data<T>() + N * N * i;
   int info;
   trtri<T>(
       /* uplo = */ &uplo,
       /* diag = */ &diag,
       /* N = */ &N,
-      /* a = */ data,
+      /* a = */ inv,
       /* lda = */ &N,
       /* info = */ &info);
 
   // zero out the other triangle
   if (upper) {
     for (int i = 0; i < N; i++) {
-      std::fill(data, data + i, 0.0f);
-      data += N;
+      std::fill(inv, inv + i, 0.0f);
+      inv += N;
     }
   } else {
     for (int i = 0; i < N; i++) {
-      std::fill(data + i + 1, data + N, 0.0f);
-      data += N;
+      std::fill(inv + i + 1, inv + N, 0.0f);
+      inv += N;
     }
   }
 
@@ -103,34 +103,53 @@ void tri_inv(array& inv, int N, int i, bool upper) {
 }
 
 template <typename T>
-void inverse_impl(const array& a, array& inv, bool tri, bool upper) {
+void inverse_impl(
+    const array& a,
+    array& inv,
+    bool tri,
+    bool upper,
+    Stream stream) {
   // 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);
+  copy(
+      a,
+      inv,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream);
 
   const int N = a.shape(-1);
   const size_t num_matrices = a.size() / (N * N);
 
-  for (int i = 0; i < num_matrices; i++) {
-    if (tri) {
-      tri_inv<T>(inv, N, i, upper);
-    } else {
-      general_inv<T>(inv, N, i);
-    }
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_output_array(inv);
+
+  auto inv_ptr = inv.data<T>();
+  if (tri) {
+    encoder.dispatch([inv_ptr, N, num_matrices, upper]() {
+      for (int i = 0; i < num_matrices; i++) {
+        tri_inv<T>(inv_ptr + N * N * i, N, upper);
+      }
+    });
+  } else {
+    encoder.dispatch([inv_ptr, N, num_matrices]() {
+      for (int i = 0; i < num_matrices; i++) {
+        general_inv<T>(inv_ptr + N * N * i, N);
+      }
+    });
   }
 }
 
 void Inverse::eval_cpu(const std::vector<array>& inputs, array& output) {
   switch (inputs[0].dtype()) {
     case float32:
-      inverse_impl<float>(inputs[0], output, tri_, upper_);
+      inverse_impl<float>(inputs[0], output, tri_, upper_, stream());
       break;
     case float64:
-      inverse_impl<double>(inputs[0], output, tri_, upper_);
+      inverse_impl<double>(inputs[0], output, tri_, upper_, stream());
       break;
     default:
       throw std::runtime_error(

mlx/backend/cpu/luf.cpp

@@ -4,15 +4,22 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 template <typename T>
-void luf_impl(const array& a, array& lu, array& pivots, array& row_indices) {
+void luf_impl(
+    const array& a,
+    array& lu,
+    array& pivots,
+    array& row_indices,
+    Stream stream) {
   int M = a.shape(-2);
   int N = a.shape(-1);
+  int K = std::min(M, N);
 
   // Copy a into lu and make it col contiguous
   auto ndim = lu.ndim();
@@ -26,57 +33,72 @@ void luf_impl(const array& a, array& lu, array& pivots, array& row_indices) {
   lu.set_data(
       allocator::malloc_or_wait(lu.nbytes()), lu.nbytes(), strides, flags);
   copy_inplace(
-      a, lu, a.shape(), a.strides(), strides, 0, 0, CopyType::GeneralGeneral);
+      a,
+      lu,
+      a.shape(),
+      a.strides(),
+      strides,
+      0,
+      0,
+      CopyType::GeneralGeneral,
+      stream);
 
   auto a_ptr = lu.data<T>();
-
   pivots.set_data(allocator::malloc_or_wait(pivots.nbytes()));
   row_indices.set_data(allocator::malloc_or_wait(row_indices.nbytes()));
   auto pivots_ptr = pivots.data<uint32_t>();
   auto row_indices_ptr = row_indices.data<uint32_t>();
-
-  int info;
   size_t num_matrices = a.size() / (M * N);
-  for (size_t i = 0; i < num_matrices; ++i) {
-    // Compute LU factorization of A
-    getrf<T>(
-        /* m */ &M,
-        /* n */ &N,
-        /* a */ a_ptr,
-        /* lda */ &M,
-        /* ipiv */ reinterpret_cast<int*>(pivots_ptr),
-        /* info */ &info);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(lu);
+  encoder.set_output_array(pivots);
+  encoder.set_output_array(row_indices);
 
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "[LUF::eval_cpu] sgetrf_ failed with code " << info
-         << ((info > 0) ? " because matrix is singular"
-                        : " because argument had an illegal value");
-      throw std::runtime_error(ss.str());
-    }
+  encoder.dispatch(
+      [a_ptr, pivots_ptr, row_indices_ptr, num_matrices, M, N, K]() mutable {
+        int info;
+        for (size_t i = 0; i < num_matrices; ++i) {
+          // Compute LU factorization of A
+          getrf<T>(
+              /* m */ &M,
+              /* n */ &N,
+              /* a */ a_ptr,
+              /* lda */ &M,
+              /* ipiv */ reinterpret_cast<int*>(pivots_ptr),
+              /* info */ &info);
 
-    // Subtract 1 to get 0-based index
-    int j = 0;
-    for (; j < pivots.shape(-1); ++j) {
-      pivots_ptr[j]--;
-      row_indices_ptr[j] = j;
-    }
-    for (; j < row_indices.shape(-1); ++j) {
-      row_indices_ptr[j] = j;
-    }
-    for (int j = pivots.shape(-1) - 1; j >= 0; --j) {
-      auto piv = pivots_ptr[j];
-      auto t1 = row_indices_ptr[piv];
-      auto t2 = row_indices_ptr[j];
-      row_indices_ptr[j] = t1;
-      row_indices_ptr[piv] = t2;
-    }
+          if (info != 0) {
+            std::stringstream ss;
+            ss << "[LUF::eval_cpu] sgetrf_ failed with code " << info
+               << ((info > 0) ? " because matrix is singular"
+                              : " because argument had an illegal value");
+            throw std::runtime_error(ss.str());
+          }
 
-    // Advance pointers to the next matrix
-    a_ptr += M * N;
-    pivots_ptr += pivots.shape(-1);
-    row_indices_ptr += pivots.shape(-1);
-  }
+          // Subtract 1 to get 0-based index
+          int j = 0;
+          for (; j < K; ++j) {
+            pivots_ptr[j]--;
+            row_indices_ptr[j] = j;
+          }
+          for (; j < M; ++j) {
+            row_indices_ptr[j] = j;
+          }
+          for (int j = K - 1; j >= 0; --j) {
+            auto piv = pivots_ptr[j];
+            auto t1 = row_indices_ptr[piv];
+            auto t2 = row_indices_ptr[j];
+            row_indices_ptr[j] = t1;
+            row_indices_ptr[piv] = t2;
+          }
+
+          // Advance pointers to the next matrix
+          a_ptr += M * N;
+          pivots_ptr += K;
+          row_indices_ptr += M;
+        }
+      });
 }
 
 void LUF::eval_cpu(
@@ -85,10 +107,10 @@ void LUF::eval_cpu(
   assert(inputs.size() == 1);
   switch (inputs[0].dtype()) {
     case float32:
-      luf_impl<float>(inputs[0], outputs[0], outputs[1], outputs[2]);
+      luf_impl<float>(inputs[0], outputs[0], outputs[1], outputs[2], stream());
       break;
     case float64:
-      luf_impl<double>(inputs[0], outputs[0], outputs[1], outputs[2]);
+      luf_impl<double>(inputs[0], outputs[0], outputs[1], outputs[2], stream());
       break;
     default:
       throw std::runtime_error(

mlx/backend/cpu/masked_mm.cpp

@@ -5,6 +5,7 @@
 #include "mlx/array.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
@@ -64,36 +65,36 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& b_pre = inputs[1];
 
   auto check_transpose =
-      [](const array& arr, bool do_copy, bool expand_all = false) {
+      [s = stream()](const array& arr, bool do_copy, bool expand_all = false) {
         auto stx = arr.strides()[arr.ndim() - 2];
         auto sty = arr.strides()[arr.ndim() - 1];
         if (!expand_all && stx == arr.shape(-1) && sty == 1) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector);
-            return std::make_tuple(false, stx, arr_copy);
+            copy(arr, arr_copy, CopyType::Vector, s);
+            return std::make_tuple(false, stx, arr_copy, true);
           }
-          return std::make_tuple(false, stx, arr);
+          return std::make_tuple(false, stx, arr, false);
         } else if (!expand_all && stx == 1 && sty == arr.shape(-2)) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector);
-            return std::make_tuple(true, sty, arr_copy);
+            copy(arr, arr_copy, CopyType::Vector, s);
+            return std::make_tuple(true, sty, arr_copy, true);
           }
-          return std::make_tuple(true, sty, arr);
+          return std::make_tuple(true, sty, arr, false);
         } else {
           array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy(arr, arr_copy, CopyType::General);
+          copy(arr, arr_copy, CopyType::General, s);
           int64_t stx = arr.shape(-1);
-          return std::make_tuple(false, stx, arr_copy);
+          return std::make_tuple(false, stx, arr_copy, true);
         }
       };
 
   bool has_op_mask = inputs.size() > 3;
   bool has_out_mask = inputs.size() == 3 || inputs.size() == 5;
-  auto [a_transposed, lda, a] =
+  auto [a_transposed, lda, a, a_copied] =
       check_transpose(a_pre, has_op_mask, inputs.back().dtype() != bool_);
-  auto [b_transposed, ldb, b] =
+  auto [b_transposed, ldb, b, b_copied] =
       check_transpose(b_pre, has_op_mask, inputs.back().dtype() != bool_);
 
   size_t M = a.shape(-2);
@@ -104,31 +105,39 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
+  auto& encoder = cpu::get_command_encoder(stream());
   if (K == 0) {
-    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
+    encoder.set_output_array(out);
+    encoder.dispatch([out_ptr = out.data<void>(), nbytes = out.nbytes()]() {
+      std::memset(out_ptr, 0, nbytes);
+    });
     return;
   }
 
-  auto mask_array = [](const array& mask,
+  auto mask_array = [](const void* mask,
                        float* data,
                        int block_size,
                        int batch_idx,
                        int X,
                        int Y,
                        size_t X_data_str,
-                       size_t Y_data_str) {
+                       size_t Y_data_str,
+                       const Shape& mask_shape,
+                       const Strides& mask_strides,
+                       bool is_bool) {
+    auto ndim = mask_shape.size();
     auto mask_offset = elem_to_loc(
-        mask.shape(-1) * mask.shape(-2) * batch_idx,
-        mask.shape(),
-        mask.strides());
+        mask_shape[ndim - 1] * mask_shape[ndim - 2] * batch_idx,
+        mask_shape,
+        mask_strides);
 
-    auto X_mask_str = mask.strides()[mask.ndim() - 2];
-    auto Y_mask_str = mask.strides()[mask.ndim() - 1];
+    auto X_mask_str = mask_strides[ndim - 2];
+    auto Y_mask_str = mask_strides[ndim - 1];
 
-    if (mask.dtype() == bool_) {
+    if (is_bool) {
       return mask_matrix(
           data,
-          mask.data<bool>(),
+          static_cast<const bool*>(mask),
           block_size,
           X,
           Y,
@@ -140,7 +149,7 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     } else {
       return mask_matrix(
           data,
-          mask.data<float>(),
+          static_cast<const float*>(mask),
           block_size,
           X,
           Y,
@@ -152,61 +161,155 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     }
   };
 
-  for (int i = 0; i < (out.size() / (M * size_t(N))); ++i) {
-    // Adjust pointer
-    float* ai =
-        a.data<float>() + elem_to_loc(M * K * i, a.shape(), a.strides());
-    float* bi =
-        b.data<float>() + elem_to_loc(K * N * i, b.shape(), b.strides());
-    float* ci = out.data<float>() + M * N * i;
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  const void* a_mask_ptr;
+  const void* b_mask_ptr;
+  const void* out_mask_ptr;
+  Shape a_mask_shape;
+  Shape b_mask_shape;
+  Shape out_mask_shape;
+  Strides a_mask_strides;
+  Strides b_mask_strides;
+  Strides out_mask_strides;
+  bool a_mask_bool;
+  bool b_mask_bool;
+  bool out_mask_bool;
+  if (has_op_mask) {
+    auto& a_mask = inputs[inputs.size() - 2];
+    auto& b_mask = inputs[inputs.size() - 1];
+    a_mask_ptr = a_mask.data<void>();
+    b_mask_ptr = b_mask.data<void>();
+    a_mask_shape = a_mask.shape();
+    b_mask_shape = b_mask.shape();
+    a_mask_strides = a_mask.strides();
+    b_mask_strides = b_mask.strides();
+    a_mask_bool = (a_mask.dtype() == bool_);
+    b_mask_bool = (b_mask.dtype() == bool_);
+    encoder.set_input_array(a_mask);
+    encoder.set_input_array(b_mask);
+  }
+  if (has_out_mask) {
+    auto& out_mask = inputs[2];
+    out_mask_ptr = out_mask.data<void>();
+    out_mask_bool = (out_mask.dtype() == bool_);
+    encoder.set_input_array(out_mask);
+    out_mask_shape = out_mask.shape();
+    out_mask_strides = out_mask.strides();
+  }
+  encoder.set_output_array(out);
+  auto a_ptr = a.data<float>();
+  auto b_ptr = b.data<float>();
+  auto out_ptr = out.data<float>();
+  size_t num_matrices = out.size() / (M * size_t(N));
+  auto ldc = out.shape(-1);
 
-    // Zero out blocks in a and b if needed
-    if (has_op_mask) {
-      auto& a_mask = inputs[inputs.size() - 2];
-      mask_array(
-          a_mask,
-          ai,
-          block_size_,
-          i,
+  encoder.dispatch([a_ptr,
+                    b_ptr,
+                    out_ptr,
+                    a_mask_ptr,
+                    b_mask_ptr,
+                    out_mask_ptr,
+                    has_op_mask,
+                    has_out_mask,
+                    block_size = block_size_,
+                    num_matrices,
+                    M,
+                    N,
+                    K,
+                    a_transposed = a_transposed,
+                    b_transposed = b_transposed,
+                    lda = lda,
+                    ldb = ldb,
+                    ldc,
+                    a_shape = a.shape(),
+                    a_strides = a.strides(),
+                    b_shape = b.shape(),
+                    b_strides = b.strides(),
+                    a_mask_shape = std::move(a_mask_shape),
+                    b_mask_shape = std::move(b_mask_shape),
+                    out_mask_shape = std::move(out_mask_shape),
+                    a_mask_strides = std::move(a_mask_strides),
+                    b_mask_strides = std::move(b_mask_strides),
+                    out_mask_strides = std::move(out_mask_strides),
+                    mask_array,
+                    a_mask_bool,
+                    b_mask_bool,
+                    out_mask_bool]() {
+    for (int i = 0; i < num_matrices; ++i) {
+      // Adjust pointer
+      float* ai = a_ptr + elem_to_loc(M * K * i, a_shape, a_strides);
+      float* bi = b_ptr + elem_to_loc(K * N * i, b_shape, b_strides);
+      float* ci = out_ptr + M * N * i;
+
+      // Zero out blocks in a and b if needed
+      if (has_op_mask) {
+        mask_array(
+            a_mask_ptr,
+            ai,
+            block_size,
+            i,
+            M,
+            K,
+            a_transposed ? 1 : lda,
+            a_transposed ? lda : 1,
+            a_mask_shape,
+            a_mask_strides,
+            a_mask_bool);
+
+        mask_array(
+            b_mask_ptr,
+            bi,
+            block_size,
+            i,
+            K,
+            N,
+            b_transposed ? 1 : ldb,
+            b_transposed ? ldb : 1,
+            b_mask_shape,
+            b_mask_strides,
+            b_mask_bool);
+      }
+
+      // Do matmul
+      cblas_sgemm(
+          CblasRowMajor,
+          a_transposed ? CblasTrans : CblasNoTrans, // transA
+          b_transposed ? CblasTrans : CblasNoTrans, // transB
           M,
-          K,
-          a_transposed ? 1 : lda,
-          a_transposed ? lda : 1);
-
-      auto& b_mask = inputs[inputs.size() - 1];
-      mask_array(
-          b_mask,
-          bi,
-          block_size_,
-          i,
-          K,
           N,
-          b_transposed ? 1 : ldb,
-          b_transposed ? ldb : 1);
-    }
+          K,
+          1.0, // alpha
+          ai,
+          lda,
+          bi,
+          ldb,
+          0.0, // beta
+          ci,
+          ldc);
 
-    // Do matmul
-    cblas_sgemm(
-        CblasRowMajor,
-        a_transposed ? CblasTrans : CblasNoTrans, // transA
-        b_transposed ? CblasTrans : CblasNoTrans, // transB
-        M,
-        N,
-        K,
-        1.0, // alpha
-        ai,
-        lda,
-        bi,
-        ldb,
-        0.0, // beta
-        ci,
-        out.shape(-1) // ldc
-    );
-
-    // Zero out blocks in out
-    if (has_out_mask) {
-      mask_array(inputs[2], ci, block_size_, i, M, N, N, 1);
+      // Zero out blocks in out
+      if (has_out_mask) {
+        mask_array(
+            out_mask_ptr,
+            ci,
+            block_size,
+            i,
+            M,
+            N,
+            N,
+            1,
+            out_mask_shape,
+            out_mask_strides,
+            out_mask_bool);
+      }
     }
+  });
+  if (a_copied) {
+    encoder.add_temporary(a);
+  }
+  if (b_copied) {
+    encoder.add_temporary(b);
   }
 }
 
@@ -220,7 +323,8 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
 
-  auto check_transpose = [](const array& arr) {
+  std::vector<array> temps;
+  auto check_transpose = [s = stream(), &temps](const array& arr) {
     auto stx = arr.strides()[arr.ndim() - 2];
     auto sty = arr.strides()[arr.ndim() - 1];
     if (stx == arr.shape(-1) && sty == 1) {
@@ -228,10 +332,10 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     } else if (stx == 1 && sty == arr.shape(-2)) {
       return std::make_tuple(true, sty, arr);
     } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy(arr, temps.back(), CopyType::General, s);
       int64_t stx = arr.shape(-1);
-      return std::make_tuple(false, stx, arr_copy);
+      return std::make_tuple(false, stx, temps.back());
     }
   };
 
@@ -246,8 +350,12 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
+  auto& encoder = cpu::get_command_encoder(stream());
   if (K == 0) {
-    std::memset(static_cast<void*>(out.data<float>()), 0, out.nbytes());
+    encoder.set_output_array(out);
+    encoder.dispatch([out_ptr = out.data<float>(), size = out.size()]() {
+      std::fill_n(out_ptr, size, 0);
+    });
     return;
   }
 
@@ -272,29 +380,61 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   const uint32_t* lhs_indices_ptr = lhs_indices.data<uint32_t>();
   const uint32_t* rhs_indices_ptr = rhs_indices.data<uint32_t>();
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(lhs_indices);
+  encoder.set_input_array(rhs_indices);
+  encoder.set_output_array(out);
+  auto ldc = out.shape(-1);
 
-  for (int i = 0; i < batch_size_out; i++) {
-    // Get index
-    uint32_t indx_A = lhs_indices_ptr[elem_to_loc(i, lhs_indices)];
-    uint32_t indx_B = rhs_indices_ptr[elem_to_loc(i, rhs_indices)];
+  encoder.dispatch([a_ptr = a.data<float>(),
+                    b_ptr = b.data<float>(),
+                    out_ptr = out.data<float>(),
+                    M,
+                    N,
+                    K,
+                    lda = lda,
+                    ldb = ldb,
+                    a_transposed = a_transposed,
+                    b_transposed = b_transposed,
+                    ldc,
+                    lhs_indices_ptr,
+                    rhs_indices_ptr,
+                    lhs_indices_shape = lhs_indices.shape(),
+                    lhs_indices_strides = lhs_indices.strides(),
+                    rhs_indices_shape = rhs_indices.shape(),
+                    rhs_indices_strides = rhs_indices.strides(),
+                    batch_size_out,
+                    matrix_stride_out,
+                    batch_shape_A = std::move(batch_shape_A),
+                    batch_shape_B = std::move(batch_shape_B),
+                    batch_strides_A = std::move(batch_strides_A),
+                    batch_strides_B = std::move(batch_strides_B)]() {
+    for (int i = 0; i < batch_size_out; i++) {
+      // Get index
+      uint32_t indx_A = lhs_indices_ptr[elem_to_loc(
+          i, lhs_indices_shape, lhs_indices_strides)];
+      uint32_t indx_B = rhs_indices_ptr[elem_to_loc(
+          i, rhs_indices_shape, rhs_indices_strides)];
 
-    cblas_sgemm(
-        CblasRowMajor,
-        a_transposed ? CblasTrans : CblasNoTrans, // transA
-        b_transposed ? CblasTrans : CblasNoTrans, // transB
-        M,
-        N,
-        K,
-        1.0f, // alpha
-        a.data<float>() + elem_to_loc(indx_A, batch_shape_A, batch_strides_A),
-        lda,
-        b.data<float>() + elem_to_loc(indx_B, batch_shape_B, batch_strides_B),
-        ldb,
-        0.0f, // beta
-        out.data<float>() + matrix_stride_out * i,
-        out.shape(-1) // ldc
-    );
-  }
+      cblas_sgemm(
+          CblasRowMajor,
+          a_transposed ? CblasTrans : CblasNoTrans, // transA
+          b_transposed ? CblasTrans : CblasNoTrans, // transB
+          M,
+          N,
+          K,
+          1.0f, // alpha
+          a_ptr + elem_to_loc(indx_A, batch_shape_A, batch_strides_A),
+          lda,
+          b_ptr + elem_to_loc(indx_B, batch_shape_B, batch_strides_B),
+          ldb,
+          0.0f, // beta
+          out_ptr + matrix_stride_out * i,
+          ldc);
+    }
+  });
+  encoder.add_temporaries(std::move(temps));
 }
 
 } // namespace mlx::core

mlx/backend/cpu/matmul.cpp

@@ -3,18 +3,76 @@
 #include <cstring>
 #include "mlx/array.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/gemm.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
+template <typename T>
+void matmul_dispatch(
+    const array& a,
+    const array& b,
+    array& out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    float alpha,
+    float beta,
+    Stream stream) {
+  const T* a_ptr = a.data<T>();
+  const T* b_ptr = b.data<T>();
+  T* out_ptr = out.data<T>();
+  size_t ldc = out.shape(-1);
+  size_t batch_size = a.size() / (a.shape(-2) * a.shape(-1));
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a_ptr,
+                    b_ptr,
+                    out_ptr,
+                    a_transposed,
+                    b_transposed,
+                    lda,
+                    ldb,
+                    ldc,
+                    alpha,
+                    beta,
+                    batch_size,
+                    a_shape = a.shape(),
+                    a_strides = a.strides(),
+                    b_shape = b.shape(),
+                    b_strides = b.strides()]() {
+    matmul<T>(
+        a_ptr,
+        b_ptr,
+        out_ptr,
+        a_transposed,
+        b_transposed,
+        lda,
+        ldb,
+        ldc,
+        alpha,
+        beta,
+        batch_size,
+        a_shape,
+        a_strides,
+        b_shape,
+        b_strides);
+  });
+}
+
 void matmul_general(
     const array& a_pre,
     const array& b_pre,
     array& out,
+    Stream stream,
     float alpha = 1.0f,
     float beta = 0.0f) {
-  auto check_transpose = [](const array& arr) {
+  std::vector<array> temps;
+  auto check_transpose = [stream, &temps](const array& arr) {
     auto stx = arr.strides()[arr.ndim() - 2];
     auto sty = arr.strides()[arr.ndim() - 1];
     if (stx == arr.shape(-1) && sty == 1) {
@@ -22,10 +80,10 @@ void matmul_general(
     } else if (stx == 1 && sty == arr.shape(-2)) {
       return std::make_tuple(true, sty, arr);
     } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy(arr, temps.back(), CopyType::General, stream);
       stx = arr.shape(-1);
-      return std::make_tuple(false, stx, arr_copy);
+      return std::make_tuple(false, stx, temps.back());
     }
   };
 
@@ -39,28 +97,34 @@ void matmul_general(
   }
 
   if (out.dtype() == float32) {
-    matmul<float>(a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta);
+    matmul_dispatch<float>(
+        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else if (out.dtype() == float16) {
-    matmul<float16_t>(
-        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta);
+    matmul_dispatch<float16_t>(
+        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else if (out.dtype() == bfloat16) {
-    matmul<bfloat16_t>(
-        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta);
+    matmul_dispatch<bfloat16_t>(
+        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else if (out.dtype() == float64) {
-    matmul<double>(
-        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta);
+    matmul_dispatch<double>(
+        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else {
     throw std::runtime_error("[Matmul::eval_cpu] Invalid type.");
   }
+  cpu::get_command_encoder(stream).add_temporaries(std::move(temps));
 }
 
 void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
   if (inputs[0].shape(-1) == 0) {
-    std::memset(out.data<void>(), 0, out.nbytes());
+    auto& encoder = cpu::get_command_encoder(stream());
+    encoder.set_output_array(out);
+    encoder.dispatch([out_ptr = out.data<void>(), nbytes = out.nbytes()]() {
+      std::memset(out_ptr, 0, nbytes);
+    });
     return;
   }
-  return matmul_general(inputs[0], inputs[1], out);
+  matmul_general(inputs[0], inputs[1], out, stream());
 }
 
 void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -74,9 +138,9 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   CopyType ctype = c.data_size() == 1
       ? CopyType::Scalar
       : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-  copy(c, out, ctype);
+  copy(c, out, ctype, stream());
 
-  return matmul_general(inputs[0], inputs[1], out, alpha_, beta_);
+  matmul_general(inputs[0], inputs[1], out, stream(), alpha_, beta_);
 }
 
 } // namespace mlx::core

mlx/backend/cpu/primitives.cpp

@@ -7,11 +7,11 @@
 #include <sstream>
 
 #include "mlx/allocator.h"
-#include "mlx/backend/common/load.h"
 #include "mlx/backend/common/slicing.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/arange.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/threefry.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
@@ -22,39 +22,58 @@ void reshape(const array& in, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
     out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    copy_inplace(in, out, CopyType::General);
+    copy_inplace(in, out, CopyType::General, out.primitive().stream());
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
 }
 
-int64_t compute_dynamic_offset(
+static std::pair<array, bool> compute_dynamic_offset(
     const array& indices,
     const Strides& strides,
-    const std::vector<int>& axes) {
-  auto compute_offset = [&strides, &axes](const auto* indices) {
-    int64_t offset = 0;
-    for (int i = 0; i < axes.size(); ++i) {
-      offset += indices[i] * strides[axes[i]];
-    }
-    return offset;
-  };
+    const std::vector<int>& axes,
+    Stream stream) {
+  array offset({1}, int64, nullptr, {});
+  bool donate = indices.is_donatable() &&
+      (indices.data_size() * indices.itemsize()) >= offset.itemsize();
+  if (donate) {
+    offset.copy_shared_buffer(indices);
+  } else {
+    offset.set_data(allocator::malloc_or_wait(offset.itemsize()));
+  }
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(indices);
+  encoder.set_output_array(offset);
+  auto compute_offset =
+      [strides, axes, offset = offset.data<int64_t>()](const auto* indices) {
+        int64_t offset_ = 0;
+        for (int i = 0; i < axes.size(); ++i) {
+          offset_ += indices[i] * strides[axes[i]];
+        }
+        offset[0] = offset_;
+      };
   switch (indices.dtype()) {
     case int8:
     case uint8:
-      return compute_offset(indices.data<uint8_t>());
+      encoder.dispatch(compute_offset, indices.data<uint8_t>());
+      break;
     case int16:
     case uint16:
-      return compute_offset(indices.data<uint16_t>());
+      encoder.dispatch(compute_offset, indices.data<uint16_t>());
+      break;
     case int32:
     case uint32:
-      return compute_offset(indices.data<uint32_t>());
+      encoder.dispatch(compute_offset, indices.data<uint32_t>());
+      break;
     case int64:
     case uint64:
-      return compute_offset(indices.data<uint64_t>());
+      encoder.dispatch(compute_offset, indices.data<uint64_t>());
+      break;
     default:
       throw std::runtime_error("Invalid indices type.");
   }
+  return {offset, donate};
 }
 
 void AsStrided::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -104,14 +123,59 @@ void Transpose::eval_cpu(const std::vector<array>& inputs, array& out) {
 }
 
 void Arange::eval_cpu(const std::vector<array>& inputs, array& out) {
-  arange(inputs, out, start_, step_);
+  assert(inputs.size() == 0);
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  switch (out.dtype()) {
+    case bool_:
+      throw std::runtime_error("Bool type unsupported for arange.");
+      break;
+    case uint8:
+      arange<uint8_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case uint16:
+      arange<uint16_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case uint32:
+      arange<uint32_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case uint64:
+      arange<uint64_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case int8:
+      arange<int8_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case int16:
+      arange<int16_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case int32:
+      arange<int32_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case int64:
+      arange<int64_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case float16:
+      arange<float16_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case float32:
+      arange<float>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case float64:
+      arange<double>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case bfloat16:
+      arange<bfloat16_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+    case complex64:
+      arange<complex64_t>(start_, start_ + step_, out, out.size(), stream());
+      break;
+  }
 }
 
 void AsType::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype);
+  copy(in, out, ctype, stream());
 }
 
 void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -134,7 +198,7 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
     size_t data_offset = strides[axis_] * sizes[i];
     out_slice.copy_shared_buffer(
         out, strides, flags, out_slice.size(), data_offset);
-    copy_inplace(inputs[i], out_slice, CopyType::GeneralGeneral);
+    copy_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
   }
 }
 
@@ -145,7 +209,7 @@ void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
       (allow_col_major_ && in.flags().col_contiguous)) {
     out.copy_shared_buffer(in);
   } else {
-    copy(in, out, CopyType::General);
+    copy(in, out, CopyType::General, stream());
   }
 }
 
@@ -169,14 +233,7 @@ void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   } else {
     ctype = CopyType::General;
   }
-  copy(in, out, ctype);
-}
-
-void Load::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 0);
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  load(out, offset_, reader_, swap_endianness_);
+  copy(in, out, ctype, stream());
 }
 
 void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -192,7 +249,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
 
   // Fill output with val
-  copy(val, out, CopyType::Scalar);
+  copy(val, out, CopyType::Scalar, stream());
 
   // Find offset for start of input values
   size_t data_offset = 0;
@@ -207,7 +264,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
       out, out.strides(), out.flags(), out_slice.size(), data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, out_slice, CopyType::GeneralGeneral);
+  copy_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
 }
 
 void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -223,39 +280,49 @@ void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   auto kptr = inputs[0].data<uint32_t>();
   auto cptr = out.data<char>();
-  size_t out_skip = (bytes_per_key + 4 - 1) / 4;
-  auto half_size = out_skip / 2;
-  bool even = out_skip % 2 == 0;
-  for (int i = 0; i < num_keys; ++i, cptr += bytes_per_key) {
-    auto ptr = reinterpret_cast<uint32_t*>(cptr);
-    // Get ith key
-    auto kidx = 2 * i;
-    auto k1_elem = elem_to_loc(kidx, keys.shape(), keys.strides());
-    auto k2_elem = elem_to_loc(kidx + 1, keys.shape(), keys.strides());
-    auto key = std::make_pair(kptr[k1_elem], kptr[k2_elem]);
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(inputs[0]);
+  encoder.set_output_array(out);
+  encoder.dispatch([kptr,
+                    cptr,
+                    bytes_per_key,
+                    num_keys,
+                    kshape = keys.shape(),
+                    kstrides = keys.strides()]() mutable {
+    size_t out_skip = (bytes_per_key + 4 - 1) / 4;
+    auto half_size = out_skip / 2;
+    bool even = out_skip % 2 == 0;
+    for (int i = 0; i < num_keys; ++i, cptr += bytes_per_key) {
+      auto ptr = reinterpret_cast<uint32_t*>(cptr);
+      // Get ith key
+      auto kidx = 2 * i;
+      auto k1_elem = elem_to_loc(kidx, kshape, kstrides);
+      auto k2_elem = elem_to_loc(kidx + 1, kshape, kstrides);
+      auto key = std::make_pair(kptr[k1_elem], kptr[k2_elem]);
 
-    std::pair<uintptr_t, uintptr_t> count{0, half_size + !even};
-    for (; count.first + 1 < half_size; count.first++, count.second++) {
-      std::tie(ptr[count.first], ptr[count.second]) =
-          random::threefry2x32_hash(key, count);
-    }
-    if (count.first < half_size) {
-      auto rb = random::threefry2x32_hash(key, count);
-      ptr[count.first++] = rb.first;
-      if (bytes_per_key % 4 > 0) {
-        std::copy(
-            reinterpret_cast<char*>(&rb.second),
-            reinterpret_cast<char*>(&rb.second) + bytes_per_key % 4,
-            cptr + 4 * count.second);
-      } else {
-        ptr[count.second] = rb.second;
+      std::pair<uintptr_t, uintptr_t> count{0, half_size + !even};
+      for (; count.first + 1 < half_size; count.first++, count.second++) {
+        std::tie(ptr[count.first], ptr[count.second]) =
+            random::threefry2x32_hash(key, count);
+      }
+      if (count.first < half_size) {
+        auto rb = random::threefry2x32_hash(key, count);
+        ptr[count.first++] = rb.first;
+        if (bytes_per_key % 4 > 0) {
+          std::copy(
+              reinterpret_cast<char*>(&rb.second),
+              reinterpret_cast<char*>(&rb.second) + bytes_per_key % 4,
+              cptr + 4 * count.second);
+        } else {
+          ptr[count.second] = rb.second;
+        }
+      }
+      if (!even) {
+        count.second = 0;
+        ptr[half_size] = random::threefry2x32_hash(key, count).first;
       }
     }
-    if (!even) {
-      count.second = 0;
-      ptr[half_size] = random::threefry2x32_hash(key, count).first;
-    }
-  }
+  });
 }
 
 void Reshape::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -269,16 +336,23 @@ void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
   }
   auto& in = inputs[0];
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  auto i_offset = compute_dynamic_offset(inputs[1], in.strides(), axes_);
+  auto [in_offset, donated] =
+      compute_dynamic_offset(inputs[1], in.strides(), axes_, stream());
   copy_inplace(
       /* const array& src = */ in,
       /* array& dst = */ out,
       /* const Shape& data_shape = */ out.shape(),
       /* const Strides& i_strides = */ in.strides(),
       /* const Strides& o_strides = */ out.strides(),
-      /* int64_t i_offset = */ i_offset,
+      /* int64_t i_offset = */ 0,
       /* int64_t o_offset = */ 0,
-      /* CopyType ctype = */ CopyType::GeneralGeneral);
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      stream(),
+      /* const std::optional<array>& dynamic_i_offset = */ in_offset,
+      /* const std::optional<array>& dynamic_o_offset = */ std::nullopt);
+  if (!donated) {
+    cpu::get_command_encoder(stream()).add_temporary(std::move(in_offset));
+  }
 }
 
 void DynamicSliceUpdate::eval_cpu(
@@ -296,9 +370,10 @@ void DynamicSliceUpdate::eval_cpu(
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype);
+  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
-  auto o_offset = compute_dynamic_offset(inputs[2], out.strides(), axes_);
+  auto [out_offset, donated] =
+      compute_dynamic_offset(inputs[2], out.strides(), axes_, stream());
   copy_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
@@ -306,8 +381,14 @@ void DynamicSliceUpdate::eval_cpu(
       /* 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 = */ o_offset,
-      /* CopyType ctype = */ CopyType::GeneralGeneral);
+      /* int64_t o_offset = */ 0,
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      stream(),
+      /* const std::optional<array>& dynamic_i_offset = */ std::nullopt,
+      /* const std::optional<array>& dynamic_o_offset = */ out_offset);
+  if (!donated) {
+    cpu::get_command_encoder(stream()).add_temporary(std::move(out_offset));
+  }
 }
 
 void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -329,7 +410,7 @@ void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype);
+  copy(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] =
@@ -344,7 +425,8 @@ void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
       /* const std::vector<stride_t>& o_strides = */ out_strides,
       /* int64_t i_offset = */ 0,
       /* int64_t o_offset = */ data_offset,
-      /* CopyType ctype = */ CopyType::GeneralGeneral);
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      stream());
 }
 
 void View::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -372,9 +454,9 @@ void View::eval_cpu(const std::vector<array>& inputs, array& out) {
     if (in.dtype() == bool_) {
       auto in_tmp = array(in.shape(), uint8, nullptr, {});
       in_tmp.copy_shared_buffer(in);
-      copy_inplace(in_tmp, tmp, CopyType::General);
+      copy_inplace(in_tmp, tmp, CopyType::General, stream());
     } else {
-      copy_inplace(in, tmp, CopyType::General);
+      copy_inplace(in, tmp, CopyType::General, stream());
     }
 
     auto flags = out.flags();
@@ -382,7 +464,7 @@ void View::eval_cpu(const std::vector<array>& inputs, array& out) {
     flags.row_contiguous = true;
     auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
     flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
-    out.move_shared_buffer(tmp, out.strides(), flags, out.size());
+    out.copy_shared_buffer(tmp, out.strides(), flags, out.size());
   }
 }
 

mlx/backend/cpu/qrf.cpp

@@ -2,20 +2,18 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 template <typename T>
-void qrf_impl(const array& a, array& q, array& r) {
+void qrf_impl(const array& a, array& q, array& r, Stream stream) {
   const int M = a.shape(-2);
   const int N = a.shape(-1);
   const int lda = M;
   size_t num_matrices = a.size() / (M * N);
-  int num_reflectors = std::min(M, N);
-  auto tau =
-      allocator::malloc_or_wait(sizeof(T) * num_matrices * num_reflectors);
 
   // Copy A to inplace input and make it col-contiguous
   array in(a.shape(), a.dtype(), nullptr, {});
@@ -29,93 +27,107 @@ void qrf_impl(const array& a, array& q, array& r) {
   strides[in.ndim() - 1] = M;
   in.set_data(
       allocator::malloc_or_wait(in.nbytes()), in.nbytes(), strides, flags);
-  copy_inplace(a, in, CopyType::GeneralGeneral);
-
-  T optimal_work;
-  int lwork = -1;
-  int info;
-
-  // Compute workspace size
-  geqrf<T>(&M, &N, nullptr, &lda, nullptr, &optimal_work, &lwork, &info);
-
-  // Update workspace size
-  lwork = optimal_work;
-  auto work = allocator::malloc_or_wait(sizeof(T) * lwork);
-
-  // Loop over matrices
-  for (int i = 0; i < num_matrices; ++i) {
-    // Solve
-    geqrf<T>(
-        &M,
-        &N,
-        in.data<T>() + M * N * i,
-        &lda,
-        static_cast<T*>(tau.raw_ptr()) + num_reflectors * i,
-        static_cast<T*>(work.raw_ptr()),
-        &lwork,
-        &info);
-  }
-  allocator::free(work);
-
+  copy_inplace(a, in, CopyType::GeneralGeneral, stream);
+  auto& encoder = cpu::get_command_encoder(stream);
+  q.set_data(allocator::malloc_or_wait(q.nbytes()));
   r.set_data(allocator::malloc_or_wait(r.nbytes()));
 
-  for (int i = 0; i < num_matrices; ++i) {
-    /// num_reflectors x N
-    for (int j = 0; j < r.shape(-2); ++j) {
-      for (int k = 0; k < j; ++k) {
-        r.data<T>()[i * N * num_reflectors + j * N + k] = 0;
-      }
-      for (int k = j; k < r.shape(-1); ++k) {
-        r.data<T>()[i * N * num_reflectors + j * N + k] =
-            in.data<T>()[i * N * M + j + k * M];
+  auto in_ptr = in.data<T>();
+  auto r_ptr = r.data<T>();
+  auto q_ptr = q.data<T>();
+
+  encoder.set_input_array(in);
+  encoder.set_output_array(q);
+  encoder.set_output_array(r);
+  encoder.dispatch([in_ptr, q_ptr, r_ptr, M, N, lda, num_matrices]() {
+    int num_reflectors = std::min(M, N);
+    auto tau =
+        allocator::malloc_or_wait(sizeof(T) * num_matrices * num_reflectors);
+
+    T optimal_work;
+    int lwork = -1;
+    int info;
+
+    // Compute workspace size
+    geqrf<T>(&M, &N, nullptr, &lda, nullptr, &optimal_work, &lwork, &info);
+
+    // Update workspace size
+    lwork = optimal_work;
+    auto work = allocator::malloc_or_wait(sizeof(T) * lwork);
+
+    // Loop over matrices
+    for (int i = 0; i < num_matrices; ++i) {
+      // Solve
+      geqrf<T>(
+          &M,
+          &N,
+          in_ptr + M * N * i,
+          &lda,
+          static_cast<T*>(tau.raw_ptr()) + num_reflectors * i,
+          static_cast<T*>(work.raw_ptr()),
+          &lwork,
+          &info);
+    }
+    allocator::free(work);
+
+    for (int i = 0; i < num_matrices; ++i) {
+      /// num_reflectors x N
+      for (int j = 0; j < num_reflectors; ++j) {
+        for (int k = 0; k < j; ++k) {
+          r_ptr[i * N * num_reflectors + j * N + k] = 0;
+        }
+        for (int k = j; k < N; ++k) {
+          r_ptr[i * N * num_reflectors + j * N + k] =
+              in_ptr[i * N * M + j + k * M];
+        }
       }
     }
-  }
 
-  // Get work size
-  lwork = -1;
-  orgqr<T>(
-      &M,
-      &num_reflectors,
-      &num_reflectors,
-      nullptr,
-      &lda,
-      nullptr,
-      &optimal_work,
-      &lwork,
-      &info);
-  lwork = optimal_work;
-  work = allocator::malloc_or_wait(sizeof(T) * lwork);
-
-  // Loop over matrices
-  for (int i = 0; i < num_matrices; ++i) {
-    // Compute Q
+    // Get work size
+    lwork = -1;
     orgqr<T>(
         &M,
         &num_reflectors,
         &num_reflectors,
-        in.data<T>() + M * N * i,
+        nullptr,
         &lda,
-        static_cast<T*>(tau.raw_ptr()) + num_reflectors * i,
-        static_cast<T*>(work.raw_ptr()),
+        nullptr,
+        &optimal_work,
         &lwork,
         &info);
-  }
+    lwork = optimal_work;
+    work = allocator::malloc_or_wait(sizeof(T) * lwork);
 
-  q.set_data(allocator::malloc_or_wait(q.nbytes()));
-  for (int i = 0; i < num_matrices; ++i) {
-    // M x num_reflectors
-    for (int j = 0; j < q.shape(-2); ++j) {
-      for (int k = 0; k < q.shape(-1); ++k) {
-        q.data<T>()[i * M * num_reflectors + j * num_reflectors + k] =
-            in.data<T>()[i * N * M + j + k * M];
+    // Loop over matrices
+    for (int i = 0; i < num_matrices; ++i) {
+      // Compute Q
+      orgqr<T>(
+          &M,
+          &num_reflectors,
+          &num_reflectors,
+          in_ptr + M * N * i,
+          &lda,
+          static_cast<T*>(tau.raw_ptr()) + num_reflectors * i,
+          static_cast<T*>(work.raw_ptr()),
+          &lwork,
+          &info);
+    }
+
+    for (int i = 0; i < num_matrices; ++i) {
+      // M x num_reflectors
+      for (int j = 0; j < M; ++j) {
+        for (int k = 0; k < num_reflectors; ++k) {
+          q_ptr[i * M * num_reflectors + j * num_reflectors + k] =
+              in_ptr[i * N * M + j + k * M];
+        }
       }
     }
-  }
 
-  // Cleanup
-  allocator::free(work);
-  allocator::free(tau);
+    // Cleanup
+    allocator::free(work);
+    allocator::free(tau);
+  });
+  encoder.add_temporary(in);
 }
 
 void QRF::eval_cpu(
@@ -123,10 +135,10 @@ void QRF::eval_cpu(
     std::vector<array>& outputs) {
   switch (inputs[0].dtype()) {
     case float32:
-      qrf_impl<float>(inputs[0], outputs[0], outputs[1]);
+      qrf_impl<float>(inputs[0], outputs[0], outputs[1], stream());
       break;
     case float64:
-      qrf_impl<double>(inputs[0], outputs[0], outputs[1]);
+      qrf_impl<double>(inputs[0], outputs[0], outputs[1], stream());
       break;
     default:
       throw std::runtime_error(

mlx/backend/cpu/quantized.cpp

@@ -3,6 +3,7 @@
 #include <cassert>
 
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
 #include "mlx/fast_primitives.h"
 #include "mlx/primitives.h"
@@ -316,7 +317,8 @@ void _qmm_dispatch_typed(
   }
 }
 
-void _qmm_dispatch(
+template <typename T>
+void _qmm_dispatch_typed(
     array& out,
     const array& x,
     const array& w,
@@ -328,63 +330,61 @@ void _qmm_dispatch(
   int K = x.shape(-1);
   int M = x.ndim() > 1 ? x.shape(-2) : 1;
   int N = out.shape(-1);
-
   int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
   int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
-
   int batch_size = x.size() / (K * M);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<T>();
+  auto biases_ptr = biases.data<T>();
   for (int i = 0; i < batch_size; i++) {
-    switch (x.dtype()) {
-      case float32:
-        _qmm_dispatch_typed<float>(
-            out.data<float>() + i * M * N,
-            x.data<float>() + elem_to_loc(i * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(i * w_els, w),
-            scales.data<float>() + elem_to_loc(i * g_els, scales),
-            biases.data<float>() + elem_to_loc(i * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case float16:
-        _qmm_dispatch_typed<float16_t>(
-            out.data<float16_t>() + i * M * N,
-            x.data<float16_t>() + elem_to_loc(i * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(i * w_els, w),
-            scales.data<float16_t>() + elem_to_loc(i * g_els, scales),
-            biases.data<float16_t>() + elem_to_loc(i * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case bfloat16:
-        _qmm_dispatch_typed<bfloat16_t>(
-            out.data<bfloat16_t>() + i * M * N,
-            x.data<bfloat16_t>() + elem_to_loc(i * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(i * w_els, w),
-            scales.data<bfloat16_t>() + elem_to_loc(i * g_els, scales),
-            biases.data<bfloat16_t>() + elem_to_loc(i * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      default:
-        throw std::invalid_argument(
-            "[quantized_matmul] only floating types are supported");
-    }
+    _qmm_dispatch_typed<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
+        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
+        biases_ptr + elem_to_loc(i * g_els, biases.shape(), biases.strides()),
+        M,
+        N,
+        K,
+        bits,
+        group_size,
+        transposed_w);
   }
 }
 
-void _bs_qmm_dispatch(
+void _qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& biases,
+    int bits,
+    int group_size,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case float32:
+      _qmm_dispatch_typed<float>(
+          out, x, w, scales, biases, bits, group_size, transposed_w);
+      break;
+    case float16:
+      _qmm_dispatch_typed<float16_t>(
+          out, x, w, scales, biases, bits, group_size, transposed_w);
+      break;
+    case bfloat16:
+      _qmm_dispatch_typed<bfloat16_t>(
+          out, x, w, scales, biases, bits, group_size, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
+template <typename T>
+void _bs_qmm_dispatch_typed(
     array& out,
     const array& x,
     const array& w,
@@ -402,60 +402,90 @@ void _bs_qmm_dispatch(
   int w_els = w.shape(-1) * w.shape(-2);
   int g_els = scales.shape(-1) * scales.shape(-2);
 
-  const uint32_t* lhs_indices_data = lhs_indices.data<uint32_t>();
-  const uint32_t* rhs_indices_data = rhs_indices.data<uint32_t>();
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<T>();
+  auto biases_ptr = biases.data<T>();
+  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
+  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
 
   for (int i = 0; i < lhs_indices.size(); i++) {
-    int x_idx = lhs_indices_data[elem_to_loc(i, lhs_indices)];
-    int w_idx = rhs_indices_data[elem_to_loc(i, rhs_indices)];
+    int x_idx = lhs_indices_ptr[elem_to_loc(
+        i, lhs_indices.shape(), lhs_indices.strides())];
+    int w_idx = rhs_indices_ptr[elem_to_loc(
+        i, rhs_indices.shape(), rhs_indices.strides())];
+    _qmm_dispatch_typed<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
+        scales_ptr +
+            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
+        biases_ptr +
+            elem_to_loc(w_idx * g_els, biases.shape(), biases.strides()),
+        M,
+        N,
+        K,
+        bits,
+        group_size,
+        transposed_w);
+  }
+}
 
-    switch (x.dtype()) {
-      case float32:
-        _qmm_dispatch_typed<float>(
-            out.data<float>() + i * M * N,
-            x.data<float>() + elem_to_loc(x_idx * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
-            scales.data<float>() + elem_to_loc(w_idx * g_els, scales),
-            biases.data<float>() + elem_to_loc(w_idx * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case float16:
-        _qmm_dispatch_typed<float16_t>(
-            out.data<float16_t>() + i * M * N,
-            x.data<float16_t>() + elem_to_loc(x_idx * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
-            scales.data<float16_t>() + elem_to_loc(w_idx * g_els, scales),
-            biases.data<float16_t>() + elem_to_loc(w_idx * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      case bfloat16:
-        _qmm_dispatch_typed<bfloat16_t>(
-            out.data<bfloat16_t>() + i * M * N,
-            x.data<bfloat16_t>() + elem_to_loc(x_idx * M * K, x),
-            w.data<uint32_t>() + elem_to_loc(w_idx * w_els, w),
-            scales.data<bfloat16_t>() + elem_to_loc(w_idx * g_els, scales),
-            biases.data<bfloat16_t>() + elem_to_loc(w_idx * g_els, biases),
-            M,
-            N,
-            K,
-            bits,
-            group_size,
-            transposed_w);
-        break;
-      default:
-        throw std::invalid_argument(
-            "[quantized_matmul] only floating types are supported");
-    }
+void _bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& biases,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    int bits,
+    int group_size,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case float32:
+      _bs_qmm_dispatch_typed<float>(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          bits,
+          group_size,
+          transposed_w);
+      break;
+    case float16:
+      _bs_qmm_dispatch_typed<float16_t>(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          bits,
+          group_size,
+          transposed_w);
+      break;
+    case bfloat16:
+      _bs_qmm_dispatch_typed<bfloat16_t>(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          bits,
+          group_size,
+          transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
   }
 }
 
@@ -469,13 +499,14 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& scales_pre = inputs[2];
   auto& biases_pre = inputs[3];
 
-  auto ensure_row_contiguous = [](const array& arr) {
+  std::vector<array> temps;
+  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      return arr_copy;
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy(arr, temps.back(), CopyType::General, s);
+      return temps.back();
     }
   };
 
@@ -485,7 +516,24 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.add_temporaries(std::move(temps));
+  encoder.set_input_array(x);
+  encoder.set_input_array(w);
+  encoder.set_input_array(scales);
+  encoder.set_input_array(biases);
+  encoder.set_output_array(out);
+  encoder.dispatch([out = array::unsafe_weak_copy(out),
+                    x = array::unsafe_weak_copy(x),
+                    w = array::unsafe_weak_copy(w),
+                    scales = array::unsafe_weak_copy(scales),
+                    biases = array::unsafe_weak_copy(biases),
+                    group_size_ = group_size_,
+                    bits_ = bits_,
+                    transpose_ = transpose_]() mutable {
+    _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
+  });
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -498,15 +546,17 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& lhs_indices = inputs[4];
   auto& rhs_indices = inputs[5];
 
-  auto ensure_row_contiguous_last_dims = [](const array& arr) {
+  std::vector<array> temps;
+  auto ensure_row_contiguous_last_dims = [s = stream(),
+                                          &temps](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      return arr_copy;
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy(arr, temps.back(), CopyType::General, s);
+      return temps.back();
     }
   };
 
@@ -516,70 +566,85 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  _bs_qmm_dispatch(
-      out,
-      x,
-      w,
-      scales,
-      biases,
-      lhs_indices,
-      rhs_indices,
-      group_size_,
-      bits_,
-      transpose_);
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.add_temporaries(std::move(temps));
+  encoder.set_input_array(x);
+  encoder.set_input_array(w);
+  encoder.set_input_array(scales);
+  encoder.set_input_array(biases);
+  encoder.set_input_array(lhs_indices);
+  encoder.set_input_array(rhs_indices);
+  encoder.set_output_array(out);
+  encoder.dispatch([out = array::unsafe_weak_copy(out),
+                    x = array::unsafe_weak_copy(x),
+                    w = array::unsafe_weak_copy(w),
+                    scales = array::unsafe_weak_copy(scales),
+                    biases = array::unsafe_weak_copy(biases),
+                    lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                    rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                    group_size_ = group_size_,
+                    bits_ = bits_,
+                    transpose_ = transpose_]() mutable {
+    _bs_qmm_dispatch(
+        out,
+        x,
+        w,
+        scales,
+        biases,
+        lhs_indices,
+        rhs_indices,
+        group_size_,
+        bits_,
+        transpose_);
+  });
 }
 
 template <typename T, typename U>
 void quantize(
-    const array& w_,
-    array& out_,
-    array& scales_,
-    array& biases_,
+    const T* w,
+    U* out,
+    T* scales,
+    T* biases,
     int bits,
-    int group_size) {
-  const T* w = w_.data<T>();
+    int group_size,
+    size_t w_size) {
+  float n_bins = (1 << bits) - 1;
+  float eps = 1e-7;
 
-  auto out = out_.data<U>();
-  T* scales = scales_.data<T>();
-  T* biases = biases_.data<T>();
-
-  T n_bins = (1 << bits) - 1;
-  T eps = 1e-7;
   bool power_of_2_bits = is_power_of_2(bits);
   int el_per_int = bits == 3 ? 8 : bits == 6 ? 4 : 32 / bits;
   // For 3/6 bits we read 3 uint8s at a time instead of 1 uint32
   int bytes_per_pack = power_of_2_bits ? 1 : 3;
   int int_per_group = group_size * bytes_per_pack / el_per_int;
-  size_t n_groups = w_.size() / group_size;
+  size_t n_groups = w_size / group_size;
 
   for (size_t i = 0; i < n_groups; ++i) {
     size_t w_idx = i * group_size;
-    T w_min = std::numeric_limits<float>::infinity();
-    T w_max = -w_min;
+    float w_min = std::numeric_limits<float>::infinity();
+    float w_max = -w_min;
     for (int j = 0; j < group_size; ++j) {
-      w_max = std::max(w_max, w[w_idx + j]);
-      w_min = std::min(w_min, w[w_idx + j]);
+      w_max = std::max(w_max, (float)w[w_idx + j]);
+      w_min = std::min(w_min, (float)w[w_idx + j]);
     }
     bool mask = std::abs(w_min) > std::abs(w_max);
-    T scale = std::max(T((w_max - w_min) / n_bins), eps);
+    float scale = std::max((w_max - w_min) / n_bins, eps);
     scale = mask ? scale : -scale;
 
-    auto edge = mask ? w_min : w_max;
-    auto q0 = std::rint(edge / scale);
-    if (q0 == 0) {
-      scales[i] = scale;
-      biases[i] = 0;
-    } else {
-      scales[i] = edge / q0;
-      biases[i] = edge;
+    float edge = mask ? w_min : w_max;
+    float q0 = std::rint(edge / scale);
+    float bias = 0;
+    if (q0 != 0) {
+      scale = edge / q0;
+      bias = edge;
     }
     size_t out_idx = i * int_per_group;
     for (int j = 0; j < int_per_group / bytes_per_pack; ++j) {
       uint32_t out_el = 0;
       for (int k = 0; k < el_per_int; ++k) {
-        T w_el = w[w_idx + j * el_per_int + k];
-        w_el = std::rint((w_el - biases[i]) / scales[i]);
-        w_el = std::min(std::max(w_el, T(0)), n_bins);
+        float w_el = w[w_idx + j * el_per_int + k];
+        w_el = std::rint((w_el - bias) / scale);
+        w_el = std::min(std::max(w_el, 0.0f), n_bins);
         out_el |= static_cast<uint32_t>(w_el) << (k * bits);
       }
       if (power_of_2_bits) {
@@ -590,23 +655,41 @@ void quantize(
         out[out_idx + bytes_per_pack * j + 2] = (out_el & 0xff0000) >> 16;
       }
     }
+    scales[i] = static_cast<T>(scale);
+    biases[i] = static_cast<T>(bias);
   }
 }
 
+template <typename T, typename U>
+void dispatch_quantize(
+    const array& w,
+    array& out,
+    array& scales,
+    array& biases,
+    int bits,
+    int group_size) {
+  auto w_ptr = w.data<T>();
+  auto out_ptr = out.data<U>();
+  auto scales_ptr = scales.data<T>();
+  auto biases_ptr = biases.data<T>();
+  quantize<T, U>(
+      w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
+}
+
 void fast::AffineQuantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
-  auto ensure_row_contiguous = [](const array& arr) {
+  auto ensure_row_contiguous = [s = stream()](const array& arr) {
     if (arr.flags().row_contiguous) {
-      return arr;
+      return std::make_pair(arr, false);
     } else {
       array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General);
-      return arr_copy;
+      copy(arr, arr_copy, CopyType::General, s);
+      return std::make_pair(arr_copy, true);
     }
   };
-  auto w = ensure_row_contiguous(inputs[0]);
 
+  auto [w, copied] = ensure_row_contiguous(inputs[0]);
   auto& out = outputs[0];
   out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
@@ -614,29 +697,49 @@ void fast::AffineQuantize::eval_cpu(
   auto& biases = outputs[2];
   scales.set_data(allocator::malloc_or_wait(scales.nbytes()));
   biases.set_data(allocator::malloc_or_wait(biases.nbytes()));
-  if (w.dtype() == float16) {
-    if (is_power_of_2(bits_)) {
-      quantize<float16_t, uint32_t>(w, out, scales, biases, bits_, group_size_);
-    } else {
-      quantize<float16_t, uint8_t>(w, out, scales, biases, bits_, group_size_);
-    }
-  } else if (w.dtype() == bfloat16) {
-    if (is_power_of_2(bits_)) {
-      quantize<bfloat16_t, uint32_t>(
-          w, out, scales, biases, bits_, group_size_);
-    } else {
-      quantize<bfloat16_t, uint8_t>(w, out, scales, biases, bits_, group_size_);
-    }
-  } else if (w.dtype() == float32) {
-    if (is_power_of_2(bits_)) {
-      quantize<float, uint32_t>(w, out, scales, biases, bits_, group_size_);
-    } else {
-      quantize<float, uint8_t>(w, out, scales, biases, bits_, group_size_);
-    }
-  } else {
-    throw std::runtime_error(
-        "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+  auto& encoder = cpu::get_command_encoder(stream());
+  if (copied) {
+    encoder.add_temporary(w);
   }
+  encoder.set_input_array(w);
+  encoder.set_input_array(scales);
+  encoder.set_input_array(biases);
+  encoder.set_output_array(out);
+  encoder.dispatch([w = array::unsafe_weak_copy(w),
+                    out = array::unsafe_weak_copy(out),
+                    scales = array::unsafe_weak_copy(scales),
+                    biases = array::unsafe_weak_copy(biases),
+                    group_size_ = group_size_,
+                    bits_ = bits_]() mutable {
+    if (w.dtype() == float16) {
+      if (is_power_of_2(bits_)) {
+        dispatch_quantize<float16_t, uint32_t>(
+            w, out, scales, biases, bits_, group_size_);
+      } else {
+        dispatch_quantize<float16_t, uint8_t>(
+            w, out, scales, biases, bits_, group_size_);
+      }
+    } else if (w.dtype() == bfloat16) {
+      if (is_power_of_2(bits_)) {
+        dispatch_quantize<bfloat16_t, uint32_t>(
+            w, out, scales, biases, bits_, group_size_);
+      } else {
+        dispatch_quantize<bfloat16_t, uint8_t>(
+            w, out, scales, biases, bits_, group_size_);
+      }
+    } else if (w.dtype() == float32) {
+      if (is_power_of_2(bits_)) {
+        dispatch_quantize<float, uint32_t>(
+            w, out, scales, biases, bits_, group_size_);
+      } else {
+        dispatch_quantize<float, uint8_t>(
+            w, out, scales, biases, bits_, group_size_);
+      }
+    } else {
+      throw std::runtime_error(
+          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+    }
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/reduce.cpp

@@ -5,6 +5,7 @@
 #include <limits>
 
 #include "mlx/backend/common/reduce.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
 #include "mlx/primitives.h"
 
@@ -139,25 +140,22 @@ void reduction_op(
     const array& x,
     array& out,
     const std::vector<int>& axes,
-    U init,
-    Op op) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+    U init) {
   ReductionPlan plan = get_reduction_plan(x, axes);
 
+  auto in_ptr = x.data<T>();
+  auto out_ptr = out.data<U>();
   if (plan.type == ContiguousAllReduce) {
-    U* out_ptr = out.data<U>();
     *out_ptr = init;
-    contiguous_reduce(x.data<T>(), out_ptr, x.size(), op, init);
+    contiguous_reduce(in_ptr, out_ptr, x.size(), Op{}, init);
     return;
   }
 
   if (plan.type == ContiguousReduce && plan.shape.size() == 1) {
     int reduction_size = plan.shape[0];
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
-    for (int i = 0; i < out.size(); i++, out_ptr++, x_ptr += reduction_size) {
+    for (int i = 0; i < out.size(); i++, out_ptr++, in_ptr += reduction_size) {
       *out_ptr = init;
-      contiguous_reduce(x_ptr, out_ptr, reduction_size, op, init);
+      contiguous_reduce(in_ptr, out_ptr, reduction_size, Op{}, init);
     }
     return;
   }
@@ -166,8 +164,6 @@ void reduction_op(
     int reduction_size = plan.shape.back();
     plan.shape.pop_back();
     plan.strides.pop_back();
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
     // Unrolling the following loop (and implementing it in order for
     // ContiguousReduce) should hold extra performance boost.
     auto [shape, strides] = shapes_without_reduction_axes(x, axes);
@@ -175,7 +171,7 @@ void reduction_op(
       for (int i = 0; i < out.size(); i++, out_ptr++) {
         int offset = elem_to_loc(i, shape, strides);
         *out_ptr = init;
-        contiguous_reduce(x_ptr + offset, out_ptr, reduction_size, op, init);
+        contiguous_reduce(in_ptr + offset, out_ptr, reduction_size, Op{}, init);
       }
     } else {
       for (int i = 0; i < out.size(); i++, out_ptr++) {
@@ -184,10 +180,10 @@ void reduction_op(
         nd_loop(
             [&](int extra_offset) {
               contiguous_reduce(
-                  x_ptr + offset + extra_offset,
+                  in_ptr + offset + extra_offset,
                   out_ptr,
                   reduction_size,
-                  op,
+                  Op{},
                   init);
             },
             plan.shape,
@@ -202,12 +198,10 @@ void reduction_op(
     size_t reduction_stride = plan.strides.back();
     plan.shape.pop_back();
     plan.strides.pop_back();
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
     for (int i = 0; i < out.size(); i += reduction_stride) {
       std::fill_n(out_ptr, reduction_stride, init);
-      strided_reduce(x_ptr, out_ptr, reduction_size, reduction_stride, op);
-      x_ptr += reduction_stride * reduction_size;
+      strided_reduce(in_ptr, out_ptr, reduction_size, reduction_stride, Op{});
+      in_ptr += reduction_stride * reduction_size;
       out_ptr += reduction_stride;
     }
     return;
@@ -219,15 +213,14 @@ void reduction_op(
     size_t reduction_stride = plan.strides.back();
     plan.shape.pop_back();
     plan.strides.pop_back();
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
     auto [shape, strides] = shapes_without_reduction_axes(x, axes);
+
     if (plan.shape.size() == 0) {
       for (int i = 0; i < out.size(); i += reduction_stride) {
         int offset = elem_to_loc(i, shape, strides);
         std::fill_n(out_ptr, reduction_stride, init);
         strided_reduce(
-            x_ptr + offset, out_ptr, reduction_size, reduction_stride, op);
+            in_ptr + offset, out_ptr, reduction_size, reduction_stride, Op{});
         out_ptr += reduction_stride;
       }
     } else {
@@ -237,11 +230,11 @@ void reduction_op(
         nd_loop(
             [&](int extra_offset) {
               strided_reduce(
-                  x_ptr + offset + extra_offset,
+                  in_ptr + offset + extra_offset,
                   out_ptr,
                   reduction_size,
                   reduction_stride,
-                  op);
+                  Op{});
             },
             plan.shape,
             plan.strides);
@@ -252,15 +245,14 @@ void reduction_op(
   }
 
   if (plan.type == GeneralReduce) {
-    const T* x_ptr = x.data<T>();
-    U* out_ptr = out.data<U>();
     auto [shape, strides] = shapes_without_reduction_axes(x, axes);
+
     for (int i = 0; i < out.size(); i++, out_ptr++) {
       int offset = elem_to_loc(i, shape, strides);
       U val = init;
       nd_loop(
           [&](int extra_offset) {
-            val = op(val, *(x_ptr + offset + extra_offset));
+            val = Op{}(val, *(in_ptr + offset + extra_offset));
           },
           plan.shape,
           plan.strides);
@@ -396,9 +388,9 @@ void reduce_dispatch_and_or(
     Reduce::ReduceType rtype,
     const std::vector<int>& axes) {
   if (rtype == Reduce::And) {
-    reduction_op<InT, bool>(in, out, axes, true, AndReduce());
+    reduction_op<InT, bool, AndReduce>(in, out, axes, true);
   } else {
-    reduction_op<InT, bool>(in, out, axes, false, OrReduce());
+    reduction_op<InT, bool, OrReduce>(in, out, axes, false);
   }
 }
 
@@ -410,15 +402,15 @@ void reduce_dispatch_sum_prod(
     const std::vector<int>& axes) {
   if (rtype == Reduce::Sum) {
     if constexpr (std::is_integral_v<InT> && sizeof(InT) <= 4) {
-      reduction_op<InT, int32_t>(in, out, axes, 0, SumReduce());
+      reduction_op<InT, int32_t, SumReduce>(in, out, axes, 0);
     } else {
-      reduction_op<InT, InT>(in, out, axes, 0, SumReduce());
+      reduction_op<InT, InT, SumReduce>(in, out, axes, 0);
     }
   } else {
     if constexpr (std::is_integral_v<InT> && sizeof(InT) <= 4) {
-      reduction_op<InT, int32_t>(in, out, axes, 1, ProdReduce());
+      reduction_op<InT, int32_t, ProdReduce>(in, out, axes, 1);
     } else {
-      reduction_op<InT, InT>(in, out, axes, 1, ProdReduce());
+      reduction_op<InT, InT, ProdReduce>(in, out, axes, 1);
     }
   }
 }
@@ -431,132 +423,141 @@ void reduce_dispatch_min_max(
     const std::vector<int>& axes) {
   if (rtype == Reduce::Max) {
     auto init = Limits<InT>::min;
-    reduction_op<InT, InT>(in, out, axes, init, MaxReduce());
+    reduction_op<InT, InT, MaxReduce>(in, out, axes, init);
   } else {
     auto init = Limits<InT>::max;
-    reduction_op<InT, InT>(in, out, axes, init, MinReduce());
+    reduction_op<InT, InT, MinReduce>(in, out, axes, init);
   }
 }
 
 void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  switch (reduce_type_) {
-    case Reduce::And:
-    case Reduce::Or: {
-      switch (in.dtype()) {
-        case bool_:
-        case uint8:
-        case int8:
-          reduce_dispatch_and_or<int8_t>(in, out, reduce_type_, axes_);
-          break;
-        case int16:
-        case uint16:
-        case float16:
-        case bfloat16:
-          reduce_dispatch_and_or<int16_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint32:
-        case int32:
-        case float32:
-          reduce_dispatch_and_or<int32_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint64:
-        case int64:
-        case float64:
-        case complex64:
-          reduce_dispatch_and_or<int64_t>(in, out, reduce_type_, axes_);
-          break;
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    reduce_type_ = reduce_type_,
+                    axes_ = axes_]() mutable {
+    switch (reduce_type_) {
+      case Reduce::And:
+      case Reduce::Or: {
+        switch (in.dtype()) {
+          case bool_:
+          case uint8:
+          case int8:
+            reduce_dispatch_and_or<int8_t>(in, out, reduce_type_, axes_);
+            break;
+          case int16:
+          case uint16:
+          case float16:
+          case bfloat16:
+            reduce_dispatch_and_or<int16_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint32:
+          case int32:
+          case float32:
+            reduce_dispatch_and_or<int32_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint64:
+          case int64:
+          case float64:
+          case complex64:
+            reduce_dispatch_and_or<int64_t>(in, out, reduce_type_, axes_);
+            break;
+        }
+        break;
       }
-      break;
-    }
-    case Reduce::Sum:
-    case Reduce::Prod: {
-      switch (in.dtype()) {
-        case bool_:
-        case uint8:
-        case int8:
-          reduce_dispatch_sum_prod<int8_t>(in, out, reduce_type_, axes_);
-          break;
-        case int16:
-        case uint16:
-          reduce_dispatch_sum_prod<int16_t>(in, out, reduce_type_, axes_);
-          break;
-        case int32:
-        case uint32:
-          reduce_dispatch_sum_prod<int32_t>(in, out, reduce_type_, axes_);
-          break;
-        case int64:
-        case uint64:
-          reduce_dispatch_sum_prod<int64_t>(in, out, reduce_type_, axes_);
-          break;
-        case float16:
-          reduce_dispatch_sum_prod<float16_t>(in, out, reduce_type_, axes_);
-          break;
-        case bfloat16:
-          reduce_dispatch_sum_prod<bfloat16_t>(in, out, reduce_type_, axes_);
-          break;
-        case float32:
-          reduce_dispatch_sum_prod<float>(in, out, reduce_type_, axes_);
-          break;
-        case float64:
-          reduce_dispatch_sum_prod<double>(in, out, reduce_type_, axes_);
-          break;
-        case complex64:
-          reduce_dispatch_sum_prod<complex64_t>(in, out, reduce_type_, axes_);
-          break;
+      case Reduce::Sum:
+      case Reduce::Prod: {
+        switch (in.dtype()) {
+          case bool_:
+          case uint8:
+          case int8:
+            reduce_dispatch_sum_prod<int8_t>(in, out, reduce_type_, axes_);
+            break;
+          case int16:
+          case uint16:
+            reduce_dispatch_sum_prod<int16_t>(in, out, reduce_type_, axes_);
+            break;
+          case int32:
+          case uint32:
+            reduce_dispatch_sum_prod<int32_t>(in, out, reduce_type_, axes_);
+            break;
+          case int64:
+          case uint64:
+            reduce_dispatch_sum_prod<int64_t>(in, out, reduce_type_, axes_);
+            break;
+          case float16:
+            reduce_dispatch_sum_prod<float16_t>(in, out, reduce_type_, axes_);
+            break;
+          case bfloat16:
+            reduce_dispatch_sum_prod<bfloat16_t>(in, out, reduce_type_, axes_);
+            break;
+          case float32:
+            reduce_dispatch_sum_prod<float>(in, out, reduce_type_, axes_);
+            break;
+          case float64:
+            reduce_dispatch_sum_prod<double>(in, out, reduce_type_, axes_);
+            break;
+          case complex64:
+            reduce_dispatch_sum_prod<complex64_t>(in, out, reduce_type_, axes_);
+            break;
+        }
+        break;
       }
-      break;
-    }
-    case Reduce::Max:
-    case Reduce::Min: {
-      switch (in.dtype()) {
-        case bool_:
-          reduce_dispatch_min_max<bool>(in, out, reduce_type_, axes_);
-          break;
-        case uint8:
-          reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint16:
-          reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint32:
-          reduce_dispatch_min_max<uint32_t>(in, out, reduce_type_, axes_);
-          break;
-        case uint64:
-          reduce_dispatch_min_max<uint64_t>(in, out, reduce_type_, axes_);
-          break;
-        case int8:
-          reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
-          break;
-        case int16:
-          reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
-          break;
-        case int32:
-          reduce_dispatch_min_max<int32_t>(in, out, reduce_type_, axes_);
-          break;
-        case int64:
-          reduce_dispatch_min_max<int64_t>(in, out, reduce_type_, axes_);
-          break;
-        case float16:
-          reduce_dispatch_min_max<float16_t>(in, out, reduce_type_, axes_);
-          break;
-        case float32:
-          reduce_dispatch_min_max<float>(in, out, reduce_type_, axes_);
-          break;
-        case float64:
-          reduce_dispatch_min_max<double>(in, out, reduce_type_, axes_);
-          break;
-        case bfloat16:
-          reduce_dispatch_min_max<bfloat16_t>(in, out, reduce_type_, axes_);
-          break;
-        case complex64:
-          reduce_dispatch_min_max<complex64_t>(in, out, reduce_type_, axes_);
-          break;
+      case Reduce::Max:
+      case Reduce::Min: {
+        switch (in.dtype()) {
+          case bool_:
+            reduce_dispatch_min_max<bool>(in, out, reduce_type_, axes_);
+            break;
+          case uint8:
+            reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint16:
+            reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint32:
+            reduce_dispatch_min_max<uint32_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint64:
+            reduce_dispatch_min_max<uint64_t>(in, out, reduce_type_, axes_);
+            break;
+          case int8:
+            reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
+            break;
+          case int16:
+            reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
+            break;
+          case int32:
+            reduce_dispatch_min_max<int32_t>(in, out, reduce_type_, axes_);
+            break;
+          case int64:
+            reduce_dispatch_min_max<int64_t>(in, out, reduce_type_, axes_);
+            break;
+          case float16:
+            reduce_dispatch_min_max<float16_t>(in, out, reduce_type_, axes_);
+            break;
+          case float32:
+            reduce_dispatch_min_max<float>(in, out, reduce_type_, axes_);
+            break;
+          case float64:
+            reduce_dispatch_min_max<double>(in, out, reduce_type_, axes_);
+            break;
+          case bfloat16:
+            reduce_dispatch_min_max<bfloat16_t>(in, out, reduce_type_, axes_);
+            break;
+          case complex64:
+            reduce_dispatch_min_max<complex64_t>(in, out, reduce_type_, axes_);
+            break;
+        }
+        break;
       }
-      break;
     }
-  }
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/scan.cpp

@@ -4,6 +4,7 @@
 
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
 #include "mlx/primitives.h"
 
@@ -153,33 +154,31 @@ void strided_scan(
 
 template <typename T, typename U, typename Op>
 void scan_op(
-    const array& input,
-    array& output,
+    const array& in,
+    array& out,
     int axis,
     bool reverse,
     bool inclusive,
     const Op& op,
     U init) {
-  output.set_data(allocator::malloc_or_wait(output.nbytes()));
-
-  if (input.flags().row_contiguous) {
-    if (input.strides()[axis] == 1) {
+  if (in.flags().row_contiguous) {
+    if (in.strides()[axis] == 1) {
       contiguous_scan(
-          input.data<T>(),
-          output.data<U>(),
-          input.size() / input.shape(axis),
-          input.shape(axis),
+          in.data<T>(),
+          out.data<U>(),
+          in.size() / in.shape(axis),
+          in.shape(axis),
           reverse,
           inclusive,
           op,
           init);
     } else {
       strided_scan(
-          input.data<T>(),
-          output.data<U>(),
-          input.size() / input.shape(axis) / input.strides()[axis],
-          input.shape(axis),
-          input.strides()[axis],
+          in.data<T>(),
+          out.data<U>(),
+          in.size() / in.shape(axis) / in.strides()[axis],
+          in.shape(axis),
+          in.strides()[axis],
           reverse,
           inclusive,
           op,
@@ -193,8 +192,8 @@ void scan_op(
 template <typename T, typename U>
 void scan_dispatch(
     Scan::ReduceType rtype,
-    const array& input,
-    array& output,
+    const array& in,
+    array& out,
     int axis,
     bool reverse,
     bool inclusive) {
@@ -202,29 +201,29 @@ void scan_dispatch(
     case Scan::Sum: {
       auto op = [](U y, T x) { return y + x; };
       auto init = static_cast<U>(0);
-      scan_op<T, U>(input, output, axis, reverse, inclusive, op, init);
+      scan_op<T, U>(in, out, axis, reverse, inclusive, op, init);
       break;
     }
     case Scan::Prod: {
       auto op = [](U y, T x) { return y * x; };
       auto init = static_cast<U>(1);
-      scan_op<T, U>(input, output, axis, reverse, inclusive, op, init);
+      scan_op<T, U>(in, out, axis, reverse, inclusive, op, init);
       break;
     }
     case Scan::Min: {
       auto op = [](U y, T x) { return x < y ? x : y; };
-      auto init = (issubdtype(input.dtype(), floating))
+      auto init = (issubdtype(in.dtype(), floating))
           ? static_cast<U>(std::numeric_limits<float>::infinity())
           : std::numeric_limits<U>::max();
-      scan_op<T, U>(input, output, axis, reverse, inclusive, op, init);
+      scan_op<T, U>(in, out, axis, reverse, inclusive, op, init);
       break;
     }
     case Scan::Max: {
       auto op = [](U y, T x) { return x < y ? y : x; };
-      auto init = (issubdtype(input.dtype(), floating))
+      auto init = (issubdtype(in.dtype(), floating))
           ? static_cast<U>(-std::numeric_limits<float>::infinity())
           : std::numeric_limits<U>::min();
-      scan_op<T, U>(input, output, axis, reverse, inclusive, op, init);
+      scan_op<T, U>(in, out, axis, reverse, inclusive, op, init);
       break;
     }
   }
@@ -235,82 +234,95 @@ void scan_dispatch(
 void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
 
+  auto& encoder = cpu::get_command_encoder(stream());
+
   // Ensure contiguity
   auto in = inputs[0];
   if (!in.flags().row_contiguous) {
     array arr_copy(in.shape(), in.dtype(), nullptr, {});
-    copy(in, arr_copy, CopyType::General);
+    copy(in, arr_copy, CopyType::General, stream());
     in = arr_copy;
+    encoder.add_temporary(arr_copy);
   }
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
 
-  switch (in.dtype()) {
-    case bool_: {
-      // We could do a full dtype x dtype switch but this is the only case
-      // where we accumulate in a different type, for now.
-      //
-      // TODO: If we add the option to accumulate floats in higher precision
-      //       floats perhaps we should add the full all-to-all dispatch.
-      if (reduce_type_ == Scan::Sum && out.dtype() == int32) {
-        scan_dispatch<bool, int32_t>(
-            reduce_type_, in, out, axis_, reverse_, inclusive_);
-      } else {
-        scan_dispatch<bool, bool>(
-            reduce_type_, in, out, axis_, reverse_, inclusive_);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    axis_ = axis_,
+                    reduce_type_ = reduce_type_,
+                    reverse_ = reverse_,
+                    inclusive_ = inclusive_]() mutable {
+    switch (in.dtype()) {
+      case bool_: {
+        // We could do a full dtype x dtype switch but this is the only case
+        // where we accumulate in a different type, for now.
+        //
+        // TODO: If we add the option to accumulate floats in higher precision
+        //       floats perhaps we should add the full all-to-all dispatch.
+        if (reduce_type_ == Scan::Sum && out.dtype() == int32) {
+          scan_dispatch<bool, int32_t>(
+              reduce_type_, in, out, axis_, reverse_, inclusive_);
+        } else {
+          scan_dispatch<bool, bool>(
+              reduce_type_, in, out, axis_, reverse_, inclusive_);
+        }
+        break;
       }
-      break;
+      case uint8:
+        scan_dispatch<uint8_t, uint8_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case uint16:
+        scan_dispatch<uint16_t, uint16_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case uint32:
+        scan_dispatch<uint32_t, uint32_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case uint64:
+        scan_dispatch<uint64_t, uint64_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case int8:
+        scan_dispatch<int8_t, int8_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case int16:
+        scan_dispatch<int16_t, int16_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case int32:
+        scan_dispatch<int32_t, int32_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case int64:
+        scan_dispatch<int64_t, int64_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case float16:
+        scan_dispatch<float16_t, float16_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case float32:
+        scan_dispatch<float, float>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case float64:
+        scan_dispatch<double, double>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case bfloat16:
+        scan_dispatch<bfloat16_t, bfloat16_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
+        break;
+      case complex64:
+        throw std::runtime_error("Scan ops do not support complex types yet");
+        break;
     }
-    case uint8:
-      scan_dispatch<uint8_t, uint8_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case uint16:
-      scan_dispatch<uint16_t, uint16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case uint32:
-      scan_dispatch<uint32_t, uint32_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case uint64:
-      scan_dispatch<uint64_t, uint64_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int8:
-      scan_dispatch<int8_t, int8_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int16:
-      scan_dispatch<int16_t, int16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int32:
-      scan_dispatch<int32_t, int32_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case int64:
-      scan_dispatch<int64_t, int64_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case float16:
-      scan_dispatch<float16_t, float16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case float32:
-      scan_dispatch<float, float>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case float64:
-      scan_dispatch<double, double>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case bfloat16:
-      scan_dispatch<bfloat16_t, bfloat16_t>(
-          reduce_type_, in, out, axis_, reverse_, inclusive_);
-      break;
-    case complex64:
-      throw std::runtime_error("Scan ops do not support complex types yet");
-      break;
-  }
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/select.cpp

@@ -16,51 +16,70 @@ void select_op(
     const array& b,
     const array& c,
     array& out,
-    Op op) {
-  switch (out.dtype()) {
-    case bool_:
-      ternary_op<bool, bool, bool, bool>(a, b, c, out, op);
-      break;
-    case uint8:
-      ternary_op<bool, uint8_t, uint8_t, uint8_t>(a, b, c, out, op);
-      break;
-    case uint16:
-      ternary_op<bool, uint16_t, uint16_t, uint16_t>(a, b, c, out, op);
-      break;
-    case uint32:
-      ternary_op<bool, uint32_t, uint32_t, uint32_t>(a, b, c, out, op);
-      break;
-    case uint64:
-      ternary_op<bool, uint64_t, uint64_t, uint64_t>(a, b, c, out, op);
-      break;
-    case int8:
-      ternary_op<bool, int8_t, int8_t, int8_t>(a, b, c, out, op);
-      break;
-    case int16:
-      ternary_op<bool, int16_t, int16_t, int16_t>(a, b, c, out, op);
-      break;
-    case int32:
-      ternary_op<bool, int32_t, int32_t, int32_t>(a, b, c, out, op);
-      break;
-    case int64:
-      ternary_op<bool, int64_t, int64_t, int64_t>(a, b, c, out, op);
-      break;
-    case float16:
-      ternary_op<bool, float16_t, float16_t, float16_t>(a, b, c, out, op);
-      break;
-    case float32:
-      ternary_op<bool, float, float, float>(a, b, c, out, op);
-      break;
-    case float64:
-      ternary_op<bool, double, double, double>(a, b, c, out, op);
-      break;
-    case bfloat16:
-      ternary_op<bool, bfloat16_t, bfloat16_t, bfloat16_t>(a, b, c, out, op);
-      break;
-    case complex64:
-      ternary_op<bool, complex64_t, complex64_t, complex64_t>(a, b, c, out, op);
-      break;
-  }
+    Op op,
+    Stream stream) {
+  TernaryOpType topt = get_ternary_op_type(a, b, c);
+  set_ternary_op_output_data(a, b, c, out, topt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    c = array::unsafe_weak_copy(c),
+                    out = array::unsafe_weak_copy(out),
+                    op,
+                    topt]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        ternary_op<bool, bool, bool, bool>(a, b, c, out, op, topt);
+        break;
+      case uint8:
+        ternary_op<bool, uint8_t, uint8_t, uint8_t>(a, b, c, out, op, topt);
+        break;
+      case uint16:
+        ternary_op<bool, uint16_t, uint16_t, uint16_t>(a, b, c, out, op, topt);
+        break;
+      case uint32:
+        ternary_op<bool, uint32_t, uint32_t, uint32_t>(a, b, c, out, op, topt);
+        break;
+      case uint64:
+        ternary_op<bool, uint64_t, uint64_t, uint64_t>(a, b, c, out, op, topt);
+        break;
+      case int8:
+        ternary_op<bool, int8_t, int8_t, int8_t>(a, b, c, out, op, topt);
+        break;
+      case int16:
+        ternary_op<bool, int16_t, int16_t, int16_t>(a, b, c, out, op, topt);
+        break;
+      case int32:
+        ternary_op<bool, int32_t, int32_t, int32_t>(a, b, c, out, op, topt);
+        break;
+      case int64:
+        ternary_op<bool, int64_t, int64_t, int64_t>(a, b, c, out, op, topt);
+        break;
+      case float16:
+        ternary_op<bool, float16_t, float16_t, float16_t>(
+            a, b, c, out, op, topt);
+        break;
+      case float32:
+        ternary_op<bool, float, float, float>(a, b, c, out, op, topt);
+        break;
+      case float64:
+        ternary_op<bool, double, double, double>(a, b, c, out, op, topt);
+        break;
+      case bfloat16:
+        ternary_op<bool, bfloat16_t, bfloat16_t, bfloat16_t>(
+            a, b, c, out, op, topt);
+        break;
+      case complex64:
+        ternary_op<bool, complex64_t, complex64_t, complex64_t>(
+            a, b, c, out, op, topt);
+        break;
+    }
+  });
 }
 
 } // namespace
@@ -70,7 +89,7 @@ void Select::eval_cpu(const std::vector<array>& inputs, array& out) {
   const auto& condition = inputs[0];
   const auto& a = inputs[1];
   const auto& b = inputs[2];
-  select_op(condition, a, b, out, detail::Select());
+  select_op(condition, a, b, out, detail::Select(), stream());
 }
 
 } // namespace mlx::core

mlx/backend/cpu/softmax.cpp

@@ -4,6 +4,7 @@
 #include <cmath>
 
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
 #include "mlx/primitives.h"
 #include "mlx/types/limits.h"
@@ -15,92 +16,100 @@ namespace {
 using namespace mlx::core::simd;
 
 template <typename T, typename AccT>
-void softmax(const array& in, array& out) {
-  constexpr bool same_t = std::is_same_v<T, AccT>;
-  constexpr int N = std::min(max_size<AccT>, max_size<T>);
+void softmax(const array& in, array& out, Stream stream) {
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
 
   const T* in_ptr = in.data<T>();
   T* out_ptr = out.data<T>();
+
   int M = in.shape().back();
   int L = in.data_size() / M;
-  const T* current_in_ptr;
-  T* current_out_ptr;
 
-  for (int i = 0; i < L; i++, in_ptr += M, out_ptr += M) {
-    // Find the maximum
-    current_in_ptr = in_ptr;
-    Simd<AccT, N> vmaximum(-numeric_limits<AccT>::infinity());
-    size_t s = M;
-    while (s >= N) {
-      Simd<AccT, N> vals = load<T, N>(current_in_ptr);
-      vmaximum = maximum(vals, vmaximum);
-      current_in_ptr += N;
-      s -= N;
-    }
+  encoder.dispatch([in_ptr, out_ptr, M, L]() mutable {
+    constexpr bool same_t = std::is_same_v<T, AccT>;
+    constexpr int N = std::min(max_size<AccT>, max_size<T>);
 
-    AccT maximum = max(vmaximum);
-    while (s-- > 0) {
-      maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
-      current_in_ptr++;
-    }
+    const T* current_in_ptr;
+    T* current_out_ptr;
 
-    // Compute the normalizer and the exponentials
-    Simd<AccT, N> vnormalizer(0.0);
-    current_out_ptr = out_ptr;
-    current_in_ptr = in_ptr;
-    s = M;
-    while (s >= N) {
-      Simd<AccT, N> vexp = load<T, N>(current_in_ptr);
-      vexp = exp(vexp - maximum);
-      if constexpr (same_t) {
-        store(current_out_ptr, vexp);
-      }
-      vnormalizer = vnormalizer + vexp;
-      current_in_ptr += N;
-      current_out_ptr += N;
-      s -= N;
-    }
-    AccT normalizer = sum(vnormalizer);
-    while (s-- > 0) {
-      AccT _exp = std::exp(*current_in_ptr - maximum);
-      if constexpr (same_t) {
-        *current_out_ptr = _exp;
-      }
-      normalizer += _exp;
-      current_in_ptr++;
-      current_out_ptr++;
-    }
-    normalizer = 1 / normalizer;
-
-    // Normalize
-    current_out_ptr = out_ptr;
-    current_in_ptr = in_ptr;
-    s = M;
-    while (s >= N) {
-      if constexpr (same_t) {
-        store(
-            current_out_ptr,
-            Simd<T, N>(load<T, N>(current_out_ptr) * normalizer));
-      } else {
-        Simd<AccT, N> vexp = load<T, N>(current_in_ptr);
-        vexp = exp(vexp - maximum) * normalizer;
-        store(current_out_ptr, Simd<T, N>(vexp));
+    for (int i = 0; i < L; i++, in_ptr += M, out_ptr += M) {
+      // Find the maximum
+      current_in_ptr = in_ptr;
+      Simd<AccT, N> vmaximum(-numeric_limits<AccT>::infinity());
+      size_t s = M;
+      while (s >= N) {
+        Simd<AccT, N> vals = load<T, N>(current_in_ptr);
+        vmaximum = maximum(vals, vmaximum);
         current_in_ptr += N;
+        s -= N;
       }
-      current_out_ptr += N;
-      s -= N;
-    }
-    while (s-- > 0) {
-      if constexpr (same_t) {
-        *current_out_ptr *= normalizer;
-      } else {
-        AccT _exp = std::exp(*current_in_ptr - maximum);
-        *current_out_ptr = static_cast<T>(_exp * normalizer);
+
+      AccT maximum = max(vmaximum);
+      while (s-- > 0) {
+        maximum = std::max(maximum, static_cast<AccT>(*current_in_ptr));
         current_in_ptr++;
       }
-      current_out_ptr++;
+
+      // Compute the normalizer and the exponentials
+      Simd<AccT, N> vnormalizer(0.0);
+      current_out_ptr = out_ptr;
+      current_in_ptr = in_ptr;
+      s = M;
+      while (s >= N) {
+        Simd<AccT, N> vexp = load<T, N>(current_in_ptr);
+        vexp = exp(vexp - maximum);
+        if constexpr (same_t) {
+          store(current_out_ptr, vexp);
+        }
+        vnormalizer = vnormalizer + vexp;
+        current_in_ptr += N;
+        current_out_ptr += N;
+        s -= N;
+      }
+      AccT normalizer = sum(vnormalizer);
+      while (s-- > 0) {
+        AccT _exp = std::exp(*current_in_ptr - maximum);
+        if constexpr (same_t) {
+          *current_out_ptr = _exp;
+        }
+        normalizer += _exp;
+        current_in_ptr++;
+        current_out_ptr++;
+      }
+      normalizer = 1 / normalizer;
+
+      // Normalize
+      current_out_ptr = out_ptr;
+      current_in_ptr = in_ptr;
+      s = M;
+      while (s >= N) {
+        if constexpr (same_t) {
+          store(
+              current_out_ptr,
+              Simd<T, N>(load<T, N>(current_out_ptr) * normalizer));
+        } else {
+          Simd<AccT, N> vexp = load<T, N>(current_in_ptr);
+          vexp = exp(vexp - maximum) * normalizer;
+          store(current_out_ptr, Simd<T, N>(vexp));
+          current_in_ptr += N;
+        }
+        current_out_ptr += N;
+        s -= N;
+      }
+      while (s-- > 0) {
+        if constexpr (same_t) {
+          *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++;
+      }
     }
-  }
+  });
 }
 
 } // namespace
@@ -109,30 +118,32 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
 
   // Make sure that the last dimension is contiguous
-  auto check_input = [](array x) {
+  auto set_output = [s = stream(), &out](const array& x) {
     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) {
+      if (x.is_donatable()) {
+        out.copy_shared_buffer(x);
+      } else {
+        out.set_data(
+            allocator::malloc_or_wait(x.data_size() * x.itemsize()),
+            x.data_size(),
+            x.strides(),
+            x.flags());
+      }
       return x;
     } else {
       array x_copy(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General);
+      copy(x, x_copy, CopyType::General, s);
+      out.copy_shared_buffer(x_copy);
       return x_copy;
     }
   };
-  array in = check_input(std::move(inputs[0]));
-  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());
-  }
+
+  auto in = set_output(inputs[0]);
 
   switch (in.dtype()) {
     case bool_:
@@ -148,24 +159,24 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
           "Softmax is defined only for floating point types");
       break;
     case float32:
-      softmax<float, float>(in, out);
+      softmax<float, float>(in, out, stream());
       break;
     case float16:
       if (precise_) {
-        softmax<float16_t, float>(in, out);
+        softmax<float16_t, float>(in, out, stream());
       } else {
-        softmax<float16_t, float16_t>(in, out);
+        softmax<float16_t, float16_t>(in, out, stream());
       }
       break;
     case bfloat16:
       if (precise_) {
-        softmax<bfloat16_t, float>(in, out);
+        softmax<bfloat16_t, float>(in, out, stream());
       } else {
-        softmax<bfloat16_t, bfloat16_t>(in, out);
+        softmax<bfloat16_t, bfloat16_t>(in, out, stream());
       }
       break;
     case float64:
-      softmax<double, double>(in, out);
+      softmax<double, double>(in, out, stream());
       break;
     case complex64:
       throw std::invalid_argument(

mlx/backend/cpu/sort.cpp

@@ -7,6 +7,7 @@
 
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 
 #include "mlx/primitives.h"
 
@@ -103,16 +104,12 @@ struct StridedIterator {
   T* ptr_;
 };
 
-template <typename T, typename IdxT = uint32_t>
-void sort(const array& in, array& out, int axis) {
-  // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype);
-
+template <typename T>
+void sort(array& out, int axis) {
   // Get axis, shape and stride info
-  axis = axis < 0 ? axis + in.ndim() : axis;
-  size_t in_size = in.flags().contiguous ? in.data_size() : in.size();
-  size_t n_rows = in_size / in.shape(axis);
+  axis = axis < 0 ? axis + out.ndim() : axis;
+  size_t in_size = out.size();
+  size_t n_rows = in_size / out.shape(axis);
 
   auto remaining_shape = out.shape();
   remaining_shape.erase(remaining_shape.begin() + axis);
@@ -126,8 +123,9 @@ void sort(const array& in, array& out, int axis) {
   // Perform sorting in place
   ContiguousIterator src_it(
       remaining_shape, remaining_strides, remaining_shape.size());
+  auto out_ptr = out.data<T>();
   for (int i = 0; i < n_rows; i++) {
-    T* data_ptr = out.data<T>() + src_it.loc;
+    T* data_ptr = out_ptr + src_it.loc;
 
     StridedIterator st(data_ptr, axis_stride, 0);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
@@ -139,9 +137,6 @@ void sort(const array& in, array& out, int axis) {
 
 template <typename T, typename IdxT = uint32_t>
 void argsort(const array& in, array& out, int axis) {
-  // Allocate output
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
   // Get axis, shape and stride info
   axis = axis < 0 ? axis + in.ndim() : axis;
   size_t n_rows = in.size() / in.shape(axis);
@@ -167,9 +162,12 @@ void argsort(const array& in, array& out, int axis) {
       in_remaining_shape, in_remaining_strides, in_remaining_shape.size());
   ContiguousIterator out_it(
       out_remaining_shape, out_remaining_strides, out_remaining_shape.size());
+  auto in_ptr = in.data<T>();
+  auto out_ptr = out.data<IdxT>();
   for (int i = 0; i < n_rows; i++) {
-    const T* data_ptr = in.data<T>() + in_it.loc;
-    IdxT* idx_ptr = out.data<IdxT>() + out_it.loc;
+    const T* data_ptr = in_ptr + in_it.loc;
+    IdxT* idx_ptr = out_ptr + out_it.loc;
+
     in_it.step();
     out_it.step();
 
@@ -191,33 +189,30 @@ void argsort(const array& in, array& out, int axis) {
   }
 }
 
-template <typename T, typename IdxT = uint32_t>
-void partition(const array& in, array& out, int axis, int kth) {
-  // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype);
-
+template <typename T>
+void partition(array& out, int axis, int kth) {
   // Get axis, shape and stride info
-  axis = axis < 0 ? axis + in.ndim() : axis;
-  size_t in_size = in.flags().contiguous ? in.data_size() : in.size();
-  size_t n_rows = in_size / in.shape(axis);
+  axis = axis < 0 ? axis + out.ndim() : axis;
+  size_t in_size = out.size();
+  size_t n_rows = in_size / out.shape(axis);
 
-  auto remaining_shape = in.shape();
+  auto remaining_shape = out.shape();
   remaining_shape.erase(remaining_shape.begin() + axis);
 
-  auto remaining_strides = in.strides();
+  auto remaining_strides = out.strides();
   remaining_strides.erase(remaining_strides.begin() + axis);
 
-  auto axis_stride = in.strides()[axis];
-  int axis_size = in.shape(axis);
+  auto axis_stride = out.strides()[axis];
+  int axis_size = out.shape(axis);
 
   kth = kth < 0 ? kth + axis_size : kth;
 
   // Perform partition in place
   ContiguousIterator src_it(
       remaining_shape, remaining_strides, remaining_shape.size());
+  auto out_ptr = out.data<T>();
   for (int i = 0; i < n_rows; i++) {
-    T* data_ptr = out.data<T>() + src_it.loc;
+    T* data_ptr = out_ptr + src_it.loc;
     src_it.step();
 
     StridedIterator st(data_ptr, axis_stride, 0);
@@ -230,9 +225,6 @@ void partition(const array& in, array& out, int axis, int kth) {
 
 template <typename T, typename IdxT = uint32_t>
 void argpartition(const array& in, array& out, int axis, int kth) {
-  // Allocate output
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
   // Get axis, shape and stride info
   axis = axis < 0 ? axis + in.ndim() : axis;
   size_t n_rows = in.size() / in.shape(axis);
@@ -260,9 +252,13 @@ void argpartition(const array& in, array& out, int axis, int kth) {
       in_remaining_shape, in_remaining_strides, in_remaining_shape.size());
   ContiguousIterator out_it(
       out_remaining_shape, out_remaining_strides, out_remaining_shape.size());
+
+  auto in_ptr = in.data<T>();
+  auto out_ptr = out.data<IdxT>();
+
   for (int i = 0; i < n_rows; i++) {
-    const T* data_ptr = in.data<T>() + in_it.loc;
-    IdxT* idx_ptr = out.data<IdxT>() + out_it.loc;
+    const T* data_ptr = in_ptr + in_it.loc;
+    IdxT* idx_ptr = out_ptr + out_it.loc;
     in_it.step();
     out_it.step();
 
@@ -291,144 +287,184 @@ void ArgSort::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
-  switch (in.dtype()) {
-    case bool_:
-      return argsort<bool>(in, out, axis_);
-    case uint8:
-      return argsort<uint8_t>(in, out, axis_);
-    case uint16:
-      return argsort<uint16_t>(in, out, axis_);
-    case uint32:
-      return argsort<uint32_t>(in, out, axis_);
-    case uint64:
-      return argsort<uint64_t>(in, out, axis_);
-    case int8:
-      return argsort<int8_t>(in, out, axis_);
-    case int16:
-      return argsort<int16_t>(in, out, axis_);
-    case int32:
-      return argsort<int32_t>(in, out, axis_);
-    case int64:
-      return argsort<int64_t>(in, out, axis_);
-    case float32:
-      return argsort<float>(in, out, axis_);
-    case float64:
-      return argsort<double>(in, out, axis_);
-    case float16:
-      return argsort<float16_t>(in, out, axis_);
-    case bfloat16:
-      return argsort<bfloat16_t>(in, out, axis_);
-    case complex64:
-      return argsort<complex64_t>(in, out, axis_);
-  }
+  // Allocate output
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_input_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    axis_ = axis_]() mutable {
+    switch (in.dtype()) {
+      case bool_:
+        return argsort<bool>(in, out, axis_);
+      case uint8:
+        return argsort<uint8_t>(in, out, axis_);
+      case uint16:
+        return argsort<uint16_t>(in, out, axis_);
+      case uint32:
+        return argsort<uint32_t>(in, out, axis_);
+      case uint64:
+        return argsort<uint64_t>(in, out, axis_);
+      case int8:
+        return argsort<int8_t>(in, out, axis_);
+      case int16:
+        return argsort<int16_t>(in, out, axis_);
+      case int32:
+        return argsort<int32_t>(in, out, axis_);
+      case int64:
+        return argsort<int64_t>(in, out, axis_);
+      case float32:
+        return argsort<float>(in, out, axis_);
+      case float64:
+        return argsort<double>(in, out, axis_);
+      case float16:
+        return argsort<float16_t>(in, out, axis_);
+      case bfloat16:
+        return argsort<bfloat16_t>(in, out, axis_);
+      case complex64:
+        return argsort<complex64_t>(in, out, axis_);
+    }
+  });
 }
 
 void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
-  switch (in.dtype()) {
-    case bool_:
-      return sort<bool>(in, out, axis_);
-    case uint8:
-      return sort<uint8_t>(in, out, axis_);
-    case uint16:
-      return sort<uint16_t>(in, out, axis_);
-    case uint32:
-      return sort<uint32_t>(in, out, axis_);
-    case uint64:
-      return sort<uint64_t>(in, out, axis_);
-    case int8:
-      return sort<int8_t>(in, out, axis_);
-    case int16:
-      return sort<int16_t>(in, out, axis_);
-    case int32:
-      return sort<int32_t>(in, out, axis_);
-    case int64:
-      return sort<int64_t>(in, out, axis_);
-    case float32:
-      return sort<float>(in, out, axis_);
-    case float64:
-      return sort<double>(in, out, axis_);
-    case float16:
-      return sort<float16_t>(in, out, axis_);
-    case bfloat16:
-      return sort<bfloat16_t>(in, out, axis_);
-    case complex64:
-      return sort<complex64_t>(in, out, axis_);
-  }
+  // Copy input to output
+  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
+  copy(in, out, ctype, stream());
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_output_array(out);
+  encoder.dispatch(
+      [out = array::unsafe_weak_copy(out), axis_ = axis_]() mutable {
+        switch (out.dtype()) {
+          case bool_:
+            return sort<bool>(out, axis_);
+          case uint8:
+            return sort<uint8_t>(out, axis_);
+          case uint16:
+            return sort<uint16_t>(out, axis_);
+          case uint32:
+            return sort<uint32_t>(out, axis_);
+          case uint64:
+            return sort<uint64_t>(out, axis_);
+          case int8:
+            return sort<int8_t>(out, axis_);
+          case int16:
+            return sort<int16_t>(out, axis_);
+          case int32:
+            return sort<int32_t>(out, axis_);
+          case int64:
+            return sort<int64_t>(out, axis_);
+          case float32:
+            return sort<float>(out, axis_);
+          case float64:
+            return sort<double>(out, axis_);
+          case float16:
+            return sort<float16_t>(out, axis_);
+          case bfloat16:
+            return sort<bfloat16_t>(out, axis_);
+          case complex64:
+            return sort<complex64_t>(out, axis_);
+        }
+      });
 }
 
 void ArgPartition::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
-  switch (in.dtype()) {
-    case bool_:
-      return argpartition<bool>(in, out, axis_, kth_);
-    case uint8:
-      return argpartition<uint8_t>(in, out, axis_, kth_);
-    case uint16:
-      return argpartition<uint16_t>(in, out, axis_, kth_);
-    case uint32:
-      return argpartition<uint32_t>(in, out, axis_, kth_);
-    case uint64:
-      return argpartition<uint64_t>(in, out, axis_, kth_);
-    case int8:
-      return argpartition<int8_t>(in, out, axis_, kth_);
-    case int16:
-      return argpartition<int16_t>(in, out, axis_, kth_);
-    case int32:
-      return argpartition<int32_t>(in, out, axis_, kth_);
-    case int64:
-      return argpartition<int64_t>(in, out, axis_, kth_);
-    case float32:
-      return argpartition<float>(in, out, axis_, kth_);
-    case float64:
-      return argpartition<double>(in, out, axis_, kth_);
-    case float16:
-      return argpartition<float16_t>(in, out, axis_, kth_);
-    case bfloat16:
-      return argpartition<bfloat16_t>(in, out, axis_, kth_);
-    case complex64:
-      return argpartition<complex64_t>(in, out, axis_, kth_);
-  }
+  // Allocate output
+  out.set_data(allocator::malloc_or_wait(out.nbytes()));
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_input_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    axis_ = axis_,
+                    kth_ = kth_]() mutable {
+    switch (in.dtype()) {
+      case bool_:
+        return argpartition<bool>(in, out, axis_, kth_);
+      case uint8:
+        return argpartition<uint8_t>(in, out, axis_, kth_);
+      case uint16:
+        return argpartition<uint16_t>(in, out, axis_, kth_);
+      case uint32:
+        return argpartition<uint32_t>(in, out, axis_, kth_);
+      case uint64:
+        return argpartition<uint64_t>(in, out, axis_, kth_);
+      case int8:
+        return argpartition<int8_t>(in, out, axis_, kth_);
+      case int16:
+        return argpartition<int16_t>(in, out, axis_, kth_);
+      case int32:
+        return argpartition<int32_t>(in, out, axis_, kth_);
+      case int64:
+        return argpartition<int64_t>(in, out, axis_, kth_);
+      case float32:
+        return argpartition<float>(in, out, axis_, kth_);
+      case float64:
+        return argpartition<double>(in, out, axis_, kth_);
+      case float16:
+        return argpartition<float16_t>(in, out, axis_, kth_);
+      case bfloat16:
+        return argpartition<bfloat16_t>(in, out, axis_, kth_);
+      case complex64:
+        return argpartition<complex64_t>(in, out, axis_, kth_);
+    }
+  });
 }
 
 void Partition::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
-  switch (in.dtype()) {
-    case bool_:
-      return partition<bool>(in, out, axis_, kth_);
-    case uint8:
-      return partition<uint8_t>(in, out, axis_, kth_);
-    case uint16:
-      return partition<uint16_t>(in, out, axis_, kth_);
-    case uint32:
-      return partition<uint32_t>(in, out, axis_, kth_);
-    case uint64:
-      return partition<uint64_t>(in, out, axis_, kth_);
-    case int8:
-      return partition<int8_t>(in, out, axis_, kth_);
-    case int16:
-      return partition<int16_t>(in, out, axis_, kth_);
-    case int32:
-      return partition<int32_t>(in, out, axis_, kth_);
-    case int64:
-      return partition<int64_t>(in, out, axis_, kth_);
-    case float32:
-      return partition<float>(in, out, axis_, kth_);
-    case float64:
-      return partition<double>(in, out, axis_, kth_);
-    case float16:
-      return partition<float16_t>(in, out, axis_, kth_);
-    case bfloat16:
-      return partition<bfloat16_t>(in, out, axis_, kth_);
-    case complex64:
-      return partition<complex64_t>(in, out, axis_, kth_);
-  }
+  // Copy input to output
+  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
+  copy(in, out, ctype, stream());
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_output_array(out);
+  encoder.dispatch([out = array::unsafe_weak_copy(out),
+                    axis_ = axis_,
+                    kth_ = kth_]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        return partition<bool>(out, axis_, kth_);
+      case uint8:
+        return partition<uint8_t>(out, axis_, kth_);
+      case uint16:
+        return partition<uint16_t>(out, axis_, kth_);
+      case uint32:
+        return partition<uint32_t>(out, axis_, kth_);
+      case uint64:
+        return partition<uint64_t>(out, axis_, kth_);
+      case int8:
+        return partition<int8_t>(out, axis_, kth_);
+      case int16:
+        return partition<int16_t>(out, axis_, kth_);
+      case int32:
+        return partition<int32_t>(out, axis_, kth_);
+      case int64:
+        return partition<int64_t>(out, axis_, kth_);
+      case float32:
+        return partition<float>(out, axis_, kth_);
+      case float64:
+        return partition<double>(out, axis_, kth_);
+      case float16:
+        return partition<float16_t>(out, axis_, kth_);
+      case bfloat16:
+        return partition<bfloat16_t>(out, axis_, kth_);
+      case complex64:
+        return partition<complex64_t>(out, axis_, kth_);
+    }
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/svd.cpp

@@ -2,13 +2,18 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 template <typename T>
-void svd_impl(const array& a, T* u_data, T* s_data, T* vt_data) {
+void svd_impl(
+    const array& a,
+    std::vector<array>& outputs,
+    bool compute_uv,
+    Stream stream) {
   // 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
@@ -22,118 +27,24 @@ void svd_impl(const array& a, T* u_data, T* s_data, T* vt_data) {
   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(), a.dtype(), nullptr, {});
-  copy(a, in, a.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+  copy(
+      a,
+      in,
+      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      stream);
 
-  auto job_u = (u_data && vt_data) ? "V" : "N";
-  auto job_vt = (u_data && vt_data) ? "V" : "N";
-  static constexpr auto range = "A";
+  // Allocate outputs.
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  auto in_ptr = in.data<T>();
+  T* u_ptr;
+  T* s_ptr;
+  T* vt_ptr;
 
-  // Will contain the number of singular values after the call has returned.
-  int ns = 0;
-  T 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 T ignored_float = 0;
-  static T ignored_output = 0;
-
-  int info;
-
-  // Compute workspace size.
-  gesvdx<T>(
-      /* 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::eval_cpu] 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(T) * lwork)};
-
-  // Loop over matrices.
-  for (int i = 0; i < num_matrices; i++) {
-    gesvdx<T>(
-        /* 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<T>() + M * N * i,
-        /* lda = */ &lda,
-        /* vl = */ &ignored_float,
-        /* vu = */ &ignored_float,
-        /* il = */ &ignored_int,
-        /* iu = */ &ignored_int,
-        /* ns = */ &ns,
-        /* s = */ s_data + K * i,
-        // According to the identity above, lapack will write Vᵀᵀ as U.
-        /* u = */ vt_data ? vt_data + N * N * i : &ignored_output,
-        /* ldu = */ &ldu,
-        // According to the identity above, lapack will write Uᵀ as Vᵀ.
-        /* vt = */ u_data ? u_data + M * M * i : &ignored_output,
-        /* ldvt = */ &ldvt,
-        /* work = */ static_cast<T*>(scratch.buffer.raw_ptr()),
-        /* lwork = */ &lwork,
-        /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "[SVD::eval_cpu] failed with code " << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    if (ns != K) {
-      std::stringstream ss;
-      ss << "[SVD::eval_cpu] expected " << K << " singular values, but " << ns
-         << " were computed.";
-      throw std::runtime_error(ss.str());
-    }
-  }
-}
-
-template <typename T>
-void compute_svd(const array& a, bool compute_uv, std::vector<array>& outputs) {
   if (compute_uv) {
     array& u = outputs[0];
     array& s = outputs[1];
@@ -143,25 +54,147 @@ void compute_svd(const array& a, bool compute_uv, std::vector<array>& outputs) {
     s.set_data(allocator::malloc_or_wait(s.nbytes()));
     vt.set_data(allocator::malloc_or_wait(vt.nbytes()));
 
-    svd_impl<T>(a, u.data<T>(), s.data<T>(), vt.data<T>());
+    encoder.set_output_array(u);
+    encoder.set_output_array(s);
+    encoder.set_output_array(vt);
+
+    s_ptr = s.data<T>();
+    u_ptr = u.data<T>();
+    vt_ptr = vt.data<T>();
   } else {
     array& s = outputs[0];
 
     s.set_data(allocator::malloc_or_wait(s.nbytes()));
 
-    svd_impl<T>(a, nullptr, s.data<T>(), nullptr);
+    encoder.set_output_array(s);
+
+    s_ptr = s.data<T>();
+    u_ptr = nullptr;
+    vt_ptr = nullptr;
   }
+
+  encoder.dispatch([in_ptr, u_ptr, s_ptr, vt_ptr, M, N, K, num_matrices]() {
+    // 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;
+
+    auto job_u = (u_ptr) ? "V" : "N";
+    auto job_vt = (u_ptr) ? "V" : "N";
+    static constexpr auto range = "A";
+
+    // Will contain the number of singular values after the call has returned.
+    int ns = 0;
+    T 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 T ignored_float = 0;
+
+    int info;
+
+    // Compute workspace size.
+    gesvdx<T>(
+        /* 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::eval_cpu] 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(T) * lwork)};
+
+    // Loop over matrices.
+    for (int i = 0; i < num_matrices; i++) {
+      gesvdx<T>(
+          /* jobu = */ job_u,
+          /* jobvt = */ job_vt,
+          /* range = */ range,
+          // M and N are swapped since lapack expects column-major.
+          /* m = */ &N,
+          /* n = */ &M,
+          /* a = */ in_ptr + M * N * i,
+          /* lda = */ &lda,
+          /* vl = */ &ignored_float,
+          /* vu = */ &ignored_float,
+          /* il = */ &ignored_int,
+          /* iu = */ &ignored_int,
+          /* ns = */ &ns,
+          /* s = */ s_ptr + K * i,
+          // According to the identity above, lapack will write Vᵀᵀ as U.
+          /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
+          /* ldu = */ &ldu,
+          // According to the identity above, lapack will write Uᵀ as Vᵀ.
+          /* vt = */ u_ptr ? u_ptr + M * M * i : nullptr,
+          /* ldvt = */ &ldvt,
+          /* work = */ static_cast<T*>(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());
+      }
+    }
+  });
+  encoder.add_temporary(in);
 }
 
+template <typename T>
+void compute_svd(
+    const array& a,
+    bool compute_uv,
+    std::vector<array>& outputs,
+    Stream stream) {}
+
 void SVD::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   switch (inputs[0].dtype()) {
     case float32:
-      compute_svd<float>(inputs[0], compute_uv_, outputs);
+      svd_impl<float>(inputs[0], outputs, compute_uv_, stream());
       break;
     case float64:
-      compute_svd<double>(inputs[0], compute_uv_, outputs);
+      svd_impl<double>(inputs[0], outputs, compute_uv_, stream());
       break;
     default:
       throw std::runtime_error(

mlx/backend/cpu/ternary.h

@@ -1,10 +1,10 @@
 // Copyright © 2023 Apple Inc.
 
 #pragma once
-#include "mlx/allocator.h"
 #include "mlx/array.h"
 #include "mlx/backend/common/ternary.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/encoder.h"
 
 namespace mlx::core {
 
@@ -53,22 +53,18 @@ void ternary_op_dims(
 
 template <typename T1, typename T2, typename T3, typename U, typename Op>
 void ternary_op_dispatch_dims(
-    const array& a,
-    const array& b,
-    const array& c,
-    array& out,
-    Op op) {
-  auto [shape, strides] = collapse_contiguous_dims(
-      a.shape(), {a.strides(), b.strides(), c.strides(), out.strides()});
+    const T1* a_ptr,
+    const T2* b_ptr,
+    const T3* c_ptr,
+    U* out_ptr,
+    Op op,
+    size_t size,
+    Shape& shape,
+    std::vector<Strides>& strides) {
   const auto& a_strides = strides[0];
   const auto& b_strides = strides[1];
   const auto& c_strides = strides[2];
   const auto& out_strides = strides[3];
-
-  const T1* a_ptr = a.data<T1>();
-  const T2* b_ptr = b.data<T2>();
-  const T3* c_ptr = c.data<T3>();
-  U* out_ptr = out.data<T3>();
   int ndim = shape.size();
   switch (ndim) {
     case 1:
@@ -105,7 +101,7 @@ void ternary_op_dispatch_dims(
   ContiguousIterator b_it(shape, b_strides, ndim - 2);
   ContiguousIterator c_it(shape, c_strides, ndim - 2);
   auto stride = out_strides[ndim - 3];
-  for (size_t elem = 0; elem < a.size(); elem += stride) {
+  for (size_t elem = 0; elem < size; elem += stride) {
     ternary_op_dims<T1, T2, T3, U, Op, 2>(
         a_ptr + a_it.loc,
         b_ptr + b_it.loc,
@@ -130,18 +126,16 @@ void ternary_op(
     const array& b,
     const array& c,
     array& out,
-    Op op) {
-  TernaryOpType topt = get_ternary_op_type(a, b, c);
-  set_ternary_op_output_data(a, b, c, out, topt);
+    Op op,
+    TernaryOpType topt) {
+  const T1* a_ptr = a.data<T1>();
+  const T2* b_ptr = b.data<T2>();
+  const T3* c_ptr = c.data<T3>();
+  U* out_ptr = out.data<U>();
 
-  // The full computation is scalar-scalar-scalar so we call the base op once.
   if (topt == TernaryOpType::ScalarScalarScalar) {
-    *(out.data<U>()) = op(*a.data<T1>(), *b.data<T2>(), *c.data<T3>());
+    *out_ptr = op(*a_ptr, *b_ptr, *c_ptr);
   } else if (topt == TernaryOpType::VectorVectorVector) {
-    const T1* a_ptr = a.data<T1>();
-    const T2* b_ptr = b.data<T2>();
-    const T3* c_ptr = c.data<T3>();
-    U* out_ptr = out.data<U>();
     for (size_t i = 0; i < out.size(); ++i) {
       *out_ptr = op(*a_ptr, *b_ptr, *c_ptr);
       a_ptr++;
@@ -150,7 +144,10 @@ void ternary_op(
       out_ptr++;
     }
   } else {
-    ternary_op_dispatch_dims<T1, T2, T3, U>(a, b, c, out, op);
+    auto [shape, strides] = collapse_contiguous_dims(
+        a.shape(), {a.strides(), b.strides(), c.strides(), out.strides()});
+    ternary_op_dispatch_dims<T1, T2, T3, U>(
+        a_ptr, b_ptr, c_ptr, out_ptr, op, out.size(), shape, strides);
   }
 }
 

mlx/backend/cpu/unary.cpp

@@ -14,88 +14,57 @@ void Abs::eval_cpu(const std::vector<array>& inputs, array& out) {
     // No-op for unsigned types
     out.copy_shared_buffer(in);
   } else {
-    auto op = detail::Abs{};
-    switch (out.dtype()) {
-      case int8:
-        unary_op<int8_t>(in, out, op);
-        break;
-      case int16:
-        unary_op<int16_t>(in, out, op);
-        break;
-      case int32:
-        unary_op<int32_t>(in, out, op);
-        break;
-      case int64:
-        unary_op<int64_t>(in, out, op);
-        break;
-      case float16:
-        unary_op<float16_t>(in, out, op);
-        break;
-      case float32:
-        unary_op<float>(in, out, op);
-        break;
-      case float64:
-        unary_op<double>(in, out, op);
-        break;
-      case bfloat16:
-        unary_op<bfloat16_t>(in, out, op);
-        break;
-      case complex64:
-        unary_op<complex64_t>(in, out, op);
-        break;
-      default:
-        throw std::runtime_error("[Abs] Called on unsigned type");
-    }
+    unary_signed(in, out, detail::Abs(), stream());
   }
 }
 
 void ArcCos::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::ArcCos());
+  unary_fp(in, out, detail::ArcCos(), stream());
 }
 
 void ArcCosh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::ArcCosh());
+  unary_fp(in, out, detail::ArcCosh(), stream());
 }
 
 void ArcSin::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::ArcSin());
+  unary_fp(in, out, detail::ArcSin(), stream());
 }
 
 void ArcSinh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::ArcSinh());
+  unary_fp(in, out, detail::ArcSinh(), stream());
 }
 
 void ArcTan::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::ArcTan());
+  unary_fp(in, out, detail::ArcTan(), stream());
 }
 
 void ArcTanh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::ArcTanh());
+  unary_fp(in, out, detail::ArcTanh(), stream());
 }
 
 void BitwiseInvert::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_int(in, out, detail::BitwiseInvert());
+  unary_int(in, out, detail::BitwiseInvert(), stream());
 }
 
 void Ceil::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   if (issubdtype(in.dtype(), inexact)) {
-    unary_fp(in, out, detail::Ceil());
+    unary_fp(in, out, detail::Ceil(), stream());
   } else {
     // No-op integer types
     out.copy_shared_buffer(in);
@@ -104,84 +73,50 @@ void Ceil::eval_cpu(const std::vector<array>& inputs, array& out) {
 
 void Conjugate::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  unary_op<complex64_t>(inputs[0], out, detail::Conjugate());
+  unary_complex(inputs[0], out, detail::Conjugate(), stream());
 }
 
 void Cos::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Cos());
+  unary_fp(in, out, detail::Cos(), stream());
 }
 
 void Cosh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Cosh());
+  unary_fp(in, out, detail::Cosh(), stream());
 }
 
 void Erf::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  switch (out.dtype()) {
-    case float32:
-      unary_op<float>(in, out, detail::Erf());
-      break;
-    case float16:
-      unary_op<float16_t>(in, out, detail::Erf());
-      break;
-    case float64:
-      unary_op<double>(in, out, detail::Erf());
-      break;
-    case bfloat16:
-      unary_op<bfloat16_t>(in, out, detail::Erf());
-      break;
-    default:
-      throw std::invalid_argument(
-          "[erf] Error function only defined for arrays"
-          " with real floating point type.");
-  }
+  unary_real_fp(in, out, detail::Erf(), stream());
 }
 
 void ErfInv::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  switch (out.dtype()) {
-    case float32:
-      unary_op<float>(in, out, detail::ErfInv());
-      break;
-    case float16:
-      unary_op<float16_t>(in, out, detail::ErfInv());
-      break;
-    case float64:
-      unary_op<double>(in, out, detail::ErfInv());
-      break;
-    case bfloat16:
-      unary_op<bfloat16_t>(in, out, detail::ErfInv());
-      break;
-    default:
-      throw std::invalid_argument(
-          "[erf_inv] Inverse error function only defined for arrays"
-          " with real floating point type.");
-  }
+  unary_real_fp(in, out, detail::ErfInv(), stream());
 }
 
 void Exp::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Exp());
+  unary_fp(in, out, detail::Exp(), stream());
 }
 
 void Expm1::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Expm1());
+  unary_fp(in, out, detail::Expm1(), stream());
 }
 
 void Floor::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   if (issubdtype(in.dtype(), inexact)) {
-    unary_fp(in, out, detail::Floor());
+    unary_fp(in, out, detail::Floor(), stream());
   } else {
     // No-op integer types
     out.copy_shared_buffer(in);
@@ -189,7 +124,7 @@ void Floor::eval_cpu(const std::vector<array>& inputs, array& out) {
 }
 
 void Imag::eval_cpu(const std::vector<array>& inputs, array& out) {
-  unary_op<complex64_t, float>(inputs[0], out, detail::Imag());
+  unary_complex_to_float(inputs[0], out, detail::Imag(), stream());
 }
 
 void Log::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -197,13 +132,13 @@ void Log::eval_cpu(const std::vector<array>& inputs, array& out) {
   const auto& in = inputs[0];
   switch (base_) {
     case Base::e:
-      unary_fp(in, out, detail::Log());
+      unary_fp(in, out, detail::Log(), stream());
       break;
     case Base::two:
-      unary_fp(in, out, detail::Log2());
+      unary_fp(in, out, detail::Log2(), stream());
       break;
     case Base::ten:
-      unary_fp(in, out, detail::Log10());
+      unary_fp(in, out, detail::Log10(), stream());
       break;
   }
 }
@@ -211,30 +146,30 @@ void Log::eval_cpu(const std::vector<array>& inputs, array& out) {
 void Log1p::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Log1p());
+  unary_fp(in, out, detail::Log1p(), stream());
 }
 
 void LogicalNot::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  unary(in, out, detail::LogicalNot());
+  unary(in, out, detail::LogicalNot(), stream());
 }
 
 void Negative::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  unary(in, out, detail::Negative());
+  unary(in, out, detail::Negative(), stream());
 }
 
 void Real::eval_cpu(const std::vector<array>& inputs, array& out) {
-  unary_op<complex64_t, float>(inputs[0], out, detail::Real());
+  unary_complex_to_float(inputs[0], out, detail::Real(), stream());
 }
 
 void Round::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   if (issubdtype(in.dtype(), inexact)) {
-    unary_fp(in, out, detail::Round());
+    unary_fp(in, out, detail::Round(), stream());
   } else {
     // No-op integer types
     out.copy_shared_buffer(in);
@@ -244,7 +179,7 @@ void Round::eval_cpu(const std::vector<array>& inputs, array& out) {
 void Sigmoid::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Sigmoid());
+  unary_fp(in, out, detail::Sigmoid(), stream());
 }
 
 void Sign::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -253,48 +188,48 @@ void Sign::eval_cpu(const std::vector<array>& inputs, array& out) {
   if (in.dtype() == bool_) {
     out.copy_shared_buffer(in);
   } else {
-    unary(in, out, detail::Sign());
+    unary(in, out, detail::Sign(), stream());
   }
 }
 
 void Sin::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Sin());
+  unary_fp(in, out, detail::Sin(), stream());
 }
 
 void Sinh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Sinh());
+  unary_fp(in, out, detail::Sinh(), stream());
 }
 
 void Square::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  unary(in, out, detail::Square());
+  unary(in, out, detail::Square(), stream());
 }
 
 void Sqrt::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   if (recip_) {
-    unary_fp(in, out, detail::Rsqrt());
+    unary_fp(in, out, detail::Rsqrt(), stream());
   } else {
-    unary_fp(in, out, detail::Sqrt());
+    unary_fp(in, out, detail::Sqrt(), stream());
   }
 }
 
 void Tan::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Tan());
+  unary_fp(in, out, detail::Tan(), stream());
 }
 
 void Tanh::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   const auto& in = inputs[0];
-  unary_fp(in, out, detail::Tanh());
+  unary_fp(in, out, detail::Tanh(), stream());
 }
 
 } // namespace mlx::core

mlx/backend/cpu/unary.h

@@ -5,174 +5,296 @@
 #include "mlx/allocator.h"
 #include "mlx/array.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
 #include "mlx/utils.h"
 
 namespace mlx::core {
 
 void set_unary_output_data(const array& in, array& out) {
-  if (is_donatable(in, out)) {
-    out.copy_shared_buffer(in);
+  if (in.flags().contiguous) {
+    if (is_donatable(in, out)) {
+      out.copy_shared_buffer(in);
+    } else {
+      auto size = in.data_size();
+      out.set_data(
+          allocator::malloc_or_wait(size * out.itemsize()),
+          size,
+          in.strides(),
+          in.flags());
+    }
   } else {
-    auto size = in.data_size();
-    out.set_data(
-        allocator::malloc_or_wait(size * out.itemsize()),
-        size,
-        in.strides(),
-        in.flags());
+    out.set_data(allocator::malloc_or_wait(out.nbytes()));
   }
 }
 
 template <typename T, typename U = T, typename Op>
-void unary_op(const T* a, U* out, Op op, size_t shape, size_t stride) {
+void unary_op(const T* a, U* out, size_t shape, size_t stride) {
   for (size_t i = 0; i < shape; i += 1) {
-    out[i] = op(*a);
+    out[i] = Op{}(*a);
     a += stride;
   }
 }
 
 template <typename T, typename U = T, typename Op>
-void unary_op(const array& a, array& out, Op op) {
-  const T* a_ptr = a.data<T>();
+void unary_op(const array& a, array& out, Op) {
+  const T* src = a.data<T>();
+  U* dst = out.data<U>();
+  auto ndim = a.ndim();
   if (a.flags().contiguous) {
-    set_unary_output_data(a, out);
-    U* dst = out.data<U>();
+    auto size = a.data_size();
     constexpr int N = simd::max_size<T>;
-    size_t size = a.data_size();
     while (size >= N) {
-      simd::store(dst, op(simd::load<T, N>(a_ptr)));
+      simd::store(dst, Op{}(simd::load<T, N>(src)));
       size -= N;
-      a_ptr += N;
+      src += N;
       dst += N;
     }
     while (size > 0) {
-      *dst = op(*a_ptr);
+      *dst = Op{}(*src);
       size--;
       dst++;
-      a_ptr++;
+      src++;
     }
   } else {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-    U* dst = out.data<U>();
-    size_t shape = a.ndim() > 0 ? a.shape(-1) : 1;
-    size_t stride = a.ndim() > 0 ? a.strides(-1) : 1;
-    if (a.ndim() <= 1) {
-      unary_op(a_ptr, dst, op, shape, stride);
+    size_t shape = ndim > 0 ? a.shape().back() : 1;
+    size_t stride = ndim > 0 ? a.strides().back() : 1;
+    if (ndim <= 1) {
+      unary_op<T, U, Op>(src, dst, shape, stride);
       return;
     }
-    ContiguousIterator it(a.shape(), a.strides(), a.ndim() - 1);
+    auto it = ContiguousIterator(a.shape(), a.strides(), ndim - 1);
     for (size_t elem = 0; elem < a.size(); elem += shape) {
-      unary_op(a_ptr + it.loc, dst + elem, op, shape, stride);
+      unary_op<T, U, Op>(src + it.loc, dst + elem, shape, stride);
       it.step();
     }
   }
 }
 
 template <typename Op>
-void unary(const array& a, array& out, Op op) {
-  switch (out.dtype()) {
-    case bool_:
-      unary_op<bool>(a, out, op);
-      break;
-    case uint8:
-      unary_op<uint8_t>(a, out, op);
-      break;
-    case uint16:
-      unary_op<uint16_t>(a, out, op);
-      break;
-    case uint32:
-      unary_op<uint32_t>(a, out, op);
-      break;
-    case uint64:
-      unary_op<uint64_t>(a, out, op);
-      break;
-    case int8:
-      unary_op<int8_t>(a, out, op);
-      break;
-    case int16:
-      unary_op<int16_t>(a, out, op);
-      break;
-    case int32:
-      unary_op<int32_t>(a, out, op);
-      break;
-    case int64:
-      unary_op<int64_t>(a, out, op);
-      break;
-    case float16:
-      unary_op<float16_t>(a, out, op);
-      break;
-    case float32:
-      unary_op<float>(a, out, op);
-      break;
-    case float64:
-      unary_op<double>(a, out, op);
-      break;
-    case bfloat16:
-      unary_op<bfloat16_t>(a, out, op);
-      break;
-    case complex64:
-      unary_op<complex64_t>(a, out, op);
-      break;
-  }
+void unary(const array& a, array& out, Op op, Stream stream) {
+  set_unary_output_data(a, out);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    out = array::unsafe_weak_copy(out),
+                    op = op]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        unary_op<bool>(a, out, op);
+        break;
+      case uint8:
+        unary_op<uint8_t>(a, out, op);
+        break;
+      case uint16:
+        unary_op<uint16_t>(a, out, op);
+        break;
+      case uint32:
+        unary_op<uint32_t>(a, out, op);
+        break;
+      case uint64:
+        unary_op<uint64_t>(a, out, op);
+        break;
+      case int8:
+        unary_op<int8_t>(a, out, op);
+        break;
+      case int16:
+        unary_op<int16_t>(a, out, op);
+        break;
+      case int32:
+        unary_op<int32_t>(a, out, op);
+        break;
+      case int64:
+        unary_op<int64_t>(a, out, op);
+        break;
+      case float16:
+        unary_op<float16_t>(a, out, op);
+        break;
+      case float32:
+        unary_op<float>(a, out, op);
+        break;
+      case float64:
+        unary_op<double>(a, out, op);
+        break;
+      case bfloat16:
+        unary_op<bfloat16_t>(a, out, op);
+        break;
+      case complex64:
+        unary_op<complex64_t>(a, out, op);
+        break;
+    }
+  });
 }
 
 template <typename Op>
-void unary_fp(const array& a, array& out, Op op) {
-  switch (out.dtype()) {
-    case bfloat16:
-      unary_op<bfloat16_t>(a, out, op);
-      break;
-    case float16:
-      unary_op<float16_t>(a, out, op);
-      break;
-    case float32:
-      unary_op<float>(a, out, op);
-      break;
-    case float64:
-      unary_op<double>(a, out, op);
-      break;
-    case complex64:
-      unary_op<complex64_t>(a, out, op);
-      break;
-    default:
-      std::ostringstream err;
-      err << "[unary_fp] Does not support " << out.dtype();
-      throw std::runtime_error(err.str());
-  }
+void unary_real_fp(const array& a, array& out, Op op, Stream stream) {
+  set_unary_output_data(a, out);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    out = array::unsafe_weak_copy(out),
+                    op = op]() mutable {
+    switch (out.dtype()) {
+      case bfloat16:
+        unary_op<bfloat16_t>(a, out, op);
+        break;
+      case float16:
+        unary_op<float16_t>(a, out, op);
+        break;
+      case float32:
+        unary_op<float>(a, out, op);
+        break;
+      case float64:
+        unary_op<double>(a, out, op);
+        break;
+      default:
+        std::ostringstream err;
+        err << "[unary_real] Does not support " << out.dtype();
+        throw std::runtime_error(err.str());
+    }
+  });
+}
+template <typename Op>
+void unary_fp(const array& a, array& out, Op op, Stream stream) {
+  set_unary_output_data(a, out);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    out = array::unsafe_weak_copy(out),
+                    op = op]() mutable {
+    switch (out.dtype()) {
+      case bfloat16:
+        unary_op<bfloat16_t>(a, out, op);
+        break;
+      case float16:
+        unary_op<float16_t>(a, out, op);
+        break;
+      case float32:
+        unary_op<float>(a, out, op);
+        break;
+      case float64:
+        unary_op<double>(a, out, op);
+        break;
+      case complex64:
+        unary_op<complex64_t>(a, out, op);
+        break;
+      default:
+        std::ostringstream err;
+        err << "[unary_fp] Does not support " << out.dtype();
+        throw std::runtime_error(err.str());
+    }
+  });
 }
 
 template <typename Op>
-void unary_int(const array& a, array& out, Op op) {
-  switch (out.dtype()) {
-    case uint8:
-      unary_op<uint8_t>(a, out, op);
-      break;
-    case uint16:
-      unary_op<uint16_t>(a, out, op);
-      break;
-    case uint32:
-      unary_op<uint32_t>(a, out, op);
-      break;
-    case uint64:
-      unary_op<uint64_t>(a, out, op);
-      break;
-    case int8:
-      unary_op<int8_t>(a, out, op);
-      break;
-    case int16:
-      unary_op<int16_t>(a, out, op);
-      break;
-    case int32:
-      unary_op<int32_t>(a, out, op);
-      break;
-    case int64:
-      unary_op<int64_t>(a, out, op);
-      break;
-    default:
-      std::ostringstream err;
-      err << "[unary_int] Does not support " << out.dtype();
-      throw std::runtime_error(err.str());
-  }
+void unary_signed(const array& a, array& out, Op op, Stream stream) {
+  set_unary_output_data(a, out);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    out = array::unsafe_weak_copy(out),
+                    op = op]() mutable {
+    switch (out.dtype()) {
+      case int8:
+        unary_op<int8_t>(a, out, op);
+        break;
+      case int16:
+        unary_op<int16_t>(a, out, op);
+        break;
+      case int32:
+        unary_op<int32_t>(a, out, op);
+        break;
+      case int64:
+        unary_op<int64_t>(a, out, op);
+        break;
+      case float16:
+        unary_op<float16_t>(a, out, op);
+        break;
+      case float32:
+        unary_op<float>(a, out, op);
+        break;
+      case float64:
+        unary_op<double>(a, out, op);
+        break;
+      case bfloat16:
+        unary_op<bfloat16_t>(a, out, op);
+        break;
+      case complex64:
+        unary_op<complex64_t>(a, out, op);
+        break;
+      default:
+        throw std::runtime_error("[Abs] Called on unsigned type");
+    }
+  });
+}
+
+template <typename Op>
+void unary_complex(const array& a, array& out, Op op, Stream stream) {
+  set_unary_output_data(a, out);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    out = array::unsafe_weak_copy(out),
+                    op = op]() mutable { unary_op<complex64_t>(a, out, op); });
+}
+
+template <typename Op>
+void unary_complex_to_float(const array& a, array& out, Op op, Stream stream) {
+  set_unary_output_data(a, out);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(out);
+  encoder.dispatch(
+      [a = array::unsafe_weak_copy(a),
+       out = array::unsafe_weak_copy(out),
+       op = op]() mutable { unary_op<complex64_t, float>(a, out, op); });
+}
+
+template <typename Op>
+void unary_int(const array& a, array& out, Op op, Stream stream) {
+  set_unary_output_data(a, out);
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    out = array::unsafe_weak_copy(out),
+                    op = op]() mutable {
+    switch (out.dtype()) {
+      case uint8:
+        unary_op<uint8_t>(a, out, op);
+        break;
+      case uint16:
+        unary_op<uint16_t>(a, out, op);
+        break;
+      case uint32:
+        unary_op<uint32_t>(a, out, op);
+        break;
+      case uint64:
+        unary_op<uint64_t>(a, out, op);
+        break;
+      case int8:
+        unary_op<int8_t>(a, out, op);
+        break;
+      case int16:
+        unary_op<int16_t>(a, out, op);
+        break;
+      case int32:
+        unary_op<int32_t>(a, out, op);
+        break;
+      case int64:
+        unary_op<int64_t>(a, out, op);
+        break;
+      default:
+        std::ostringstream err;
+        err << "[unary_int] Does not support " << out.dtype();
+        throw std::runtime_error(err.str());
+    }
+  });
 }
 
 } // namespace mlx::core

mlx/backend/metal/binary.cpp

@@ -102,16 +102,9 @@ void binary_op_gpu_inplace(
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(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.
-  // - If there is only one output only one of a and b will be donated.
-  bool donate_a = a.data_shared_ptr() == nullptr;
-  bool donate_b = b.data_shared_ptr() == nullptr;
   int arg_idx = 0;
-  compute_encoder.set_input_array(donate_a ? outputs[0] : a, arg_idx++);
-  compute_encoder.set_input_array(
-      donate_b ? (donate_a ? outputs[1] : outputs[0]) : b, arg_idx++);
+  compute_encoder.set_input_array(a, arg_idx++);
+  compute_encoder.set_input_array(b, arg_idx++);
   compute_encoder.set_output_array(outputs[0], arg_idx++);
   if (outputs.size() == 2) {
     compute_encoder.set_output_array(outputs[1], arg_idx++);
@@ -164,8 +157,8 @@ void binary_op_gpu(
   auto& a = inputs[0];
   auto& b = inputs[1];
   auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, outputs[0], bopt, true);
-  set_binary_op_output_data(a, b, outputs[1], bopt, true);
+  set_binary_op_output_data(a, b, outputs[0], bopt);
+  set_binary_op_output_data(a, b, outputs[1], bopt);
   binary_op_gpu_inplace(inputs, outputs, op, s);
 }
 
@@ -195,7 +188,7 @@ void binary_op_gpu(
   auto& a = inputs[0];
   auto& b = inputs[1];
   auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt, true);
+  set_binary_op_output_data(a, b, out, bopt);
   binary_op_gpu_inplace(inputs, out, op, s);
 }
 

mlx/backend/metal/compiled.cpp

@@ -457,7 +457,7 @@ void Compiled::eval_gpu(
   }
 
   compiled_allocate_outputs(
-      inputs, outputs, inputs_, constant_ids_, contiguous, true);
+      inputs, outputs, inputs_, constant_ids_, contiguous);
 
   // Put the outputs in
   for (auto& x : outputs) {

mlx/backend/metal/copy.cpp

@@ -14,25 +14,11 @@ namespace mlx::core {
 constexpr int MAX_COPY_SPECIALIZED_DIMS = 3;
 
 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()),
-          in.data_size(),
-          in.strides(),
-          in.flags());
-    }
-  } else {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
+  bool donated = set_copy_output_data(in, out, ctype);
+  if (donated && in.dtype() == out.dtype()) {
+    // If the output has the same type as the input then there is nothing to
+    // copy, just use the buffer.
+    return;
   }
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;

mlx/backend/metal/device.cpp

@@ -168,9 +168,10 @@ void CommandEncoder::set_output_array(
   register_output_array(a);
 }
 
-void CommandEncoder::register_output_array(array& a) {
+void CommandEncoder::register_output_array(const array& a) {
   all_outputs_.insert(a.buffer().ptr());
-  auto buf = static_cast<MTL::Resource*>(a.buffer().ptr());
+
+  auto buf = static_cast<MTL::Resource*>(const_cast<void*>(a.buffer().ptr()));
   if (concurrent_) {
     concurrent_outputs_.insert(buf);
   } else {

mlx/backend/metal/device.h

@@ -62,7 +62,7 @@ struct CommandEncoder {
 
   void set_input_array(const array& a, int idx, int64_t offset = 0);
   void set_output_array(array& a, int idx, int64_t offset = 0);
-  void register_output_array(array& a);
+  void register_output_array(const array& a);
   void dispatch_threadgroups(MTL::Size grid_dims, MTL::Size group_dims);
   void dispatch_threads(MTL::Size grid_dims, MTL::Size group_dims);
   void maybeInsertBarrier();

mlx/backend/metal/distributed.cpp

@@ -4,149 +4,30 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/event.h"
-#include "mlx/backend/metal/fence.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/distributed/ops.h"
 #include "mlx/distributed/primitives.h"
+#include "mlx/fence.h"
 #include "mlx/scheduler.h"
 
 namespace mlx::core::distributed {
 
-void signal_and_wait(const Event& e_signal, const Event& e_wait) {
-  if (e_signal.valid()) {
-    encode_signal(e_signal);
-  }
-  encode_wait(e_wait);
+void AllReduce::eval_gpu(const std::vector<array>&, std::vector<array>&) {
+  throw std::runtime_error("[AllReduce::eval_gpu] has no GPU implementation.");
 }
 
-void AllReduce::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 1);
-  assert(outputs.size() == 1);
-
-  auto& in = inputs[0];
-
-  Fence f{stream()};
-
-  if (in.event().valid()) {
-    f.update_gpu(in);
-  }
-
-  auto& out = outputs[0];
-  if (in.is_donatable()) {
-    out.move_shared_buffer(in);
-  } else {
-    out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  }
-  f.wait_gpu(out);
-
-  auto task = [in = in,
-               out = unsafe_weak_copy(out),
-               f = std::move(f),
-               reduce_type = reduce_type_,
-               group = group()]() mutable {
-    if (in.event().valid()) {
-      f.wait();
-    }
-    switch (reduce_type) {
-      case Sum:
-        distributed::detail::all_sum(
-            group, in.data_shared_ptr() == nullptr ? out : in, out);
-        break;
-      default:
-        throw std::runtime_error("Only all reduce sum is supported for now");
-    }
-    f.update();
-  };
-  scheduler::enqueue(detail::communication_stream(), std::move(task));
+void AllGather::eval_gpu(const std::vector<array>&, std::vector<array>&) {
+  throw std::runtime_error("[AllGather::eval_gpu] has no GPU implementation.");
 }
 
-void AllGather::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 1);
-  assert(outputs.size() == 1);
-  auto& in = inputs[0];
-  auto& out = outputs[0];
-
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  Fence f{stream()};
-
-  if (in.event().valid()) {
-    f.update_gpu(in);
-  }
-  f.wait_gpu(out);
-
-  auto task = [in = in,
-               out = unsafe_weak_copy(out),
-               f = std::move(f),
-               group = group()]() mutable {
-    if (in.event().valid()) {
-      f.wait();
-    }
-    distributed::detail::all_gather(group, in, out);
-    f.update();
-  };
-  scheduler::enqueue(detail::communication_stream(), std::move(task));
+void Send::eval_gpu(const std::vector<array>&, std::vector<array>&) {
+  throw std::runtime_error("[Send::eval_gpu] has no GPU implementation.");
 }
 
-void Send::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 1);
-  assert(outputs.size() == 1);
-
-  auto& in = inputs[0];
-
-  // Encode a signal event for the input
-  Fence f{stream()};
-  if (in.event().valid()) {
-    f.update_gpu(in);
-  }
-
-  auto& out = outputs[0];
-  move_or_copy(in, out);
-
-  // Schedule an async send on the comm stream
-  auto task = [in = in,
-               out = unsafe_weak_copy(out),
-               f = std::move(f),
-               group = group(),
-               dst = dst_]() mutable {
-    if (in.event().valid()) {
-      f.wait();
-    }
-    distributed::detail::send(group, out, dst);
-  };
-  scheduler::enqueue(detail::communication_stream(), std::move(task));
-}
-
-void Recv::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 0);
-  assert(outputs.size() == 1);
-
-  auto& out = outputs[0];
-
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-
-  Fence f{stream()};
-  f.wait_gpu(out);
-
-  // Schedule an async recv on the comm stream
-  auto task = [out = unsafe_weak_copy(out),
-               f = std::move(f),
-               group = group(),
-               src = src_]() mutable {
-    distributed::detail::recv(group, out, src);
-    f.update();
-  };
-  scheduler::enqueue(detail::communication_stream(), std::move(task));
+void Recv::eval_gpu(const std::vector<array>&, std::vector<array>&) {
+  throw std::runtime_error("[Recv::eval_gpu] has no GPU implementation.");
 }
 
 } // namespace mlx::core::distributed

mlx/backend/metal/event.cpp

@@ -3,52 +3,58 @@
 #include "mlx/event.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/metal_impl.h"
+#include "mlx/scheduler.h"
 
 namespace mlx::core {
 
-void encode_wait(Event e) {
-  auto& d = metal::device(e.stream().device);
-  d.end_encoding(e.stream().index);
-  auto command_buffer = d.get_command_buffer(e.stream().index);
-  command_buffer->encodeWait(
-      static_cast<MTL::Event*>(e.raw_event().get()), e.value());
-  command_buffer->addCompletedHandler(
-      [e = std::move(e)](MTL::CommandBuffer* cbuf) {});
-}
-
-void encode_signal(Event e) {
-  auto& d = metal::device(e.stream().device);
-  d.end_encoding(e.stream().index);
-  auto command_buffer = d.get_command_buffer(e.stream().index);
-  command_buffer->encodeSignalEvent(
-      static_cast<MTL::Event*>(e.raw_event().get()), e.value());
-  command_buffer->addCompletedHandler(
-      [e = std::move(e)](MTL::CommandBuffer* cbuf) {});
-}
-
-Event::Event(const Stream& stream) : stream_(stream) {
+Event::Event(Stream stream) : stream_(stream) {
   auto dtor = [](void* ptr) {
     auto p = metal::new_scoped_memory_pool();
     static_cast<MTL::SharedEvent*>(ptr)->release();
   };
   auto p = metal::new_scoped_memory_pool();
   event_ = std::shared_ptr<void>(
-      metal::device(stream.device).mtl_device()->newSharedEvent(), dtor);
+      metal::device(Device::gpu).mtl_device()->newSharedEvent(), dtor);
 }
 
 void Event::wait() {
-  if (!static_cast<MTL::SharedEvent*>(raw_event().get())
+  if (!static_cast<MTL::SharedEvent*>(event_.get())
            ->waitUntilSignaledValue(value(), -1)) {
     throw std::runtime_error("[Event::wait] Timed out");
   }
 }
 
 void Event::signal() {
-  static_cast<MTL::SharedEvent*>(raw_event().get())->setSignaledValue(value());
+  static_cast<MTL::SharedEvent*>(event_.get())->setSignaledValue(value());
+}
+
+void Event::wait(Stream stream) {
+  if (stream.device == Device::cpu) {
+    scheduler::enqueue(stream, [*this]() mutable { wait(); });
+  } else {
+    auto& d = metal::device(stream.device);
+    d.end_encoding(stream.index);
+    auto command_buffer = d.get_command_buffer(stream.index);
+    command_buffer->encodeWait(static_cast<MTL::Event*>(event_.get()), value());
+    command_buffer->addCompletedHandler([*this](MTL::CommandBuffer*) {});
+  }
+}
+
+void Event::signal(Stream stream) {
+  if (stream.device == Device::cpu) {
+    scheduler::enqueue(stream, [*this]() mutable { signal(); });
+  } else {
+    auto& d = metal::device(stream.device);
+    d.end_encoding(stream.index);
+    auto command_buffer = d.get_command_buffer(stream.index);
+    command_buffer->encodeSignalEvent(
+        static_cast<MTL::Event*>(event_.get()), value());
+    command_buffer->addCompletedHandler([*this](MTL::CommandBuffer*) {});
+  }
 }
 
 bool Event::is_signaled() const {
-  return static_cast<MTL::SharedEvent*>(raw_event().get())->signaledValue() >=
+  return static_cast<MTL::SharedEvent*>(event_.get())->signaledValue() >=
       value();
 }
 

mlx/backend/metal/event.h

@@ -1,10 +0,0 @@
-// Copyright © 2024 Apple Inc.
-#pragma once
-
-namespace mlx::core {
-
-void encode_wait(Event e);
-
-void encode_signal(Event e);
-
-} // namespace mlx::core

mlx/backend/metal/fence.cpp

@@ -1,49 +1,81 @@
 // Copyright © 2024 Apple Inc.
-
-#include "mlx/backend/metal/fence.h"
+#include "mlx/fence.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/event.h"
 #include "mlx/backend/metal/metal_impl.h"
+#include "mlx/scheduler.h"
 #include "mlx/utils.h"
 
 namespace mlx::core {
 
-Fence::Fence(const Stream& stream) : stream_(stream) {
-  auto d = metal::device(stream.device).mtl_device();
-  if (!d->supportsFamily(MTL::GPUFamilyMetal3)) {
-    use_fast_ = false;
-  } else if (__builtin_available(macOS 15, iOS 18, *)) {
-    use_fast_ = env::metal_fast_synch();
-  }
-  if (!use_fast_) {
-    // Wraps Metal SharedEvent
-    auto dtor = [](void* ptr) {
+struct FenceImpl {
+  FenceImpl() {
+    auto d = metal::device(Device::gpu).mtl_device();
+    if (!d->supportsFamily(MTL::GPUFamilyMetal3)) {
+      use_fast = false;
+    } else if (__builtin_available(macOS 15, iOS 18, *)) {
+      use_fast = env::metal_fast_synch();
+    }
+
+    if (!use_fast) {
       auto p = metal::new_scoped_memory_pool();
-      static_cast<MTL::SharedEvent*>(ptr)->release();
-    };
-    auto p = metal::new_scoped_memory_pool();
-    fence_ = std::shared_ptr<void>(
-        metal::device(stream.device).mtl_device()->newSharedEvent(), dtor);
-  } else {
-    auto dtor = [](void* buf) {
-      allocator::free(static_cast<MTL::Buffer*>(buf));
-    };
-    auto buf = allocator::malloc_or_wait(sizeof(uint32_t)).ptr();
-    fence_ = std::shared_ptr<void>(buf, dtor);
-    cpu_value()[0] = 0;
+      fence = static_cast<void*>(d->newSharedEvent());
+    } else {
+      auto buf = allocator::malloc_or_wait(sizeof(uint32_t)).ptr();
+      fence = static_cast<void*>(buf);
+      cpu_value()[0] = 0;
+    }
   }
+
+  ~FenceImpl() {
+    if (!use_fast) {
+      // Wraps Metal SharedEvent
+      auto p = metal::new_scoped_memory_pool();
+      static_cast<MTL::SharedEvent*>(fence)->release();
+    } else {
+      allocator::free(static_cast<MTL::Buffer*>(fence));
+    }
+  }
+  bool use_fast{false};
+  uint32_t count{0};
+  void* fence;
+
+  std::atomic_uint* cpu_value() {
+    return static_cast<std::atomic_uint*>(
+        static_cast<MTL::Buffer*>(fence)->contents());
+  }
+};
+
+Fence::Fence(Stream) {
+  auto dtor = [](void* ptr) { delete static_cast<FenceImpl*>(ptr); };
+  fence_ = std::shared_ptr<void>(new FenceImpl{}, dtor);
 }
 
-void Fence::wait_gpu(array& x) {
-  gpu_count_++;
-  auto& d = metal::device(stream_.device);
-  auto idx = stream_.index;
+void Fence::wait(Stream stream, const array& x) {
+  auto& f = *static_cast<FenceImpl*>(fence_.get());
 
-  if (!use_fast_) {
+  if (stream.device == Device::cpu) {
+    scheduler::enqueue(stream, [fence_ = fence_, count = f.count]() mutable {
+      auto& f = *static_cast<FenceImpl*>(fence_.get());
+      if (!f.use_fast) {
+        if (!static_cast<MTL::SharedEvent*>(f.fence)->waitUntilSignaledValue(
+                count, -1)) {
+          throw std::runtime_error("[Fence::wait] Timed out");
+        }
+        return;
+      }
+      while (f.cpu_value()[0] < count) {
+      }
+    });
+    return;
+  }
+
+  auto& d = metal::device(stream.device);
+  auto idx = stream.index;
+
+  if (!f.use_fast) {
     d.end_encoding(idx);
     auto command_buffer = d.get_command_buffer(idx);
-    command_buffer->encodeWait(
-        static_cast<MTL::Event*>(fence_.get()), gpu_count_);
+    command_buffer->encodeWait(static_cast<MTL::Event*>(f.fence), f.count);
     command_buffer->addCompletedHandler(
         [fence_ = fence_](MTL::CommandBuffer* cbuf) {});
     return;
@@ -51,7 +83,7 @@ void Fence::wait_gpu(array& x) {
 
   auto& compute_encoder = d.get_command_encoder(idx);
 
-  // Register the output to ensure that no kernels which depends on the
+  // Register outputs to ensure that no kernels which depends on the
   // output starts before this one is done
   compute_encoder.register_output_array(x);
 
@@ -59,36 +91,49 @@ void Fence::wait_gpu(array& x) {
   MTL::Size kernel_dims = MTL::Size(1, 1, 1);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  auto buf = static_cast<MTL::Buffer*>(fence_.get());
+  auto buf = static_cast<MTL::Buffer*>(f.fence);
   compute_encoder.set_buffer(buf, 0);
-  compute_encoder.set_bytes(gpu_count_, 1);
+  compute_encoder.set_bytes(f.count, 1);
   compute_encoder.dispatch_threads(kernel_dims, kernel_dims);
 
   d.get_command_buffer(idx)->addCompletedHandler(
-      [fence = fence_](MTL::CommandBuffer* cbuf) {});
+      [fence_ = fence_](MTL::CommandBuffer* cbuf) {});
 }
 
-void Fence::update_gpu(const array& x) {
-  gpu_count_++;
-  auto& d = metal::device(stream_.device);
-  auto idx = stream_.index;
+void Fence::update(Stream stream, const array& x) {
+  auto& f = *static_cast<FenceImpl*>(fence_.get());
+  f.count++;
 
-  if (!use_fast_) {
+  if (stream.device == Device::cpu) {
+    scheduler::enqueue(stream, [fence_ = fence_, count = f.count]() mutable {
+      auto& f = *static_cast<FenceImpl*>(fence_.get());
+      if (!f.use_fast) {
+        static_cast<MTL::SharedEvent*>(f.fence)->setSignaledValue(count);
+        return;
+      }
+
+      f.cpu_value()[0] = count;
+    });
+    return;
+  }
+
+  auto& d = metal::device(stream.device);
+  auto idx = stream.index;
+  if (!f.use_fast) {
     d.end_encoding(idx);
     auto command_buffer = d.get_command_buffer(idx);
     command_buffer->encodeSignalEvent(
-        static_cast<MTL::Event*>(fence_.get()), gpu_count_);
+        static_cast<MTL::Event*>(f.fence), f.count);
     command_buffer->addCompletedHandler(
         [fence_ = fence_](MTL::CommandBuffer* cbuf) {});
     return;
   }
 
-  // Launch input visibility kernel
+  // Launch input visibility kernels
   auto& compute_encoder = d.get_command_encoder(idx);
   auto kernel = d.get_kernel("input_coherent");
   uint32_t nthreads =
       (x.data_size() * x.itemsize() + sizeof(uint32_t) - 1) / sizeof(uint32_t);
-
   MTL::Size group_dims = MTL::Size(1024, 1, 1);
   MTL::Size grid_dims = MTL::Size((nthreads + 1024 - 1) / 1024, 1, 1);
   compute_encoder.set_compute_pipeline_state(kernel);
@@ -96,7 +141,7 @@ void Fence::update_gpu(const array& x) {
   compute_encoder.set_bytes(nthreads, 1);
   compute_encoder.dispatch_threadgroups(group_dims, grid_dims);
 
-  // Barrier on previous kernel
+  // Barrier on previous kernels
   compute_encoder.barrier();
 
   // Launch value update kernel
@@ -104,42 +149,13 @@ void Fence::update_gpu(const array& x) {
   MTL::Size kernel_dims = MTL::Size(1, 1, 1);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  auto buf = static_cast<MTL::Buffer*>(fence_.get());
+  auto buf = static_cast<MTL::Buffer*>(f.fence);
   compute_encoder.set_buffer(buf, 0);
-  compute_encoder.set_bytes(gpu_count_, 1);
+  compute_encoder.set_bytes(f.count, 1);
   compute_encoder.dispatch_threads(kernel_dims, kernel_dims);
 
   d.get_command_buffer(idx)->addCompletedHandler(
-      [fence = fence_](MTL::CommandBuffer* cbuf) {});
-}
-
-void Fence::wait() {
-  cpu_count_++;
-  if (!use_fast_) {
-    if (!static_cast<MTL::SharedEvent*>(fence_.get())
-             ->waitUntilSignaledValue(cpu_count_, -1)) {
-      throw std::runtime_error("[Event::wait] Timed out");
-    }
-    return;
-  }
-  while (cpu_value()[0] < cpu_count_) {
-  }
-}
-
-void Fence::update() {
-  cpu_count_++;
-
-  if (!use_fast_) {
-    static_cast<MTL::SharedEvent*>(fence_.get())->setSignaledValue(cpu_count_);
-    return;
-  }
-
-  cpu_value()[0] = cpu_count_;
-}
-
-std::atomic_uint* Fence::cpu_value() {
-  return static_cast<std::atomic_uint*>(
-      static_cast<MTL::Buffer*>(fence_.get())->contents());
+      [fence_ = fence_](MTL::CommandBuffer* cbuf) {});
 }
 
 } // namespace mlx::core

mlx/backend/metal/fence.h

@@ -1,40 +0,0 @@
-// Copyright © 2024 Apple Inc.
-
-#include "mlx/backend/metal/device.h"
-
-namespace mlx::core {
-
-/* A fence to be used for synchronizing work between the CPU and GPU
- *
- * Calls to `update_gpu` should be paired with calls to `wait`. This ensures
- * that the array passed to `update_gpu` is computed and visible to the CPU
- * after the call to `wait` returns.
- *
- * Calls to `update` should be paired with calls to `wait_gpu`. This ensures
- * that the array passed to `wait_gpu` will not be read by the GPU until the CPU
- * has called `update`.
- *
- * The fence supports slow (default) and fast mode. Fast mode requires setting
- * the environment variable `MLX_METAL_FAST_SYNCH=1`. Fast mode also requires
- * Metal 3.2+ (macOS 15+, iOS 18+).
- */
-class Fence {
- public:
-  Fence(const Stream& stream);
-
-  void update_gpu(const array& x);
-  void wait_gpu(array& x);
-
-  void wait();
-  void update();
-
- private:
-  Stream stream_;
-  std::shared_ptr<void> fence_;
-  uint32_t cpu_count_{0};
-  uint32_t gpu_count_{0};
-  bool use_fast_;
-  std::atomic_uint* cpu_value();
-};
-
-} // namespace mlx::core

mlx/backend/metal/hadamard.cpp

@@ -82,7 +82,7 @@ void Hadamard::eval_gpu(const std::vector<array>& inputs, array& out) {
   const array& in_contiguous = check_input(in);
 
   if (in_contiguous.is_donatable()) {
-    out.move_shared_buffer(in_contiguous);
+    out.copy_shared_buffer(in_contiguous);
   } else {
     out.set_data(allocator::malloc_or_wait(out.nbytes()));
   }

mlx/backend/metal/jit_kernels.cpp

@@ -176,8 +176,13 @@ MTL::ComputePipelineState* get_copy_kernel(
     auto out_type = get_type_string(out.dtype());
     kernel_source +=
         get_template_definition("s_" + lib_name, "copy_s", in_type, out_type);
+    kernel_source +=
+        get_template_definition("s2_" + lib_name, "copy_s2", in_type, out_type);
     kernel_source +=
         get_template_definition("v_" + lib_name, "copy_v", in_type, out_type);
+    kernel_source +=
+        get_template_definition("v2_" + lib_name, "copy_v2", in_type, out_type);
+
     kernel_source += get_template_definition(
         "g1_" + lib_name, "copy_g_nd1", in_type, out_type, "int");
     kernel_source += get_template_definition(

mlx/backend/metal/kernels/gemv.metal

@@ -23,7 +23,8 @@ template <
     const int SN, /* Simdgroup cols (in threads) */
     const int TM, /* Thread rows (in elements) */
     const int TN, /* Thread cols (in elements) */
-    const bool kDoAxpby> /* Do out = alpha * out + beta * bias */
+    const bool kDoAxpby, /* Do out = alpha * out + beta * bias */
+    typename AccT = float>
 struct GEMVKernel {
   MLX_MTL_CONST int threadsM = BM * SM;
   MLX_MTL_CONST int threadsN = BN * SN;
@@ -60,28 +61,32 @@ struct GEMVKernel {
   MLX_MTL_CONST short tgp_mem_size = BN > 1 ? BN*(blockM + TM) : 0;
   MLX_MTL_CONST bool needs_tgp_reduction = BN > 1;
 
+  template <typename U = T>
   static METAL_FUNC void
-  load_unsafe(const device T* src, thread T dst[TN], const int src_offset = 0) {
+  load_unsafe(const device T* src, thread U dst[TN], const int src_offset = 0) {
     MLX_MTL_PRAGMA_UNROLL
     for (int tn = 0; tn < TN; tn++) {
-      dst[tn] = src[src_offset + tn];
+      dst[tn] = static_cast<U>(src[src_offset + tn]);
     }
   }
 
+  template <typename U = T>
   static METAL_FUNC void load_safe(
       const device T* src,
-      thread T dst[TN],
+      thread U dst[TN],
       const int src_offset = 0,
       const int src_size = TN) {
     if (src_offset + TN <= src_size) {
       MLX_MTL_PRAGMA_UNROLL
       for (int tn = 0; tn < TN; tn++) {
-        dst[tn] = src[src_offset + tn];
+        dst[tn] = static_cast<U>(src[src_offset + tn]);
       }
     } else { // Edgecase
       MLX_MTL_PRAGMA_UNROLL
       for (int tn = 0; tn < TN; tn++) {
-        dst[tn] = src_offset + tn < src_size ? src[src_offset + tn] : 0;
+        dst[tn] = src_offset + tn < src_size
+            ? static_cast<U>(src[src_offset + tn])
+            : U(0);
       }
     }
   }
@@ -97,7 +102,7 @@ struct GEMVKernel {
       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)]],
+      threadgroup AccT* tgp_memory [[threadgroup(0)]],
       uint3 tid [[threadgroup_position_in_grid]],
       uint3 lid [[thread_position_in_threadgroup]],
       uint simd_gid [[simdgroup_index_in_threadgroup]],
@@ -106,9 +111,9 @@ struct GEMVKernel {
     (void)lid;
 
     // Thread local accumulation results
-    thread T result[TM] = {0};
+    thread AccT result[TM] = {0};
     thread T inter[TN];
-    thread T v_coeff[TN];
+    thread AccT v_coeff[TN];
 
     const int thrM = SN != 32 ? simd_lid / SN : 0;
     const int thrN = SN != 32 ? simd_lid % SN : int(simd_lid);
@@ -142,7 +147,7 @@ struct GEMVKernel {
 
     // Loop over in_vec in blocks of blockN
     for (int i = 0; i < n_iter; ++i) {
-      load_unsafe(in_vec, v_coeff, bn);
+      load_unsafe<AccT>(in_vec, v_coeff, bn);
 
       // Per thread work loop
       int mat_offset = 0;
@@ -164,7 +169,7 @@ struct GEMVKernel {
     }
 
     if (leftover > 0) {
-      load_safe(in_vec, v_coeff, bn, in_size);
+      load_safe<AccT>(in_vec, v_coeff, bn, in_size);
 
       // Per thread work loop
       MLX_MTL_PRAGMA_UNROLL
@@ -191,7 +196,7 @@ struct GEMVKernel {
 
     // Threadgroup accumulation results
     if (needs_tgp_reduction) {
-      threadgroup T* tgp_results = tgp_memory + sgN * (blockM + TM) + bm;
+      threadgroup AccT* tgp_results = tgp_memory + sgN * (blockM + TM) + bm;
       if (thrN == 0) {
         MLX_MTL_PRAGMA_UNROLL
         for (int tm = 0; tm < TM; tm++) {
@@ -217,10 +222,11 @@ struct GEMVKernel {
       MLX_MTL_PRAGMA_UNROLL
       for (int tm = 0; tm < TM; tm++) {
         if (kDoAxpby) {
-          out_vec[out_row + tm] = static_cast<T>(alpha) * result[tm] +
+          out_vec[out_row + tm] =
+              static_cast<T>(alpha) * static_cast<T>(result[tm]) +
               static_cast<T>(beta) * bias[(out_row + tm) * bias_stride];
         } else {
-          out_vec[out_row + tm] = result[tm];
+          out_vec[out_row + tm] = static_cast<T>(result[tm]);
         }
       }
     }
@@ -239,7 +245,8 @@ template <
     const int SN, /* Simdgroup cols (in threads) */
     const int TM, /* Thread rows (in elements) */
     const int TN, /* Thread cols (in elements) */
-    const bool kDoAxpby> /* Do out = alpha * out + beta * bias */
+    const bool kDoAxpby, /* Do out = alpha * out + beta * bias */
+    typename AccT = float>
 struct GEMVTKernel {
   MLX_MTL_CONST int threadsM = BM * SM;
   MLX_MTL_CONST int threadsN = BN * SN;
@@ -282,7 +289,7 @@ struct GEMVTKernel {
       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)]],
+      threadgroup AccT* tgp_memory [[threadgroup(0)]],
       uint3 tid [[threadgroup_position_in_grid]],
       uint3 lid [[thread_position_in_threadgroup]],
       uint simd_gid [[simdgroup_index_in_threadgroup]],
@@ -291,10 +298,9 @@ struct GEMVTKernel {
     (void)lid;
 
     // Thread local accumulation results
-    T result[TN] = {0};
+    AccT result[TN] = {0};
     T inter[TN];
-    T v_coeff[TM];
-
+    AccT v_coeff[TM];
     const int thrM = SN != 32 ? simd_lid / SN : 0;
     const int thrN = SN != 32 ? simd_lid % SN : int(simd_lid);
 
@@ -330,16 +336,17 @@ struct GEMVTKernel {
 
         MLX_MTL_PRAGMA_UNROLL
         for (int tm = 0; tm < TM; tm++) {
-          v_coeff[tm] = in_vec[bm + tm];
+          v_coeff[tm] = static_cast<AccT>(in_vec[bm + tm]);
         }
 
         MLX_MTL_PRAGMA_UNROLL
         for (int tm = 0; tm < TM; tm++) {
+          auto vc = float(v_coeff[tm]);
           for (int tn = 0; tn < TN; 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];
+            result[tn] += vc * inter[tn];
           }
         }
 
@@ -348,7 +355,7 @@ struct GEMVTKernel {
 
       if (leftover > 0) {
         for (int tm = 0; tm < TM && bm + tm < in_vec_size; tm++) {
-          v_coeff[tm] = in_vec[bm + tm];
+          v_coeff[tm] = static_cast<AccT>(in_vec[bm + tm]);
 
           MLX_MTL_PRAGMA_UNROLL
           for (int tn = 0; tn < TN; tn++) {
@@ -374,7 +381,7 @@ struct GEMVTKernel {
 
     // Threadgroup accumulation results
     if (needs_tgp_reduction) {
-      threadgroup T* tgp_results = tgp_memory + sgM * (blockN + TN) + bn;
+      threadgroup AccT* tgp_results = tgp_memory + sgM * (blockN + TN) + bn;
       if (thrM == 0) {
         MLX_MTL_PRAGMA_UNROLL
         for (int tn = 0; tn < TN; tn++) {
@@ -400,10 +407,11 @@ struct GEMVTKernel {
       MLX_MTL_PRAGMA_UNROLL
       for (int j = 0; j < TN; j++) {
         if (kDoAxpby) {
-          out_vec[out_col + j] = static_cast<T>(alpha) * result[j] +
+          out_vec[out_col + j] =
+              static_cast<T>(alpha) * static_cast<T>(result[j]) +
               static_cast<T>(beta) * bias[(out_col + j) * bias_stride];
         } else {
-          out_vec[out_col + j] = result[j];
+          out_vec[out_col + j] = static_cast<T>(result[j]);
         }
       }
     }
@@ -445,7 +453,7 @@ template <
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel = GEMVKernel<T, BM, BN, SM, SN, TM, TN, kDoAxpby>;
-  threadgroup T tgp_memory
+  threadgroup float tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   // Update batch offsets
@@ -553,7 +561,7 @@ template <
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel = GEMVKernel<T, BM, BN, SM, SN, TM, TN, false>;
-  threadgroup T tgp_memory
+  threadgroup float tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   uint32_t indx_vec;
@@ -660,7 +668,7 @@ template <
     uint simd_gid [[simdgroup_index_in_threadgroup]],
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel = GEMVTKernel<T, BM, BN, SM, SN, TM, TN, kDoAxpby>;
-  threadgroup T tgp_memory
+  threadgroup float tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   // Update batch offsets
@@ -761,8 +769,8 @@ template <
     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, SM, SN, TM, TN, false>;
-  threadgroup T tgp_memory
+  using gemv_kernel = GEMVTKernel<T, BM, BN, SM, SN, TM, TN, false, float>;
+  threadgroup float tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   uint32_t indx_vec;

mlx/backend/metal/kernels/gemv_masked.h

@@ -44,7 +44,8 @@ template <
     const int SM, /* Simdgroup rows (in threads) */
     const int SN, /* Simdgroup cols (in threads) */
     const int TM, /* Thread rows (in elements) */
-    const int TN> /* Thread cols (in elements) */
+    const int TN, /* Thread cols (in elements) */
+    typename AccT = float>
 struct GEMVKernel {
   MLX_MTL_CONST int threadsM = BM * SM;
   MLX_MTL_CONST int threadsN = BN * SN;
@@ -91,28 +92,32 @@ struct GEMVKernel {
   MLX_MTL_CONST short tgp_mem_size = BN > 1 ? BN*(blockM + TM) : 0;
   MLX_MTL_CONST bool needs_tgp_reduction = BN > 1;
 
+  template <typename U = T>
   static METAL_FUNC void
-  load_unsafe(const device T* src, thread T dst[TN], const int src_offset = 0) {
+  load_unsafe(const device T* src, thread U dst[TN], const int src_offset = 0) {
     MLX_MTL_PRAGMA_UNROLL
     for (int tn = 0; tn < TN; tn++) {
-      dst[tn] = src[src_offset + tn];
+      dst[tn] = static_cast<U>(src[src_offset + tn]);
     }
   }
 
+  template <typename U = T>
   static METAL_FUNC void load_safe(
       const device T* src,
-      thread T dst[TN],
+      thread U dst[TN],
       const int src_offset = 0,
       const int src_size = TN) {
     if (src_offset + TN <= src_size) {
       MLX_MTL_PRAGMA_UNROLL
       for (int tn = 0; tn < TN; tn++) {
-        dst[tn] = src[src_offset + tn];
+        dst[tn] = static_cast<U>(src[src_offset + tn]);
       }
     } else { // Edgecase
       MLX_MTL_PRAGMA_UNROLL
       for (int tn = 0; tn < TN; tn++) {
-        dst[tn] = src_offset + tn < src_size ? src[src_offset + tn] : 0;
+        dst[tn] = src_offset + tn < src_size
+            ? static_cast<U>(src[src_offset + tn])
+            : U(0);
       }
     }
   }
@@ -128,7 +133,7 @@ struct GEMVKernel {
       const device op_mask_t* mat_mask [[buffer(21)]],
       const device op_mask_t* vec_mask [[buffer(22)]],
       const constant int* mask_strides [[buffer(23)]],
-      threadgroup T* tgp_memory [[threadgroup(0)]],
+      threadgroup AccT* tgp_memory [[threadgroup(0)]],
       uint3 tid [[threadgroup_position_in_grid]],
       uint3 lid [[thread_position_in_threadgroup]],
       uint simd_gid [[simdgroup_index_in_threadgroup]],
@@ -137,9 +142,9 @@ struct GEMVKernel {
     (void)lid;
 
     // Thread local accumulation results
-    thread T result[TM] = {0};
+    thread AccT result[TM] = {0};
     thread T inter[TN];
-    thread T v_coeff[TN];
+    thread AccT v_coeff[TN];
 
     const int thrM = SN != 32 ? simd_lid / SN : 0;
     const int thrN = SN != 32 ? simd_lid % SN : int(simd_lid);
@@ -225,7 +230,7 @@ struct GEMVKernel {
               T(mat_mask[mat_mask_offset]) * T(vec_mask[vec_mask_offset]);
         }
 
-        load_unsafe(in_vec, v_coeff, bn);
+        load_unsafe<AccT>(in_vec, v_coeff, bn);
 
         // Apply scale
         if (has_mul_operand_mask) {
@@ -267,7 +272,7 @@ struct GEMVKernel {
             T(mat_mask[mat_mask_offset]) * T(vec_mask[vec_mask_offset]);
       }
 
-      load_safe(in_vec, v_coeff, bn, in_size);
+      load_safe<AccT>(in_vec, v_coeff, bn, in_size);
 
       // Apply scale
       if (has_mul_operand_mask) {
@@ -310,7 +315,7 @@ struct GEMVKernel {
 
     // Threadgroup accumulation results
     if (needs_tgp_reduction) {
-      threadgroup T* tgp_results = tgp_memory + sgN * (blockM + TM) + bm;
+      threadgroup AccT* tgp_results = tgp_memory + sgN * (blockM + TM) + bm;
       if (thrN == 0) {
         MLX_MTL_PRAGMA_UNROLL
         for (int tm = 0; tm < TM; tm++) {
@@ -335,7 +340,7 @@ struct GEMVKernel {
     if (simdN == 0 && thrN == 0) {
       MLX_MTL_PRAGMA_UNROLL
       for (int tm = 0; tm < TM; tm++) {
-        out_vec[out_row + tm] = result[tm];
+        out_vec[out_row + tm] = static_cast<T>(result[tm]);
       }
     }
   }
@@ -354,7 +359,8 @@ template <
     const int SM, /* Simdgroup rows (in threads) */
     const int SN, /* Simdgroup cols (in threads) */
     const int TM, /* Thread rows (in elements) */
-    const int TN> /* Thread cols (in elements) */
+    const int TN, /* Thread cols (in elements) */
+    typename AccT = float>
 struct GEMVTKernel {
   MLX_MTL_CONST int threadsM = BM * SM;
   MLX_MTL_CONST int threadsN = BN * SN;
@@ -405,7 +411,7 @@ struct GEMVTKernel {
       const device op_mask_t* mat_mask [[buffer(21)]],
       const device op_mask_t* vec_mask [[buffer(22)]],
       const constant int* mask_strides [[buffer(23)]],
-      threadgroup T* tgp_memory [[threadgroup(0)]],
+      threadgroup AccT* tgp_memory [[threadgroup(0)]],
       uint3 tid [[threadgroup_position_in_grid]],
       uint3 lid [[thread_position_in_threadgroup]],
       uint simd_gid [[simdgroup_index_in_threadgroup]],
@@ -414,9 +420,9 @@ struct GEMVTKernel {
     (void)lid;
 
     // Thread local accumulation results
-    T result[TN] = {0};
+    AccT result[TN] = {0};
     T inter[TN];
-    T v_coeff[TM];
+    AccT v_coeff[TM];
 
     const int thrM = SN != 32 ? simd_lid / SN : 0;
     const int thrN = SN != 32 ? simd_lid % SN : int(simd_lid);
@@ -511,7 +517,7 @@ struct GEMVTKernel {
 
           MLX_MTL_PRAGMA_UNROLL
           for (int tm = 0; tm < TM; tm++) {
-            v_coeff[tm] = in_vec[bm + tm];
+            v_coeff[tm] = static_cast<AccT>(in_vec[bm + tm]);
           }
 
           // Apply scale
@@ -549,7 +555,7 @@ struct GEMVTKernel {
         }
 
         for (int tm = 0; tm < TM && bm + tm < in_vec_size; tm++) {
-          v_coeff[tm] = in_vec[bm + tm];
+          v_coeff[tm] = static_cast<AccT>(in_vec[bm + tm]);
 
           if (has_mul_operand_mask) {
             v_coeff[tm] *= block_scale;
@@ -587,7 +593,7 @@ struct GEMVTKernel {
 
     // Threadgroup accumulation results
     if (needs_tgp_reduction) {
-      threadgroup T* tgp_results = tgp_memory + sgM * (blockN + TN) + bn;
+      threadgroup AccT* tgp_results = tgp_memory + sgM * (blockN + TN) + bn;
       if (thrM == 0) {
         MLX_MTL_PRAGMA_UNROLL
         for (int tn = 0; tn < TN; tn++) {
@@ -612,7 +618,7 @@ struct GEMVTKernel {
     if (cm == 0 && out_col < out_vec_size) {
       MLX_MTL_PRAGMA_UNROLL
       for (int j = 0; j < TN; j++) {
-        out_vec[out_col + j] = result[j];
+        out_vec[out_col + j] = static_cast<T>(result[j]);
       }
     }
   }
@@ -655,7 +661,7 @@ template <
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel =
       GEMVKernel<T, out_mask_t, op_mask_t, BM, BN, SM, SN, TM, TN>;
-  threadgroup T tgp_memory
+  threadgroup float tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   constexpr bool has_operand_mask = !metal::is_same_v<op_mask_t, nomask_t>;
@@ -755,7 +761,7 @@ template <
     uint simd_lid [[thread_index_in_simdgroup]]) {
   using gemv_kernel =
       GEMVTKernel<T, out_mask_t, op_mask_t, BM, BN, SM, SN, TM, TN>;
-  threadgroup T tgp_memory
+  threadgroup float tgp_memory
       [gemv_kernel::tgp_mem_size == 0 ? 1 : gemv_kernel::tgp_mem_size];
 
   constexpr bool has_operand_mask = !metal::is_same_v<op_mask_t, nomask_t>;

mlx/backend/metal/kernels/quantized.h

@@ -2015,9 +2015,9 @@ template <typename T, const int group_size, const int bits>
     device T* biases [[buffer(3)]],
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  constexpr T eps = T(1e-7);
+  constexpr float eps = 1e-7;
   constexpr int simd_size = 32;
-  constexpr T n_bins = (1 << bits) - 1;
+  constexpr float n_bins = (1 << bits) - 1;
   constexpr int packs_per_int = bits == 3 ? 8 : bits == 6 ? 4 : 8 / bits;
   constexpr int values_per_reduce = group_size / simd_size;
   constexpr int writes_per_reduce = packs_per_int / values_per_reduce;
@@ -2036,13 +2036,13 @@ template <typename T, const int group_size, const int bits>
       ? offset * writes_per_pack
       : offset * bytes_per_pack / writes_per_reduce;
 
-  T w_thread[values_per_reduce];
-  T w_min = Limits<T>::max;
-  T w_max = 0;
+  float w_thread[values_per_reduce];
+  float w_min = Limits<T>::max;
+  float w_max = 0;
 
 #pragma clang loop unroll(full)
   for (int i = 0; i < values_per_reduce; i++) {
-    T val = w[in_index + i];
+    float val = w[in_index + i];
     w_thread[i] = val;
     w_min = min(w_min, val);
     w_max = max(w_max, val);
@@ -2051,20 +2051,20 @@ template <typename T, const int group_size, const int bits>
   w_min = simd_min(w_min);
   w_max = simd_max(w_max);
 
-  T scale = max((w_max - w_min) / n_bins, eps);
+  float scale = max((w_max - w_min) / n_bins, eps);
   bool side = abs(w_min) > abs(w_max);
   scale = side ? scale : -scale;
-  T edge = side ? w_min : w_max;
-  T q0 = round(edge / scale);
+  float edge = side ? w_min : w_max;
+  float q0 = round(edge / scale);
   bool at_zero = q0 == 0.0f;
   scale = at_zero ? scale : edge / q0;
-  T bias = at_zero ? T(0) : edge;
+  float bias = at_zero ? 0 : edge;
 
   // Write out the scales and biases
   size_t gindex = in_index / group_size;
   if (in_index % group_size == 0) {
-    scales[gindex] = scale;
-    biases[gindex] = bias;
+    scales[gindex] = static_cast<T>(scale);
+    biases[gindex] = static_cast<T>(bias);
   }
 
   // We accumulate 3 bytes worth for 3/6 bit so we need a uint32_t

mlx/backend/metal/kernels/sdpa_vector.h

@@ -15,8 +15,10 @@ template <typename T, int D, int V = D>
     device T* out [[buffer(3)]],
     const constant int& gqa_factor,
     const constant int& N,
-    const constant size_t& k_stride,
-    const constant size_t& v_stride,
+    const constant size_t& k_head_stride,
+    const constant size_t& k_seq_stride,
+    const constant size_t& v_head_stride,
+    const constant size_t& v_seq_stride,
     const constant float& scale,
     const device bool* mask [[function_constant(has_mask)]],
     const constant int& mask_kv_seq_stride [[function_constant(has_mask)]],
@@ -30,8 +32,8 @@ template <typename T, int D, int V = D>
   constexpr int BD = 32;
   constexpr int qk_per_thread = D / BD;
   constexpr int v_per_thread = V / BD;
-  constexpr int inner_k_stride = BN * D;
-  constexpr int inner_v_stride = BN * V;
+  int inner_k_stride = BN * int(k_seq_stride);
+  int inner_v_stride = BN * int(v_seq_stride);
 
   typedef float U;
 
@@ -51,8 +53,10 @@ template <typename T, int D, int V = D>
   const int q_offset =
       query_transposed ? o_offset : head_idx * tpg.y + q_seq_idx;
   queries += q_offset * D + simd_lid * qk_per_thread;
-  keys += kv_head_idx * k_stride + simd_gid * D + simd_lid * qk_per_thread;
-  values += kv_head_idx * v_stride + simd_gid * V + simd_lid * v_per_thread;
+  keys += kv_head_idx * k_head_stride + simd_gid * k_seq_stride +
+      simd_lid * qk_per_thread;
+  values += kv_head_idx * v_head_stride + simd_gid * v_seq_stride +
+      simd_lid * v_per_thread;
   if (has_mask) {
     mask += head_idx * mask_head_stride + simd_gid * mask_kv_seq_stride +
         q_seq_idx * mask_q_seq_stride;
@@ -147,8 +151,10 @@ template <typename T, int D, int V = D>
     device float* maxs [[buffer(5)]],
     const constant int& gqa_factor,
     const constant int& N,
-    const constant size_t& k_stride,
-    const constant size_t& v_stride,
+    const constant size_t& k_head_stride,
+    const constant size_t& k_seq_stride,
+    const constant size_t& v_head_stride,
+    const constant size_t& v_seq_stride,
     const constant float& scale,
     const device bool* mask [[function_constant(has_mask)]],
     const constant int& mask_kv_seq_stride [[function_constant(has_mask)]],
@@ -162,8 +168,8 @@ template <typename T, int D, int V = D>
   constexpr int BD = 32;
   constexpr int qk_per_thread = D / BD;
   constexpr int v_per_thread = V / BD;
-  constexpr int inner_k_stride = BN * D;
-  constexpr int inner_v_stride = BN * V;
+  int inner_k_stride = BN * int(k_seq_stride);
+  int inner_v_stride = BN * int(v_seq_stride);
   constexpr int blocks = 32;
 
   typedef float U;
@@ -186,10 +192,10 @@ template <typename T, int D, int V = D>
   const int kv_head_idx = head_idx / gqa_factor;
 
   queries += q_offset * D + simd_lid * qk_per_thread;
-  keys += kv_head_idx * k_stride + (block_idx * BN + simd_gid) * D +
-      simd_lid * qk_per_thread;
-  values += kv_head_idx * v_stride + (block_idx * BN + simd_gid) * V +
-      simd_lid * v_per_thread;
+  keys += kv_head_idx * k_head_stride +
+      (block_idx * BN + simd_gid) * k_seq_stride + simd_lid * qk_per_thread;
+  values += kv_head_idx * v_head_stride +
+      (block_idx * BN + simd_gid) * v_seq_stride + simd_lid * v_per_thread;
   out += o_offset * blocks * V + block_idx * V + simd_lid * v_per_thread;
   if (has_mask) {
     mask += head_idx * mask_head_stride +

mlx/backend/metal/metal.cpp

@@ -2,7 +2,6 @@
 #include <memory>
 
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/event.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/primitives.h"
 #include "mlx/scheduler.h"
@@ -23,88 +22,78 @@ inline void check_error(MTL::CommandBuffer* cbuf) {
   }
 }
 
-std::function<void()> make_task(array arr, bool signal) {
-  auto task = [arr = std::move(arr), signal]() mutable {
-    auto pool = new_scoped_memory_pool();
-    auto s = arr.primitive().stream();
-    auto& d = metal::device(s.device);
-    auto command_buffer = d.get_command_buffer(s.index);
+void eval(array& arr) {
+  auto pool = new_scoped_memory_pool();
+  auto s = arr.primitive().stream();
+  auto& d = metal::device(s.device);
+  auto command_buffer = d.get_command_buffer(s.index);
 
-    for (auto& input : arr.inputs()) {
-      if (input.event().valid() &&
-          input.event().stream() != arr.primitive().stream()) {
-        input.event().wait();
-      }
+  auto outputs = arr.outputs();
+  {
+    // If the array is a tracer hold a reference
+    // to its inputs so they don't get donated
+    std::vector<array> inputs;
+    if (arr.is_tracer()) {
+      inputs = arr.inputs();
     }
 
-    auto outputs = arr.outputs();
-    {
-      // If the array is a tracer hold a reference
-      // to its inputs so they don't get donated
-      std::vector<array> inputs;
-      if (arr.is_tracer()) {
-        inputs = arr.inputs();
-      }
+    debug_set_primitive_buffer_label(command_buffer, arr.primitive());
+    arr.primitive().eval_gpu(arr.inputs(), outputs);
+  }
+  std::unordered_set<std::shared_ptr<array::Data>> buffers;
+  for (auto& in : arr.inputs()) {
+    buffers.insert(in.data_shared_ptr());
+  }
+  for (auto& s : arr.siblings()) {
+    buffers.insert(s.data_shared_ptr());
+  }
+  // Remove the output if it was donated to by an input
+  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+    buffers.erase(it);
+  }
 
-      debug_set_primitive_buffer_label(command_buffer, arr.primitive());
-      try {
-        arr.primitive().eval_gpu(arr.inputs(), outputs);
-      } catch (const std::exception& error) {
-        abort_with_exception(error);
-      }
-    }
-    std::vector<std::shared_ptr<array::Data>> buffers;
-    for (auto& in : arr.inputs()) {
-      buffers.push_back(in.data_shared_ptr());
-    }
-    for (auto& s : arr.siblings()) {
-      buffers.push_back(s.data_shared_ptr());
-    }
-    if (!arr.is_tracer()) {
-      arr.detach();
-    }
-    for (auto& out : outputs) {
-      out.set_status(array::Status::evaluated);
-    }
-
-    if (signal || d.command_buffer_needs_commit(s.index)) {
-      if (signal) {
-        encode_signal(arr.event());
-      }
-      d.end_encoding(s.index);
-      scheduler::notify_new_task(s);
-      command_buffer->addCompletedHandler(
-          [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
-            scheduler::notify_task_completion(s);
-            check_error(cbuf);
-          });
-      d.commit_command_buffer(s.index);
-      d.get_command_buffer(s.index);
-    } else {
-      command_buffer->addCompletedHandler(
-          [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
-            check_error(cbuf);
-          });
-    }
-  };
-  return task;
+  if (d.command_buffer_needs_commit(s.index)) {
+    d.end_encoding(s.index);
+    scheduler::notify_new_task(s);
+    command_buffer->addCompletedHandler(
+        [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
+          scheduler::notify_task_completion(s);
+          check_error(cbuf);
+        });
+    d.commit_command_buffer(s.index);
+    d.get_command_buffer(s.index);
+  } else {
+    command_buffer->addCompletedHandler(
+        [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
+          check_error(cbuf);
+        });
+  }
 }
 
-std::function<void()> make_synchronize_task(
-    Stream s,
-    std::shared_ptr<std::promise<void>> p) {
-  return [s, p = std::move(p)]() {
-    auto pool = new_scoped_memory_pool();
-    auto& d = metal::device(s.device);
-    auto cb = d.get_command_buffer(s.index);
-    cb->retain();
-    d.end_encoding(s.index);
-    d.commit_command_buffer(s.index);
-    cb->waitUntilCompleted();
-    check_error(cb);
-    cb->release();
-    p->set_value();
-  };
+void finalize(Stream s) {
+  auto pool = new_scoped_memory_pool();
+  auto& d = metal::device(s.device);
+  auto cb = d.get_command_buffer(s.index);
+  d.end_encoding(s.index);
+  scheduler::notify_new_task(s);
+  cb->addCompletedHandler([s](MTL::CommandBuffer* cbuf) {
+    scheduler::notify_task_completion(s);
+    check_error(cbuf);
+  });
+  d.commit_command_buffer(s.index);
+  d.get_command_buffer(s.index);
+}
+
+void synchronize(Stream s) {
+  auto pool = new_scoped_memory_pool();
+  auto& d = metal::device(s.device);
+  auto cb = d.get_command_buffer(s.index);
+  cb->retain();
+  d.end_encoding(s.index);
+  d.commit_command_buffer(s.index);
+  cb->waitUntilCompleted();
+  check_error(cb);
+  cb->release();
 }
 
 void start_capture(std::string path, id object) {

mlx/backend/metal/metal_impl.h

@@ -14,10 +14,8 @@ void new_stream(Stream stream);
 
 std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool();
 
-std::function<void()> make_task(array arr, bool signal);
-
-std::function<void()> make_synchronize_task(
-    Stream s,
-    std::shared_ptr<std::promise<void>> p);
+void eval(array& arr);
+void finalize(Stream s);
+void synchronize(Stream s);
 
 } // namespace mlx::core::metal

mlx/backend/metal/normalization.cpp

@@ -18,33 +18,33 @@ void RMSNorm::eval_gpu(
   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& {
+  auto set_output = [&s, &out](const array& x) {
     bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
     if (no_copy && x.ndim() > 1) {
       auto s = x.strides()[x.ndim() - 2];
       no_copy &= (s == 0 || s == x.shape().back());
     }
     if (no_copy) {
+      if (x.is_donatable()) {
+        out.copy_shared_buffer(x);
+      } else {
+        out.set_data(
+            allocator::malloc_or_wait(x.data_size() * x.itemsize()),
+            x.data_size(),
+            x.strides(),
+            x.flags());
+      }
       return x;
     } else {
-      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
-      copy_gpu(x, copies.back(), CopyType::General, s);
-      return copies.back();
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy_gpu(x, x_copy, CopyType::General, s);
+      out.copy_shared_buffer(x_copy);
+      return x_copy;
     }
   };
-  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());
-  }
+  const array x = set_output(inputs[0]);
+  const array& w = inputs[1];
 
   auto axis_size = static_cast<uint32_t>(x.shape().back());
   int n_rows = x.data_size() / axis_size;
@@ -88,8 +88,6 @@ void RMSNorm::eval_gpu(
     compute_encoder.set_bytes(w_stride, 5);
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   }
-
-  d.add_temporaries(std::move(copies), s.index);
 }
 
 void RMSNormVJP::eval_gpu(
@@ -101,20 +99,22 @@ void RMSNormVJP::eval_gpu(
   // 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.
-  auto check_input = [&d, &s](const array& x) -> array {
+  auto check_input = [&d, &s](const array& x) -> std::pair<array, bool> {
     if (x.flags().row_contiguous) {
-      return x;
+      return {x, false};
     }
 
     array x_copy(x.shape(), x.dtype(), nullptr, {});
     copy_gpu(x, x_copy, CopyType::General, s);
-    d.add_temporary(x_copy, s.index);
-
-    return x_copy;
+    return {x_copy, true};
   };
-  const array& x = check_input(inputs[0]);
+  bool donate_x = inputs[0].is_donatable();
+  bool donate_g = inputs[2].is_donatable();
+  auto [x, copied] = check_input(inputs[0]);
+  donate_x |= copied;
   const array& w = inputs[1];
-  const array& g = check_input(inputs[2]);
+  auto [g, g_copied] = check_input(inputs[2]);
+  donate_g |= g_copied;
   array& gx = outputs[0];
   array& gw = outputs[1];
 
@@ -122,17 +122,18 @@ void RMSNormVJP::eval_gpu(
   bool has_w = w.ndim() != 0;
 
   // 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;
+    gx.copy_shared_buffer(x);
   } else if (g.is_donatable()) {
-    gx.move_shared_buffer(g);
+    gx.copy_shared_buffer(g);
     g_in_gx = true;
   } else {
     gx.set_data(allocator::malloc_or_wait(gx.nbytes()));
   }
+  if (g_copied && !g_in_gx) {
+    d.add_temporary(g, s.index);
+  }
 
   auto axis_size = static_cast<uint32_t>(x.shape().back());
   int n_rows = x.data_size() / axis_size;
@@ -141,11 +142,9 @@ void RMSNormVJP::eval_gpu(
   // gradients before they are accumulated.
   array gw_temp =
       (has_w) ? array({n_rows, x.shape().back()}, gw.dtype(), nullptr, {}) : w;
-  bool g_in_gw = false;
   if (has_w) {
-    if (!g_in_gx && g.is_donatable()) {
-      gw_temp.move_shared_buffer(g);
-      g_in_gw = true;
+    if (!g_in_gx && donate_g) {
+      gw_temp.copy_shared_buffer(g);
     } else {
       gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes()));
       d.add_temporary(gw_temp, s.index);
@@ -189,9 +188,9 @@ void RMSNormVJP::eval_gpu(
 
     uint32_t w_stride = (w.ndim() == 1) ? w.strides()[0] : 0;
     compute_encoder.set_compute_pipeline_state(kernel);
-    compute_encoder.set_input_array(x_in_gx ? gx : x, 0);
+    compute_encoder.set_input_array(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_input_array(g, 2);
     compute_encoder.set_output_array(gx, 3);
     compute_encoder.set_output_array(gw_temp, 4);
     compute_encoder.set_bytes(eps_, 5);
@@ -216,35 +215,35 @@ void LayerNorm::eval_gpu(
   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& {
+  auto set_output = [&s, &out](const array& x) {
     bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
     if (no_copy && x.ndim() > 1) {
       auto s = x.strides()[x.ndim() - 2];
       no_copy &= (s == 0 || s == x.shape().back());
     }
     if (no_copy) {
+      if (x.is_donatable()) {
+        out.copy_shared_buffer(x);
+      } else {
+        out.set_data(
+            allocator::malloc_or_wait(x.data_size() * x.itemsize()),
+            x.data_size(),
+            x.strides(),
+            x.flags());
+      }
       return x;
     } else {
-      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
-      copy_gpu(x, copies.back(), CopyType::General, s);
-      return copies.back();
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy_gpu(x, x_copy, CopyType::General, s);
+      out.copy_shared_buffer(x_copy);
+      return x_copy;
     }
   };
-  const array& x = check_input(inputs[0]);
+
+  const array x = set_output(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;
 
@@ -290,8 +289,6 @@ void LayerNorm::eval_gpu(
     compute_encoder.set_bytes(b_stride, 7);
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   }
-
-  d.add_temporaries(std::move(copies), s.index);
 }
 
 void LayerNormVJP::eval_gpu(
@@ -303,21 +300,22 @@ void LayerNormVJP::eval_gpu(
   // 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.
-  auto check_input = [&d, &s](const array& x) -> array {
+  auto check_input = [&s](const array& x) -> std::pair<array, bool> {
     if (x.flags().row_contiguous) {
-      return x;
+      return {x, false};
     }
-
     array x_copy(x.shape(), x.dtype(), nullptr, {});
     copy_gpu(x, x_copy, CopyType::General, s);
-    d.add_temporary(x_copy, s.index);
-
-    return x_copy;
+    return {x_copy, true};
   };
-  const array& x = check_input(inputs[0]);
+  bool donate_x = inputs[0].is_donatable();
+  bool donate_g = inputs[3].is_donatable();
+  auto [x, copied] = check_input(inputs[0]);
+  donate_x |= copied;
   const array& w = inputs[1];
   const array& b = inputs[2];
-  const array& g = check_input(inputs[3]);
+  auto [g, g_copied] = check_input(inputs[3]);
+  donate_g |= g_copied;
   array& gx = outputs[0];
   array& gw = outputs[1];
   array& gb = outputs[2];
@@ -326,17 +324,18 @@ void LayerNormVJP::eval_gpu(
   bool has_w = w.ndim() != 0;
 
   // 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);
+  if (donate_x) {
+    gx.copy_shared_buffer(x);
+  } else if (donate_g) {
+    gx.copy_shared_buffer(g);
     g_in_gx = true;
   } else {
     gx.set_data(allocator::malloc_or_wait(gx.nbytes()));
   }
+  if (g_copied && !g_in_gx) {
+    d.add_temporary(g, s.index);
+  }
 
   auto axis_size = static_cast<uint32_t>(x.shape().back());
   int n_rows = x.data_size() / axis_size;
@@ -345,15 +344,13 @@ void LayerNormVJP::eval_gpu(
   // gradient accumulators.
   array gw_temp =
       (has_w) ? array({n_rows, x.shape().back()}, gw.dtype(), nullptr, {}) : w;
-  bool g_in_gw = false;
   if (has_w) {
-    if (!g_in_gx && g.is_donatable()) {
-      gw_temp.move_shared_buffer(g);
-      g_in_gw = true;
+    if (!g_in_gx && donate_g) {
+      gw_temp.copy_shared_buffer(g);
     } else {
       gw_temp.set_data(allocator::malloc_or_wait(gw_temp.nbytes()));
+      d.add_temporary(gw_temp, s.index);
     }
-    d.add_temporary(gw_temp, s.index);
   }
   gw.set_data(allocator::malloc_or_wait(gw.nbytes()));
   gb.set_data(allocator::malloc_or_wait(gb.nbytes()));
@@ -364,14 +361,7 @@ void LayerNormVJP::eval_gpu(
     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);
+        g, gb, "sum", plan, {0}, compute_encoder, d, s);
   }
 
   const int simd_size = 32;
@@ -409,9 +399,9 @@ void LayerNormVJP::eval_gpu(
 
     uint32_t w_stride = (w.ndim() == 1) ? w.strides()[0] : 0;
     compute_encoder.set_compute_pipeline_state(kernel);
-    compute_encoder.set_input_array(x_in_gx ? gx : x, 0);
+    compute_encoder.set_input_array(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_input_array(g, 2);
     compute_encoder.set_output_array(gx, 3);
     compute_encoder.set_output_array(gw_temp, 4);
     compute_encoder.set_bytes(eps_, 5);

mlx/backend/metal/primitives.cpp

@@ -5,12 +5,10 @@
 #include <sstream>
 
 #include "mlx/backend/common/compiled.h"
-#include "mlx/backend/common/load.h"
 #include "mlx/backend/common/slicing.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/event.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/slicing.h"
 #include "mlx/backend/metal/utils.h"
@@ -45,7 +43,8 @@ void reshape(const array& in, array& out, Stream s) {
     shared_buffer_reshape(in, out_strides, out);
   }
 }
-array compute_dynamic_offset(
+
+static array compute_dynamic_offset(
     const array& indices,
     const Strides& strides,
     const std::vector<int>& axes,
@@ -57,7 +56,7 @@ array compute_dynamic_offset(
   bool donate = indices.is_donatable() &&
       (indices.data_size() * indices.itemsize()) >= offset.itemsize();
   if (donate) {
-    offset.move_shared_buffer(indices);
+    offset.copy_shared_buffer(indices);
   } else {
     offset.set_data(allocator::malloc_or_wait(offset.itemsize()));
   }
@@ -88,7 +87,7 @@ array compute_dynamic_offset(
 
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
-  compute_encoder.set_input_array(donate ? offset : indices, 0);
+  compute_encoder.set_input_array(indices, 0);
   compute_encoder.set_output_array(offset, 1);
   compute_encoder.set_vector_bytes(strides, 2);
   compute_encoder.set_vector_bytes(axes, 3);
@@ -254,7 +253,7 @@ void Contiguous::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
   if (in.flags().row_contiguous ||
       (allow_col_major_ && in.flags().col_contiguous)) {
-    move_or_copy(in, out);
+    out.copy_shared_buffer(in);
   } else {
     copy_gpu(in, out, CopyType::General);
   }
@@ -302,31 +301,7 @@ void Unflatten::eval_gpu(const std::vector<array>& inputs, array& out) {
 }
 
 void Load::eval_gpu(const std::vector<array>& inputs, array& out) {
-  out.set_data(allocator::malloc_or_wait(out.nbytes()));
-  auto read_task = [out = unsafe_weak_copy(out),
-                    offset = offset_,
-                    reader = reader_,
-                    swap_endianness = swap_endianness_]() mutable {
-    load(out, offset, reader, swap_endianness);
-  };
-
-  // Limit the size that the command buffer will wait on to avoid timing out
-  // on the event (<4 seconds).
-  if (out.nbytes() > (1 << 28)) {
-    read_task();
-    return;
-  }
-
-  auto fut = io::thread_pool().enqueue(std::move(read_task)).share();
-
-  auto e = Event(stream());
-  e.set_value(1);
-  encode_wait(e);
-  auto signal_task = [e = std::move(e), fut = std::move(fut)]() mutable {
-    fut.wait();
-    e.signal();
-  };
-  scheduler::enqueue(io_stream(), std::move(signal_task));
+  throw std::runtime_error("[Load::eval_gpu] Not implemented.");
 }
 
 void NumberOfElements::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -452,7 +427,7 @@ void DynamicSliceUpdate::eval_gpu(
   auto& start_indices = inputs[2];
 
   if (upd.size() == 0) {
-    move_or_copy(in, out);
+    out.copy_shared_buffer(in);
     return;
   }
 
@@ -491,7 +466,7 @@ void SliceUpdate::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& upd = inputs[1];
 
   if (upd.size() == 0) {
-    move_or_copy(in, out);
+    out.copy_shared_buffer(in);
     return;
   }
 
@@ -575,8 +550,8 @@ void View::eval_gpu(const std::vector<array>& inputs, array& out) {
       strides[i] *= ibytes;
       strides[i] /= obytes;
     }
-    move_or_copy(
-        in, out, strides, in.flags(), in.data_size() * ibytes / obytes);
+    out.copy_shared_buffer(
+        in, strides, in.flags(), in.data_size() * ibytes / obytes);
   } else {
     auto tmp = array(in.shape(), in.dtype(), nullptr, {});
     tmp.set_data(allocator::malloc_or_wait(tmp.nbytes()));
@@ -587,7 +562,7 @@ void View::eval_gpu(const std::vector<array>& inputs, array& out) {
     flags.row_contiguous = true;
     auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
     flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
-    out.move_shared_buffer(tmp, out.strides(), flags, out.size());
+    out.copy_shared_buffer(tmp, out.strides(), flags, out.size());
   }
 }
 

mlx/backend/metal/rope.cpp

@@ -41,7 +41,7 @@ void RoPE::eval_gpu(
   } else if (in.flags().row_contiguous) {
     if (in.is_donatable()) {
       donated = true;
-      out.move_shared_buffer(in);
+      out.copy_shared_buffer(in);
     } else {
       out.set_data(allocator::malloc_or_wait(out.nbytes()));
     }

mlx/backend/metal/scaled_dot_product_attention.cpp

@@ -131,8 +131,11 @@ void sdpa_vector(
   int gqa_factor = q.shape(1) / k.shape(1);
   int N = k.shape(2);
   int B = q.shape(0) * q.shape(1);
-  size_t k_stride = k.strides()[1];
-  size_t v_stride = v.strides()[1];
+  size_t k_head_stride = k.strides()[1];
+  size_t k_seq_stride = k.strides()[2];
+  size_t v_head_stride = v.strides()[1];
+  size_t v_seq_stride = v.strides()[2];
+
   MTL::Size group_dims(1024, 1, 1);
   MTL::Size grid_dims(B, q.shape(2), 1);
 
@@ -152,26 +155,29 @@ void sdpa_vector(
   compute_encoder.set_compute_pipeline_state(kernel);
 
   // Set its arguments
-  compute_encoder.set_input_array(q.data_shared_ptr() == nullptr ? out : q, 0);
+  compute_encoder.set_input_array(q, 0);
   compute_encoder.set_input_array(k, 1);
   compute_encoder.set_input_array(v, 2);
   compute_encoder.set_output_array(out, 3);
   compute_encoder.set_bytes(gqa_factor, 4);
   compute_encoder.set_bytes(N, 5);
-  compute_encoder.set_bytes(k_stride, 6);
-  compute_encoder.set_bytes(v_stride, 7);
-  compute_encoder.set_bytes(scale, 8);
+  compute_encoder.set_bytes(k_head_stride, 6);
+  compute_encoder.set_bytes(k_seq_stride, 7);
+  compute_encoder.set_bytes(v_head_stride, 8);
+  compute_encoder.set_bytes(v_seq_stride, 9);
+
+  compute_encoder.set_bytes(scale, 10);
   if (has_mask) {
     auto& m = *mask;
-    compute_encoder.set_input_array(m, 9);
+    compute_encoder.set_input_array(m, 11);
     auto nd = m.ndim();
     int32_t kv_seq_stride =
         nd >= 1 && m.shape(-1) > 1 ? m.strides()[nd - 1] : 0;
     int32_t q_seq_stride = nd >= 2 && m.shape(-2) > 1 ? m.strides()[nd - 2] : 0;
     int32_t head_stride = nd >= 3 && m.shape(-3) > 1 ? m.strides()[nd - 3] : 0;
-    compute_encoder.set_bytes(kv_seq_stride, 10);
-    compute_encoder.set_bytes(q_seq_stride, 11);
-    compute_encoder.set_bytes(head_stride, 12);
+    compute_encoder.set_bytes(kv_seq_stride, 12);
+    compute_encoder.set_bytes(q_seq_stride, 13);
+    compute_encoder.set_bytes(head_stride, 14);
   }
 
   // Launch
@@ -202,8 +208,10 @@ void sdpa_vector_2pass(
   int N = k.shape(2);
   int blocks = 32;
   int B = q.shape(0) * q.shape(1);
-  auto k_stride = k.strides()[1];
-  auto v_stride = v.strides()[1];
+  size_t k_head_stride = k.strides()[1];
+  size_t k_seq_stride = k.strides()[2];
+  size_t v_head_stride = v.strides()[1];
+  size_t v_seq_stride = v.strides()[2];
   MTL::Size group_dims(8 * 32, 1, 1);
   MTL::Size grid_dims(B, q.shape(2), blocks);
 
@@ -242,7 +250,7 @@ void sdpa_vector_2pass(
   compute_encoder.set_compute_pipeline_state(kernel);
 
   // Set its arguments
-  compute_encoder.set_input_array(q.data_shared_ptr() == nullptr ? out : q, 0);
+  compute_encoder.set_input_array(q, 0);
   compute_encoder.set_input_array(k, 1);
   compute_encoder.set_input_array(v, 2);
   compute_encoder.set_output_array(intermediate, 3);
@@ -250,20 +258,22 @@ void sdpa_vector_2pass(
   compute_encoder.set_output_array(maxs, 5);
   compute_encoder.set_bytes(gqa_factor, 6);
   compute_encoder.set_bytes(N, 7);
-  compute_encoder.set_bytes(k_stride, 8);
-  compute_encoder.set_bytes(v_stride, 9);
-  compute_encoder.set_bytes(scale, 10);
+  compute_encoder.set_bytes(k_head_stride, 8);
+  compute_encoder.set_bytes(k_seq_stride, 9);
+  compute_encoder.set_bytes(v_head_stride, 10);
+  compute_encoder.set_bytes(v_seq_stride, 11);
+  compute_encoder.set_bytes(scale, 12);
   if (has_mask) {
     auto& m = *mask;
-    compute_encoder.set_input_array(m, 11);
+    compute_encoder.set_input_array(m, 13);
     auto nd = m.ndim();
     int32_t kv_seq_stride =
         nd >= 1 && m.shape(-1) > 1 ? m.strides()[nd - 1] : 0;
     int32_t q_seq_stride = nd >= 2 && m.shape(-2) > 1 ? m.strides()[nd - 2] : 0;
     int32_t head_stride = nd >= 3 && m.shape(-3) > 1 ? m.strides()[nd - 3] : 0;
-    compute_encoder.set_bytes(kv_seq_stride, 12);
-    compute_encoder.set_bytes(q_seq_stride, 13);
-    compute_encoder.set_bytes(head_stride, 14);
+    compute_encoder.set_bytes(kv_seq_stride, 14);
+    compute_encoder.set_bytes(q_seq_stride, 15);
+    compute_encoder.set_bytes(head_stride, 16);
   }
 
   // Launch
@@ -334,15 +344,6 @@ void ScaledDotProductAttention::eval_gpu(
         (strides[1] == shape[3]) && (strides[0] == strides[2] * shape[2]);
   };
 
-  // Returns true if the array is row contiguous except the sequence length
-  // dimension that can be sliced but with step=1.
-  auto is_contiguous_except_seq_len = [](const array& arr) {
-    auto& strides = arr.strides();
-    auto& shape = arr.shape();
-    return strides[3] == 1 && strides[2] == shape[3] &&
-        strides[0] == strides[1] * shape[1];
-  };
-
   // Checks that the headdim dimension has stride 1.
   auto is_matrix_contiguous = [](const array& arr) {
     return arr.strides(3) == 1;
@@ -351,13 +352,13 @@ void ScaledDotProductAttention::eval_gpu(
   // We are in vector mode ie single query
   if (q_pre.shape(2) <= 8) {
     const auto& q = copy_unless(is_contiguous_or_head_seq_transposed, q_pre);
-    const auto& k = copy_unless(is_contiguous_except_seq_len, k_pre);
-    const auto& v = copy_unless(is_contiguous_except_seq_len, v_pre);
+    const auto& k = copy_unless(is_matrix_contiguous, k_pre);
+    const auto& v = copy_unless(is_matrix_contiguous, v_pre);
 
     // Donate the query if possible
     if (q.is_donatable() && (q.shape(2) == 1 || !q.flags().row_contiguous) &&
         q.size() == o.size()) {
-      o.move_shared_buffer(q);
+      o.copy_shared_buffer(q);
     } else {
       if (o.shape(2) == 1) {
         o.set_data(allocator::malloc_or_wait(o.nbytes()));

mlx/backend/metal/scan.cpp

@@ -21,7 +21,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto in = inputs[0];
   if (in.flags().contiguous && in.strides()[axis_] != 0) {
     if (donate && in.itemsize() == out.itemsize()) {
-      out.move_shared_buffer(in);
+      out.copy_shared_buffer(in);
     } else {
       out.set_data(
           allocator::malloc_or_wait(in.data_size() * out.itemsize()),
@@ -33,7 +33,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
     array arr_copy(in.shape(), in.dtype(), nullptr, {});
     copy_gpu(in, arr_copy, CopyType::General, s);
     in = std::move(arr_copy);
-    out.move_shared_buffer(in);
+    out.copy_shared_buffer(in);
   }
 
   bool contiguous = in.strides()[axis_] == 1;
@@ -68,8 +68,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
   if (contiguous) {
     auto& compute_encoder = d.get_command_encoder(s.index);
     compute_encoder.set_compute_pipeline_state(kernel);
-    compute_encoder.set_input_array(
-        in.data_shared_ptr() == nullptr ? out : in, 0);
+    compute_encoder.set_input_array(in, 0);
     compute_encoder.set_output_array(out, 1);
     size_t size = in.shape(axis_);
     compute_encoder.set_bytes(size, 2);

mlx/backend/metal/softmax.cpp

@@ -22,31 +22,32 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& d = metal::device(s.device);
 
   // Make sure that the last dimension is contiguous
-  std::vector<array> copies;
-  auto check_input = [&copies, &s](const array& x) -> const array& {
+  auto set_output = [&s, &out](const array& x) {
     bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
     if (no_copy && x.ndim() > 1) {
       auto s = x.strides()[x.ndim() - 2];
       no_copy &= (s == 0 || s == x.shape().back());
     }
     if (no_copy) {
+      if (x.is_donatable()) {
+        out.copy_shared_buffer(x);
+      } else {
+        out.set_data(
+            allocator::malloc_or_wait(x.data_size() * x.itemsize()),
+            x.data_size(),
+            x.strides(),
+            x.flags());
+      }
       return x;
     } else {
-      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
-      copy_gpu(x, copies.back(), CopyType::General, s);
-      return copies.back();
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy_gpu(x, x_copy, CopyType::General, s);
+      out.copy_shared_buffer(x_copy);
+      return x_copy;
     }
   };
-  const array& in = check_input(inputs[0]);
-  if (in.is_donatable()) {
-    out.move_shared_buffer(in);
-  } else {
-    out.set_data(
-        allocator::malloc_or_wait(in.data_size() * in.itemsize()),
-        in.data_size(),
-        in.strides(),
-        in.flags());
-  }
+
+  const array in = set_output(inputs[0]);
 
   int axis_size = in.shape().back();
   int n_rows = in.data_size() / axis_size;
@@ -82,14 +83,11 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
     }
 
     compute_encoder.set_compute_pipeline_state(kernel);
-    compute_encoder.set_input_array(
-        in.data_shared_ptr() == nullptr ? out : in, 0);
+    compute_encoder.set_input_array(in, 0);
     compute_encoder.set_output_array(out, 1);
     compute_encoder.set_bytes(axis_size, 2);
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   }
-
-  d.add_temporaries(std::move(copies), s.index);
 }
 
 } // namespace mlx::core

mlx/backend/metal/ternary.cpp

@@ -71,12 +71,9 @@ void ternary_op_gpu_inplace(
 
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
-  bool donate_a = a.data_shared_ptr() == nullptr;
-  bool donate_b = b.data_shared_ptr() == nullptr;
-  bool donate_c = c.data_shared_ptr() == nullptr;
-  compute_encoder.set_input_array(donate_a ? out : a, 0);
-  compute_encoder.set_input_array(donate_b ? out : b, 1);
-  compute_encoder.set_input_array(donate_c ? out : c, 2);
+  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);
 
   auto thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
@@ -129,7 +126,7 @@ void ternary_op_gpu(
   auto& b = inputs[1];
   auto& c = inputs[2];
   TernaryOpType topt = get_ternary_op_type(a, b, c);
-  set_ternary_op_output_data(a, b, c, out, topt, true /* donate_with_move */);
+  set_ternary_op_output_data(a, b, c, out, topt);
   ternary_op_gpu_inplace(inputs, out, op, s);
 }
 

mlx/backend/metal/unary.cpp

@@ -57,8 +57,7 @@ void unary_op_gpu_inplace(
   auto thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
-  compute_encoder.set_input_array(
-      in.data_shared_ptr() == nullptr ? out : in, 0);
+  compute_encoder.set_input_array(in, 0);
   compute_encoder.set_output_array(out, 1);
   if (!contig) {
     // Launch up to 3D grid of threads
@@ -95,7 +94,7 @@ void unary_op_gpu(
   bool contig = in.flags().contiguous;
   if (contig) {
     if (in.is_donatable() && in.itemsize() == out.itemsize()) {
-      out.move_shared_buffer(in);
+      out.copy_shared_buffer(in);
     } else {
       out.set_data(
           allocator::malloc_or_wait(in.data_size() * out.itemsize()),
@@ -169,7 +168,7 @@ void Round::eval_gpu(const std::vector<array>& inputs, array& out) {
     unary_op_gpu(inputs, out, get_primitive_string(this));
   } else {
     // No-op integer types
-    move_or_copy(in, out);
+    out.copy_shared_buffer(in);
   }
 }
 

mlx/backend/metal/utils.h

@@ -69,19 +69,4 @@ void concatenate(std::string& acc, T first, Args... args) {
   concatenate(acc, args...);
 }
 
-/**
- * Get a new array that refers to the same data but has a non-owning pointer to
- * them.
- */
-inline array unsafe_weak_copy(const array& x) {
-  return array(
-      x.buffer(),
-      x.shape(),
-      x.dtype(),
-      x.strides(),
-      x.data_size(),
-      x.flags(),
-      [](auto b) {});
-}
-
 } // namespace mlx::core

mlx/backend/no_cpu/CMakeLists.txt

@@ -1,2 +1,6 @@
-target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
-                           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp)
+target_sources(
+  mlx
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/../cpu/eval.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/../cpu/encoder.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp)

mlx/backend/no_cpu/primitives.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2024 Apple Inc.
 
 #include "mlx/primitives.h"
+#include "mlx/distributed/primitives.h"
 #include "mlx/fast_primitives.h"
 
 #define NO_CPU_MULTI(func)                                             \
@@ -75,7 +76,6 @@ NO_CPU(Hadamard)
 NO_CPU(Imag)
 NO_CPU(Less)
 NO_CPU(LessEqual)
-NO_CPU(Load)
 NO_CPU(Log)
 NO_CPU(Log1p)
 NO_CPU(LogicalNot)
@@ -129,4 +129,11 @@ namespace fast {
 NO_CPU_MULTI(AffineQuantize)
 } // namespace fast
 
+namespace distributed {
+NO_CPU_MULTI(AllReduce)
+NO_CPU_MULTI(AllGather)
+NO_CPU_MULTI(Send)
+NO_CPU_MULTI(Recv)
+} // namespace distributed
+
 } // namespace mlx::core

mlx/backend/no_metal/CMakeLists.txt

@@ -2,5 +2,6 @@ target_sources(
   mlx
   PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/metal.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp)

mlx/backend/no_metal/event.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2024 Apple Inc.
 
 #include "mlx/event.h"
+#include "mlx/scheduler.h"
 
 #include <condition_variable>
 #include <mutex>
@@ -13,13 +14,13 @@ struct EventCounter {
   std::condition_variable cv;
 };
 
-Event::Event(const Stream& stream) : stream_(stream) {
+Event::Event(Stream stream) : stream_(stream) {
   auto dtor = [](void* ptr) { delete static_cast<EventCounter*>(ptr); };
   event_ = std::shared_ptr<void>(new EventCounter{}, dtor);
 }
 
 void Event::wait() {
-  auto ec = static_cast<EventCounter*>(raw_event().get());
+  auto ec = static_cast<EventCounter*>(event_.get());
   std::unique_lock<std::mutex> lk(ec->mtx);
   if (ec->value >= value()) {
     return;
@@ -28,7 +29,7 @@ void Event::wait() {
 }
 
 void Event::signal() {
-  auto ec = static_cast<EventCounter*>(raw_event().get());
+  auto ec = static_cast<EventCounter*>(event_.get());
   {
     std::lock_guard<std::mutex> lk(ec->mtx);
     ec->value = value();
@@ -36,11 +37,19 @@ void Event::signal() {
   ec->cv.notify_all();
 }
 
+void Event::wait(Stream stream) {
+  scheduler::enqueue(stream, [*this]() mutable { wait(); });
+}
+
+void Event::signal(Stream stream) {
+  scheduler::enqueue(stream, [*this]() mutable { signal(); });
+}
+
 bool Event::is_signaled() const {
-  auto ec = static_cast<EventCounter*>(raw_event().get());
+  auto ec = static_cast<EventCounter*>(event_.get());
   {
     std::lock_guard<std::mutex> lk(ec->mtx);
-    return (ec->value > value());
+    return (ec->value >= value());
   }
 }
 } // namespace mlx::core

mlx/backend/no_metal/fence.cpp

@@ -0,0 +1,54 @@
+// Copyright © 2024 Apple Inc.
+
+#include <condition_variable>
+#include <mutex>
+
+#include "mlx/fence.h"
+#include "mlx/scheduler.h"
+
+namespace mlx::core {
+
+struct FenceImpl {
+  uint32_t count{0};
+  uint32_t value{0};
+  std::mutex mtx;
+  std::condition_variable cv;
+};
+
+Fence::Fence(Stream) {
+  auto dtor = [](void* ptr) { delete static_cast<FenceImpl*>(ptr); };
+  fence_ = std::shared_ptr<void>(new FenceImpl{}, dtor);
+}
+
+void Fence::wait(Stream stream, const array&) {
+  auto& f = *static_cast<FenceImpl*>(fence_.get());
+  if (stream.device == Device::cpu) {
+    scheduler::enqueue(stream, [count = f.count, fence_ = fence_]() mutable {
+      auto& f = *static_cast<FenceImpl*>(fence_.get());
+      std::unique_lock<std::mutex> lk(f.mtx);
+      if (f.value >= count) {
+        return;
+      }
+      f.cv.wait(lk, [&f, count] { return f.value >= count; });
+    });
+  } else {
+    throw std::runtime_error("[Fence::wait] Invalid stream.");
+  }
+}
+
+void Fence::update(Stream stream, const array&) {
+  auto& f = *static_cast<FenceImpl*>(fence_.get());
+  f.count++;
+  if (stream.device == Device::cpu) {
+    scheduler::enqueue(stream, [count = f.count, fence_ = fence_]() mutable {
+      auto& f = *static_cast<FenceImpl*>(fence_.get());
+      std::unique_lock<std::mutex> lk(f.mtx);
+      f.value = count;
+      f.cv.notify_all();
+    });
+  } else {
+    throw std::runtime_error("[Fence::update] Invalid stream.");
+  }
+}
+
+} // namespace mlx::core

mlx/backend/no_metal/metal.cpp

@@ -16,17 +16,19 @@ std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool() {
   return nullptr;
 }
 
-std::function<void()> make_task(array, bool) {
+void eval(array&) {
   throw std::runtime_error(
-      "[metal::make_task] Cannot make GPU task without metal backend");
+      "[metal::eval] Cannot eval on GPU without metal backend");
 }
 
-std::function<void()> make_synchronize_task(
-    Stream,
-    std::shared_ptr<std::promise<void>>) {
+void finalize(Stream) {
   throw std::runtime_error(
-      "[metal::make_synchronize_task] Cannot synchronize GPU"
-      " without metal backend");
+      "[metal::finalize] Cannot finalize GPU without metal backend");
+}
+
+void synchronize(Stream) {
+  throw std::runtime_error(
+      "[metal::synchronize] Cannot synchronize GPU without metal backend");
 }
 
 // No-ops when Metal is not available.

mlx/distributed/distributed.cpp

@@ -6,31 +6,25 @@
 #include "mlx/distributed/distributed_impl.h"
 #include "mlx/distributed/mpi/mpi.h"
 #include "mlx/distributed/ring/ring.h"
-#include "mlx/scheduler.h"
 
 namespace mlx::core::distributed {
 
 namespace detail {
 
-Stream communication_stream() {
-  static Stream comm_stream = new_stream(Device::cpu);
-  return comm_stream;
+void all_sum(Group group, const array& input, array& output, Stream stream) {
+  group.raw_group()->all_sum(input, output, stream);
 }
 
-void all_sum(Group group, const array& input, array& output) {
-  group.raw_group()->all_sum(input, output);
+void all_gather(Group group, const array& input, array& output, Stream stream) {
+  group.raw_group()->all_gather(input, output, stream);
 }
 
-void all_gather(Group group, const array& input, array& output) {
-  group.raw_group()->all_gather(input, output);
+void send(Group group, const array& input, int dst, Stream stream) {
+  group.raw_group()->send(input, dst, stream);
 }
 
-void send(Group group, const array& input, int dst) {
-  group.raw_group()->send(input, dst);
-}
-
-void recv(Group group, array& out, int src) {
-  group.raw_group()->recv(out, src);
+void recv(Group group, array& out, int src, Stream stream) {
+  group.raw_group()->recv(out, src, stream);
 }
 
 class EmptyGroup : public GroupImpl {
@@ -47,19 +41,19 @@ class EmptyGroup : public GroupImpl {
     throw std::runtime_error("Cannot split the distributed group further.");
   }
 
-  void all_sum(const array& input, array& output) override {
+  void all_sum(const array&, array&, Stream) override {
     throw std::runtime_error(
         "Communication not implemented in an empty distributed group.");
   }
-  void all_gather(const array& input, array& output) override {
+  void all_gather(const array&, array&, Stream) override {
     throw std::runtime_error(
         "Communication not implemented in an empty distributed group.");
   }
-  void send(const array& input, int dst) override {
+  void send(const array&, int, Stream) override {
     throw std::runtime_error(
         "Communication not implemented in an empty distributed group.");
   }
-  void recv(array& out, int src) override {
+  void recv(array&, int, Stream) override {
     throw std::runtime_error(
         "Communication not implemented in an empty distributed group.");
   }
@@ -122,10 +116,6 @@ Group init(bool strict /* = false */, const std::string& bk /* = "any" */) {
     backends.insert({"any", group});
   }
   backends.insert({std::move(bk_), group});
-
-  // Ensure the communication stream is alive before
-  // the graph is evaluated
-  detail::communication_stream();
   return Group(group);
 }
 

mlx/distributed/distributed_impl.h

@@ -17,25 +17,22 @@ class GroupImpl {
   virtual int size() = 0;
   virtual std::shared_ptr<GroupImpl> split(int color, int key = -1) = 0;
 
-  virtual void all_sum(const array& input, array& output) = 0;
-  virtual void all_gather(const array& input, array& output) = 0;
-  virtual void send(const array& input, int dst) = 0;
-  virtual void recv(array& out, int src) = 0;
+  virtual void all_sum(const array& input, array& output, Stream stream) = 0;
+  virtual void all_gather(const array& input, array& output, Stream stream) = 0;
+  virtual void send(const array& input, int dst, Stream stream) = 0;
+  virtual void recv(array& out, int src, Stream stream) = 0;
 };
 
-/* Return the communication stream. */
-Stream communication_stream();
-
 /* Perform an all reduce sum operation */
-void all_sum(Group group, const array& input, array& output);
+void all_sum(Group group, const array& input, array& output, Stream stream);
 
 /* Perform an all gather operation */
-void all_gather(Group group, const array& input, array& output);
+void all_gather(Group group, const array& input, array& output, Stream stream);
 
 /** Send an array to the dst rank */
-void send(Group group, const array& input, int dst);
+void send(Group group, const array& input, int dst, Stream stream);
 
 /** Recv an array from the src rank */
-void recv(Group group, array& out, int src);
+void recv(Group group, array& out, int src, Stream stream);
 
 } // namespace mlx::core::distributed::detail

mlx/distributed/mpi/mpi.cpp

@@ -3,11 +3,10 @@
 #include <dlfcn.h>
 #include <mpi.h>
 
-#include "mlx/backend/cpu/copy.h"
+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/distributed/distributed.h"
 #include "mlx/distributed/distributed_impl.h"
 #include "mlx/distributed/mpi/mpi.h"
-#include "mlx/scheduler.h"
 
 #define LOAD_SYMBOL(symbol, variable)                              \
   {                                                                \
@@ -25,16 +24,6 @@ using GroupImpl = mlx::core::distributed::detail::GroupImpl;
 
 namespace {
 
-array ensure_row_contiguous(const array& arr) {
-  if (arr.flags().row_contiguous) {
-    return arr;
-  } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy(arr, arr_copy, CopyType::General);
-    return arr_copy;
-  }
-}
-
 template <typename T>
 void simple_sum(
     void* input,
@@ -281,9 +270,12 @@ class MPIGroup : public GroupImpl {
     return std::make_shared<MPIGroup>(new_comm, false);
   }
 
-  void all_sum(const array& input_, array& output) override {
-    array input = ensure_row_contiguous(input_);
-    mpi().all_reduce(
+  void all_sum(const array& input, array& output, Stream stream) override {
+    auto& encoder = cpu::get_command_encoder(stream);
+    encoder.set_input_array(input);
+    encoder.set_output_array(output);
+    encoder.dispatch(
+        mpi().all_reduce,
         (input.data<void>() == output.data<void>()) ? MPI_IN_PLACE
                                                     : input.data<void>(),
         output.data<void>(),
@@ -293,9 +285,12 @@ class MPIGroup : public GroupImpl {
         comm_);
   }
 
-  void all_gather(const array& input_, array& output) override {
-    array input = ensure_row_contiguous(input_);
-    mpi().all_gather(
+  void all_gather(const array& input, array& output, Stream stream) override {
+    auto& encoder = cpu::get_command_encoder(stream);
+    encoder.set_input_array(input);
+    encoder.set_output_array(output);
+    encoder.dispatch(
+        mpi().all_gather,
         input.data<void>(),
         input.size(),
         mpi().datatype(input),
@@ -305,22 +300,30 @@ class MPIGroup : public GroupImpl {
         comm_);
   }
 
-  void send(const array& input_, int dst) override {
-    array input = ensure_row_contiguous(input_);
-    mpi().send(
-        input.data<void>(), input.size(), mpi().datatype(input), dst, 0, comm_);
+  void send(const array& input, int dst, Stream stream) override {
+    auto& encoder = cpu::get_command_encoder(stream);
+    encoder.set_input_array(input);
+    encoder.dispatch(
+        mpi().send,
+        input.data<void>(),
+        input.size(),
+        mpi().datatype(input),
+        dst,
+        0,
+        comm_);
   }
 
-  void recv(array& out, int src) override {
-    MPI_Status status;
-    mpi().recv(
-        out.data<void>(),
-        out.size(),
-        mpi().datatype(out),
-        src,
-        MPI_ANY_TAG,
-        comm_,
-        &status);
+  void recv(array& out, int src, Stream stream) override {
+    auto& encoder = cpu::get_command_encoder(stream);
+    encoder.set_output_array(out);
+    encoder.dispatch([out_ptr = out.data<void>(),
+                      out_size = out.size(),
+                      out_type = mpi().datatype(out),
+                      src,
+                      comm = comm_]() {
+      MPI_Status status;
+      mpi().recv(out_ptr, out_size, out_type, src, MPI_ANY_TAG, comm, &status);
+    });
   }
 
  private:

mlx/distributed/ops.cpp

@@ -28,11 +28,11 @@ array all_sum(
   if (group.size() == 1) {
     return x;
   }
-
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(to_stream(s), group, AllReduce::Sum),
+      std::make_shared<AllReduce>(
+          to_stream(s, Device::cpu), group, AllReduce::Sum),
       {x});
 }
 
@@ -55,7 +55,7 @@ array all_gather(
   return array(
       std::move(result_shape),
       x.dtype(),
-      std::make_shared<AllGather>(to_stream(s), group),
+      std::make_shared<AllGather>(to_stream(s, Device::cpu), group),
       {x});
 }
 
@@ -80,7 +80,7 @@ array send(
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<Send>(to_stream(s), group, dst),
+      std::make_shared<Send>(to_stream(s, Device::cpu), group, dst),
       {x});
 }
 
@@ -105,7 +105,7 @@ array recv(
   return array(
       std::move(shape),
       std::move(dtype),
-      std::make_shared<Recv>(to_stream(s), group, src),
+      std::make_shared<Recv>(to_stream(s, Device::cpu), group, src),
       std::vector<array>{});
 }