fix cuda pypi package (#2423)

* fix cuda pypi package

* patch bump

.circleci/config.yml

@@ -7,18 +7,9 @@ parameters:
   nightly_build:
     type: boolean
     default: false
-  weekly_build:
-    type: boolean
-    default: false
   test_release:
     type: boolean
     default: false
-  linux_release:
-    type: boolean
-    default: false
-  cuda_release:
-    type: boolean
-    default: false
 
 jobs:
   build_documentation:
@@ -41,7 +32,7 @@ jobs:
             pip install --upgrade pip
             pip install --upgrade cmake
             pip install -r docs/requirements.txt
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install . -v
+            pip install . -v
       - when:
           condition:
             not: << parameters.upload-docs >>
@@ -73,9 +64,9 @@ jobs:
                  git push -f origin gh-pages
 
   linux_build_and_test:
-    docker:
-      - image: cimg/python:3.9
-
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
@@ -87,21 +78,17 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
             sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get upgrade -y
+            pip install --upgrade cmake
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
             sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py develop
+            pip install -e ".[dev]"
       - run:
           name: Generate package stubs
           command: |
@@ -111,13 +98,14 @@ jobs:
       - run:
           name: Run Python tests
           command: |
-            python3 -m unittest discover python/tests -v
+            python -m unittest discover python/tests -v
             mpirun --bind-to none -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build CPP only
           command: |
-            mkdir -p build && cd build 
+            mkdir -p build && cd build
             cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
             make -j `nproc`
       - run:
@@ -157,8 +145,7 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-            CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               pip install -e . -v
       - run:
           name: Generate package stubs
@@ -173,7 +160,8 @@ jobs:
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
             mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build example extension
           command: |
@@ -208,16 +196,19 @@ jobs:
           name: Run Python tests with JIT
           command: |
             source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
+            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
               pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               python -m xmlrunner discover -v python/tests -o test-results/gpu_jit
 
   cuda_build_and_test:
+    parameters:
+      image_date:
+        type: string
+        default: "2023.11.1"
     machine:
-      image: linux-cuda-12:default
+      image: "linux-cuda-12:<< parameters.image_date >>"
       resource_class: gpu.nvidia.small.gen2
     steps:
       - checkout
@@ -225,11 +216,11 @@ jobs:
           name: Install Python package
           command: |
             sudo apt-get update
+            sudo apt-get install libcudnn9-dev-cuda-12
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            python -m venv env
+            python3 -m venv env
             source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               pip install -e ".[dev]"
       - run:
           name: Run Python tests
@@ -278,7 +269,6 @@ jobs:
           command: |
             source env/bin/activate
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               pip install . -v
       - run:
           name: Generate package stubs
@@ -290,9 +280,18 @@ jobs:
           name: Build Python package
           command: |
             source env/bin/activate
-            << parameters.build_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              python -m build -w
+            python setup.py clean --all
+            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
+      - when:
+          condition:
+            equal: ["3.9", << parameters.python_version >>]
+          steps:
+            - run:
+                name: Build common package
+                command: |
+                  source env/bin/activate
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
           condition: << parameters.build_env >>
           steps:
@@ -309,94 +308,104 @@ jobs:
       python_version:
         type: string
         default: "3.9"
-      extra_env:
+      build_env:
         type: string
-        default: "DEV_RELEASE=1"
-    docker:
-      - image: ubuntu:20.04
+        default: ""
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
           name: Build wheel
           command: |
             PYTHON=python<< parameters.python_version >>
-            apt-get update
-            apt-get upgrade -y
-            DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata
-            apt-get install -y apt-utils
-            apt-get install -y software-properties-common
-            add-apt-repository -y ppa:deadsnakes/ppa
-            apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            apt-get install -y build-essential git
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
+            sudo apt-get update
+            sudo apt-get upgrade -y
+            TZ=Etc/UTC sudo apt-get -y install tzdata
+            sudo apt-get install -y apt-utils
+            sudo apt-get install -y software-properties-common
+            sudo add-apt-repository -y ppa:deadsnakes/ppa
+            sudo apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install -y build-essential git
             $PYTHON -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              pip install . -v
+            << parameters.build_env >> pip install ".[dev]" -v
             pip install typing_extensions
             python setup.py generate_stubs
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python -m build --wheel
-            auditwheel show dist/*
-            auditwheel repair dist/* --plat manylinux_2_31_x86_64
-      - run:
-          name: Upload package
-          command: |
-            source env/bin/activate
-            twine upload wheelhouse/*
+            python setup.py clean --all
+            MLX_BUILD_STAGE=1 << parameters.build_env >> python -m build -w
+            bash python/scripts/repair_linux.sh
+      - when:
+          condition:
+            equal: ["3.9", << parameters.python_version >>]
+          steps:
+            - run:
+                name: Build common package
+                command: |
+                  source env/bin/activate
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 \
+                    python -m build -w
+                  auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_x86_64
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload packages
+                command: |
+                  source env/bin/activate
+                  twine upload wheelhouse/*.whl
       - store_artifacts:
           path: wheelhouse/
 
   build_cuda_release:
     parameters:
-      python_version:
+      build_env:
         type: string
-        default: "3.9"
-      extra_env:
-        type: string
-        default: "DEV_RELEASE=1"
+        default: ""
     machine:
-      image: linux-cuda-12:default
-      resource_class: gpu.nvidia.small.gen2
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
           name: Build wheel
           command: |
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
             sudo apt-get update
+            sudo apt-get install cuda-toolkit-12-9 libcudnn9-dev-cuda-12
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            python -m venv env
-            source env/bin/activate
+            sudo apt-get install zip
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+            export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
+            export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
+            << parameters.build_env >> MLX_BUILD_STAGE=2 \
               CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              pip install ".[dev]" -v
-            python setup.py generate_stubs
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              python -m build --wheel
+              python -m build -w
             bash python/scripts/repair_cuda.sh
-      - run:
-          name: Upload package
-          command: |
-            source env/bin/activate
-            twine upload wheelhouse/*.whl
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload package
+                command: |
+                  twine upload wheelhouse/*.whl
       - store_artifacts:
           path: wheelhouse/
 
@@ -408,7 +417,6 @@ workflows:
             pattern: "^(?!pull/)[-\\w]+$"
             value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
         - not: << pipeline.parameters.test_release >>
     jobs:
       - mac_build_and_test:
@@ -416,14 +424,16 @@ workflows:
             parameters:
               macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test
-      - cuda_build_and_test 
+      - cuda_build_and_test:
+          matrix:
+            parameters:
+              image_date: ["2023.11.1", "2025.05.1"]
       - build_documentation 
 
   build_pypi_release:
     when:
       and:
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
         - not: << pipeline.parameters.test_release >>
     jobs:
       - build_release:
@@ -506,6 +516,25 @@ workflows:
             branches:
               ignore: /.*/
           upload-docs: true
+      - build_linux_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              build_env: ["PYPI_RELEASE=1"]
+      - build_cuda_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              build_env: ["PYPI_RELEASE=1"]
 
   prb:
     when:
@@ -584,11 +613,17 @@ workflows:
               - macosx_deployment_target: "15.0"
                 xcode_version: "15.0.0"
                 python_version: "3.13"
-  weekly_build:
+      - build_linux_release:
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+      - build_cuda_release
+
+  build_dev_release:
     when:
       and:
         - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.weekly_build >>
+        - << pipeline.parameters.test_release >>
     jobs:
       - build_release:
           matrix:
@@ -658,25 +693,12 @@ workflows:
                 xcode_version: "15.0.0"
                 python_version: "3.13"
                 build_env: "DEV_RELEASE=1"
-  linux_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.linux_release >>
-    jobs:
       - build_linux_release:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]
-  cuda_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.cuda_release >>
-    jobs:
+              build_env: ["DEV_RELEASE=1"]
       - build_cuda_release:
           matrix:
             parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]
+              build_env: ["DEV_RELEASE=1"]

ACKNOWLEDGMENTS.md

@@ -19,6 +19,7 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />

CMakeLists.txt

@@ -64,10 +64,8 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
       message(WARNING "Building for x86_64 arch is not officially supported.")
     endif()
   endif()
-
 else()
   set(MLX_BUILD_METAL OFF)
-  message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
 endif()
 
 # ----------------------------- Lib -----------------------------

README.md

@@ -11,10 +11,10 @@ brought to you by Apple machine learning research.
 
 Some key features of MLX include:
 
- - **Familiar APIs**: MLX has a Python API that closely follows NumPy.  MLX
+ - **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
    also has fully featured C++, [C](https://github.com/ml-explore/mlx-c), and
    [Swift](https://github.com/ml-explore/mlx-swift/) APIs, which closely mirror
-   the Python API.  MLX has higher-level packages like `mlx.nn` and
+   the Python API. MLX has higher-level packages like `mlx.nn` and
    `mlx.optimizers` with APIs that closely follow PyTorch to simplify building
    more complex models.
 
@@ -68,18 +68,23 @@ in the documentation.
 
 ## Installation
 
-MLX is available on [PyPI](https://pypi.org/project/mlx/). To install the Python API, run:
+MLX is available on [PyPI](https://pypi.org/project/mlx/). To install MLX on
+macOS, run:
 
-**With `pip`**:
-
-```
+```bash
 pip install mlx
 ```
 
-**With `conda`**:
+To install the CUDA backend on Linux, run:
 
+```bash
+pip install "mlx[cuda]"
 ```
-conda install -c conda-forge mlx
+
+To install a CPU-only Linux package, run:
+
+```bash
+pip install "mlx[cpu]"
 ```
 
 Checkout the

benchmarks/cpp/single_ops.cpp

@@ -203,6 +203,11 @@ void time_reductions() {
   TIME(max_along_0);
   auto max_along_1 = [&b]() { return mx::max(b, 1, false); };
   TIME(max_along_1);
+
+  auto min_along_0 = [&b]() { return mx::min(b, 0, false); };
+  TIME(min_along_0);
+  auto min_along_1 = [&b]() { return mx::min(b, 1, false); };
+  TIME(min_along_1);
 }
 
 void time_gather_scatter() {

benchmarks/python/single_ops.py

@@ -58,6 +58,13 @@ def time_max():
     time_fn(mx.max, a, 0)
 
 
+def time_min():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    a[1, 1] = mx.nan
+    mx.eval(a)
+    time_fn(mx.min, a, 0)
+
+
 def time_negative():
     a = mx.random.uniform(shape=(10000, 1000))
     mx.eval(a)
@@ -115,6 +122,7 @@ if __name__ == "__main__":
 
     time_add()
     time_matmul()
+    time_min()
     time_max()
     time_maximum()
     time_exp()

docs/src/dev/extensions.rst

@@ -138,13 +138,13 @@ more concrete:
         * representing the vectorized computation and the axis which
         * corresponds to the output vectorized dimension.
         */
-        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
+        std::pair<std::vector<array>, std::vector<int>> vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) override;
 
-        /** Print the primitive. */
-        void print(std::ostream& os) override {
-            os << "Axpby";
+        /** The name of primitive. */
+        const char* name() const override {
+          return "Axpby";
         }
 
         /** Equivalence check **/

docs/src/install.rst

@@ -13,7 +13,7 @@ silicon computer is
 
     pip install mlx
 
-To install from PyPI you must meet the following requirements:
+To install from PyPI your system must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
 - Using a native Python >= 3.9
@@ -23,22 +23,39 @@ To install from PyPI you must meet the following requirements:
     MLX is only available on devices running macOS >= 13.5
     It is highly recommended to use macOS 14 (Sonoma)
 
-
-MLX is also available on conda-forge. To install MLX with conda do:
-
-.. code-block:: shell
-
-   conda install conda-forge::mlx
-
 CUDA
 ^^^^
 
-MLX has a CUDA backend which you can use on any Linux platform with CUDA 12
-and SM 7.0 (Volta) and up. To install MLX with CUDA support, run:
+MLX has a CUDA backend which you can install with:
 
 .. code-block:: shell
 
-    pip install mlx-cuda
+    pip install "mlx[cuda]"
+
+To install the CUDA package from PyPi your system must meet the following
+requirements:
+
+- Nvidia architecture >= SM 7.0 (Volta)
+- Nvidia driver >= 550.54.14
+- CUDA toolkit >= 12.0
+- Linux distribution with glibc >= 2.35
+- Python >= 3.9
+
+
+CPU-only (Linux)
+^^^^^^^^^^^^^^^^
+
+For a CPU-only version of MLX that runs on Linux use:
+
+.. code-block:: shell
+
+    pip install "mlx[cpu]"
+
+To install the CPU-only package from PyPi your system must meet the following
+requirements:
+
+- Linux distribution with glibc >= 2.35
+- Python >= 3.9
 
 
 Troubleshooting
@@ -88,20 +105,20 @@ Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
+  pip install .
 
 For developing, install the package with development dependencies, and use an
 editable install:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
+  pip install -e ".[dev]"
 
 Once the development dependencies are installed, you can build faster with:
 
 .. code-block:: shell
 
- CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
+ python setup.py build_ext --inplace
 
 Run the tests with:
 
@@ -262,7 +279,7 @@ When building either the Python or C++ APIs make sure to pass the cmake flag
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON" pip install -e ".[dev]"
+  CMAKE_ARGS="-DMLX_BUILD_CUDA=ON" pip install -e ".[dev]"
 
 To build the C++ package run:
 

docs/src/python/optimizers/common_optimizers.rst

@@ -19,3 +19,4 @@ Common Optimizers
    Adamax
    Lion
    MultiOptimizer
+   Muon

examples/extensions/axpby/axpby.h

@@ -74,9 +74,9 @@ class Axpby : public mx::Primitive {
       const std::vector<mx::array>& inputs,
       const std::vector<int>& axes) override;
 
-  /** Print the primitive. */
-  void print(std::ostream& os) override {
-    os << "Axpby";
+  /** The name of primitive. */
+  const char* name() const override {
+    return "Axpby";
   }
 
   /** Equivalence check **/

mlx/backend/common/utils.cpp

@@ -1,14 +1,20 @@
 // Copyright © 2023-2024 Apple Inc.
 
+#include <dlfcn.h>
+
 #include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive) {
-  std::ostringstream op_t;
-  primitive->print(op_t);
-  return op_t.str();
+std::filesystem::path current_binary_dir() {
+  static std::filesystem::path binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path();
+  }();
+  return binary_dir;
 }
 
 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(

mlx/backend/common/utils.h

@@ -2,6 +2,7 @@
 
 #pragma once
 
+#include <filesystem>
 #include <tuple>
 #include <vector>
 
@@ -9,7 +10,8 @@
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive);
+// Return the directory that contains current shared library.
+std::filesystem::path current_binary_dir();
 
 inline int64_t
 elem_to_loc(int elem, const Shape& shape, const Strides& strides) {

mlx/backend/cpu/cholesky.cpp

@@ -20,7 +20,7 @@ void cholesky_impl(const array& a, array& factor, bool upper, Stream stream) {
 
   // The decomposition is computed in place, so just copy the input to the
   // output.
-  copy(
+  copy_cpu(
       a,
       factor,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/compiled.cpp

@@ -231,7 +231,7 @@ inline void build_kernel(
       os << "static_cast<" << get_type_string(x.dtype()) << ">(tmp_"
          << namer.get_name(x.inputs()[0]) << ");" << std::endl;
     } else {
-      x.primitive().print(os);
+      os << x.primitive().name();
       os << "()(";
       for (int i = 0; i < x.inputs().size() - 1; i++) {
         os << "tmp_" << namer.get_name(x.inputs()[i]) << ", ";

mlx/backend/cpu/conv.cpp

@@ -883,7 +883,7 @@ void explicit_gemm_conv_1D_cpu(
   // Fill with zeros
   std::vector<array> temps;
   temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = padding_lo[0] * in_padded.strides()[1];
@@ -895,7 +895,7 @@ void explicit_gemm_conv_1D_cpu(
       in_padded_slice.size(),
       data_offset);
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -920,7 +920,7 @@ void explicit_gemm_conv_1D_cpu(
   // Materialize strided view
   Shape strided_reshape = {N * oH, wH * C};
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -938,13 +938,13 @@ void explicit_gemm_conv_1D_cpu(
         wt.size(),
         0);
     gemm_wt = array(wt_transpose.shape(), float32, nullptr, {});
-    copy(wt_transpose, gemm_wt, CopyType::General, stream);
+    copy_cpu(wt_transpose, gemm_wt, CopyType::General, stream);
     temps.push_back(gemm_wt);
   } else if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
@@ -991,7 +991,7 @@ void explicit_gemm_conv_1D_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }
@@ -1029,7 +1029,7 @@ void explicit_gemm_conv_2D_cpu(
   // Fill with zeros
   std::vector<array> temps;
   temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = padding_lo[0] * in_padded.strides()[1] +
@@ -1044,7 +1044,7 @@ void explicit_gemm_conv_2D_cpu(
   temps.push_back(in_padded_slice);
 
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
 
   // Make strided view
   Shape strided_shape = {N, oH, oW, wH, wW, C};
@@ -1065,7 +1065,7 @@ void explicit_gemm_conv_2D_cpu(
   // Materialize strided view
   Shape strided_reshape = {N * oH * oW, wH * wW * C};
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -1076,7 +1076,7 @@ void explicit_gemm_conv_2D_cpu(
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
@@ -1116,7 +1116,7 @@ void explicit_gemm_conv_2D_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }
@@ -1156,7 +1156,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Fill with zeros
   std::vector<array> temps = {array(0, conv_dtype)};
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = 0;
@@ -1173,7 +1173,7 @@ void explicit_gemm_conv_ND_cpu(
       data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -1212,7 +1212,7 @@ void explicit_gemm_conv_ND_cpu(
   }
 
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -1223,13 +1223,13 @@ void explicit_gemm_conv_ND_cpu(
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
   if (flip) {
     auto gemm_wt_ = array(gemm_wt.shape(), float32, nullptr, {});
-    copy(gemm_wt, gemm_wt_, CopyType::Vector, stream);
+    copy_cpu(gemm_wt, gemm_wt_, CopyType::Vector, stream);
     temps.push_back(gemm_wt_);
 
     // Calculate the total size of the spatial dimensions
@@ -1284,7 +1284,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }

mlx/backend/cpu/copy.cpp

@@ -295,7 +295,11 @@ inline void copy_inplace_dispatch(
 
 } // namespace
 
-void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu_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);
@@ -305,7 +309,7 @@ void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
        ctype]() mutable { copy_inplace_dispatch(src, dst, ctype); });
 }
 
-void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu(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
@@ -315,10 +319,10 @@ void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;
   }
-  copy_inplace(src, dst, ctype, stream);
+  copy_cpu_inplace(src, dst, ctype, stream);
 }
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,
@@ -373,4 +377,10 @@ void copy_inplace(
       });
 }
 
+array contiguous_copy_cpu(const array& arr, Stream stream) {
+  array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+  copy_cpu(arr, arr_copy, CopyType::General, stream);
+  return arr_copy;
+}
+
 } // namespace mlx::core

mlx/backend/cpu/copy.h

@@ -10,10 +10,14 @@
 
 namespace mlx::core {
 
-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_cpu(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_cpu_inplace(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Stream stream);
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,
@@ -26,4 +30,7 @@ void copy_inplace(
     const std::optional<array>& dynamic_i_offset = std::nullopt,
     const std::optional<array>& dynamic_o_offset = std::nullopt);
 
+// Return a contiguous array with same shape that copies the data of |arr|.
+array contiguous_copy_cpu(const array& arr, Stream stream);
+
 } // namespace mlx::core

mlx/backend/cpu/distributed.cpp

@@ -13,9 +13,7 @@ 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};
+    return {contiguous_copy_cpu(arr, stream), true};
   }
 };
 
@@ -34,8 +32,7 @@ void AllReduce::eval_cpu(
       }
       return in;
     } else {
-      array arr_copy(in.shape(), in.dtype(), nullptr, {});
-      copy(in, arr_copy, CopyType::General, s);
+      array arr_copy = contiguous_copy_cpu(in, s);
       out.copy_shared_buffer(arr_copy);
       return arr_copy;
     }

mlx/backend/cpu/eig.cpp

@@ -135,7 +135,7 @@ void Eig::eval_cpu(
       : array(a.shape(), complex64, nullptr, {});
 
   auto a_copy = array(a.shape(), a.dtype(), nullptr, {});
-  copy(
+  copy_cpu(
       a,
       a_copy,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/eigh.cpp

@@ -196,7 +196,7 @@ void Eigh::eval_cpu(
 
   values.set_data(allocator::malloc(values.nbytes()));
 
-  copy(
+  copy_cpu(
       a,
       vectors,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/hadamard.cpp

@@ -96,7 +96,7 @@ void Hadamard::eval_cpu(const std::vector<array>& inputs, array& out) {
   if (in.flags().row_contiguous && in.is_donatable()) {
     out.copy_shared_buffer(in);
   } else {
-    copy(
+    copy_cpu(
         in,
         out,
         in.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/indexing.cpp

@@ -517,7 +517,7 @@ void Scatter::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, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   std::vector<array> inds;
@@ -686,7 +686,7 @@ 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, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(idx);

mlx/backend/cpu/inverse.cpp

@@ -115,7 +115,7 @@ void inverse_impl(
   //   (A⁻¹)ᵀ = (Aᵀ)⁻¹
 
   // The inverse is computed in place, so just copy the input to the output.
-  copy(
+  copy_cpu(
       a,
       inv,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/logsumexp.cpp

@@ -87,8 +87,7 @@ void LogSumExp::eval_cpu(const std::vector<array>& inputs, array& out) {
     if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_cpu(x, s);
       encoder.add_temporary(x_copy);
       return x_copy;
     }

mlx/backend/cpu/luf.cpp

@@ -31,7 +31,7 @@ void luf_impl(
   strides[ndim - 1] = M;
   strides[ndim - 2] = 1;
   lu.set_data(allocator::malloc(lu.nbytes()), lu.nbytes(), strides, flags);
-  copy_inplace(
+  copy_cpu_inplace(
       a,
       lu,
       a.shape(),

mlx/backend/cpu/masked_mm.cpp

@@ -6,6 +6,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/gemm.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
@@ -52,6 +53,58 @@ inline void mask_matrix(
   }
 }
 
+template <typename T>
+inline void segmented_mm(
+    const T* a,
+    const T* b,
+    const uint32_t* segments,
+    T* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides,
+    size_t num_segments,
+    const Shape& segments_shape,
+    const Strides& segments_strides) {
+  int ndim = a_shape.size();
+  Shape a_copy = a_shape;
+  Shape b_copy = b_shape;
+  int32_t M = a_copy[ndim - 2];
+  int32_t N = b_copy[ndim - 1];
+  for (int i = 0; i < num_segments; i++) {
+    uint32_t k_start =
+        segments[elem_to_loc(2 * i, segments_shape, segments_strides)];
+    uint32_t k_end =
+        segments[elem_to_loc(2 * i + 1, segments_shape, segments_strides)];
+    if (k_end <= k_start) {
+      std::fill_n(out + i * M * N, M * N, T(0));
+      continue;
+    }
+    a_copy[ndim - 1] = k_end - k_start;
+    b_copy[ndim - 2] = k_end - k_start;
+    matmul<T>(
+        a + k_start * a_strides[ndim - 1],
+        b + k_start * b_strides[ndim - 2],
+        out + i * M * N,
+        a_transposed,
+        b_transposed,
+        lda,
+        ldb,
+        N,
+        1.0,
+        0.0,
+        1,
+        a_copy,
+        a_strides,
+        b_copy,
+        b_strides);
+  }
+}
+
 } // namespace
 
 void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -71,21 +124,20 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
         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, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(false, stx, arr_copy, true);
           }
           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, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(true, sty, arr_copy, true);
           }
           return std::make_tuple(true, sty, arr, false);
         } else {
-          array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy(arr, arr_copy, CopyType::General, s);
           int64_t stx = arr.shape(-1);
+          array arr_copy = contiguous_copy_cpu(arr, s);
           return std::make_tuple(false, stx, arr_copy, true);
         }
       };
@@ -333,7 +385,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       int64_t stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -437,4 +489,121 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   encoder.add_temporaries(std::move(temps));
 }
 
+void SegmentedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto check_transpose = [&s, &encoder](const array& x) {
+    auto stx = x.strides()[x.ndim() - 2];
+    auto sty = x.strides()[x.ndim() - 1];
+    if (stx == x.shape(-1) && sty == 1) {
+      return std::make_tuple(false, stx, x);
+    } else if (stx == 1 && sty == x.shape(-2)) {
+      return std::make_tuple(true, sty, x);
+    } else {
+      array xc(x.shape(), x.dtype(), nullptr, {});
+      copy_cpu(x, xc, CopyType::General, s);
+      encoder.add_temporary(xc);
+      int64_t stx = x.shape(-1);
+      return std::make_tuple(false, stx, xc);
+    }
+  };
+
+  auto [a_transposed, lda, a] = check_transpose(inputs[0]);
+  auto [b_transposed, ldb, b] = check_transpose(inputs[1]);
+  auto& segments = inputs[2];
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(segments);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    segments = array::unsafe_weak_copy(segments),
+                    out_ptr = out.data<void>(),
+                    a_transposed = a_transposed,
+                    b_transposed = b_transposed,
+                    lda = lda,
+                    ldb = ldb]() {
+    switch (a.dtype()) {
+      case float64:
+        segmented_mm<double>(
+            a.data<double>(),
+            b.data<double>(),
+            segments.data<uint32_t>(),
+            static_cast<double*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float32:
+        segmented_mm<float>(
+            a.data<float>(),
+            b.data<float>(),
+            segments.data<uint32_t>(),
+            static_cast<float*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float16:
+        segmented_mm<float16_t>(
+            a.data<float16_t>(),
+            b.data<float16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<float16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case bfloat16:
+        segmented_mm<bfloat16_t>(
+            a.data<bfloat16_t>(),
+            b.data<bfloat16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<bfloat16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      default:
+        throw std::invalid_argument(
+            "Segmented mm supports only real float types.");
+    }
+  });
+}
+
 } // namespace mlx::core

mlx/backend/cpu/matmul.cpp

@@ -81,7 +81,7 @@ void matmul_general(
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, stream);
+      copy_cpu(arr, temps.back(), CopyType::General, stream);
       stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -142,7 +142,7 @@ 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, stream());
+  copy_cpu(c, out, ctype, stream());
   if (inputs[0].shape(-1) == 0) {
     return;
   }

mlx/backend/cpu/primitives.cpp

@@ -22,7 +22,7 @@ void reshape(const array& in, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
     out.set_data(allocator::malloc(out.nbytes()));
-    copy_inplace(in, out, CopyType::General, out.primitive().stream());
+    copy_cpu_inplace(in, out, CopyType::General, out.primitive().stream());
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -175,7 +175,7 @@ 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, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -198,7 +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, stream());
+    copy_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
   }
 }
 
@@ -211,7 +211,7 @@ 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, stream());
+    copy_cpu(in, out, CopyType::General, stream());
   }
 }
 
@@ -235,7 +235,7 @@ void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   } else {
     ctype = CopyType::General;
   }
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -251,7 +251,7 @@ 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, stream());
+  copy_cpu(val, out, CopyType::Scalar, stream());
 
   // Find offset for start of input values
   size_t data_offset = 0;
@@ -266,7 +266,7 @@ 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, stream());
+  copy_cpu_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
 }
 
 void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -340,7 +340,7 @@ void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc(out.nbytes()));
   auto [in_offset, donated] =
       compute_dynamic_offset(inputs[1], in.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ in,
       /* array& dst = */ out,
       /* const Shape& data_shape = */ out.shape(),
@@ -372,11 +372,11 @@ 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, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   auto [out_offset, donated] =
       compute_dynamic_offset(inputs[2], out.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -412,14 +412,14 @@ 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, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   // Calculate out strides, initial offset and if copy needs to be made
   auto [data_offset, out_strides] =
       prepare_slice(out, start_indices_, strides_);
 
   // Do copy
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -456,9 +456,9 @@ 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, stream());
+      copy_cpu_inplace(in_tmp, tmp, CopyType::General, stream());
     } else {
-      copy_inplace(in, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in, tmp, CopyType::General, stream());
     }
 
     auto flags = out.flags();

mlx/backend/cpu/qrf.cpp

@@ -26,7 +26,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
   strides[in.ndim() - 2] = 1;
   strides[in.ndim() - 1] = M;
   in.set_data(allocator::malloc(in.nbytes()), in.nbytes(), strides, flags);
-  copy_inplace(a, in, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(a, in, CopyType::GeneralGeneral, stream);
   auto& encoder = cpu::get_command_encoder(stream);
   q.set_data(allocator::malloc(q.nbytes()));
   r.set_data(allocator::malloc(r.nbytes()));

mlx/backend/cpu/quantized.cpp

@@ -529,7 +529,7 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
       return arr;
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       return temps.back();
     }
   };
@@ -579,7 +579,7 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return arr;
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       return temps.back();
     }
   };
@@ -712,9 +712,7 @@ void fast::AffineQuantize::eval_cpu(
     if (arr.flags().row_contiguous) {
       return std::make_pair(arr, false);
     } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General, s);
-      return std::make_pair(arr_copy, true);
+      return std::make_pair(contiguous_copy_cpu(arr, s), true);
     }
   };
 

mlx/backend/cpu/reduce.cpp

@@ -350,7 +350,15 @@ struct MinReduce {
   };
 
   template <int N, typename T>
-  T operator()(simd::Simd<T, N> x) {
+  std::enable_if_t<std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    return simd::min(x);
+  };
+
+  template <int N, typename T>
+  std::enable_if_t<!std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    if (simd::any(x != x)) {
+      return static_cast<T>(NAN);
+    }
     return simd::min(x);
   };
 };

mlx/backend/cpu/scan.cpp

@@ -250,10 +250,8 @@ void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
   // 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, stream());
-    in = arr_copy;
-    encoder.add_temporary(arr_copy);
+    in = contiguous_copy_cpu(in, stream());
+    encoder.add_temporary(in);
   }
   out.set_data(allocator::malloc(out.nbytes()));
 

mlx/backend/cpu/softmax.cpp

@@ -131,8 +131,7 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
       }
       return x;
     } else {
-      array x_copy(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_cpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }

mlx/backend/cpu/sort.cpp

@@ -334,8 +334,10 @@ void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);
@@ -426,8 +428,10 @@ void Partition::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);

mlx/backend/cpu/svd.cpp

@@ -31,7 +31,7 @@ void svd_impl(
 
   // lapack clobbers the input, so we have to make a copy.
   array in(a.shape(), a.dtype(), nullptr, {});
-  copy(
+  copy_cpu(
       a,
       in,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cuda/CMakeLists.txt

@@ -15,11 +15,14 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_dynamic.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_input.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/gemv.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/jit_module.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/kernel_utils.cu
@@ -35,14 +38,24 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/sort.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.9.0)
+  target_sources(
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_batched_gemm_12_9.cu)
+else()
+  target_sources(
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_batched_gemm_12_0.cpp)
+endif()
+
 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
 
 # Embed kernel sources in binary for JIT compilation.
@@ -67,6 +80,11 @@ target_include_directories(mlx PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/gen")
 target_compile_options(mlx
                        PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>")
 
+# Enable calling host constexpr functions from device. This is needed because
+# the constexpr version of isnan is host only.
+target_compile_options(
+  mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>")
+
 # CUDA 12.8 emits warning #20280-D for copy kernels which is a false positive.
 # Explicitly pass this flag to suppress the warning, it is safe to set it to
 # true but the warning wouldn't be suppressed.
@@ -80,6 +98,13 @@ endif()
 target_compile_options(
   mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--Wno-deprecated-gpu-targets>")
 
+# Use stronger binaries compression. This feature was introduced in CUDA 12.8
+# and requires drivers released after CUDA 12.4.
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
+  target_compile_options(
+    mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--compress-mode=size>")
+endif()
+
 # Compute capability 7 is required for synchronization between CPU/GPU with
 # managed memory. TODO: Add more architectures for potential performance gain.
 set(MLX_CUDA_ARCHITECTURES
@@ -116,6 +141,27 @@ target_link_libraries(mlx PRIVATE CUDA::cublasLt)
 # Use NVRTC and driver APIs.
 target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 
+# Use the frontend APIs of cuDNN.
+FetchContent_Declare(
+  cudnn
+  GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
+  GIT_TAG v1.12.1
+  GIT_SHALLOW TRUE
+  EXCLUDE_FROM_ALL)
+set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)
+set(CUDNN_FRONTEND_BUILD_SAMPLES OFF)
+set(CUDNN_FRONTEND_BUILD_TESTS OFF)
+set(CUDNN_FRONTEND_BUILD_PYTHON_BINDINGS OFF)
+FetchContent_MakeAvailable(cudnn)
+target_link_libraries(mlx PRIVATE cudnn_frontend)
+# Link with the actual cuDNN libraries.
+include(${cudnn_frontend_SOURCE_DIR}/cmake/cuDNN.cmake)
+target_link_libraries(mlx PRIVATE CUDNN::cudnn_all)
+
 # Suppress nvcc warnings on MLX headers.
 target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
                                    --diag_suppress=997>)
+
+# Install CCCL headers for JIT.
+install(DIRECTORY ${cccl_SOURCE_DIR}/include/cuda
+        DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/cccl)

mlx/backend/cuda/allocator.cpp

@@ -2,7 +2,6 @@
 
 #include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/utils.h"
-#include "mlx/backend/cuda/worker.h"
 #include "mlx/utils.h"
 
 #include <cuda_runtime.h>
@@ -17,14 +16,66 @@ namespace cu {
 
 constexpr int page_size = 16384;
 
+// Any allocations smaller than this will try to use the small pool
+constexpr int small_block_size = 8;
+
+// The small pool size in bytes. This should be a multiple of the host page
+// size and small_block_size.
+constexpr int small_pool_size = 4 * page_size;
+
+SmallSizePool::SmallSizePool() {
+  auto num_blocks = small_pool_size / small_block_size;
+  buffer_ = new Block[num_blocks];
+
+  next_free_ = buffer_;
+
+  CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
+  CHECK_CUDA_ERROR(
+      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, 0));
+
+  auto curr = next_free_;
+  for (size_t i = 1; i < num_blocks; ++i) {
+    curr->next = buffer_ + i;
+    curr = curr->next;
+  }
+  curr->next = nullptr;
+}
+
+SmallSizePool::~SmallSizePool() {
+  CHECK_CUDA_ERROR(cudaFree(data_));
+  delete[] buffer_;
+}
+
+CudaBuffer* SmallSizePool::malloc() {
+  if (next_free_ == nullptr) {
+    return nullptr;
+  }
+  Block* b = next_free_;
+  uint64_t i = next_free_ - buffer_;
+  next_free_ = next_free_->next;
+  b->buf.data = static_cast<char*>(data_) + i * small_block_size;
+  b->buf.size = small_block_size;
+  return &b->buf;
+}
+
+void SmallSizePool::free(CudaBuffer* buf) {
+  auto b = reinterpret_cast<Block*>(buf);
+  b->next = next_free_;
+  next_free_ = b;
+}
+
+bool SmallSizePool::in_pool(CudaBuffer* buf) {
+  constexpr int num_blocks = (small_pool_size / small_block_size);
+  auto b = reinterpret_cast<Block*>(buf);
+  int64_t block_num = b - buffer_;
+  return block_num >= 0 && block_num < num_blocks;
+}
+
 CudaAllocator::CudaAllocator()
     : buffer_cache_(
           page_size,
           [](CudaBuffer* buf) { return buf->size; },
-          [this](CudaBuffer* buf) {
-            cuda_free(buf->data);
-            delete buf;
-          }) {
+          [this](CudaBuffer* buf) { cuda_free(buf); }) {
   // TODO: Set memory limit for multi-device.
   size_t free, total;
   CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
@@ -36,7 +87,9 @@ Buffer CudaAllocator::malloc(size_t size) {
   // Find available buffer from cache.
   auto orig_size = size;
   std::unique_lock lock(mutex_);
-  if (size < page_size) {
+  if (size <= small_block_size) {
+    size = 8;
+  } else if (size < page_size) {
     size = next_power_of_2(size);
   } else {
     size = page_size * ((size + page_size - 1) / page_size);
@@ -44,19 +97,25 @@ Buffer CudaAllocator::malloc(size_t size) {
 
   CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
   if (!buf) {
-    // If we have a lot of memory pressure or are over the maximum cache size,
-    // try to reclaim memory from the cache.
-    size_t mem_required = get_active_memory() + get_cache_memory() + size;
-    if (mem_required >= memory_limit_) {
-      buffer_cache_.release_cached_buffers(mem_required - memory_limit_);
+    // If we have a lot of memory pressure try to reclaim memory from the cache.
+    int64_t mem_to_free =
+        get_active_memory() + get_cache_memory() + size - memory_limit_;
+    if (mem_to_free > 0) {
+      buffer_cache_.release_cached_buffers(mem_to_free);
     }
 
+    // Try the scalar pool first
+    if (size <= small_block_size) {
+      buf = scalar_pool_.malloc();
+    }
     lock.unlock();
-    buf = new CudaBuffer{nullptr, size};
-    cudaError_t err = cudaMallocManaged(&buf->data, size);
-    if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
-      throw std::runtime_error(fmt::format(
-          "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+    if (!buf) {
+      buf = new CudaBuffer{nullptr, size};
+      cudaError_t err = cudaMallocManaged(&buf->data, size);
+      if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
+        throw std::runtime_error(fmt::format(
+            "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+      }
     }
     lock.lock();
   }
@@ -67,7 +126,6 @@ Buffer CudaAllocator::malloc(size_t size) {
   if (get_cache_memory() > max_pool_size_) {
     buffer_cache_.release_cached_buffers(get_cache_memory() - max_pool_size_);
   }
-
   return Buffer{buf};
 }
 
@@ -82,9 +140,7 @@ void CudaAllocator::free(Buffer buffer) {
   if (get_cache_memory() < max_pool_size_) {
     buffer_cache_.recycle_to_cache(buf);
   } else {
-    lock.unlock();
-    cuda_free(buf->data);
-    delete buf;
+    cuda_free(buf);
   }
 }
 
@@ -96,27 +152,14 @@ size_t CudaAllocator::size(Buffer buffer) const {
   return buf->size;
 }
 
-void CudaAllocator::register_this_thread() {
-  std::lock_guard lock(worker_mutex_);
-  allowed_threads_.insert(std::this_thread::get_id());
-}
-
-void CudaAllocator::cuda_free(void* buf) {
-  // If cuda_free() is called from a unregistered thread, reschedule the call to
-  // worker.
-  {
-    std::lock_guard lock(worker_mutex_);
-    if (allowed_threads_.count(std::this_thread::get_id()) == 0) {
-      if (!worker_) {
-        worker_.reset(new Worker);
-      }
-      worker_->add_task([this, buf]() { this->cuda_free(buf); });
-      worker_->end_batch();
-      worker_->commit();
-      return;
-    }
+// This must be called with mutex_ aquired
+void CudaAllocator::cuda_free(CudaBuffer* buf) {
+  if (scalar_pool_.in_pool(buf)) {
+    scalar_pool_.free(buf);
+  } else {
+    cudaFree(buf->data);
+    delete buf;
   }
-  cudaFree(buf);
 }
 
 size_t CudaAllocator::get_active_memory() const {

mlx/backend/cuda/allocator.h

@@ -7,13 +7,10 @@
 
 #include <mutex>
 #include <set>
-#include <thread>
 #include <utility>
 
 namespace mlx::core::cu {
 
-class Worker;
-
 using allocator::Buffer;
 
 // Stores cuda-managed unified memory.
@@ -22,21 +19,35 @@ struct CudaBuffer {
   size_t size;
 };
 
+class SmallSizePool {
+ private:
+  union Block {
+    Block* next;
+    CudaBuffer buf;
+  };
+
+  Block* buffer_{nullptr};
+  void* data_{nullptr};
+  Block* next_free_{nullptr};
+
+ public:
+  SmallSizePool();
+  ~SmallSizePool();
+
+  SmallSizePool(const SmallSizePool&) = delete;
+  SmallSizePool& operator=(const SmallSizePool&) = delete;
+
+  CudaBuffer* malloc();
+  void free(CudaBuffer* buf);
+  bool in_pool(CudaBuffer* buf);
+};
+
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
   void free(Buffer buffer) override;
   size_t size(Buffer buffer) const override;
 
-  // Register current thread as safe to free buffers.
-  // In cuda freeing a buffer implicitly synchronizes stream, and for threads
-  // that may be waited by gpu stream (for example cpu stream threads), freeing
-  // buffers there would result in dead lock.
-  void register_this_thread();
-
-  // Call cudaFree in the safe thread.
-  void cuda_free(void* buf);
-
   size_t get_active_memory() const;
   size_t get_peak_memory() const;
   void reset_peak_memory();
@@ -47,19 +58,18 @@ class CudaAllocator : public allocator::Allocator {
   void clear_cache();
 
  private:
+  void cuda_free(CudaBuffer* buf);
+
   CudaAllocator();
   friend CudaAllocator& allocator();
 
-  std::mutex worker_mutex_;
-  std::unique_ptr<Worker> worker_;
-  std::set<std::thread::id> allowed_threads_;
-
   std::mutex mutex_;
   size_t memory_limit_;
   size_t max_pool_size_;
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
+  SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();

mlx/backend/cuda/arg_reduce.cu

@@ -1,7 +1,8 @@
 // Copyright © 2025 Apple Inc.
+
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
+#include "mlx/backend/cuda/device/fp16_math.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -114,7 +115,7 @@ __global__ void arg_reduce_general(
     T vals[N_READS];
     auto tid = r * BLOCK_DIM + block.thread_index().x;
     cub::LoadDirectBlocked(
-        tid, strided_iterator(in + in_idx, axis_stride), vals, axis_size, init);
+        tid, StridedIterator(in + in_idx, axis_stride), vals, axis_size, init);
     best = op.reduce_many(best, vals, tid * N_READS);
   }
 

mlx/backend/cuda/binary.cu

@@ -3,7 +3,6 @@
 #include "mlx/backend/common/binary.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/device/binary_ops.cuh"
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -17,35 +16,86 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_ss(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[0], b[0]);
+
+  if ((index + 1) * N_READS > size) {
+    for (int i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a[0], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_sv(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[0], b[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[i]);
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a[0], b_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_vs(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[0]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[0]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[i]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -78,7 +128,7 @@ __global__ void binary_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_4d(
+    auto [a_idx, b_idx] = elem_to_loc(
         index, shape.data(), a_strides.data(), b_strides.data(), ndim);
     out[index] = Op{}(a[a_idx], b[b_idx]);
   }
@@ -126,7 +176,7 @@ template <typename Op>
 void binary_op_gpu_inplace(
     const std::vector<array>& inputs,
     array& out,
-    std::string_view op,
+    const char* op,
     const Stream& s) {
   assert(inputs.size() > 1);
   const auto& a = inputs[0];
@@ -196,18 +246,25 @@ void binary_op_gpu_inplace(
                 }
               });
         } else {
-          dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
+          dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
             using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            auto kernel = cu::binary_ss<Op, InType, OutType, IdxT>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::binary_ss<Op, InType, OutType, IdxT, N_READS>;
             if (bopt == BinaryOpType::ScalarVector) {
-              kernel = cu::binary_sv<Op, InType, OutType, IdxT>;
+              kernel = cu::binary_sv<Op, InType, OutType, IdxT, N_READS>;
             } else if (bopt == BinaryOpType::VectorScalar) {
-              kernel = cu::binary_vs<Op, InType, OutType, IdxT>;
+              kernel = cu::binary_vs<Op, InType, OutType, IdxT, N_READS>;
             } else if (bopt == BinaryOpType::VectorVector) {
-              kernel = cu::binary_vv<Op, InType, OutType, IdxT>;
+              kernel = cu::binary_vv<Op, InType, OutType, IdxT, N_READS>;
             }
             auto [num_blocks, block_dims] = get_launch_args(
-                kernel, out.data_size(), out.shape(), out.strides(), large());
+                kernel,
+                out.data_size(),
+                out.shape(),
+                out.strides(),
+                large(),
+                N_READS);
             encoder.add_kernel_node(
                 kernel,
                 num_blocks,
@@ -233,7 +290,7 @@ template <typename Op>
 void binary_op_gpu(
     const std::vector<array>& inputs,
     array& out,
-    std::string_view op,
+    const char* op,
     const Stream& s) {
   auto& a = inputs[0];
   auto& b = inputs[1];
@@ -242,11 +299,11 @@ void binary_op_gpu(
   binary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 
-#define BINARY_GPU(func)                                                 \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) {    \
-    nvtx3::scoped_range r(#func "::eval_gpu");                           \
-    auto& s = out.primitive().stream();                                  \
-    binary_op_gpu<cu::func>(inputs, out, get_primitive_string(this), s); \
+#define BINARY_GPU(func)                                              \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    nvtx3::scoped_range r(#func "::eval_gpu");                        \
+    auto& s = out.primitive().stream();                               \
+    binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
   }
 
 BINARY_GPU(Add)
@@ -270,33 +327,31 @@ BINARY_GPU(Subtract)
 void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
   nvtx3::scoped_range r("Equal::eval_gpu");
   auto& s = out.primitive().stream();
-  auto op = get_primitive_string(this);
   if (equal_nan_) {
-    binary_op_gpu<cu::NaNEqual>(inputs, out, op, s);
+    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
   } else {
-    binary_op_gpu<cu::Equal>(inputs, out, op, s);
+    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
   }
 }
 
 void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
   nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
   auto& s = out.primitive().stream();
-  auto op = get_primitive_string(this);
   switch (op_) {
     case BitwiseBinary::And:
-      binary_op_gpu<cu::BitwiseAnd>(inputs, out, op, s);
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
       break;
     case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, op, s);
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
       break;
     case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, op, s);
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
       break;
     case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, op, s);
+      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
       break;
     case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, op, s);
+      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
       break;
   }
 }

mlx/backend/cuda/binary_two.cu

@@ -3,7 +3,6 @@
 #include "mlx/backend/common/binary.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/device/binary_ops.cuh"
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -17,52 +16,119 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void
-binary_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+binary_two_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto out = Op{}(a[0], b[0]);
-    out_a[0] = out[0];
-    out_b[0] = out[1];
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void
-binary_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+binary_two_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto out = Op{}(a[0], b[index]);
-    out_a[index] = out[0];
-    out_b[index] = out[1];
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b_vec.val[i]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void
-binary_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+binary_two_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto out = Op{}(a[index], b[0]);
-    out_a[index] = out[0];
-    out_b[index] = out[1];
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec.val[i], b[0]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void
-binary_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto out = Op{}(a[index], b[index]);
-    out_a[index] = out[0];
-    out_b[index] = out[1];
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec.val[i], b_vec.val[i]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
   }
 }
 
 template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
-__global__ void binary_g_nd(
+__global__ void binary_two_g_nd(
     const In* a,
     const In* b,
     Out* out_a,
@@ -82,7 +148,7 @@ __global__ void binary_g_nd(
 }
 
 template <typename Op, typename In, typename Out, typename IdxT>
-__global__ void binary_g(
+__global__ void binary_two_g(
     const In* a,
     const In* b,
     Out* out_a,
@@ -94,7 +160,7 @@ __global__ void binary_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_4d(
+    auto [a_idx, b_idx] = elem_to_loc(
         index, shape.data(), a_strides.data(), b_strides.data(), ndim);
     auto out = Op{}(a[a_idx], b[b_idx]);
     out_a[index] = out[0];
@@ -103,7 +169,7 @@ __global__ void binary_g(
 }
 
 template <typename Op, typename In, typename Out>
-constexpr bool supports_binary_op() {
+constexpr bool supports_binary_two_op() {
   if (std::is_same_v<Op, DivMod>) {
     return std::is_same_v<In, Out> &&
         (std::is_integral_v<Out> || is_floating_v<Out>);
@@ -114,10 +180,10 @@ constexpr bool supports_binary_op() {
 } // namespace cu
 
 template <typename Op>
-void binary_op_gpu_inplace(
+void binary_two_op_gpu_inplace(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    std::string_view op,
+    const char* op,
     const Stream& s) {
   assert(inputs.size() > 1);
   const auto& a = inputs[0];
@@ -141,7 +207,7 @@ void binary_op_gpu_inplace(
     dispatch_all_types(out_a.dtype(), [&](auto out_type_tag) {
       using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
       using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
-      if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+      if constexpr (cu::supports_binary_two_op<Op, CTYPE_IN, CTYPE_OUT>()) {
         using InType = cuda_type_t<CTYPE_IN>;
         using OutType = cuda_type_t<CTYPE_OUT>;
 
@@ -161,8 +227,12 @@ void binary_op_gpu_inplace(
                 int ndim = shape.size();
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::
-                        binary_g_nd<Op, InType, OutType, IdxT, dims_constant()>;
+                    auto kernel = cu::binary_two_g_nd<
+                        Op,
+                        InType,
+                        OutType,
+                        IdxT,
+                        dims_constant()>;
                     auto [num_blocks, block_dims] =
                         get_launch_args(kernel, out_a, large());
                     encoder.add_kernel_node(
@@ -179,7 +249,7 @@ void binary_op_gpu_inplace(
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
+                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT>;
                   auto [num_blocks, block_dims] =
                       get_launch_args(kernel, out_a, large());
                   encoder.add_kernel_node(
@@ -198,22 +268,25 @@ void binary_op_gpu_inplace(
                 }
               });
         } else {
-          dispatch_bool(out_a.data_size() > INT32_MAX, [&](auto large) {
+          dispatch_bool(out_a.data_size() > UINT32_MAX, [&](auto large) {
             using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            auto kernel = cu::binary_ss<Op, InType, OutType, IdxT>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::binary_two_ss<Op, InType, OutType, IdxT, N_READS>;
             if (bopt == BinaryOpType::ScalarVector) {
-              kernel = cu::binary_sv<Op, InType, OutType, IdxT>;
+              kernel = cu::binary_two_sv<Op, InType, OutType, IdxT, N_READS>;
             } else if (bopt == BinaryOpType::VectorScalar) {
-              kernel = cu::binary_vs<Op, InType, OutType, IdxT>;
+              kernel = cu::binary_two_vs<Op, InType, OutType, IdxT, N_READS>;
             } else if (bopt == BinaryOpType::VectorVector) {
-              kernel = cu::binary_vv<Op, InType, OutType, IdxT>;
+              kernel = cu::binary_two_vv<Op, InType, OutType, IdxT, N_READS>;
             }
             auto [num_blocks, block_dims] = get_launch_args(
                 kernel,
                 out_a.data_size(),
                 out_a.shape(),
                 out_a.strides(),
-                large());
+                large(),
+                N_READS);
             encoder.add_kernel_node(
                 kernel,
                 num_blocks,
@@ -237,17 +310,17 @@ void binary_op_gpu_inplace(
 }
 
 template <typename Op>
-void binary_op_gpu(
+void binary_two_op_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    std::string_view op,
+    const char* op,
     const Stream& s) {
   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);
   set_binary_op_output_data(a, b, outputs[1], bopt);
-  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
+  binary_two_op_gpu_inplace<Op>(inputs, outputs, op, s);
 }
 
 void DivMod::eval_gpu(
@@ -255,7 +328,7 @@ void DivMod::eval_gpu(
     std::vector<array>& outputs) {
   nvtx3::scoped_range r("DivMod::eval_gpu");
   auto& s = outputs[0].primitive().stream();
-  binary_op_gpu<cu::DivMod>(inputs, outputs, get_primitive_string(this), s);
+  binary_two_op_gpu<cu::DivMod>(inputs, outputs, name(), s);
 }
 
 } // namespace mlx::core

mlx/backend/cuda/compiled.cpp

@@ -53,9 +53,10 @@ struct FusedKernelBuilder {
 
     // Build function signature.
     if (contiguous) {
-      os += "template <typename IdxT = uint32_t>\n";
+      os += "template <typename IdxT = uint32_t, int work_per_thread = 1>\n";
     } else {
-      os += "template <int NDIM, typename IdxT = uint32_t>\n";
+      os +=
+          "template <int NDIM, typename IdxT = uint32_t, int work_per_thread = 1>\n";
     }
     os += fmt::format("__global__ void {}(\n", kernel_name + name);
     for (size_t i = 0; i < params.size(); ++i) {
@@ -67,12 +68,46 @@ struct FusedKernelBuilder {
     }
     os += ") {\n";
 
-    // Index.
+    // Index. For non contiguous kernels we create a separate index
+    // variable per variable otherwise everyone uses `index`.
     os +=
-        "  IdxT index = cg::this_grid().thread_rank();\n"
+        "  IdxT index = cg::this_grid().thread_rank() * work_per_thread;\n"
         "  if (index >= size) {\n"
         "    return;\n"
         "  }\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "  IdxT " + xname + "_idx = 0;\n";
+      }
+      os += "  {\n";
+      os += "    IdxT loc = index;\n";
+      os +=
+          "    #pragma unroll\n"
+          "    for (int i = NDIM - 1; i >= 0; i--) {\n";
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "      " + xname + "_idx += (loc \% shape[i]) * IdxT(" + xname +
+            "_strides[i]);\n";
+      }
+      os +=
+          "      loc /= shape[i];\n"
+          "    }\n"
+          "  }\n";
+    }
+
+    // Work loop
+    os +=
+        "\n"
+        "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
 
     // Read inputs.
     for (size_t i = 0; i < inputs.size(); ++i) {
@@ -89,12 +124,9 @@ struct FusedKernelBuilder {
       } else if (contiguous) {
         value = fmt::format("{}[index]", xname);
       } else {
-        std::string index = fmt::format(
-            "elem_to_loc_nd<NDIM>(index, shape.data(), {}_strides.data())",
-            xname);
-        value = fmt::format("{}[{}]", xname, index);
+        value = fmt::format("{}[{}_idx]", xname, xname);
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write tape.
@@ -106,23 +138,37 @@ struct FusedKernelBuilder {
         value = fmt::format(
             "static_cast<{}>(tmp_{})", type, namer.get_name(x.inputs()[0]));
       } else {
-        std::ostringstream ss;
-        x.primitive().print(ss);
-        value = ss.str();
+        value = x.primitive().name();
         value += "{}(";
         for (size_t i = 0; i < x.inputs().size() - 1; ++i) {
           value += fmt::format("tmp_{}, ", namer.get_name(x.inputs()[i]));
         }
         value += fmt::format("tmp_{})", namer.get_name(x.inputs().back()));
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write output.
     for (const auto& x : outputs) {
-      os += fmt::format("  {0}[index] = tmp_{0};\n", namer.get_name(x));
+      os += fmt::format("    {0}[index] = tmp_{0};\n", namer.get_name(x));
     }
 
+    // End of work loop
+    os +=
+        "\n"
+        "    index++;\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "    " + xname + "_idx += " + xname + "_strides[NDIM - 1];\n";
+      }
+    }
+    os += "  }\n";
+
     os += "}\n";
   }
 };
@@ -158,15 +204,28 @@ void Compiled::eval_gpu(
     builder.build("_strided", false);
     builder.os += "\n} // namespace mlx::core::cu\n";
     // Build kernel names.
-    std::vector<std::string> kernel_names = {
-        fmt::format("mlx::core::cu::{}_contiguous<uint32_t>", lib_name()),
-        fmt::format("mlx::core::cu::{}_contiguous<int64_t>", lib_name()),
-    };
-    for (int i = 1; i <= MAX_NDIM; ++i) {
+    std::vector<std::string> kernel_names;
+    for (auto work_per_thread : std::array<int, 2>{1, 4}) {
       kernel_names.push_back(fmt::format(
-          "mlx::core::cu::{}_strided<{}, uint32_t>", lib_name(), i));
-      kernel_names.push_back(
-          fmt::format("mlx::core::cu::{}_strided<{}, int64_t>", lib_name(), i));
+          "mlx::core::cu::{}_contiguous<uint32_t, {}>",
+          lib_name(),
+          work_per_thread));
+      kernel_names.push_back(fmt::format(
+          "mlx::core::cu::{}_contiguous<int64_t, {}>",
+          lib_name(),
+          work_per_thread));
+      for (int i = 1; i <= MAX_NDIM; ++i) {
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, uint32_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, int64_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+      }
     }
     return std::make_pair(std::move(builder.os), std::move(kernel_names));
   });
@@ -209,13 +268,21 @@ void Compiled::eval_gpu(
     args.append<uint32_t>(outputs[0].data_size());
   }
 
+  // Choose work per thread
+  int work_per_thread = 4;
+  if (!contiguous && shape.back() % work_per_thread != 0) {
+    work_per_thread = 1;
+  }
+
   // Launch kernel.
   const char* index_type = large ? "int64_t" : "uint32_t";
   std::string kernel_name = fmt::format("mlx::core::cu::{}", lib_name());
   if (contiguous) {
-    kernel_name += fmt::format("_contiguous<{}>", index_type);
+    kernel_name +=
+        fmt::format("_contiguous<{}, {}>", index_type, work_per_thread);
   } else {
-    kernel_name += fmt::format("_strided<{}, {}>", shape.size(), index_type);
+    kernel_name += fmt::format(
+        "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
   }
   auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
@@ -226,7 +293,8 @@ void Compiled::eval_gpu(
   }
 
   auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(kernel, outputs[0], large);
+  auto [num_blocks, block_dims] =
+      get_launch_args(kernel, outputs[0], large, work_per_thread);
   encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
 }
 

mlx/backend/cuda/conv.cpp

@@ -0,0 +1,340 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/config.h"
+#include "mlx/backend/cuda/lru_cache.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+// cudnn_frontend.h redefines this macro.
+#undef CHECK_CUDA_ERROR
+
+#include <cudnn_frontend.h>
+#include <cudnn_frontend_find_plan.h>
+#include <fmt/format.h>
+#include <nvtx3/nvtx3.hpp>
+
+#include <cassert>
+#include <numeric>
+
+namespace mlx::core {
+
+namespace {
+
+// Not all engines support it so can not use this API now.
+#define MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API 0
+
+struct ConvCacheKey {
+  int device_id;
+  cudnnBackendDescriptorType_t backend_type;
+  cudnnDataType_t cudnn_type;
+  std::array<int, MAX_NDIM> input_shape;
+  std::array<int, MAX_NDIM> filter_shape;
+  std::array<int, MAX_NDIM> padding_lo;
+  std::array<int, MAX_NDIM> padding_hi;
+  std::array<int, MAX_NDIM> stride;
+  std::array<int, MAX_NDIM> dilation;
+  int groups;
+  uint8_t input_alignment;
+  uint8_t filter_alignment;
+  uint8_t output_alignment;
+};
+
+auto& conv_cache() {
+  static LRUBytesKeyCache<ConvCacheKey, cudnn_frontend::ExecutionPlan> cache(
+      /* capacity */ 128);
+  return cache;
+}
+
+template <typename T, typename U>
+inline std::vector<T> convert_vector(const std::vector<U>& vec) {
+  return std::vector<T>(vec.begin(), vec.end());
+}
+
+template <typename T>
+inline std::array<T, MAX_NDIM> fixed_vector(const std::vector<T>& vec) {
+  if (vec.size() > MAX_NDIM) {
+    throw std::runtime_error(
+        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
+  }
+  std::array<T, MAX_NDIM> result = {};
+  std::copy_n(vec.begin(), vec.size(), result.begin());
+  return result;
+}
+
+auto nhwc_to_nchw(const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  shape.insert(shape.begin() + 1, shape.back());
+  shape.erase(shape.end() - 1);
+  auto strides = convert_vector<int64_t>(x.strides());
+  strides.insert(strides.begin() + 1, strides.back());
+  strides.erase(strides.end() - 1);
+  return std::make_tuple(shape, strides);
+}
+
+inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return CUDNN_DATA_INT8;
+    case int32:
+      return CUDNN_DATA_INT32;
+    case uint8:
+      return CUDNN_DATA_UINT8;
+    case float16:
+      return CUDNN_DATA_HALF;
+    case bfloat16:
+      return CUDNN_DATA_BFLOAT16;
+    case float32:
+      return CUDNN_DATA_FLOAT;
+    case float64:
+      return CUDNN_DATA_DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+  }
+}
+
+inline uint8_t get_alignment(const array& x) {
+  uint8_t alignment = 1;
+  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
+  for (; alignment < 32; alignment *= 2) {
+    if (address % (alignment * 2)) {
+      return alignment;
+    }
+  }
+  return alignment;
+}
+
+inline cudnn_frontend::Tensor build_tensor(int64_t id, const array& x) {
+  auto [shape, strides] = nhwc_to_nchw(x);
+  return cudnn_frontend::TensorBuilder()
+      .setDim(shape.size(), shape.data())
+      .setStrides(strides.size(), strides.data())
+      .setId(id)
+      .setAlignment(get_alignment(x))
+      .setDataType(dtype_to_cudnn_type(x.dtype()))
+      .build();
+}
+
+cudnn_frontend::EngineConfigList get_engine_configs(
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph,
+    bool use_fallback = false) {
+  cudnn_frontend::GeneratorSource source;
+  if (use_fallback) {
+    source = [&backend_type](cudnn_frontend::OperationGraph& op_graph) {
+      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
+                          .setOperationGraph(op_graph)
+                          .setOperation(backend_type)
+                          .build();
+      return fallback.getFallbackList();
+    };
+  } else {
+    source = [](cudnn_frontend::OperationGraph& op_graph) {
+      auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
+                            .setOperationGraph(op_graph)
+                            .setHeurMode(CUDNN_HEUR_MODE_A)
+                            .build();
+      return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
+    };
+  }
+
+  cudnn_frontend::EngineConfigGenerator generator(1, &source);
+  auto configs = generator.generate_engine_config(op_graph);
+
+  cudnn_frontend::EngineConfigList filtered_configs;
+  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
+    if (cudnn_frontend::hasNumericalNote<
+            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
+      return true;
+    }
+    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
+        dtype == float32 && !env::enable_tf32()) {
+      return true;
+    }
+    return false;
+  });
+  return filtered_configs;
+}
+
+bool execute_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    const array& in,
+    const array& wt,
+    array& out) {
+  int workspace_size = plan.getWorkspaceSize();
+  array workspace(allocator::malloc(workspace_size), {workspace_size}, uint8);
+
+  int64_t uids[3] = {'x', 'w', 'y'};
+  void* data_ptrs[3] = {
+      const_cast<void*>(in.data<void>()),
+      const_cast<void*>(wt.data<void>()),
+      out.data<void>(),
+  };
+
+  auto variantPack = cudnn_frontend::VariantPackBuilder()
+                         .setWorkspacePointer(workspace.data<void>())
+                         .setDataPointers(3, data_ptrs)
+                         .setUids(3, uids)
+                         .build();
+
+  auto handle = encoder.device().cudnn_handle();
+  cudnnSetStream(handle, encoder.stream());
+
+#if CUDNN_VERSION >= 90500 && MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API
+  cudaGraph_t graph;
+  cudaGraphCreate(&graph, 0);
+  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
+      &graph, [](cudaGraph_t* p) { cudaGraphDestroy(*p); });
+  if (cudnnBackendPopulateCudaGraph(
+          handle, plan.get_raw_desc(), variantPack.get_raw_desc(), graph) !=
+      CUDNN_STATUS_SUCCESS) {
+    return false;
+  }
+  encoder.add_graph_node(graph);
+#else
+  auto capture = encoder.capture_context();
+  if (cudnnBackendExecute(
+          handle, plan.get_raw_desc(), variantPack.get_raw_desc()) !=
+      CUDNN_STATUS_SUCCESS) {
+    // Discard the captured graph when failed.
+    capture.discard = true;
+    return false;
+  }
+#endif
+
+  encoder.add_temporary(workspace);
+  return true;
+}
+
+bool try_engines(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::EngineConfigList& configs,
+    const ConvCacheKey& cache_key,
+    const std::string& op_graph_tag,
+    const array& in,
+    const array& wt,
+    array& out) {
+  for (auto& config : configs) {
+    try {
+      auto plan = cudnn_frontend::ExecutionPlanBuilder()
+                      .setHandle(encoder.device().cudnn_handle())
+                      .setEngineConfig(config, op_graph_tag)
+                      .build();
+      if (execute_plan(encoder, plan, in, wt, out)) {
+        conv_cache().emplace(cache_key, std::move(plan));
+        return true;
+      }
+    } catch (cudnn_frontend::cudnnException&) {
+    }
+  }
+  return false;
+}
+
+} // namespace
+
+void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Convolution::eval_gpu");
+  if (out.size() == 0) {
+    return;
+  }
+
+  assert(inputs.size() == 2);
+  array in = inputs[0];
+  array wt = inputs[1];
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  // cuDNN requires contiguous input.
+  // TODO: Handle NCHW format specially.
+  if (!in.flags().row_contiguous) {
+    in = contiguous_copy_gpu(in, s);
+    encoder.add_temporary(in);
+  }
+  if (!wt.flags().row_contiguous) {
+    wt = contiguous_copy_gpu(wt, s);
+    encoder.add_temporary(wt);
+  }
+
+  encoder.set_input_array(in);
+  encoder.set_input_array(wt);
+  encoder.set_output_array(out);
+
+  auto backend_type = CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR;
+  auto cudnn_type = dtype_to_cudnn_type(in.dtype());
+
+  // Search cache.
+  ConvCacheKey cache_key{
+      encoder.device().cuda_device(),
+      backend_type,
+      cudnn_type,
+      fixed_vector(in.shape()),
+      fixed_vector(wt.shape()),
+      fixed_vector(padding_lo_),
+      fixed_vector(padding_hi_),
+      fixed_vector(kernel_strides_),
+      fixed_vector(kernel_dilation_),
+      groups_,
+      get_alignment(in),
+      get_alignment(wt),
+      get_alignment(out)};
+  if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
+    if (!execute_plan(encoder, it->second, in, wt, out)) {
+      throw std::runtime_error("Cached convolution plan failed to execute.");
+    }
+    return;
+  }
+
+  // Build operation graph.
+  auto compute_data_type = (in.dtype() == float16 || in.dtype() == bfloat16)
+      ? CUDNN_DATA_FLOAT
+      : cudnn_type;
+
+  auto stride = convert_vector<int64_t>(kernel_strides_);
+  auto padding_lo = convert_vector<int64_t>(padding_lo_);
+  auto padding_hi = convert_vector<int64_t>(padding_hi_);
+  auto dilation = convert_vector<int64_t>(kernel_dilation_);
+
+  auto conv_desc = cudnn_frontend::ConvDescBuilder()
+                       .setDataType(compute_data_type)
+                       .setMathMode(CUDNN_CROSS_CORRELATION)
+                       .setNDims(stride.size())
+                       .setStrides(stride.size(), stride.data())
+                       .setPrePadding(padding_lo.size(), padding_lo.data())
+                       .setPostPadding(padding_hi.size(), padding_hi.data())
+                       .setDilation(dilation.size(), dilation.data())
+                       .build();
+
+  auto op = cudnn_frontend::OperationBuilder(backend_type)
+                .setxDesc(build_tensor('x', in))
+                .setwDesc(build_tensor('w', wt))
+                .setyDesc(build_tensor('y', out))
+                .setcDesc(conv_desc)
+                .build();
+
+  std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
+  auto op_graph = cudnn_frontend::OperationGraphBuilder()
+                      .setHandle(encoder.device().cudnn_handle())
+                      .setOperationGraph(ops.size(), ops.data())
+                      .build();
+
+  // Try to run plans based on heuristics.
+  auto configs = get_engine_configs(backend_type, in.dtype(), op_graph);
+  auto op_graph_tag = op_graph.getTag();
+  if (try_engines(encoder, configs, cache_key, op_graph_tag, in, wt, out)) {
+    return;
+  }
+  // Then try fallback plans.
+  configs = get_engine_configs(backend_type, in.dtype(), op_graph);
+  if (try_engines(encoder, configs, cache_key, op_graph_tag, in, wt, out)) {
+    return;
+  }
+  throw std::runtime_error("Unable to find an engine for convolution.");
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy/copy_contiguous.cu

@@ -10,19 +10,43 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_s(const In* in, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = CastOp<In, Out>{}(in[0]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = cast_to<Out>(in[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_v(const In* in, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = CastOp<In, Out>{}(in[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[i]);
+    }
+  } else {
+    auto in_vec = load_vector<N_READS>(in, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = cast_to<Out>(in_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -41,12 +65,19 @@ void copy_contiguous(
         using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
         using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
         using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-        auto kernel = cu::copy_s<InType, OutType, IdxT>;
+        // TODO: Choose optimized value based on type size.
+        constexpr int N_READS = 4;
+        auto kernel = cu::copy_s<InType, OutType, IdxT, N_READS>;
         if (ctype == CopyType::Vector) {
-          kernel = cu::copy_v<InType, OutType, IdxT>;
+          kernel = cu::copy_v<InType, OutType, IdxT, N_READS>;
         }
         auto [num_blocks, block_dims] = get_launch_args(
-            kernel, out.data_size(), out.shape(), out.strides(), large());
+            kernel,
+            out.data_size(),
+            out.shape(),
+            out.strides(),
+            large(),
+            N_READS);
         encoder.add_kernel_node(
             kernel,
             num_blocks,

mlx/backend/cuda/copy/copy_general.cu

@@ -37,7 +37,7 @@ __global__ void copy_gg(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_4d(
+    auto [idx_in, idx_out] = elem_to_loc(
         index, shape.data(), strides_in.data(), strides_out.data(), ndim);
     out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
   }

mlx/backend/cuda/copy/copy_general_dynamic.cu

@@ -41,7 +41,7 @@ __global__ void copy_gg_dynamic(
     const int64_t* offset_out) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_4d(
+    auto [idx_in, idx_out] = elem_to_loc(
         index, shape.data(), strides_in.data(), strides_out.data(), ndim);
     out[idx_out + *offset_out] = CastOp<In, Out>{}(in[idx_in + *offset_in]);
   }

mlx/backend/cuda/copy/copy_general_input.cu

@@ -34,7 +34,7 @@ __global__ void copy_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    IdxT idx_in = elem_to_loc_4d(index, shape.data(), strides_in.data(), ndim);
+    IdxT idx_in = elem_to_loc(index, shape.data(), strides_in.data(), ndim);
     out[index] = CastOp<In, Out>{}(in[idx_in]);
   }
 }

mlx/backend/cuda/device.cpp

@@ -9,12 +9,23 @@
 #include <future>
 #include <unordered_set>
 
-namespace mlx::core {
+namespace mlx::core::cu {
+
+namespace {
 
 // Can be tuned with MLX_MAX_OPS_PER_BUFFER
 // This should be less than 255
 constexpr int default_max_nodes_per_graph = 20;
 
+#define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
+
+void check_cudnn_error(const char* name, cudnnStatus_t err) {
+  if (err != CUDNN_STATUS_SUCCESS) {
+    throw std::runtime_error(
+        fmt::format("{} failed: {}.", name, cudnnGetErrorString(err)));
+  }
+}
+
 int cuda_graph_cache_size() {
   static int cache_size = []() {
     return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 100);
@@ -22,7 +33,7 @@ int cuda_graph_cache_size() {
   return cache_size;
 }
 
-namespace cu {
+} // namespace
 
 Device::Device(int device) : device_(device) {
   CHECK_CUDA_ERROR(cudaDeviceGetAttribute(
@@ -40,11 +51,14 @@ Device::Device(int device) : device_(device) {
   }
   // The cublasLt handle is used by matmul.
   make_current();
-  cublasLtCreate(&lt_);
+  CHECK_CUBLAS_ERROR(cublasLtCreate(&lt_));
+  // The cudnn handle is used by Convolution.
+  CHECK_CUDNN_ERROR(cudnnCreate(&cudnn_));
 }
 
 Device::~Device() {
-  cublasLtDestroy(lt_);
+  CHECK_CUDNN_ERROR(cudnnDestroy(cudnn_));
+  CHECK_CUBLAS_ERROR(cublasLtDestroy(lt_));
 }
 
 void Device::make_current() {
@@ -57,31 +71,45 @@ void Device::make_current() {
   }
 }
 
+CommandEncoder& Device::get_command_encoder(Stream s) {
+  auto it = encoders_.find(s.index);
+  if (it == encoders_.end()) {
+    it = encoders_.try_emplace(s.index, *this).first;
+  }
+  return it->second;
+}
+
 CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
-  CHECK_CUDA_ERROR(cudaGraphCreate(&graph, 0));
+  enc.device().make_current();
   CHECK_CUDA_ERROR(
       cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
 }
 
 CommandEncoder::CaptureContext::~CaptureContext() {
   CHECK_CUDA_ERROR(cudaStreamEndCapture(enc.stream(), &graph));
+  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
+      &graph, [](cudaGraph_t* p) { CHECK_CUDA_ERROR(cudaGraphDestroy(*p)); });
+  if (discard) {
+    return;
+  }
+
+  // Extract and add as single kernel node when possible.
   size_t num_nodes;
   CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, NULL, &num_nodes));
   if (num_nodes == 1) {
     cudaGraphNode_t captured_node;
     CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, &captured_node, &num_nodes));
-    CUDA_KERNEL_NODE_PARAMS params;
-    CHECK_CUDA_ERROR(cuGraphKernelNodeGetParams(captured_node, &params));
-    cudaGraphNode_t node;
-    CHECK_CUDA_ERROR(cuGraphAddKernelNode(&node, enc.graph_, NULL, 0, &params));
-    enc.insert_graph_dependencies(GraphNode{node, 'K'});
-  } else {
-    cudaGraphNode_t node;
-    CHECK_CUDA_ERROR(
-        cudaGraphAddChildGraphNode(&node, enc.graph_, NULL, 0, graph));
-    enc.insert_graph_dependencies(GraphNode{node, 'G'});
+    cudaGraphNodeType type;
+    CHECK_CUDA_ERROR(cudaGraphNodeGetType(captured_node, &type));
+    if (type == cudaGraphNodeTypeKernel) {
+      CUDA_KERNEL_NODE_PARAMS params;
+      CHECK_CUDA_ERROR(cuGraphKernelNodeGetParams(captured_node, &params));
+      enc.add_kernel_node(params);
+      return;
+    }
   }
-  CHECK_CUDA_ERROR(cudaGraphDestroy(graph));
+  // Otherwise add the captured graph as subgraph.
+  enc.add_graph_node(graph);
 }
 
 CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
@@ -168,15 +196,7 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
   }
 }
 
-CommandEncoder& Device::get_command_encoder(Stream s) {
-  auto it = encoders_.find(s.index);
-  if (it == encoders_.end()) {
-    it = encoders_.try_emplace(s.index, *this).first;
-  }
-  return it->second;
-}
-
-CommandEncoder::CommandEncoder(Device& d) : stream_(d) {
+CommandEncoder::CommandEncoder(Device& d) : device_(d), stream_(d) {
   CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
 }
 
@@ -222,10 +242,7 @@ void CommandEncoder::add_kernel_node(
   kernel_params.gridDim = grid_dim;
   kernel_params.blockDim = block_dim;
   kernel_params.kernelParams = params;
-  cudaGraphNode_t node;
-  CHECK_CUDA_ERROR(
-      cudaGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
-  insert_graph_dependencies(GraphNode{node, 'K'});
+  add_kernel_node(kernel_params);
 }
 
 void CommandEncoder::add_kernel_node(
@@ -242,12 +259,27 @@ void CommandEncoder::add_kernel_node(
   kernel_params.blockDimY = block_dim.y;
   kernel_params.blockDimZ = block_dim.z;
   kernel_params.kernelParams = params;
-  CUgraphNode node;
-  CHECK_CUDA_ERROR(
-      cuGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
+  add_kernel_node(kernel_params);
+}
+
+void CommandEncoder::add_kernel_node(const cudaKernelNodeParams& params) {
+  cudaGraphNode_t node;
+  CHECK_CUDA_ERROR(cudaGraphAddKernelNode(&node, graph_, NULL, 0, &params));
   insert_graph_dependencies(GraphNode{node, 'K'});
 }
 
+void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
+  CUgraphNode node;
+  CHECK_CUDA_ERROR(cuGraphAddKernelNode(&node, graph_, NULL, 0, &params));
+  insert_graph_dependencies(GraphNode{node, 'K'});
+}
+
+void CommandEncoder::add_graph_node(cudaGraph_t child) {
+  cudaGraphNode_t node;
+  CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
+  insert_graph_dependencies(GraphNode{node, 'G'});
+}
+
 void CommandEncoder::commit() {
   if (!temporaries_.empty()) {
     add_completed_handler([temporaries = std::move(temporaries_)]() {});
@@ -264,22 +296,30 @@ void CommandEncoder::commit() {
     graph_key_ += std::to_string(graph_node_count_);
     graph_key_ += ".";
     graph_key_ += std::to_string(empty_node_count_);
-    auto [it, _] = graph_cache_.emplace(graph_key_, nullptr);
-    auto& graph_exec = it->second;
 
-    if (graph_exec != NULL) {
-      cudaGraphExecUpdateResultInfo update_result;
-      cudaGraphExecUpdate(graph_exec, graph_, &update_result);
-      if (update_result.result != cudaGraphExecUpdateSuccess) {
-        cudaGetLastError();
+    cudaGraphExec_t& graph_exec = graph_cache_[graph_key_];
+
+    if (graph_exec != nullptr) {
+      cudaGraphExecUpdateResult update_result;
+#if CUDART_VERSION >= 12000
+      cudaGraphExecUpdateResultInfo info;
+      cudaGraphExecUpdate(graph_exec, graph_, &info);
+      update_result = info.result;
+#else
+      cudaGraphNode_t error_node;
+      cudaGraphExecUpdate(graph_exec, graph_, &error_node, &update_result);
+#endif // CUDART_VERSION >= 12000
+      if (update_result != cudaGraphExecUpdateSuccess) {
+        cudaGetLastError(); // reset error
         CHECK_CUDA_ERROR(cudaGraphExecDestroy(graph_exec));
-        graph_exec = NULL;
+        graph_exec = nullptr;
       }
     }
-    if (graph_exec == NULL) {
+    if (graph_exec == nullptr) {
       CHECK_CUDA_ERROR(
           cudaGraphInstantiate(&graph_exec, graph_, NULL, NULL, 0));
     }
+    device_.make_current();
     CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
 
     // TODO smarter cache policy
@@ -299,7 +339,6 @@ void CommandEncoder::commit() {
   }
 
   // Put completion handlers in a batch.
-  worker_.end_batch();
   worker_.commit(stream_);
 }
 
@@ -308,7 +347,6 @@ void CommandEncoder::synchronize() {
   auto p = std::make_shared<std::promise<void>>();
   std::future<void> f = p->get_future();
   add_completed_handler([p = std::move(p)]() { p->set_value(); });
-  worker_.end_batch();
   commit();
   f.wait();
 }
@@ -326,6 +364,4 @@ CommandEncoder& get_command_encoder(Stream s) {
   return device(s.device).get_command_encoder(s);
 }
 
-} // namespace cu
-
-} // namespace mlx::core
+} // namespace mlx::core::cu

mlx/backend/cuda/device.h

@@ -8,6 +8,7 @@
 
 #include <cublasLt.h>
 #include <cuda.h>
+#include <cudnn.h>
 #include <thrust/execution_policy.h>
 
 #include <unordered_map>
@@ -21,6 +22,7 @@ class CommandEncoder {
     ~CaptureContext();
     cudaGraph_t graph;
     CommandEncoder& enc;
+    bool discard{false};
   };
   struct ConcurrentContext {
     ConcurrentContext(CommandEncoder& enc);
@@ -65,6 +67,11 @@ class CommandEncoder {
   void
   add_kernel_node(void* func, dim3 grid_dim, dim3 block_dim, void** params);
 
+  // Low-level graph helpers.
+  void add_kernel_node(const cudaKernelNodeParams& params);
+  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
+  void add_graph_node(cudaGraph_t child);
+
   void add_temporary(const array& arr) {
     temporaries_.push_back(arr.data_shared_ptr());
   }
@@ -73,6 +80,10 @@ class CommandEncoder {
   void maybe_commit();
   void commit();
 
+  Device& device() {
+    return device_;
+  }
+
   CudaStream& stream() {
     return stream_;
   }
@@ -93,6 +104,7 @@ class CommandEncoder {
   void insert_graph_dependencies(GraphNode node);
   void insert_graph_dependencies(std::vector<GraphNode> nodes);
 
+  Device& device_;
   CudaStream stream_;
   cudaGraph_t graph_;
   Worker worker_;
@@ -136,12 +148,16 @@ class Device {
   cublasLtHandle_t lt_handle() const {
     return lt_;
   }
+  cudnnHandle_t cudnn_handle() const {
+    return cudnn_;
+  }
 
  private:
   int device_;
   int compute_capability_major_;
   int compute_capability_minor_;
   cublasLtHandle_t lt_;
+  cudnnHandle_t cudnn_;
   std::unordered_map<int, CommandEncoder> encoders_;
 };
 

mlx/backend/cuda/device/atomic_ops.cuh

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
+#include "mlx/backend/cuda/device/complex.cuh"
 #include "mlx/backend/cuda/device/fp16_math.cuh"
 
 #include <cuda/atomic>
@@ -48,7 +48,7 @@ inline __device__ void atomic_add(__half* out, __half val) {
   atomicAdd(out, val);
 }
 
-inline __device__ void atomic_add(cuComplex* out, cuComplex val) {
+inline __device__ void atomic_add(complex64_t* out, complex64_t val) {
 #if __CUDA_ARCH__ < 900
   atomic_add_general(out, val);
 #else
@@ -58,12 +58,7 @@ inline __device__ void atomic_add(cuComplex* out, cuComplex val) {
 
 inline __device__ void atomic_add(__nv_bfloat16* out, __nv_bfloat16 val) {
 #if __CUDA_ARCH__ < 800
-#if CCCL_VERSION >= 2008000
   atomic_add_general(out, val);
-#else
-  bool cccl_version_too_old_for_bfloat16_atomic_add = false;
-  assert(cccl_version_too_old_for_bfloat16_atomic_add);
-#endif
 #else
   atomicAdd(out, val);
 #endif

mlx/backend/cuda/device/binary_ops.cuh

@@ -1,10 +1,7 @@
 // Copyright © 2025 Apple Inc.
 
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
-#include "mlx/backend/cuda/device/fp16_math.cuh"
-#include "mlx/backend/cuda/device/utils.cuh"
+#include "mlx/backend/cuda/device/unary_ops.cuh"
 
-#include <cuComplex.h>
 #include <cuda/std/array>
 
 namespace mlx::core::cu {
@@ -47,7 +44,7 @@ struct Remainder {
       } else {
         return x % y;
       }
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+    } else if constexpr (is_complex_v<T>) {
       return x % y;
     } else {
       T r = fmod(x, y);
@@ -69,14 +66,12 @@ struct Equal {
 struct NaNEqual {
   template <typename T>
   __device__ bool operator()(T x, T y) {
-    if constexpr (std::is_same_v<T, cuComplex>) {
+    if constexpr (is_complex_v<T>) {
       return x == y ||
-          (isnan(cuCrealf(x)) && isnan(cuCrealf(y)) && isnan(cuCimagf(x)) &&
-           isnan(cuCimagf(y))) ||
-          (cuCrealf(x) == cuCrealf(y) && isnan(cuCimagf(x)) &&
-           isnan(cuCimagf(y))) ||
-          (isnan(cuCrealf(x)) && isnan(cuCrealf(y)) &&
-           cuCimagf(x) == cuCimagf(y));
+          (isnan(x.real()) && isnan(y.real()) && isnan(x.imag()) &&
+           isnan(y.imag())) ||
+          (x.real() == y.real() && isnan(x.imag()) && isnan(y.imag())) ||
+          (isnan(x.real()) && isnan(y.real()) && x.imag() == y.imag());
     } else {
       return x == y || (isnan(x) && isnan(y));
     }
@@ -114,36 +109,38 @@ struct LessEqual {
 struct LogAddExp {
   template <typename T>
   __device__ T operator()(T x, T y) {
-    if (isnan(x) || isnan(y)) {
-      return cuda::std::numeric_limits<T>::quiet_NaN();
+    if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag()) || isnan(y.real()) ||
+          isnan(y.imag())) {
+        return {
+            cuda::std::numeric_limits<float>::quiet_NaN(),
+            cuda::std::numeric_limits<float>::quiet_NaN()};
+      }
+      auto max = x.real() > y.real() ? x : y;
+      auto min = x.real() < y.real() ? x : y;
+      auto min_real = min.real();
+      auto max_real = max.real();
+      if (!isfinite(min_real) && (min_real == max_real)) {
+        if (min_real < 0) {
+          return min;
+        } else {
+          return Log{}(Exp{}(min) + Exp{}(max));
+        }
+      } else {
+        return Log1p{}(Exp{}(min - max)) + max;
+      }
+    } else {
+      if (isnan(x) || isnan(y)) {
+        return cuda::std::numeric_limits<T>::quiet_NaN();
+      }
+      T maxval = max(x, y);
+      T minval = min(x, y);
+      return (minval == -cuda::std::numeric_limits<T>::infinity() ||
+              maxval == cuda::std::numeric_limits<T>::infinity())
+          ? maxval
+          : T(float(maxval) + log1p(expf(minval - maxval)));
     }
-    T maxval = max(x, y);
-    T minval = min(x, y);
-    return (minval == -cuda::std::numeric_limits<T>::infinity() ||
-            maxval == cuda::std::numeric_limits<T>::infinity())
-        ? maxval
-        : T(float(maxval) + log1p(expf(minval - maxval)));
   };
-
-  __device__ cuComplex operator()(cuComplex x, cuComplex y) {
-    if (isnan(cuCrealf(x)) || isnan(cuCimagf(x)) || isnan(cuCrealf(y)) ||
-        isnan(cuCimagf(y))) {
-      return {
-          cuda::std::numeric_limits<float>::quiet_NaN(),
-          cuda::std::numeric_limits<float>::quiet_NaN()};
-    }
-    float inf = cuda::std::numeric_limits<float>::infinity();
-    auto maxval = x > y ? x : y;
-    auto minval = x < y ? x : y;
-    if (cuCrealf(minval) == -inf || cuCrealf(maxval) == inf)
-      return maxval;
-    float m = exp(cuCrealf(minval) - cuCrealf(maxval));
-    cuComplex dexp{
-        m * cos(cuCimagf(minval) - cuCimagf(maxval)),
-        m * sin(cuCimagf(minval) - cuCimagf(maxval)),
-    };
-    return maxval + log1p(dexp);
-  }
 };
 
 struct Maximum {
@@ -151,8 +148,8 @@ struct Maximum {
   __device__ T operator()(T x, T y) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return max(x, y);
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (isnan(cuCrealf(x)) || isnan(cuCimagf(x))) {
+    } else if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag())) {
         return x;
       }
       return x > y ? x : y;
@@ -170,8 +167,8 @@ struct Minimum {
   __device__ T operator()(T x, T y) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return min(x, y);
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (isnan(cuCrealf(x)) || isnan(cuCimagf(x))) {
+    } else if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag())) {
         return x;
       }
       return x < y ? x : y;
@@ -194,8 +191,8 @@ struct Multiply {
 struct NotEqual {
   template <typename T>
   __device__ bool operator()(T x, T y) {
-    if constexpr (std::is_same_v<T, cuComplex>) {
-      return cuCrealf(x) != cuCrealf(y) || cuCimagf(x) != cuCimagf(y);
+    if constexpr (is_complex_v<T>) {
+      return x.real() != y.real() || x.imag() != y.imag();
     } else {
       return x != y;
     }
@@ -215,19 +212,8 @@ struct Power {
         base *= base;
       }
       return res;
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (base.y == 0 && base.x == 0) {
-        if (isnan(exp.x) || isnan(exp.y)) {
-          auto nan = cuda::std::numeric_limits<float>::quiet_NaN();
-          return make_cuFloatComplex(nan, nan);
-        }
-        return make_cuFloatComplex(0.0, 0.0);
-      }
-      auto x_theta = atan2f(base.y, base.x);
-      auto x_ln_r = 0.5 * logf(base.x * base.x + base.y * base.y);
-      auto mag = expf(exp.x * x_ln_r - exp.y * x_theta);
-      auto phase = exp.y * x_ln_r + exp.x * x_theta;
-      return make_cuFloatComplex(mag * cosf(phase), mag * sinf(phase));
+    } else if constexpr (is_complex_v<T>) {
+      return pow(base, exp);
     } else {
       return powf(base, exp);
     }

mlx/backend/cuda/device/cast_op.cuh

@@ -2,7 +2,10 @@
 
 #pragma once
 
-#include <cuComplex.h>
+#include "mlx/backend/cuda/device/complex.cuh"
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
 #include <thrust/iterator/transform_iterator.h>
 
 namespace mlx::core::cu {
@@ -17,34 +20,48 @@ struct CastOp {
   }
 };
 
-// Converting a complex number to real number discards the imaginary part.
-template <typename DstT>
-struct CastOp<
-    cuComplex,
-    DstT,
-    cuda::std::enable_if_t<!cuda::std::is_same_v<cuComplex, DstT>>> {
-  static constexpr bool is_castable = cuda::std::is_convertible_v<float, DstT>;
+// Castings between complex and boolean.
+template <typename T>
+struct CastOp<complex_t<T>, bool> {
+  static constexpr bool is_castable = true;
 
-  __device__ DstT operator()(cuComplex x) {
-    static_assert(!cuda::std::is_same_v<cuComplex, DstT>);
-    return static_cast<DstT>(cuCrealf(x));
+  __device__ bool operator()(complex_t<T> x) {
+    return x.real() != 0 && x.imag() != 0;
+  }
+};
+
+template <typename T>
+struct CastOp<bool, complex_t<T>> {
+  static constexpr bool is_castable = true;
+
+  __device__ complex_t<T> operator()(bool x) {
+    return x ? complex_t<T>{1, 1} : complex_t<T>{0, 0};
+  }
+};
+
+// Converting a complex number to real number discards the imaginary part.
+template <typename T, typename DstT>
+struct CastOp<complex_t<T>, DstT, cuda::std::enable_if_t<!is_complex_v<DstT>>> {
+  static constexpr bool is_castable = cuda::std::is_convertible_v<T, DstT>;
+
+  __device__ DstT operator()(complex_t<T> x) {
+    static_assert(!is_complex_v<DstT>);
+    return static_cast<DstT>(x.real());
   }
 };
 
 // Allow converting a real number to complex number.
-template <typename SrcT>
-struct CastOp<
-    SrcT,
-    cuComplex,
-    cuda::std::enable_if_t<!cuda::std::is_same_v<SrcT, cuComplex>>> {
-  static constexpr bool is_castable = cuda::std::is_convertible_v<SrcT, float>;
+template <typename SrcT, typename T>
+struct CastOp<SrcT, complex_t<T>, cuda::std::enable_if_t<!is_complex_v<SrcT>>> {
+  static constexpr bool is_castable = cuda::std::is_convertible_v<SrcT, T>;
 
-  __device__ cuComplex operator()(SrcT x) {
-    static_assert(!cuda::std::is_same_v<SrcT, cuComplex>);
-    return cuComplex{static_cast<float>(x), 0};
+  __device__ complex_t<T> operator()(SrcT x) {
+    static_assert(!is_complex_v<SrcT>);
+    return complex_t<T>{static_cast<T>(x), 0};
   }
 };
 
+// Do nothing when no casting is needed.
 template <typename SrcT, typename DstT>
 struct CastOp<
     SrcT,
@@ -57,9 +74,51 @@ struct CastOp<
   }
 };
 
+// In CUDA 11 the half types do not define conversions between some types,
+// provide fallbacks here.
+#if CUDART_VERSION < 12000
+template <typename SrcT, typename DstT>
+struct CastOp<
+    SrcT,
+    DstT,
+    cuda::std::enable_if_t<
+        !cuda::std::is_convertible_v<SrcT, DstT> && !is_complex_v<SrcT> &&
+        (cuda::std::is_same_v<DstT, __half> ||
+         cuda::std::is_same_v<DstT, __nv_bfloat16>)>> {
+  static constexpr bool is_castable = true;
+
+  __device__ DstT operator()(SrcT x) {
+    return DstT(static_cast<float>(x));
+  }
+};
+
+template <typename SrcT, typename DstT>
+struct CastOp<
+    SrcT,
+    DstT,
+    cuda::std::enable_if_t<
+        !cuda::std::is_convertible_v<SrcT, DstT> && !is_complex_v<SrcT> &&
+        !cuda::std::is_same_v<DstT, __half> &&
+        !cuda::std::is_same_v<DstT, __nv_bfloat16> &&
+        (cuda::std::is_same_v<SrcT, __half> ||
+         cuda::std::is_same_v<SrcT, __nv_bfloat16>)>> {
+  static constexpr bool is_castable = true;
+
+  __device__ DstT operator()(SrcT x) {
+    return DstT(static_cast<float>(x));
+  }
+};
+#endif // CUDART_VERSION < 12000
+
+// Helper to deduce the SrcT.
+template <typename DstT, typename SrcT>
+inline __host__ __device__ auto cast_to(SrcT x) {
+  return CastOp<SrcT, DstT>{}(x);
+}
+
 // Return an iterator that cast the value to DstT using CastOp.
 template <typename DstT, typename Iterator>
-__host__ __device__ auto make_cast_iterator(Iterator it) {
+inline __host__ __device__ auto make_cast_iterator(Iterator it) {
   using SrcT = typename cuda::std::iterator_traits<Iterator>::value_type;
   if constexpr (std::is_same_v<SrcT, DstT>) {
     return it;

mlx/backend/cuda/device/complex.cuh

@@ -0,0 +1,60 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+// Make multiplication and division faster.
+#define LIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS
+
+#include <cuda/std/complex>
+#include <cuda/std/type_traits>
+
+namespace mlx::core::cu {
+
+// TODO: Consider using a faster implementation as cuda::std::complex has to
+// conform to C++ standard.
+template <typename T>
+using complex_t = cuda::std::complex<T>;
+
+using complex64_t = complex_t<float>;
+using complex128_t = complex_t<double>;
+
+template <typename T>
+struct is_complex : cuda::std::false_type {};
+
+template <typename T>
+struct is_complex<cuda::std::complex<T>> : cuda::std::true_type {};
+
+template <typename T>
+inline constexpr bool is_complex_v = is_complex<T>::value;
+
+// cuda::std::complex is missing some operators.
+template <typename T>
+inline __host__ __device__ complex_t<T> operator%(
+    complex_t<T> a,
+    complex_t<T> b) {
+  T r = a.real() - floor(a.real() / b.real()) * b.real();
+  T i = a.imag() - floor(a.imag() / b.imag()) * b.imag();
+  return complex_t<T>{r, i};
+}
+
+template <typename T>
+inline __host__ __device__ bool operator>(complex_t<T> a, complex_t<T> b) {
+  return (a.real() > b.real()) || (a.real() == b.real() && a.imag() > b.imag());
+}
+
+template <typename T>
+inline __host__ __device__ bool operator<(complex_t<T> a, complex_t<T> b) {
+  return operator>(b, a);
+}
+
+template <typename T>
+inline __host__ __device__ bool operator<=(complex_t<T> a, complex_t<T> b) {
+  return !(a > b);
+}
+
+template <typename T>
+inline __host__ __device__ bool operator>=(complex_t<T> a, complex_t<T> b) {
+  return !(a < b);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/device/cucomplex_math.cuh

@@ -1,240 +0,0 @@
-// Copyright © 2025 Apple Inc.
-// Copyright © 2017-2024 The Simons Foundation, Inc.
-//
-// FINUFFT is licensed under the Apache License, Version 2.0 (the
-// "License"); you may not use this file except in compliance with the
-// License.  You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//
-// Forked from
-// https://github.com/flatironinstitute/finufft/blob/main/include/cufinufft/contrib/helper_math.h
-
-#pragma once
-
-#include <cuComplex.h>
-
-// This header provides some helper functions for cuComplex types.
-// It mainly wraps existing CUDA implementations to provide operator overloads
-// e.g. cuAdd, cuSub, cuMul, cuDiv, cuCreal, cuCimag, cuCabs, cuCarg, cuConj are
-// all provided by CUDA
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator+(const cuDoubleComplex& a, const cuDoubleComplex& b) {
-  return cuCadd(a, b);
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator-(const cuDoubleComplex& a, const cuDoubleComplex& b) {
-  return cuCsub(a, b);
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator*(const cuDoubleComplex& a, const cuDoubleComplex& b) {
-  return cuCmul(a, b);
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator/(const cuDoubleComplex& a, const cuDoubleComplex& b) {
-  return cuCdiv(a, b);
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator%(const cuDoubleComplex& a, const cuDoubleComplex& b) {
-  double r = cuCreal(a) - (floorf(cuCreal(a) / cuCreal(b)) * cuCreal(b));
-  double i = cuCimag(a) - (floorf(cuCimag(a) / cuCimag(b)) * cuCimag(b));
-  return make_cuDoubleComplex(r, i);
-}
-
-__forceinline__ __host__ __device__ bool operator==(
-    const cuDoubleComplex& a,
-    const cuDoubleComplex& b) {
-  return cuCreal(a) == cuCreal(b) && cuCimag(a) == cuCimag(b);
-}
-
-__forceinline__ __host__ __device__ bool operator!=(
-    const cuDoubleComplex& a,
-    const cuDoubleComplex& b) {
-  return !(a == b);
-}
-
-__forceinline__ __host__ __device__ bool operator>(
-    const cuDoubleComplex& a,
-    const cuDoubleComplex& b) {
-  double mag_a = sqrt(cuCreal(a) * cuCreal(a) + cuCimag(a) * cuCimag(a));
-  double mag_b = sqrt(cuCreal(b) * cuCreal(b) + cuCimag(b) * cuCimag(b));
-  return mag_a > mag_b;
-}
-
-__forceinline__ __host__ __device__ bool operator>=(
-    const cuDoubleComplex& a,
-    const cuDoubleComplex& b) {
-  return a > b || a == b;
-}
-
-__forceinline__ __host__ __device__ bool operator<(
-    const cuDoubleComplex& a,
-    const cuDoubleComplex& b) {
-  return b > a;
-}
-
-__forceinline__ __host__ __device__ bool operator<=(
-    const cuDoubleComplex& a,
-    const cuDoubleComplex& b) {
-  return b > a || a == b;
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator+(const cuDoubleComplex& a, double b) {
-  return make_cuDoubleComplex(cuCreal(a) + b, cuCimag(a));
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator+(double a, const cuDoubleComplex& b) {
-  return make_cuDoubleComplex(a + cuCreal(b), cuCimag(b));
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator-(const cuDoubleComplex& a, double b) {
-  return make_cuDoubleComplex(cuCreal(a) - b, cuCimag(a));
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator-(double a, const cuDoubleComplex& b) {
-  return make_cuDoubleComplex(a - cuCreal(b), -cuCimag(b));
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator*(const cuDoubleComplex& a, double b) {
-  return make_cuDoubleComplex(cuCreal(a) * b, cuCimag(a) * b);
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator*(double a, const cuDoubleComplex& b) {
-  return make_cuDoubleComplex(a * cuCreal(b), a * cuCimag(b));
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator/(const cuDoubleComplex& a, double b) {
-  return make_cuDoubleComplex(cuCreal(a) / b, cuCimag(a) / b);
-}
-
-__forceinline__ __host__ __device__ cuDoubleComplex
-operator/(double a, const cuDoubleComplex& b) {
-  double denom = cuCreal(b) * cuCreal(b) + cuCimag(b) * cuCimag(b);
-  return make_cuDoubleComplex(
-      (a * cuCreal(b)) / denom, (-a * cuCimag(b)) / denom);
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator+(const cuFloatComplex& a, const cuFloatComplex& b) {
-  return cuCaddf(a, b);
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator-(const cuFloatComplex& a, const cuFloatComplex& b) {
-  return cuCsubf(a, b);
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator*(const cuFloatComplex& a, const cuFloatComplex& b) {
-  return cuCmulf(a, b);
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator/(const cuFloatComplex& a, const cuFloatComplex& b) {
-  return cuCdivf(a, b);
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator%(const cuFloatComplex& a, const cuFloatComplex& b) {
-  float r = cuCrealf(a) - (floorf(cuCrealf(a) / cuCrealf(b)) * cuCrealf(b));
-  float i = cuCimagf(a) - (floorf(cuCimagf(a) / cuCimagf(b)) * cuCimagf(b));
-  return make_cuFloatComplex(r, i);
-}
-
-__forceinline__ __host__ __device__ bool operator==(
-    const cuFloatComplex& a,
-    const cuFloatComplex& b) {
-  return cuCrealf(a) == cuCrealf(b) && cuCimagf(a) == cuCimagf(b);
-}
-
-__forceinline__ __host__ __device__ bool operator!=(
-    const cuFloatComplex& a,
-    const cuFloatComplex& b) {
-  return !(a == b);
-}
-
-__forceinline__ __host__ __device__ bool operator>(
-    const cuFloatComplex& a,
-    const cuFloatComplex& b) {
-  float mag_a = sqrt(cuCrealf(a) * cuCrealf(a) + cuCimagf(a) * cuCimagf(a));
-  float mag_b = sqrt(cuCrealf(b) * cuCrealf(b) + cuCimagf(b) * cuCimagf(b));
-  return mag_a > mag_b;
-}
-
-__forceinline__ __host__ __device__ bool operator>=(
-    const cuFloatComplex& a,
-    const cuFloatComplex& b) {
-  return a > b || a == b;
-}
-
-__forceinline__ __host__ __device__ bool operator<(
-    const cuFloatComplex& a,
-    const cuFloatComplex& b) {
-  return b > a;
-}
-
-__forceinline__ __host__ __device__ bool operator<=(
-    const cuFloatComplex& a,
-    const cuFloatComplex& b) {
-  return b > a || a == b;
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator+(const cuFloatComplex& a, float b) {
-  return make_cuFloatComplex(cuCrealf(a) + b, cuCimagf(a));
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator+(float a, const cuFloatComplex& b) {
-  return make_cuFloatComplex(a + cuCrealf(b), cuCimagf(b));
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator-(const cuFloatComplex& a, float b) {
-  return make_cuFloatComplex(cuCrealf(a) - b, cuCimagf(a));
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator-(float a, const cuFloatComplex& b) {
-  return make_cuFloatComplex(a - cuCrealf(b), -cuCimagf(b));
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator*(const cuFloatComplex& a, float b) {
-  return make_cuFloatComplex(cuCrealf(a) * b, cuCimagf(a) * b);
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator*(float a, const cuFloatComplex& b) {
-  return make_cuFloatComplex(a * cuCrealf(b), a * cuCimagf(b));
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator/(const cuFloatComplex& a, float b) {
-  return make_cuFloatComplex(cuCrealf(a) / b, cuCimagf(a) / b);
-}
-
-__forceinline__ __host__ __device__ cuFloatComplex
-operator/(float a, const cuFloatComplex& b) {
-  float denom = cuCrealf(b) * cuCrealf(b) + cuCimagf(b) * cuCimagf(b);
-  return make_cuFloatComplex(
-      (a * cuCrealf(b)) / denom, (-a * cuCimagf(b)) / denom);
-}

mlx/backend/cuda/device/unary_ops.cuh

@@ -14,8 +14,6 @@ struct Abs {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_unsigned_v<T>) {
       return x;
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {sqrt(cuCrealf(x) * cuCrealf(x) + cuCimagf(x) * cuCimagf(x)), 0};
     } else {
       return abs(x);
     }
@@ -27,8 +25,6 @@ struct ArcCos {
   __device__ T operator()(T x) {
     return acos(x);
   }
-
-  __device__ cuComplex operator()(cuComplex x);
 };
 
 struct ArcCosh {
@@ -43,8 +39,6 @@ struct ArcSin {
   __device__ T operator()(T x) {
     return asin(x);
   }
-
-  __device__ cuComplex operator()(cuComplex x);
 };
 
 struct ArcSinh {
@@ -59,8 +53,6 @@ struct ArcTan {
   __device__ T operator()(T x) {
     return atan(x);
   }
-
-  __device__ cuComplex operator()(cuComplex x);
 };
 
 struct ArcTanh {
@@ -82,6 +74,8 @@ struct Ceil {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return x;
+    } else if constexpr (is_complex_v<T>) {
+      return T{ceil(x.real()), ceil(x.imag())};
     } else {
       return ceil(x);
     }
@@ -89,34 +83,23 @@ struct Ceil {
 };
 
 struct Conjugate {
-  __device__ cuComplex operator()(cuComplex x) {
-    return {cuCrealf(x), -cuCimagf(x)};
+  template <typename T>
+  __device__ complex_t<T> operator()(complex_t<T> x) {
+    return conj(x);
   }
 };
 
 struct Cos {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          cos(cuCrealf(x)) * cosh(cuCimagf(x)),
-          -sin(cuCrealf(x)) * sinh(cuCimagf(x))};
-    } else {
-      return cos(x);
-    }
+    return cos(x);
   }
 };
 
 struct Cosh {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          cosh(cuCrealf(x)) * cos(cuCimagf(x)),
-          sinh(cuCrealf(x)) * sin(cuCimagf(x))};
-    } else {
-      return cosh(x);
-    }
+    return cosh(x);
   }
 };
 
@@ -149,12 +132,7 @@ struct ErfInv {
 struct Exp {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      auto m = exp(cuCrealf(x));
-      return {m * cos(cuCimagf(x)), m * sinh(cuCimagf(x))};
-    } else {
-      return exp(x);
-    }
+    return exp(x);
   }
 };
 
@@ -176,6 +154,8 @@ struct Floor {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return x;
+    } else if constexpr (is_complex_v<T>) {
+      return T{floor(x.real()), floor(x.imag())};
     } else {
       return floor(x);
     }
@@ -183,30 +163,25 @@ struct Floor {
 };
 
 struct Imag {
-  __device__ float operator()(cuComplex x) {
-    return cuCimagf(x);
+  template <typename T>
+  __device__ auto operator()(complex_t<T> x) {
+    return x.imag();
   }
 };
 
 struct Log {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      auto r = log(cuCrealf(Abs{}(x)));
-      auto i = atan2f(cuCimagf(x), cuCrealf(x));
-      return {r, i};
-    } else {
-      return log(x);
-    }
+    return log(x);
   }
 };
 
 struct Log2 {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+    if constexpr (is_complex_v<T>) {
       auto y = Log{}(x);
-      return {cuCrealf(y) / CUDART_LN2_F, cuCimagf(y) / CUDART_LN2_F};
+      return {y.real() / CUDART_LN2_F, y.imag() / CUDART_LN2_F};
     } else {
       return log2(x);
     }
@@ -216,20 +191,31 @@ struct Log2 {
 struct Log10 {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      auto y = Log{}(x);
-      return {cuCrealf(y) / CUDART_LNT_F, cuCimagf(y) / CUDART_LNT_F};
-      return y;
-    } else {
-      return log10(x);
-    }
+    return log10(x);
   }
 };
 
 struct Log1p {
   template <typename T>
-  __device__ T operator()(T x) {
-    return log1p(x);
+  __device__ T operator()(T z) {
+    if constexpr (is_complex_v<T>) {
+      float x = z.real();
+      float y = z.imag();
+      float zabs = Abs{}(z).real();
+      float theta = atan2f(y, x + 1);
+      if (zabs < 0.5f) {
+        float r = x * (2 + x) + y * y;
+        if (r == 0) { // handle underflow
+          return {x, theta};
+        }
+        return {0.5f * log1pf(r), theta};
+      } else {
+        float z0 = hypotf(x + 1, y);
+        return {logf(z0), theta};
+      }
+    } else {
+      return log1p(z);
+    }
   }
 };
 
@@ -242,8 +228,8 @@ struct LogicalNot {
 struct Negative {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return 0 - x;
+    if constexpr (is_complex_v<T>) {
+      return T{0, 0} - x;
     } else {
       return -x;
     }
@@ -251,16 +237,17 @@ struct Negative {
 };
 
 struct Real {
-  __device__ float operator()(cuComplex x) {
-    return cuCrealf(x);
+  template <typename T>
+  __device__ auto operator()(complex_t<T> x) {
+    return x.real();
   }
 };
 
 struct Round {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {rint(cuCrealf(x)), rint(cuCimagf(x))};
+    if constexpr (is_complex_v<T>) {
+      return {rint(x.real()), rint(x.imag())};
     } else {
       return rint(x);
     }
@@ -280,8 +267,8 @@ struct Sign {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_unsigned_v<T>) {
       return x != 0;
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (cuCrealf(x) == 0 && cuCimagf(x) == 0) {
+    } else if constexpr (is_complex_v<T>) {
+      if (x.real() == 0 && x.imag() == 0) {
         return x;
       } else {
         return x / Abs()(x);
@@ -297,26 +284,14 @@ struct Sign {
 struct Sin {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          sin(cuCrealf(x)) * cosh(cuCimagf(x)),
-          cos(cuCrealf(x)) * sinh(cuCimagf(x))};
-    } else {
-      return sin(x);
-    }
+    return sin(x);
   }
 };
 
 struct Sinh {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          sinh(cuCrealf(x)) * cos(cuCimagf(x)),
-          cosh(cuCrealf(x)) * sin(cuCimagf(x))};
-    } else {
-      return sinh(x);
-    }
+    return sinh(x);
   }
 };
 
@@ -332,77 +307,31 @@ struct Sqrt {
   __device__ T operator()(T x) {
     return sqrt(x);
   }
-
-  __device__ cuComplex operator()(cuComplex x) {
-    auto xr = cuCrealf(x);
-    auto xi = cuCimagf(x);
-    if (xr == 0.0f && xi == 0.0f) {
-      return {0.0f, 0.0f};
-    }
-    auto r = cuCrealf(Abs{}(x));
-    auto a = sqrt((r + xr) / 2.0f);
-    auto b_abs = sqrt((r - xr) / 2.0f);
-    auto b = copysign(b_abs, xi);
-    return {a, b};
-  }
 };
 
 struct Rsqrt {
   template <typename T>
   __device__ T operator()(T x) {
-    return rsqrt(x);
-  }
-  __device__ cuComplex operator()(cuComplex x) {
-    return 1.0f / Sqrt{}(x);
+    if constexpr (is_complex_v<T>) {
+      return 1.0f / Sqrt{}(x);
+    } else {
+      return rsqrt(x);
+    }
   }
 };
 
 struct Tan {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      float tan_a = tan(cuCrealf(x));
-      float tanh_b = tanh(cuCimagf(x));
-      float t1 = tan_a * tanh_b;
-      float denom = 1. + t1 * t1;
-      return {(tan_a - tanh_b * t1) / denom, (tanh_b + tan_a * t1) / denom};
-    } else {
-      return tan(x);
-    }
+    return tan(x);
   }
 };
 
 struct Tanh {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      float tanh_a = tanh(cuCrealf(x));
-      float tan_b = tan(cuCimagf(x));
-      float t1 = tanh_a * tan_b;
-      float denom = 1. + t1 * t1;
-      return {(tanh_a + tan_b * t1) / denom, (tan_b - tanh_a * t1) / denom};
-    } else {
-      return tanh(x);
-    }
+    return tanh(x);
   }
 };
 
-__device__ cuComplex ArcCos::operator()(cuComplex x) {
-  auto i = cuComplex{0.0, 1.0};
-  auto y = Log{}(x + i * Sqrt{}(1.0 - x * x));
-  return {cuCimagf(y), -cuCrealf(y)};
-};
-
-__device__ cuComplex ArcSin::operator()(cuComplex x) {
-  auto i = cuComplex{0.0f, 1.0f};
-  auto y = Log{}(i * x + Sqrt{}(1.0f - x * x));
-  return {cuCimagf(y), -cuCrealf(y)};
-};
-
-__device__ cuComplex ArcTan::operator()(cuComplex x) {
-  auto i = cuComplex{0.0f, 1.0f};
-  auto ix = i * x;
-  return (1.0f / cuComplex{0.0f, 2.0f}) * Log{}((1.0f + ix) / (1.0f - ix));
-};
-
 } // namespace mlx::core::cu

mlx/backend/cuda/device/utils.cuh

@@ -8,9 +8,9 @@
 
 #pragma once
 
+#include "mlx/backend/cuda/device/complex.cuh"
 #include "mlx/backend/cuda/device/config.h"
 
-#include <cuComplex.h>
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 #include <cuda/std/array>
@@ -28,6 +28,41 @@ namespace mlx::core::cu {
 using Shape = cuda::std::array<int32_t, MAX_NDIM>;
 using Strides = cuda::std::array<int64_t, MAX_NDIM>;
 
+// Vectorized load/store.
+template <typename T, int N>
+struct alignas(sizeof(T) * N) AlignedVector {
+  T val[N];
+};
+
+template <int N, typename T>
+inline __device__ AlignedVector<T, N> load_vector(
+    const T* ptr,
+    uint32_t offset) {
+  auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
+  return from[offset];
+}
+
+template <int N, typename T>
+inline __device__ void
+store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
+  auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
+  to[offset] = vec;
+}
+
+// Helper for accessing strided data.
+template <typename T>
+struct StridedIterator {
+  T it;
+  int64_t stride;
+
+  __host__ __device__ StridedIterator(T it, int64_t stride)
+      : it(it), stride(stride) {}
+
+  __host__ __device__ auto operator[](int i) const {
+    return it[i * stride];
+  }
+};
+
 ///////////////////////////////////////////////////////////////////////////////
 // Type limits utils
 ///////////////////////////////////////////////////////////////////////////////
@@ -78,20 +113,20 @@ struct Limits<
     return cuda::std::numeric_limits<T>::infinity();
   }
   static constexpr __host__ __device__ T min() {
-#if defined(__CUDA_ARCH__) || CUDART_VERSION >= 12000
-    return -cuda::std::numeric_limits<T>::infinity();
-#else
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
     return -cuda::std::numeric_limits<float>::infinity();
+#else
+    return -cuda::std::numeric_limits<T>::infinity();
 #endif
   }
   static constexpr __host__ __device__ T finite_max() {
     return cuda::std::numeric_limits<T>::max();
   }
   static constexpr __host__ __device__ T finite_min() {
-#if defined(__CUDA_ARCH__) || CUDART_VERSION >= 12000
-    return cuda::std::numeric_limits<T>::lowest();
-#else
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
     return cuda::std::numeric_limits<float>::lowest();
+#else
+    return cuda::std::numeric_limits<T>::lowest();
 #endif
   }
 };
@@ -106,13 +141,13 @@ struct Limits<bool> {
   }
 };
 
-template <>
-struct Limits<cuComplex> {
-  static constexpr __host__ __device__ cuComplex max() {
-    return {Limits<float>::max(), Limits<float>::max()};
+template <typename T>
+struct Limits<complex_t<T>> {
+  static constexpr __host__ __device__ complex_t<T> max() {
+    return {Limits<T>::max(), Limits<T>::max()};
   }
-  static constexpr __host__ __device__ cuComplex min() {
-    return {Limits<float>::min(), Limits<float>::min()};
+  static constexpr __host__ __device__ complex_t<T> min() {
+    return {Limits<T>::min(), Limits<T>::min()};
   }
 };
 
@@ -183,20 +218,8 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_nd(
   return cuda::std::make_tuple(a_loc, b_loc, c_loc);
 }
 
-// Optimized version when ndim is larger than 4.
 template <typename IdxT = int64_t>
-inline __host__ __device__ IdxT
-elem_to_loc_4d(IdxT elem, const int* shape, const int64_t* strides, int ndim) {
-  IdxT loc = 0;
-  for (int i = ndim - 1; i >= 0; --i) {
-    loc += (elem % shape[i]) * IdxT(strides[i]);
-    elem /= shape[i];
-  }
-  return loc;
-}
-
-template <typename IdxT = int64_t>
-inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_4d(
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc(
     IdxT elem,
     const int* shape,
     const int64_t* a_strides,
@@ -214,7 +237,7 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_4d(
 }
 
 template <typename IdxT = int64_t>
-inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_4d(
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc(
     IdxT elem,
     const int* shape,
     const int64_t* a_strides,
@@ -338,21 +361,4 @@ struct LoopedElemToLoc<1, false, OffsetT> {
   }
 };
 
-inline __device__ cuComplex log1p(cuComplex in) {
-  float x = cuCrealf(in);
-  float y = cuCimagf(in);
-  float zabs = sqrt(x * x + y * y);
-  float theta = atan2f(y, x + 1);
-  if (zabs < 0.5f) {
-    float r = x * (2 + x) + y * y;
-    if (r == 0) { // handle underflow
-      return {x, theta};
-    }
-    return {0.5f * log1pf(r), theta};
-  } else {
-    auto z0 = sqrt((x + 1) * (x + 1) + y * y);
-    return {log(z0), theta};
-  }
-}
-
 } // namespace mlx::core::cu

mlx/backend/cuda/eval.cpp

@@ -19,8 +19,6 @@ void new_stream(Stream s) {
   cudaFree(nullptr);
   // Ensure the static stream objects get created.
   cu::get_command_encoder(s);
-  // The main thread is safe to free buffers.
-  cu::allocator().register_this_thread();
 }
 
 void eval(array& arr) {

mlx/backend/cuda/event.cu

@@ -90,8 +90,6 @@ bool CudaEvent::completed() const {
 // SharedEvent implementations
 ///////////////////////////////////////////////////////////////////////////////
 
-namespace {
-
 __host__ __device__ void event_wait(SharedEvent::Atomic* ac, uint64_t value) {
   uint64_t current;
   while ((current = ac->load()) < value) {
@@ -112,26 +110,26 @@ __global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
   event_signal(ac, value);
 }
 
-} // namespace
+SharedEvent::Atomic* to_atomic(std::shared_ptr<Buffer> buf) {
+  return static_cast<SharedEvent::Atomic*>(buf->raw_ptr());
+}
 
 SharedEvent::SharedEvent() {
-  // Allocate cuda::atomic on managed memory.
-  Atomic* ac;
-  CHECK_CUDA_ERROR(cudaMallocManaged(&ac, sizeof(Atomic)));
-  new (ac) Atomic(0);
-  ac_ = std::shared_ptr<Atomic>(ac, [](Atomic* ptr) {
-    ptr->~Atomic();
-    allocator().cuda_free(ptr);
-  });
+  buf_ = std::shared_ptr<Buffer>(
+      new Buffer{allocator().malloc(sizeof(Atomic))}, [](Buffer* ptr) {
+        allocator().free(*ptr);
+        delete ptr;
+      });
+  *static_cast<uint64_t*>(buf_->raw_ptr()) = 0;
 }
 
 void SharedEvent::wait(uint64_t value) {
   nvtx3::scoped_range r("cu::SharedEvent::wait");
-  event_wait(ac_.get(), value);
+  event_wait(to_atomic(buf_), value);
 }
 
 void SharedEvent::wait(cudaStream_t stream, uint64_t value) {
-  event_wait_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
+  event_wait_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
 }
 
 void SharedEvent::wait(Stream s, uint64_t value) {
@@ -142,17 +140,17 @@ void SharedEvent::wait(Stream s, uint64_t value) {
     auto& encoder = get_command_encoder(s);
     encoder.commit();
     wait(encoder.stream(), value);
-    encoder.add_completed_handler([ac = ac_]() {});
+    encoder.add_completed_handler([buf = buf_]() {});
   }
 }
 
 void SharedEvent::signal(uint64_t value) {
   nvtx3::scoped_range r("cu::SharedEvent::signal");
-  event_signal(ac_.get(), value);
+  event_signal(to_atomic(buf_), value);
 }
 
 void SharedEvent::signal(cudaStream_t stream, uint64_t value) {
-  event_signal_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
+  event_signal_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
 }
 
 void SharedEvent::signal(Stream s, uint64_t value) {
@@ -166,18 +164,18 @@ void SharedEvent::signal(Stream s, uint64_t value) {
     auto& encoder = get_command_encoder(s);
     encoder.commit();
     signal(encoder.stream(), value);
-    encoder.add_completed_handler([ac = ac_]() {});
+    encoder.add_completed_handler([buf = buf_]() {});
   }
 }
 
 bool SharedEvent::is_signaled(uint64_t value) const {
   nvtx3::scoped_range r("cu::SharedEvent::is_signaled");
-  return ac_->load() >= value;
+  return to_atomic(buf_)->load() >= value;
 }
 
 uint64_t SharedEvent::value() const {
   nvtx3::scoped_range r("cu::SharedEvent::value");
-  return ac_->load();
+  return to_atomic(buf_)->load();
 }
 
 } // namespace cu

mlx/backend/cuda/event.h

@@ -2,6 +2,7 @@
 
 #pragma once
 
+#include "mlx/allocator.h"
 #include "mlx/stream.h"
 
 #include <cuda_runtime.h>
@@ -55,12 +56,8 @@ class SharedEvent {
   bool is_signaled(uint64_t value) const;
   uint64_t value() const;
 
-  const std::shared_ptr<Atomic>& atomic() const {
-    return ac_;
-  }
-
  private:
-  std::shared_ptr<Atomic> ac_;
+  std::shared_ptr<mlx::core::allocator::Buffer> buf_;
 };
 
 } // namespace mlx::core::cu

mlx/backend/cuda/gemms/cublas_batched_gemm_12_0.cpp

@@ -0,0 +1,73 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+
+namespace mlx::core::cu {
+
+void Matmul::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const mlx::core::Shape& batch_shape,
+    const mlx::core::Strides& a_batch_strides,
+    const mlx::core::Strides& b_batch_strides) {
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  auto nbatch = out.size() / (M_ * N_ * batch_shape.back());
+  ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
+  ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
+  auto concurrent = encoder.concurrent_context();
+  for (size_t i = 0; i < nbatch; ++i) {
+    run_impl(
+        encoder,
+        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M_ * N_,
+        a.data<int8_t>() + a.itemsize() * a_it.loc,
+        b.data<int8_t>() + b.itemsize() * b_it.loc,
+        nullptr);
+    a_it.step();
+    b_it.step();
+  }
+}
+
+void Matmul::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const array& c,
+    const mlx::core::Shape& batch_shape,
+    const mlx::core::Strides& a_batch_strides,
+    const mlx::core::Strides& b_batch_strides,
+    const mlx::core::Strides& c_batch_strides,
+    float alpha,
+    float beta) {
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  auto nbatch = out.size() / (M_ * N_ * batch_shape.back());
+  ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
+  ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
+  ContiguousIterator c_it(batch_shape, c_batch_strides, batch_shape.size() - 1);
+  auto concurrent = encoder.concurrent_context();
+  for (size_t i = 0; i < nbatch; ++i) {
+    run_impl(
+        encoder,
+        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M_ * N_,
+        a.data<int8_t>() + a.itemsize() * a_it.loc,
+        b.data<int8_t>() + b.itemsize() * b_it.loc,
+        c.data<int8_t>() + c.itemsize() * c_it.loc,
+        alpha,
+        beta);
+    a_it.step();
+    b_it.step();
+    c_it.step();
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/cublas_batched_gemm_12_9.cu

@@ -0,0 +1,206 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core::cu {
+
+namespace cg = cooperative_groups;
+
+__global__ void set_mm_device_pointers(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides a_batch_strides,
+    const __grid_constant__ Strides b_batch_strides,
+    int64_t batch_stride,
+    int batch_ndim,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset] = elem_to_loc(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      batch_ndim);
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+__global__ void set_addmm_device_pointers(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* c_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides a_batch_strides,
+    const __grid_constant__ Strides b_batch_strides,
+    const __grid_constant__ Strides c_batch_strides,
+    int64_t batch_stride,
+    int batch_ndim,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset, c_offset] = elem_to_loc(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      c_batch_strides.data(),
+      batch_ndim);
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] = c_start + item_size * c_offset;
+  pointers[index + 3 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+void set_pointer_mode(cublasLtMatrixLayout_t desc, int batch_count) {
+  auto batch_mode = CUBLASLT_BATCH_MODE_POINTER_ARRAY;
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc,
+      CUBLASLT_MATRIX_LAYOUT_BATCH_MODE,
+      &batch_mode,
+      sizeof(batch_mode)));
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc, CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT, &batch_count, sizeof(int32_t)));
+}
+
+void Matmul::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const mlx::core::Shape& batch_shape,
+    const mlx::core::Strides& a_batch_strides,
+    const mlx::core::Strides& b_batch_strides) {
+  auto batch_count = out.size() / (M_ * N_);
+  set_pointer_mode(a_desc_, batch_count);
+  set_pointer_mode(b_desc_, batch_count);
+  set_pointer_mode(out_desc_, batch_count);
+
+  // Launch kernel to set device offsets
+  auto pointers = array(
+      allocator::malloc(batch_count * sizeof(uint64_t) * 3),
+      {static_cast<int>(batch_count * 3)},
+      uint64);
+
+  encoder.add_temporary(pointers);
+  int block_size = 512;
+  encoder.set_output_array(pointers);
+
+  encoder.add_kernel_node(
+      cu::set_mm_device_pointers,
+      cuda::ceil_div(pointers.size(), block_size),
+      block_size,
+      pointers.data<int8_t*>(),
+      a.data<int8_t>(),
+      b.data<int8_t>(),
+      out.data<int8_t>(),
+      static_cast<int>(out.dtype().size()),
+      const_param(batch_shape),
+      const_param(a_batch_strides),
+      const_param(b_batch_strides),
+      static_cast<int64_t>(M_) * N_,
+      static_cast<int>(batch_shape.size()),
+      batch_count);
+
+  // Run matmul
+  encoder.set_input_array(pointers);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+
+  auto a_pointers = pointers.data<int8_t*>();
+  auto b_pointers = a_pointers + batch_count;
+  auto out_pointers = b_pointers + batch_count;
+  run_impl(
+      encoder,
+      reinterpret_cast<void*>(out_pointers),
+      reinterpret_cast<void*>(a_pointers),
+      reinterpret_cast<void*>(b_pointers),
+      nullptr);
+}
+
+void Matmul::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const array& c,
+    const mlx::core::Shape& batch_shape,
+    const mlx::core::Strides& a_batch_strides,
+    const mlx::core::Strides& b_batch_strides,
+    const mlx::core::Strides& c_batch_strides,
+    float alpha,
+    float beta) {
+  auto batch_count = out.size() / (M_ * N_);
+  set_pointer_mode(a_desc_, batch_count);
+  set_pointer_mode(b_desc_, batch_count);
+  set_pointer_mode(c_desc_, batch_count);
+  set_pointer_mode(out_desc_, batch_count);
+
+  // Launch kernel to set device offsets
+  auto pointers = array(
+      allocator::malloc(batch_count * sizeof(uint64_t) * 4),
+      {static_cast<int>(batch_count * 4)},
+      uint64);
+
+  encoder.add_temporary(pointers);
+  int block_size = 512;
+  encoder.set_output_array(pointers);
+  encoder.add_kernel_node(
+      cu::set_addmm_device_pointers,
+      cuda::ceil_div(pointers.size(), block_size),
+      block_size,
+      pointers.data<int8_t*>(),
+      a.data<int8_t>(),
+      b.data<int8_t>(),
+      c.data<int8_t>(),
+      out.data<int8_t>(),
+      static_cast<int>(out.dtype().size()),
+      const_param(batch_shape),
+      const_param(a_batch_strides),
+      const_param(b_batch_strides),
+      const_param(c_batch_strides),
+      static_cast<int64_t>(M_) * N_,
+      static_cast<int>(batch_shape.size()),
+      batch_count);
+
+  // Run matmul
+  encoder.set_input_array(pointers);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  auto a_pointers = pointers.data<int8_t*>();
+  auto b_pointers = a_pointers + batch_count;
+  auto c_pointers = b_pointers + batch_count;
+  auto out_pointers = c_pointers + batch_count;
+  run_impl(
+      encoder,
+      reinterpret_cast<void*>(out_pointers),
+      reinterpret_cast<void*>(a_pointers),
+      reinterpret_cast<void*>(b_pointers),
+      reinterpret_cast<void*>(c_pointers),
+      alpha,
+      beta);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -0,0 +1,282 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/utils.h"
+
+#include <fmt/format.h>
+
+namespace mlx::core::cu {
+
+struct CublasPreference {
+  CublasPreference(Device& device) {
+    // The recommended cublas workspace size is 4 MiB for pre-Hopper and 32 MiB
+    // for Hopper+:
+    // https://docs.nvidia.com/cuda/cublas/#cublassetworkspace
+    uint64_t MiB = 1024 * 1024;
+    uint64_t workspace_size =
+        device.compute_capability_major() >= 9 ? 32 * MiB : 4 * MiB;
+
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceCreate(&pref_));
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceSetAttribute(
+        pref_,
+        CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
+        &workspace_size,
+        sizeof(uint64_t)));
+  }
+
+  ~CublasPreference() {
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceDestroy(pref_));
+  }
+
+  cublasLtMatmulPreference_t pref_{nullptr};
+};
+
+cublasLtMatmulPreference_t cublas_preference(Device& device) {
+  static CublasPreference pref(device);
+  return pref.pref_;
+}
+
+cublasComputeType_t dtype_to_compute_type(Dtype dtype) {
+  switch (dtype) {
+    case float16:
+      return CUBLAS_COMPUTE_32F;
+    case bfloat16:
+      return CUBLAS_COMPUTE_32F;
+    case float32:
+      return mlx::core::env::enable_tf32() ? CUBLAS_COMPUTE_32F_FAST_TF32
+                                           : CUBLAS_COMPUTE_32F;
+    case float64:
+    case complex64:
+      return CUBLAS_COMPUTE_64F;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Matmul: {}.", dtype_to_string(dtype)));
+  }
+}
+
+cudaDataType_t dtype_to_cublas_type(Dtype dtype) {
+  switch (dtype) {
+    case float16:
+      return CUDA_R_16F;
+    case bfloat16:
+      return CUDA_R_16BF;
+    case float32:
+      return CUDA_R_32F;
+    case float64:
+      return CUDA_R_64F;
+    case complex64:
+      return CUDA_C_32F;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Matmul: {}.", dtype_to_string(dtype)));
+  }
+}
+
+cublasLtMatrixLayout_t create_matrix_layout(
+    cudaDataType_t type,
+    uint64_t rows,
+    uint64_t cols,
+    bool transposed,
+    int64_t ld,
+    int32_t batch_count,
+    int64_t batch_stride) {
+  cublasLtMatrixLayout_t desc;
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutCreate(&desc, type, rows, cols, ld));
+  cublasLtOrder_t order = transposed ? CUBLASLT_ORDER_COL : CUBLASLT_ORDER_ROW;
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc, CUBLASLT_MATRIX_LAYOUT_ORDER, &order, sizeof(cublasLtOrder_t)));
+  if (batch_count > 1) {
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+        desc,
+        CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT,
+        &batch_count,
+        sizeof(int32_t)));
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+        desc,
+        CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET,
+        &batch_stride,
+        sizeof(int64_t)));
+  }
+  return desc;
+}
+
+Matmul::Matmul(
+    Device& device,
+    Dtype dtype,
+    bool a_transposed,
+    uint64_t a_rows,
+    uint64_t a_cols,
+    int64_t lda,
+    bool b_transposed,
+    uint64_t b_rows,
+    uint64_t b_cols,
+    int64_t ldb,
+    int32_t batch_count,
+    int64_t a_batch_stride,
+    int64_t b_batch_stride)
+    : handle_(device.lt_handle()),
+      pref_(cublas_preference(device)),
+      M_(a_rows),
+      N_(b_cols) {
+  heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
+
+  auto scale_type = dtype_to_cublas_type(dtype);
+  if (dtype == bfloat16 || dtype == float16) {
+    scale_type = CUDA_R_32F;
+  }
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescCreate(
+      &matmul_desc_, dtype_to_compute_type(dtype), scale_type));
+  int32_t pointer_mode = CUBLASLT_POINTER_MODE_HOST;
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_POINTER_MODE,
+      &pointer_mode,
+      sizeof(int32_t)));
+  cublasOperation_t op = CUBLAS_OP_N;
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_TRANSA,
+      &op,
+      sizeof(cublasOperation_t)));
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_TRANSB,
+      &op,
+      sizeof(cublasOperation_t)));
+
+  auto type = dtype_to_cublas_type(dtype);
+  a_desc_ = create_matrix_layout(
+      type, a_rows, a_cols, a_transposed, lda, batch_count, a_batch_stride);
+  b_desc_ = create_matrix_layout(
+      type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
+  out_desc_ = create_matrix_layout(
+      type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
+}
+
+Matmul::Matmul(
+    Device& device,
+    Dtype dtype,
+    bool a_transposed,
+    uint64_t a_rows,
+    uint64_t a_cols,
+    int64_t lda,
+    bool b_transposed,
+    uint64_t b_rows,
+    uint64_t b_cols,
+    int64_t ldb,
+    int64_t ldc,
+    int32_t batch_count,
+    int64_t a_batch_stride,
+    int64_t b_batch_stride,
+    int64_t c_batch_stride)
+    : Matmul(
+          device,
+          dtype,
+          a_transposed,
+          a_rows,
+          a_cols,
+          lda,
+          b_transposed,
+          b_rows,
+          b_cols,
+          ldb,
+          batch_count,
+          a_batch_stride,
+          b_batch_stride) {
+  auto type = dtype_to_cublas_type(dtype);
+  c_desc_ = create_matrix_layout(
+      type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
+}
+
+Matmul::~Matmul() {
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(a_desc_));
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(b_desc_));
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(c_desc_));
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(out_desc_));
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescDestroy(matmul_desc_));
+}
+
+void Matmul::run_impl(
+    cu::CommandEncoder& encoder,
+    void* out,
+    const void* a,
+    const void* b,
+    const void* c,
+    float alpha /* = 1 */,
+    float beta /* = 0 */) {
+  if (heuristic_.state != CUBLAS_STATUS_SUCCESS) {
+    int ret = 0;
+    CHECK_CUBLAS_ERROR(cublasLtMatmulAlgoGetHeuristic(
+        handle_,
+        matmul_desc_,
+        a_desc_,
+        b_desc_,
+        out_desc_, // TODO should that be c_desc is it's set?
+        out_desc_,
+        pref_,
+        1,
+        &heuristic_,
+        &ret));
+    if (ret == 0) {
+      throw std::runtime_error("Can not find algorithm for matmul.");
+    }
+  }
+
+  void* workspace_ptr = nullptr;
+  if (heuristic_.workspaceSize > 0) {
+    array workspace(
+        allocator::malloc(heuristic_.workspaceSize),
+        {static_cast<int>(heuristic_.workspaceSize)},
+        int8);
+    encoder.add_temporary(workspace);
+    workspace_ptr = workspace.data<void>();
+  }
+
+  auto capture = encoder.capture_context();
+  CHECK_CUBLAS_ERROR(cublasLtMatmul(
+      handle_,
+      matmul_desc_,
+      &alpha,
+      a,
+      a_desc_,
+      b,
+      b_desc_,
+      &beta,
+      c ? c : out,
+      c ? c_desc_ : out_desc_,
+      out,
+      out_desc_,
+      &heuristic_.algo,
+      workspace_ptr,
+      heuristic_.workspaceSize,
+      encoder.stream()));
+}
+
+void Matmul::run(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const std::optional<array>& c /* = std::nullopt */,
+    float alpha /* = 1 */,
+    float beta /* = 0 */) {
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  if (c) {
+    encoder.set_input_array(*c);
+  }
+  encoder.set_output_array(out);
+
+  run_impl(
+      encoder,
+      out.data<void>(),
+      a.data<void>(),
+      b.data<void>(),
+      c ? c->data<void>() : nullptr,
+      alpha,
+      beta);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/cublas_gemm.h

@@ -0,0 +1,100 @@
+// Copyright © 2025 Apple Inc.
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/cuda/device.h"
+
+#include <cublasLt.h>
+#include <optional>
+
+namespace mlx::core::cu {
+class Matmul {
+ public:
+  Matmul(
+      Device& device,
+      Dtype dtype,
+      bool a_transposed,
+      uint64_t a_rows,
+      uint64_t a_cols,
+      int64_t lda,
+      bool b_transposed,
+      uint64_t b_rows,
+      uint64_t b_cols,
+      int64_t ldb,
+      int32_t batch_count,
+      int64_t a_batch_stride,
+      int64_t b_batch_stride);
+
+  Matmul(
+      Device& device,
+      Dtype dtype,
+      bool a_transposed,
+      uint64_t a_rows,
+      uint64_t a_cols,
+      int64_t lda,
+      bool b_transposed,
+      uint64_t b_rows,
+      uint64_t b_cols,
+      int64_t ldb,
+      int64_t ldc,
+      int32_t batch_count,
+      int64_t a_batch_stride,
+      int64_t b_batch_stride,
+      int64_t c_batch_stride);
+
+  ~Matmul();
+
+  void run(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const std::optional<array>& c = std::nullopt,
+      float alpha = 1,
+      float beta = 0);
+
+  void run_batched(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const mlx::core::Shape& batch_shape,
+      const mlx::core::Strides& a_batch_strides,
+      const mlx::core::Strides& b_batch_strides);
+
+  void run_batched(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const array& c,
+      const mlx::core::Shape& batch_shape,
+      const mlx::core::Strides& a_batch_strides,
+      const mlx::core::Strides& b_batch_strides,
+      const mlx::core::Strides& c_batch_strides,
+      float alpha,
+      float beta);
+
+ private:
+  void run_impl(
+      cu::CommandEncoder& encoder,
+      void* out,
+      const void* a,
+      const void* b,
+      const void* c,
+      float alpha = 1,
+      float beta = 0);
+
+  uint64_t M_;
+  uint64_t N_;
+  cublasLtMatmulPreference_t pref_{nullptr};
+  cublasLtHandle_t handle_{nullptr};
+  cublasLtMatmulDesc_t matmul_desc_{nullptr};
+  cublasLtMatrixLayout_t a_desc_{nullptr};
+  cublasLtMatrixLayout_t b_desc_{nullptr};
+  cublasLtMatrixLayout_t c_desc_{nullptr};
+  cublasLtMatrixLayout_t out_desc_{nullptr};
+  cublasLtMatmulHeuristicResult_t heuristic_;
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/gemv.cu

@@ -0,0 +1,147 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/gemms/gemv.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+
+namespace mlx::core::cu {
+
+namespace cg = cooperative_groups;
+
+static constexpr int n_per_thread = 4;
+static constexpr int rows_per_block = 8;
+
+template <typename T, int rows_per_block, int n_per_thread>
+__device__ void
+gemv_impl(const T* mat, const T* vec, T* out, int rows, int cols) {
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  auto g_idx = block.group_index();
+  auto t_idx = block.thread_index();
+  int row = g_idx.x * rows_per_block + t_idx.y;
+
+  if (row < rows) {
+    float sum = 0.0f;
+    for (int col = n_per_thread * warp.thread_rank(); col < cols;
+         col += (WARP_SIZE * n_per_thread)) {
+      auto local_mat = load_vector<n_per_thread>(mat + row * cols + col, 0);
+      auto local_vec = load_vector<n_per_thread>(vec + col, 0);
+#pragma unroll
+      for (int j = 0; j < n_per_thread; ++j) {
+        sum += static_cast<float>(local_mat.val[j]) *
+            static_cast<float>(local_vec.val[j]);
+      }
+    }
+
+    sum = cg::reduce(warp, sum, cg::plus<float>{});
+    if (warp.thread_rank() == 0) {
+      out[row] = static_cast<T>(sum);
+    }
+  }
+}
+
+template <typename T, int rows_per_block, int n_per_thread>
+__global__ void
+gemv_single(const T* mat, const T* vec, T* out, int rows, int cols) {
+  gemv_impl<T, rows_per_block, n_per_thread>(mat, vec, out, rows, cols);
+}
+
+template <typename T, int rows_per_block, int n_per_thread>
+__global__ void gemv_batched(
+    const T* mat,
+    const T* vec,
+    T* out,
+    int rows,
+    int cols,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides mat_batch_strides,
+    const __grid_constant__ Strides vec_batch_strides,
+    int batch_ndim) {
+  auto block = cg::this_thread_block();
+  auto batch_idx = block.group_index().y;
+  auto [vec_offset, mat_offset] = elem_to_loc(
+      batch_idx,
+      batch_shape.data(),
+      vec_batch_strides.data(),
+      mat_batch_strides.data(),
+      batch_ndim);
+  gemv_impl<T, rows_per_block, n_per_thread>(
+      mat + mat_offset, vec + vec_offset, out + batch_idx * rows, rows, cols);
+}
+
+bool can_use_gemv(int M, int N, int K, bool a_transposed, bool b_transposed) {
+  return K % (WARP_SIZE * n_per_thread) == 0 &&
+      ((M == 1 && b_transposed) || (N == 1 && !a_transposed));
+}
+
+void gemv(
+    const array& a,
+    const array& b,
+    array& out,
+    int M,
+    int N,
+    int K,
+    uint32_t batch_count,
+    const mlx::core::Shape& batch_shape,
+    const mlx::core::Strides& a_batch_strides,
+    const mlx::core::Strides& b_batch_strides,
+    CommandEncoder& encoder) {
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  dispatch_float_types(out.dtype(), "gemv", [&](auto type_tag) {
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    dim3 block_dims{WARP_SIZE, rows_per_block};
+    const DataType* mat;
+    const DataType* vec;
+    int rows;
+    int cols = K;
+    auto mat_strides = const_param(a_batch_strides);
+    auto vec_strides = const_param(b_batch_strides);
+
+    if (M == 1) {
+      mat = b.data<DataType>();
+      vec = a.data<DataType>();
+      rows = N;
+      std::swap(mat_strides, vec_strides);
+    } else {
+      mat = a.data<DataType>();
+      vec = b.data<DataType>();
+      rows = M;
+    }
+    uint32_t num_blocks_x = (rows + rows_per_block - 1) / rows_per_block;
+    if (batch_count == 1) {
+      auto kernel = gemv_single<DataType, rows_per_block, n_per_thread>;
+      encoder.add_kernel_node(
+          kernel,
+          num_blocks_x,
+          block_dims,
+          mat,
+          vec,
+          out.data<DataType>(),
+          rows,
+          cols);
+    } else {
+      auto kernel = gemv_batched<DataType, rows_per_block, n_per_thread>;
+      encoder.add_kernel_node(
+          kernel,
+          dim3{num_blocks_x, batch_count},
+          block_dims,
+          mat,
+          vec,
+          out.data<DataType>(),
+          rows,
+          cols,
+          const_param(batch_shape),
+          mat_strides,
+          vec_strides,
+          batch_shape.size());
+    }
+  });
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/gemv.h

@@ -0,0 +1,24 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core::cu {
+
+bool can_use_gemv(int M, int N, int K, bool a_transposed, bool b_transposed);
+
+void gemv(
+    const array& a,
+    const array& b,
+    array& out,
+    int M,
+    int N,
+    int K,
+    uint32_t batch_count,
+    const mlx::core::Shape& batch_shape,
+    const mlx::core::Strides& a_batch_strides,
+    const mlx::core::Strides& b_batch_strides,
+    CommandEncoder& encoder);
+
+} // namespace mlx::core::cu

mlx/backend/cuda/iterators/general_iterator.cuh

@@ -1,121 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include <thrust/iterator/iterator_adaptor.h>
-#include <cuda/std/utility>
-
-#include "mlx/backend/cuda/kernel_utils.cuh"
-
-namespace mlx::core::cu {
-
-// Iterating non-contiguous array.
-template <typename Iterator, typename IdxT = int64_t>
-class general_iterator
-    : public thrust::
-          iterator_adaptor<general_iterator<Iterator, IdxT>, Iterator> {
- public:
-  using super_t =
-      thrust::iterator_adaptor<general_iterator<Iterator, IdxT>, Iterator>;
-
-  using reference = typename super_t::reference;
-  using difference_type = typename super_t::difference_type;
-
-  __host__ __device__ general_iterator(
-      Iterator it,
-      IdxT index,
-      int ndim,
-      Shape shape,
-      Strides strides)
-      : super_t(it),
-        index_(index),
-        ndim_(ndim),
-        shape_(cuda::std::move(shape)),
-        strides_(cuda::std::move(strides)) {}
-
-  __host__ __device__ IdxT index() const {
-    return index_;
-  }
-
-  __host__ __device__ const Shape& shape() const {
-    return shape_;
-  }
-
-  __host__ __device__ const Strides& strides() const {
-    return strides_;
-  }
-
- private:
-  friend class thrust::iterator_core_access;
-
-  __host__ __device__ bool equal(const general_iterator& other) const {
-    return this->base() == other.base() && this->index() == other.index();
-  }
-
-  __host__ __device__ void advance(difference_type n) {
-    this->index_ += n;
-  }
-
-  __host__ __device__ void increment() {
-    this->index_ += 1;
-  }
-
-  __host__ __device__ void decrement() {
-    this->index_ -= 1;
-  }
-
-  __host__ __device__ difference_type
-  distance_to(const general_iterator& other) const {
-    _CCCL_ASSERT(
-        this->base() == other.base(),
-        "Underlying iterator must point to same base iterator");
-    return other.index() - this->index();
-  }
-
-  // The dereference is device-only to avoid accidental running in host.
-  __device__ typename super_t::reference dereference() const {
-    IdxT offset = elem_to_loc(index_, shape_.data(), strides_.data(), ndim_);
-    return *(this->base() + offset);
-  }
-
-  IdxT index_;
-  int ndim_;
-  Shape shape_;
-  Strides strides_;
-};
-
-template <typename IdxT, typename Iterator>
-__host__ __device__ auto make_general_iterator(
-    Iterator it,
-    IdxT index,
-    int ndim,
-    Shape shape,
-    Strides strides) {
-  return general_iterator<Iterator, IdxT>(
-      it, index, ndim, cuda::std::move(shape), cuda::std::move(strides));
-}
-
-template <typename IdxT, typename Iterator>
-auto make_general_iterator(
-    Iterator it,
-    const std::vector<int32_t>& shape,
-    const std::vector<int64_t>& strides) {
-  return make_general_iterator<IdxT>(
-      it, 0, shape.size(), const_param(shape), const_param(strides));
-}
-
-template <typename IdxT, typename Iterator>
-auto make_general_iterators(
-    Iterator it,
-    IdxT size,
-    const std::vector<int32_t>& shape,
-    const std::vector<int64_t>& strides) {
-  auto ndim = shape.size();
-  auto shape_arg = const_param(shape);
-  auto strides_arg = const_param(strides);
-  return std::make_pair(
-      make_general_iterator<IdxT>(it, 0, ndim, shape_arg, strides_arg),
-      make_general_iterator<IdxT>(it, size, ndim, shape_arg, strides_arg));
-}
-
-} // namespace mlx::core::cu

mlx/backend/cuda/iterators/strided_iterator.cuh

@@ -1,60 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include <thrust/iterator/iterator_adaptor.h>
-#include <thrust/iterator/iterator_facade.h>
-
-namespace mlx::core::cu {
-
-// RandomAccessIterator for strided access to array entries.
-template <typename Iterator, typename Stride = int64_t>
-class strided_iterator
-    : public thrust::
-          iterator_adaptor<strided_iterator<Iterator, Stride>, Iterator> {
- public:
-  using super_t =
-      thrust::iterator_adaptor<strided_iterator<Iterator, Stride>, Iterator>;
-
-  using reference = typename super_t::reference;
-  using difference_type = typename super_t::difference_type;
-
-  __host__ __device__ strided_iterator(Iterator it, Stride stride)
-      : super_t(it), stride_(stride) {}
-
-  __host__ __device__ Stride stride() const {
-    return stride_;
-  }
-
- private:
-  friend class thrust::iterator_core_access;
-
-  __host__ __device__ bool equal(const strided_iterator& other) const {
-    return this->base() == other.base();
-  }
-
-  __host__ __device__ void advance(difference_type n) {
-    this->base_reference() += n * stride_;
-  }
-
-  __host__ __device__ void increment() {
-    this->base_reference() += stride_;
-  }
-
-  __host__ __device__ void decrement() {
-    this->base_reference() -= stride_;
-  }
-
-  __host__ __device__ difference_type
-  distance_to(const strided_iterator& other) const {
-    const difference_type dist = other.base() - this->base();
-    _CCCL_ASSERT(
-        dist % stride() == 0,
-        "Underlying iterator difference must be divisible by the stride");
-    return dist / stride();
-  }
-
-  Stride stride_;
-};
-
-} // namespace mlx::core::cu

mlx/backend/cuda/jit_module.cpp

@@ -2,6 +2,7 @@
 
 #include "mlx/backend/cuda/jit_module.h"
 #include "mlx/backend/cuda/device.h"
+#include "mlx/version.h"
 
 #include "cuda_jit_sources.h"
 
@@ -12,6 +13,7 @@
 
 #include <fmt/format.h>
 #include <nvrtc.h>
+#include <unistd.h>
 
 namespace mlx::core::cu {
 
@@ -49,14 +51,41 @@ const std::string& cuda_home() {
   return home;
 }
 
+// Return the location of CCCL headers shipped with the distribution.
+const std::string& cccl_dir() {
+  static std::string dir = []() {
+    std::filesystem::path path;
+#if defined(MLX_CCCL_DIR)
+    // First search the install dir if defined.
+    path = MLX_CCCL_DIR;
+    if (std::filesystem::exists(path)) {
+      return path.string();
+    }
+#endif
+    // Then search dynamically from the dir of libmlx.so file.
+    path = current_binary_dir().parent_path() / "include" / "cccl";
+    if (std::filesystem::exists(path)) {
+      return path.string();
+    }
+    // Finally check the environment variable.
+    path = std::getenv("MLX_CCCL_DIR");
+    if (!path.empty() && std::filesystem::exists(path)) {
+      return path.string();
+    }
+    return std::string();
+  }();
+  return dir;
+}
+
 // Get the cache directory for storing compiled results.
 const std::filesystem::path& ptx_cache_dir() {
   static std::filesystem::path cache = []() -> std::filesystem::path {
     std::filesystem::path cache;
-    if (auto c = std::getenv("MLX_PTX_CACHE"); c) {
+    if (auto c = std::getenv("MLX_PTX_CACHE_DIR"); c) {
       cache = c;
     } else {
-      cache = std::filesystem::temp_directory_path() / "mlx" / "ptx";
+      cache =
+          std::filesystem::temp_directory_path() / "mlx" / version() / "ptx";
     }
     if (!std::filesystem::exists(cache)) {
       std::error_code error;
@@ -108,7 +137,8 @@ void write_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
     const std::vector<char>& ptx,
-    const std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
+    const std::string& source_code) {
   if (cache_dir.empty()) {
     return;
   }
@@ -121,6 +151,9 @@ void write_cached_ptx(
   for (const auto& [name, mangled] : ptx_kernels) {
     txt_file << name << "\t" << mangled << std::endl;
   }
+
+  std::ofstream source_file(cache_dir / (module_name + ".cu"));
+  source_file << source_code;
 }
 
 // Return if |device|'s version is not newer than |major|.|minor| version.
@@ -160,7 +193,7 @@ constexpr const char* g_include_names[] = {
     INCLUDE_PREFIX "binary_ops.cuh",
     INCLUDE_PREFIX "cast_op.cuh",
     INCLUDE_PREFIX "config.h",
-    INCLUDE_PREFIX "cucomplex_math.cuh",
+    INCLUDE_PREFIX "complex.cuh",
     INCLUDE_PREFIX "fp16_math.cuh",
     INCLUDE_PREFIX "indexing.cuh",
     INCLUDE_PREFIX "scatter_ops.cuh",
@@ -176,7 +209,7 @@ constexpr const char* g_headers[] = {
     jit_source_binary_ops,
     jit_source_cast_op,
     jit_source_config,
-    jit_source_cucomplex_math,
+    jit_source_complex,
     jit_source_fp16_math,
     jit_source_indexing,
     jit_source_scatter_ops,
@@ -213,16 +246,24 @@ JitModule::JitModule(
     }
 
     // Compile program.
+    std::vector<const char*> args;
     bool use_sass = compiler_supports_device_sass(device);
     std::string compute = fmt::format(
         "--gpu-architecture={}_{}{}",
         use_sass ? "sm" : "compute",
         device.compute_capability_major(),
         device.compute_capability_minor());
-    std::string include = fmt::format("--include-path={}/include", cuda_home());
-    const char* args[] = {compute.c_str(), include.c_str()};
+    args.push_back(compute.c_str());
+    std::string cccl_include = cccl_dir();
+    if (!cccl_include.empty()) {
+      cccl_include = fmt::format("--include-path={}", cccl_include);
+      args.push_back(cccl_include.c_str());
+    }
+    std::string cuda_include =
+        fmt::format("--include-path={}/include", cuda_home());
+    args.push_back(cuda_include.c_str());
     nvrtcResult compile_result =
-        nvrtcCompileProgram(prog, std::size(args), args);
+        nvrtcCompileProgram(prog, args.size(), args.data());
     if (compile_result != NVRTC_SUCCESS) {
       size_t log_size;
       CHECK_NVRTC_ERROR(nvrtcGetProgramLogSize(prog, &log_size));
@@ -252,7 +293,8 @@ JitModule::JitModule(
     } else {
       CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
     }
-    write_cached_ptx(ptx_cache_dir(), module_name, ptx, ptx_kernels);
+    write_cached_ptx(
+        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
   }
 
   // Load module.

mlx/backend/cuda/kernel_utils.cuh

@@ -11,7 +11,6 @@
 #include "mlx/array.h"
 #include "mlx/backend/cuda/device/utils.cuh"
 
-#include <cuComplex.h>
 #include <cuda.h>
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
@@ -79,7 +78,7 @@ struct CTypeToCudaType<bfloat16_t> {
 
 template <>
 struct CTypeToCudaType<complex64_t> {
-  using type = cuComplex;
+  using type = cu::complex64_t;
 };
 
 template <typename T>
@@ -91,10 +90,14 @@ inline constexpr bool is_floating_v =
     cuda::std::is_same_v<T, float> || cuda::std::is_same_v<T, double> ||
     cuda::std::is_same_v<T, float16_t> || cuda::std::is_same_v<T, bfloat16_t>;
 
+// Type traits for detecting complex numbers.
+template <typename T>
+inline constexpr bool is_complex_v = cuda::std::is_same_v<T, complex64_t> ||
+    cuda::std::is_same_v<T, complex128_t>;
+
 // Type traits for detecting complex or real floating point numbers.
 template <typename T>
-inline constexpr bool is_inexact_v =
-    is_floating_v<T> || cuda::std::is_same_v<T, complex64_t>;
+inline constexpr bool is_inexact_v = is_floating_v<T> || is_complex_v<T>;
 
 // Utility to copy data from vector to array in host.
 template <int NDIM = MAX_NDIM, typename T = int32_t>

mlx/backend/cuda/layer_norm.cu

@@ -1,7 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 #include "mlx/backend/gpu/copy.h"
@@ -105,8 +104,8 @@ __global__ void layer_norm(
     T wn[N_READS];
     T bn[N_READS];
     cub::LoadDirectBlocked(index, x, xn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(b, b_stride), bn, axis_size);
+    cub::LoadDirectBlocked(index, StridedIterator(w, w_stride), wn, axis_size);
+    cub::LoadDirectBlocked(index, StridedIterator(b, b_stride), bn, axis_size);
     for (int i = 0; i < N_READS; ++i) {
       float norm = (static_cast<float>(xn[i]) - mean) * normalizer;
       xn[i] = wn[i] * static_cast<T>(norm) + bn[i];
@@ -162,7 +161,7 @@ __global__ void layer_norm_vjp(
     auto index = r * BLOCK_DIM + block.thread_rank();
     cub::LoadDirectBlocked(index, x, xn, axis_size, mean);
     cub::LoadDirectBlocked(index, g, gn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    cub::LoadDirectBlocked(index, StridedIterator(w, w_stride), wn, axis_size);
     for (int i = 0; i < N_READS; i++) {
       float t = static_cast<float>(xn[i]) - mean;
       float wi = wn[i];
@@ -185,7 +184,7 @@ __global__ void layer_norm_vjp(
     T gn[N_READS];
     cub::LoadDirectBlocked(index, x, xn, axis_size);
     cub::LoadDirectBlocked(index, g, gn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    cub::LoadDirectBlocked(index, StridedIterator(w, w_stride), wn, axis_size);
     for (int i = 0; i < N_READS; i++) {
       float xi = (static_cast<float>(xn[i]) - mean) * normalizer;
       float wi = wn[i];
@@ -237,8 +236,7 @@ void LayerNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -295,9 +293,7 @@ void LayerNormVJP::eval_gpu(
       return x;
     }
     copied = true;
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
-    copy_gpu(x, x_copy, CopyType::General, s);
-    return x_copy;
+    return contiguous_copy_gpu(x, s);
   };
   bool donate_x = inputs[0].is_donatable();
   bool donate_g = inputs[3].is_donatable();

mlx/backend/cuda/logsumexp.cu

@@ -108,8 +108,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
     if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       encoder.add_temporary(x_copy);
       return x_copy;
     }

mlx/backend/cuda/lru_cache.h

@@ -0,0 +1,146 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <list>
+#include <unordered_map>
+#include <utility>
+
+namespace mlx::core {
+
+template <
+    typename K,
+    typename V,
+    template <typename...> typename M = std::unordered_map>
+class LRUCache {
+ public:
+  using value_type = std::pair<K, V>;
+  using list_type = std::list<value_type>;
+  using iterator = typename list_type::iterator;
+  using const_iterator = typename list_type::const_iterator;
+  using map_type = M<K, iterator>;
+
+  explicit LRUCache(size_t capacity) : capacity_(capacity) {}
+
+  size_t size() const {
+    return map_.size();
+  }
+  size_t capacity() const {
+    return capacity_;
+  }
+  bool empty() const {
+    return vlist_.empty();
+  }
+
+  void resize(size_t new_capacity) {
+    capacity_ = new_capacity;
+    trim();
+  }
+
+  iterator begin() {
+    return vlist_.begin();
+  }
+  const_iterator begin() const {
+    return vlist_.begin();
+  }
+  iterator end() {
+    return vlist_.end();
+  }
+  const_iterator end() const {
+    return vlist_.end();
+  }
+
+  void clear() {
+    map_.clear();
+    vlist_.clear();
+  }
+
+  iterator find(const K& key) {
+    auto it = map_.find(key);
+    if (it == map_.end())
+      return end();
+    vlist_.splice(vlist_.begin(), vlist_, it->second);
+    return it->second;
+  }
+
+  template <typename U>
+  std::pair<iterator, bool> emplace(const K& key, U&& value) {
+    auto it = map_.find(key);
+    if (it != map_.end()) {
+      vlist_.splice(vlist_.begin(), vlist_, it->second);
+      return {it->second, false};
+    }
+
+    vlist_.emplace_front(key, std::forward<U>(value));
+    map_[key] = vlist_.begin();
+
+    trim();
+
+    return {vlist_.begin(), true};
+  }
+
+  iterator erase(iterator pos) {
+    map_.erase(pos->first);
+    return vlist_.erase(pos);
+  }
+
+ private:
+  void trim() {
+    while (map_.size() > capacity_) {
+      auto last = std::prev(vlist_.end());
+      map_.erase(last->first);
+      vlist_.pop_back();
+    }
+  }
+
+  list_type vlist_;
+  map_type map_;
+  size_t capacity_;
+};
+
+// Turn a POD struct into a container key by doing bytes compare.
+template <typename T>
+struct BytesKey {
+  T pod;
+  static_assert(std::is_standard_layout_v<T>, "T is not POD");
+
+  BytesKey(T pod) : pod(std::move(pod)) {}
+
+  BytesKey(const BytesKey& other) {
+    memcpy(&pod, &other.pod, sizeof(T));
+  }
+
+  BytesKey(BytesKey&& other) {
+    memcpy(&pod, &other.pod, sizeof(T));
+  }
+
+  bool operator==(const BytesKey& other) const {
+    auto* ptr1 = reinterpret_cast<const uint8_t*>(&pod);
+    auto* ptr2 = reinterpret_cast<const uint8_t*>(&other.pod);
+    return memcmp(ptr1, ptr2, sizeof(T)) == 0;
+  }
+};
+
+// Compute hash according to the bytes value of T.
+template <typename T>
+struct BytesHash {
+  static_assert(std::is_standard_layout_v<T>, "T is not POD");
+
+  size_t operator()(const T& pod) const {
+    auto* ptr = reinterpret_cast<const uint8_t*>(&pod);
+    uint32_t value = 0x811C9DC5;
+    for (int i = 0; i < sizeof(T); ++i) {
+      value ^= ptr[i];
+      value *= 0x01000193;
+    }
+    return value;
+  }
+};
+
+template <typename K, typename V>
+using BytesKeyHashMap = std::unordered_map<K, V, BytesHash<K>>;
+
+template <typename K, typename V>
+using LRUBytesKeyCache = LRUCache<BytesKey<K>, V, BytesKeyHashMap>;
+
+} // namespace mlx::core

mlx/backend/cuda/matmul.cpp

@@ -2,275 +2,15 @@
 
 #include "mlx/backend/common/matmul.h"
 #include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/gemms/gemv.h"
 #include "mlx/backend/gpu/copy.h"
-#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
-#include "mlx/utils.h"
 
-#include <cublasLt.h>
-#include <fmt/format.h>
 #include <nvtx3/nvtx3.hpp>
-
 #include <numeric>
 
 namespace mlx::core {
-
-namespace cu {
-
-#define CHECK_CUBLAS_ERROR(cmd) check_cublas_error(#cmd, (cmd))
-
-void check_cublas_error(const char* name, cublasStatus_t err) {
-  if (err != CUBLAS_STATUS_SUCCESS) {
-    // TODO: Use cublasGetStatusString when it is widely available.
-    throw std::runtime_error(
-        fmt::format("{} failed with code: {}.", name, static_cast<int>(err)));
-  }
-}
-
-class MatMul {
- public:
-  MatMul(
-      Device& device,
-      Dtype dtype,
-      bool a_transposed,
-      uint64_t a_rows,
-      uint64_t a_cols,
-      int64_t lda,
-      bool b_transposed,
-      uint64_t b_rows,
-      uint64_t b_cols,
-      int64_t ldb,
-      int32_t batch_count,
-      int64_t a_batch_stride,
-      int64_t b_batch_stride)
-      : handle_(device.lt_handle()) {
-    heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
-
-    auto scale_type = dtype_to_cuda_type(dtype);
-    if (dtype == bfloat16 || dtype == float16) {
-      scale_type = CUDA_R_32F;
-    }
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescCreate(
-        &matmul_desc_, dtype_to_compute_type(dtype), scale_type));
-    int32_t pointer_mode = CUBLASLT_POINTER_MODE_HOST;
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
-        matmul_desc_,
-        CUBLASLT_MATMUL_DESC_POINTER_MODE,
-        &pointer_mode,
-        sizeof(int32_t)));
-    cublasOperation_t op = CUBLAS_OP_N;
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
-        matmul_desc_,
-        CUBLASLT_MATMUL_DESC_TRANSA,
-        &op,
-        sizeof(cublasOperation_t)));
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
-        matmul_desc_,
-        CUBLASLT_MATMUL_DESC_TRANSB,
-        &op,
-        sizeof(cublasOperation_t)));
-
-    auto type = dtype_to_cuda_type(dtype);
-    a_desc_ = create_matrix_layout(
-        type, a_rows, a_cols, a_transposed, lda, batch_count, a_batch_stride);
-    b_desc_ = create_matrix_layout(
-        type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
-    out_desc_ = create_matrix_layout(
-        type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
-
-    // The recommended cublas workspace size is 4 MiB for pre-Hopper and 32 MiB
-    // for Hopper+:
-    // https://docs.nvidia.com/cuda/cublas/#cublassetworkspace
-    uint64_t MiB = 1024 * 1024;
-    uint64_t workspace_size =
-        device.compute_capability_major() >= 9 ? 32 * MiB : 4 * MiB;
-
-    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceCreate(&pref_));
-    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceSetAttribute(
-        pref_,
-        CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
-        &workspace_size,
-        sizeof(uint64_t)));
-  }
-
-  MatMul(
-      Device& device,
-      Dtype dtype,
-      bool a_transposed,
-      uint64_t a_rows,
-      uint64_t a_cols,
-      int64_t lda,
-      bool b_transposed,
-      uint64_t b_rows,
-      uint64_t b_cols,
-      int64_t ldb,
-      bool c_transposed,
-      int64_t ldc,
-      int32_t batch_count,
-      int64_t a_batch_stride,
-      int64_t b_batch_stride,
-      int64_t c_batch_stride)
-      : MatMul(
-            device,
-            dtype,
-            a_transposed,
-            a_rows,
-            a_cols,
-            lda,
-            b_transposed,
-            b_rows,
-            b_cols,
-            ldb,
-            batch_count,
-            a_batch_stride,
-            b_batch_stride) {
-    auto type = dtype_to_cuda_type(dtype);
-    c_desc_ = create_matrix_layout(
-        type, a_rows, b_cols, c_transposed, ldc, batch_count, c_batch_stride);
-  }
-
-  ~MatMul() {
-    cublasLtMatrixLayoutDestroy(a_desc_);
-    cublasLtMatrixLayoutDestroy(b_desc_);
-    cublasLtMatrixLayoutDestroy(c_desc_);
-    cublasLtMatrixLayoutDestroy(out_desc_);
-    cublasLtMatmulDescDestroy(matmul_desc_);
-  }
-
-  void run(
-      cu::CommandEncoder& encoder,
-      void* out,
-      void* a,
-      void* b,
-      void* c = nullptr,
-      float alpha = 1,
-      float beta = 0) {
-    if (heuristic_.state != CUBLAS_STATUS_SUCCESS) {
-      int ret = 0;
-      CHECK_CUBLAS_ERROR(cublasLtMatmulAlgoGetHeuristic(
-          handle_,
-          matmul_desc_,
-          a_desc_,
-          b_desc_,
-          out_desc_,
-          out_desc_,
-          pref_,
-          1,
-          &heuristic_,
-          &ret));
-      if (ret == 0) {
-        throw std::runtime_error("Can not find algorithm for matmul.");
-      }
-    }
-
-    void* workspace_ptr = nullptr;
-    if (heuristic_.workspaceSize > 0) {
-      array workspace(
-          allocator::malloc(heuristic_.workspaceSize),
-          {static_cast<int>(heuristic_.workspaceSize)},
-          int8);
-      encoder.add_temporary(workspace);
-      workspace_ptr = workspace.data<void>();
-    }
-
-    auto capture = encoder.capture_context();
-    CHECK_CUBLAS_ERROR(cublasLtMatmul(
-        handle_,
-        matmul_desc_,
-        &alpha,
-        a,
-        a_desc_,
-        b,
-        b_desc_,
-        &beta,
-        c ? c : out,
-        c ? c_desc_ : out_desc_,
-        out,
-        out_desc_,
-        &heuristic_.algo,
-        workspace_ptr,
-        heuristic_.workspaceSize,
-        encoder.stream()));
-  }
-
- private:
-  cublasComputeType_t dtype_to_compute_type(Dtype dtype) {
-    switch (dtype) {
-      case float16:
-        return CUBLAS_COMPUTE_32F;
-      case bfloat16:
-        return CUBLAS_COMPUTE_32F;
-      case float32:
-        return mlx::core::env::enable_tf32() ? CUBLAS_COMPUTE_32F_FAST_TF32
-                                             : CUBLAS_COMPUTE_32F;
-      case float64:
-      case complex64:
-        return CUBLAS_COMPUTE_64F;
-      default:
-        throw std::runtime_error(fmt::format(
-            "Unsupported dtype in MatMul: {}.", dtype_to_string(dtype)));
-    }
-  }
-
-  cudaDataType_t dtype_to_cuda_type(Dtype dtype) {
-    switch (dtype) {
-      case float16:
-        return CUDA_R_16F;
-      case bfloat16:
-        return CUDA_R_16BF;
-      case float32:
-        return CUDA_R_32F;
-      case float64:
-        return CUDA_R_64F;
-      case complex64:
-        return CUDA_C_32F;
-      default:
-        throw std::runtime_error(fmt::format(
-            "Unsupported dtype in MatMul: {}.", dtype_to_string(dtype)));
-    }
-  }
-
-  cublasLtMatrixLayout_t create_matrix_layout(
-      cudaDataType_t type,
-      uint64_t rows,
-      uint64_t cols,
-      bool transposed,
-      int64_t ld,
-      int32_t batch_count,
-      int64_t batch_stride) {
-    cublasLtMatrixLayout_t desc;
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutCreate(&desc, type, rows, cols, ld));
-    cublasLtOrder_t order =
-        transposed ? CUBLASLT_ORDER_COL : CUBLASLT_ORDER_ROW;
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-        desc, CUBLASLT_MATRIX_LAYOUT_ORDER, &order, sizeof(cublasLtOrder_t)));
-    if (batch_count > 1) {
-      CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-          desc,
-          CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT,
-          &batch_count,
-          sizeof(int32_t)));
-      CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-          desc,
-          CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET,
-          &batch_stride,
-          sizeof(int64_t)));
-    }
-    return desc;
-  }
-
-  cublasLtHandle_t handle_{nullptr};
-  cublasLtMatmulDesc_t matmul_desc_{nullptr};
-  cublasLtMatmulPreference_t pref_{nullptr};
-  cublasLtMatrixLayout_t a_desc_{nullptr};
-  cublasLtMatrixLayout_t b_desc_{nullptr};
-  cublasLtMatrixLayout_t c_desc_{nullptr};
-  cublasLtMatrixLayout_t out_desc_{nullptr};
-  cublasLtMatmulHeuristicResult_t heuristic_;
-};
-
-} // namespace cu
-
 namespace {
 
 std::tuple<bool, int64_t, array>
@@ -282,8 +22,7 @@ check_transpose(cu::CommandEncoder& enc, const Stream& s, const array& arr) {
   } else if (stx == 1 && sty == arr.shape(-2)) {
     return std::make_tuple(true, sty, arr);
   } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy_gpu(arr, arr_copy, CopyType::General, s);
+    array arr_copy = contiguous_copy_gpu(arr, s);
     enc.add_temporary(arr_copy);
     return std::make_tuple(false, arr.shape(-1), arr_copy);
   }
@@ -340,10 +79,25 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     batch_shape = {1};
   }
 
+  if (cu::can_use_gemv(M, N, K, a_transposed, b_transposed)) {
+    cu::gemv(
+        a,
+        b,
+        out,
+        M,
+        N,
+        K,
+        batch_count,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        encoder);
+    return;
+  }
+
   /////////////////////////////////////////////////////////////////////////////
   // Invoke cublasLt
-
-  cu::MatMul matmul(
+  cu::Matmul matmul(
       cu::device(s.device),
       a.dtype(),
       a_transposed,
@@ -358,27 +112,13 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
       a_batch_strides.back(),
       b_batch_strides.back());
 
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_output_array(out);
-  auto nbatch = batch_count / batch_shape.back();
-  if (nbatch == 1) {
-    matmul.run(encoder, out.data<int8_t>(), a.data<int8_t>(), b.data<int8_t>());
+  if ((batch_count / batch_shape.back()) == 1) {
+    matmul.run(encoder, out, a, b);
     return;
   }
 
-  ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
-  ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
-  auto concurrent = encoder.concurrent_context();
-  for (size_t i = 0; i < nbatch; ++i) {
-    matmul.run(
-        encoder,
-        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,
-        a.data<int8_t>() + a.itemsize() * a_it.loc,
-        b.data<int8_t>() + b.itemsize() * b_it.loc);
-    a_it.step();
-    b_it.step();
-  }
+  matmul.run_batched(
+      encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides);
 }
 
 void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -389,9 +129,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 3);
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
-  auto& c_pre = inputs[2];
-
-  out.set_data(allocator::malloc(out.nbytes()));
+  auto c = inputs[2];
 
   /////////////////////////////////////////////////////////////////////////////
   // Init checks and prep
@@ -404,7 +142,24 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   // the arrays
   auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
   auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
-  auto [c_transposed, ldc, c] = check_transpose(encoder, s, c_pre);
+
+  int64_t ldc;
+  {
+    auto stx = c.strides()[c.ndim() - 2];
+    auto sty = c.strides()[c.ndim() - 1];
+    if (sty == 1 && stx == c.shape(-1)) {
+      ldc = stx;
+      out.set_data(allocator::malloc(out.nbytes()));
+    } else if (sty == 1 && stx == 0) {
+      ldc = 0;
+      out.set_data(allocator::malloc(out.nbytes()));
+    } else {
+      // Copy C into out and set C to out
+      ldc = c.shape(-1);
+      copy_gpu(c, out, CopyType::General, s);
+      c = out;
+    }
+  }
 
   /////////////////////////////////////////////////////////////////////////////
   // Check and collapse batch dimensions
@@ -431,7 +186,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   /////////////////////////////////////////////////////////////////////////////
   // Invoke cublasLt
 
-  cu::MatMul matmul(
+  cu::Matmul matmul(
       cu::device(s.device),
       a.dtype(),
       a_transposed,
@@ -442,48 +197,28 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       K,
       N,
       ldb,
-      c_transposed,
       ldc,
       batch_shape.back(),
       a_batch_strides.back(),
       b_batch_strides.back(),
       c_batch_strides.back());
 
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_input_array(c);
-  encoder.set_output_array(out);
-
-  auto nbatch = batch_count / batch_shape.back();
-  if (nbatch == 1) {
-    matmul.run(
-        encoder,
-        out.data<int8_t>(),
-        a.data<int8_t>(),
-        b.data<int8_t>(),
-        c.data<int8_t>(),
-        alpha_,
-        beta_);
+  if ((batch_count / batch_shape.back()) == 1) {
+    matmul.run(encoder, out, a, b, c, alpha_, beta_);
     return;
   }
-
-  ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
-  ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
-  ContiguousIterator c_it(batch_shape, c_batch_strides, batch_shape.size() - 1);
-  auto concurrent = encoder.concurrent_context();
-  for (size_t i = 0; i < nbatch; ++i) {
-    matmul.run(
-        encoder,
-        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,
-        a.data<int8_t>() + a.itemsize() * a_it.loc,
-        b.data<int8_t>() + b.itemsize() * b_it.loc,
-        c.data<int8_t>() + c.itemsize() * c_it.loc,
-        alpha_,
-        beta_);
-    a_it.step();
-    b_it.step();
-    c_it.step();
-  }
+  matmul.run_batched(
+      encoder,
+      out,
+      a,
+      b,
+      c,
+      batch_shape,
+      a_batch_strides,
+      b_batch_strides,
+      c_batch_strides,
+      alpha_,
+      beta_);
 }
 
 } // namespace mlx::core

mlx/backend/cuda/primitives.cu

@@ -71,7 +71,6 @@ bool fast::ScaledDotProductAttention::use_fallback(
   }
 
 NO_GPU(BlockMaskedMM)
-NO_GPU(Convolution)
 NO_GPU(DynamicSlice)
 NO_GPU(DynamicSliceUpdate)
 NO_GPU(FFT)
@@ -82,7 +81,7 @@ NO_GPU(Load)
 NO_GPU_MULTI(LUF)
 NO_GPU_MULTI(QRF)
 NO_GPU(QuantizedMatmul)
-NO_GPU(Scan)
+NO_GPU(SegmentedMM)
 NO_GPU_MULTI(SVD)
 NO_GPU(Inverse)
 NO_GPU(Cholesky)
@@ -91,7 +90,6 @@ NO_GPU_MULTI(Eigh)
 
 namespace fast {
 NO_GPU(ScaledDotProductAttention)
-NO_GPU_MULTI(AffineQuantize)
 NO_GPU_MULTI(CustomKernel)
 } // namespace fast
 

mlx/backend/cuda/quantized.cu

@@ -0,0 +1,386 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/fast_primitives.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <int bits, int wsize = 8>
+inline constexpr __device__ short get_pack_factor() {
+  return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
+}
+
+template <int bits, int wsize = 8>
+inline constexpr __device__ short get_bytes_per_pack() {
+  constexpr int power_of_2_bits = (bits & (bits - 1)) == 0;
+  return power_of_2_bits ? (wsize / 8) : (bits == 5 ? 5 : 3);
+}
+
+template <typename T, int group_size, int bits>
+__global__ void
+affine_quantize(const T* w, uint8_t* out, T* scales, T* biases, size_t size) {
+  auto block_size = cg::this_thread_block().dim_threads();
+  auto block_idx = cg::this_thread_block().group_index();
+  auto idx_in_block = cg::this_thread_block().thread_index();
+
+  auto tidx = block_idx.x * block_size.x + idx_in_block.x;
+  auto tidy = block_idx.y * block_size.y + idx_in_block.y;
+
+  auto grid_dim_x =
+      cg::this_grid().dim_blocks().x * cg::this_grid().block_index().x;
+  constexpr float eps = 1e-7;
+  constexpr int simd_size = WARP_SIZE;
+  constexpr float n_bins = (1 << bits) - 1;
+  constexpr int pack_factor = get_pack_factor<bits, 8>();
+  constexpr int bytes_per_pack = get_bytes_per_pack<bits>();
+  constexpr int values_per_reduce = group_size / simd_size;
+  constexpr int writes_per_reduce = pack_factor / values_per_reduce;
+  constexpr int writes_per_pack =
+      writes_per_reduce > 1 ? 1 : values_per_reduce / pack_factor;
+  constexpr int power_of_2_bits = (bits & (bits - 1)) == 0;
+
+  size_t offset = tidx + grid_dim_x * size_t(tidy);
+  size_t in_index = offset * values_per_reduce;
+  if (in_index >= size) {
+    return;
+  }
+  size_t out_index = power_of_2_bits
+      ? offset * writes_per_pack
+      : offset * bytes_per_pack / writes_per_reduce;
+
+  float w_thread[values_per_reduce];
+  float w_min = Limits<float>::max();
+  float w_max = 0;
+
+#pragma clang loop unroll(full)
+  for (int i = 0; i < values_per_reduce; i++) {
+    float val = w[in_index + i];
+    w_thread[i] = val;
+    w_min = min(w_min, val);
+    w_max = max(w_max, val);
+  }
+
+  cg::greater<float> max_op;
+  cg::less<float> min_op;
+  auto warp = cg::tiled_partition<WARP_SIZE>(cg::this_thread_block());
+
+  w_min = cg::reduce(warp, w_min, min_op);
+  w_max = cg::reduce(warp, w_max, max_op);
+
+  float scale = max((w_max - w_min) / n_bins, eps);
+  bool side = abs(w_min) > abs(w_max);
+  scale = side ? scale : -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;
+  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] = static_cast<T>(scale);
+    biases[gindex] = static_cast<T>(bias);
+  }
+
+  using OutType = std::conditional_t<bits == 5, uint64_t, uint32_t>;
+  OutType output = 0;
+
+#pragma clang loop unroll(full)
+  for (int i = 0; i < values_per_reduce; i++) {
+    uint8_t val = min(round((w_thread[i] - bias) / scale), n_bins);
+    if (bits == 8) {
+      output = val;
+    } else {
+      output |= val << (bits * (i % pack_factor));
+    }
+
+    if (pack_factor < values_per_reduce && i % pack_factor == pack_factor - 1) {
+      out[out_index + i / pack_factor] = output;
+      output = 0;
+    } else {
+#pragma clang loop unroll(full)
+      for (int j = 1; j < writes_per_reduce; j++) {
+        uint8_t sval = warp.shfl_down(val, j);
+        output |= static_cast<OutType>(sval)
+            << (bits * (j * values_per_reduce + i));
+      }
+    }
+  }
+  if constexpr (bits == 3 || bits == 6) {
+    if (in_index % pack_factor == 0 && out_index % bytes_per_pack == 0) {
+      out[out_index] = output & 0xff;
+      out[out_index + 1] = (output & 0xff00) >> 8;
+      out[out_index + 2] = (output & 0xff0000) >> 16;
+    }
+  } else if constexpr (bits == 5) {
+    if (in_index % pack_factor == 0 && out_index % bytes_per_pack == 0) {
+      out[out_index] = output & 0xff;
+      out[out_index + 1] = (output & 0xff00) >> 8;
+      out[out_index + 2] = (output & 0xff0000) >> 16;
+      out[out_index + 3] = (output & 0xff000000) >> 24;
+      out[out_index + 4] = (output & 0xff00000000) >> 32;
+    }
+  } else {
+    if constexpr (writes_per_reduce > 0) {
+      if (out_index % writes_per_reduce == 0) {
+        out[out_index / writes_per_reduce] = output;
+      }
+    }
+  }
+}
+
+template <typename T, int group_size, int bits>
+__global__ void affine_dequantize(
+    const uint8_t* w,
+    const T* scales,
+    const T* biases,
+    T* out,
+    size_t size) {
+  auto block_size = cg::this_thread_block().dim_threads();
+  auto block_idx = cg::this_thread_block().group_index();
+  auto idx_in_block = cg::this_thread_block().thread_index();
+
+  auto tidx = block_idx.x * block_size.x + idx_in_block.x;
+  auto tidy = block_idx.y * block_size.y + idx_in_block.y;
+
+  auto grid_dim_x =
+      cg::this_grid().dim_blocks().x * cg::this_grid().block_index().x;
+
+  constexpr int pack_factor = get_pack_factor<bits, 8>();
+  constexpr int bytes_per_pack = get_bytes_per_pack<bits>();
+
+  size_t offset = tidx + grid_dim_x * size_t(tidy);
+  size_t oindex = offset * pack_factor;
+
+  if (oindex >= size) {
+    return;
+  }
+
+  size_t gindex = oindex / group_size;
+  T scale = scales[gindex];
+  T bias = biases[gindex];
+  out += oindex;
+
+  if constexpr (bits == 3) {
+    w += offset * bytes_per_pack;
+    out[0] = static_cast<T>(w[0] & 0x7) * scale + bias;
+    out[1] = static_cast<T>((w[0] & 0x38) >> 3) * scale + bias;
+    out[2] = (static_cast<T>((w[0] & 0xc0) >> 6) +
+              static_cast<T>((w[1] & 0x1) << 2)) *
+            scale +
+        bias;
+    out[3] = static_cast<T>((w[1] & 0xe) >> 1) * scale + bias;
+    out[4] = static_cast<T>((w[1] & 0x70) >> 4) * scale + bias;
+    out[5] = (static_cast<T>((w[1] & 0x80) >> 7) +
+              static_cast<T>((w[2] & 0x3) << 1)) *
+            scale +
+        bias;
+    out[6] = static_cast<T>((w[2] & 0x1c) >> 2) * scale + bias;
+    out[7] = static_cast<T>((w[2] & 0xe0) >> 5) * scale + bias;
+  } else if constexpr (bits == 5) {
+    w += offset * bytes_per_pack;
+    out[0] = static_cast<T>(w[0] & 0x1f) * scale + bias;
+    out[1] = (static_cast<T>((w[0] & 0xe0) >> 5) +
+              static_cast<T>((w[1] & 0x3) << 3)) *
+            scale +
+        bias;
+    out[2] = static_cast<T>((w[1] & 0x7c) >> 2) * scale + bias;
+    out[3] = (static_cast<T>((w[1] & 0x80) >> 7) +
+              static_cast<T>((w[2] & 0xf) << 1)) *
+            scale +
+        bias;
+    out[4] = (static_cast<T>((w[2] & 0xf0) >> 4) +
+              static_cast<T>((w[3] & 0x1) << 4)) *
+            scale +
+        bias;
+    out[5] = static_cast<T>((w[3] & 0x3e) >> 1) * scale + bias;
+    out[6] = (static_cast<T>((w[3] & 0xc0) >> 6) +
+              static_cast<T>((w[4] & 0x7) << 2)) *
+            scale +
+        bias;
+    out[7] = static_cast<T>((w[4] & 0xf8) >> 3) * scale + bias;
+  } else if constexpr (bits == 6) {
+    w += offset * bytes_per_pack;
+    out[0] = static_cast<T>(w[0] & 0x3f) * scale + bias;
+    out[1] = (static_cast<T>((w[0] >> 6) & 0x03) +
+              static_cast<T>((w[1] & 0x0f) << 2)) *
+            scale +
+        bias;
+    out[2] = (static_cast<T>((w[1] >> 4) & 0x0f) +
+              static_cast<T>((w[2] & 0x03) << 4)) *
+            scale +
+        bias;
+    out[3] = static_cast<T>((w[2] >> 2) & 0x3f) * scale + bias;
+  } else {
+    uint val = w[offset];
+#pragma clang loop unroll(full)
+    for (int i = 0; i < pack_factor; i++) {
+      uint8_t d;
+      if (bits == 2) {
+        d = (val >> (bits * i)) & 0x03;
+      } else if (bits == 4) {
+        d = (val >> (bits * i)) & 0x0f;
+      } else if (bits == 8) {
+        d = val;
+      }
+      out[i] = scale * static_cast<T>(d) + bias;
+    }
+  }
+}
+
+} // namespace cu
+namespace {
+
+inline array ensure_row_contiguous(
+    const array& x,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  if (!x.flags().row_contiguous) {
+    array x_copy = contiguous_copy_gpu(x, s);
+    enc.add_temporary(x_copy);
+    return x_copy;
+  } else {
+    return x;
+  }
+}
+
+} // namespace
+
+template <typename F>
+void dispatch_groups(int group_size, F&& f) {
+  switch (group_size) {
+    case 32:
+      f(std::integral_constant<int, 32>{});
+      break;
+    case 64:
+      f(std::integral_constant<int, 64>{});
+      break;
+    case 128:
+      f(std::integral_constant<int, 128>{});
+      break;
+  }
+}
+
+template <typename F>
+void dispatch_bits(int bits, F&& f) {
+  switch (bits) {
+    case 2:
+      f(std::integral_constant<int, 2>{});
+      break;
+    case 3:
+      f(std::integral_constant<int, 3>{});
+      break;
+    case 4:
+      f(std::integral_constant<int, 4>{});
+      break;
+    case 5:
+      f(std::integral_constant<int, 5>{});
+      break;
+    case 6:
+      f(std::integral_constant<int, 6>{});
+      break;
+    case 8:
+      f(std::integral_constant<int, 8>{});
+      break;
+  }
+}
+
+void fast::AffineQuantize::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& w_pre = inputs[0];
+  auto& out = outputs[0];
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& d = cu::device(s.device);
+  auto& enc = d.get_command_encoder(s);
+
+  auto w = ensure_row_contiguous(w_pre, enc, s);
+  enc.set_input_array(w);
+  if (dequantize_) {
+    auto scales = ensure_row_contiguous(inputs[1], enc, s);
+    auto biases = ensure_row_contiguous(inputs[2], enc, s);
+    enc.set_input_array(scales);
+    enc.set_input_array(biases);
+    enc.set_output_array(out);
+  } else {
+    auto& scales = outputs[1];
+    auto& biases = outputs[2];
+    scales.set_data(allocator::malloc(scales.nbytes()));
+    biases.set_data(allocator::malloc(biases.nbytes()));
+    enc.set_output_array(out);
+    enc.set_output_array(scales);
+    enc.set_output_array(biases);
+  }
+
+  auto dtype = dequantize_ ? outputs[0].dtype() : inputs[0].dtype();
+
+  // Treat uint32 as uint8 in kernel
+  int uint8_per_uint32 = 4;
+  int packs_per_int = (bits_ == 3 || bits_ == 5) ? 8
+      : bits_ == 6                               ? 4
+                                                 : 8 / bits_;
+  int per_thread = dequantize_ ? packs_per_int : group_size_ / WARP_SIZE;
+  size_t size =
+      dequantize_ ? out.size() / packs_per_int : w.size() / per_thread;
+
+  bool large = size > UINT_MAX;
+  auto grid_shape = w.shape();
+
+  if (dequantize_) {
+    grid_shape.back() *= uint8_per_uint32;
+  } else {
+    grid_shape.back() /= per_thread;
+  }
+
+  dispatch_float_types(dtype, "affine_quantize", [&](auto type_tag) {
+    dispatch_groups(group_size_, [&](auto group_size) {
+      dispatch_bits(bits_, [&](auto bits) {
+        using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        if (dequantize_) {
+          auto kernel =
+              cu::affine_dequantize<DataType, group_size.value, bits.value>;
+          auto [num_blocks, block_dims] =
+              get_launch_args(kernel, size, grid_shape, w.strides(), large);
+          enc.add_kernel_node(
+              kernel,
+              num_blocks,
+              block_dims,
+              w.data<uint8_t>(),
+              inputs[1].data<DataType>(),
+              inputs[2].data<DataType>(),
+              out.data<DataType>(),
+              out.size());
+        } else {
+          auto kernel =
+              cu::affine_quantize<DataType, group_size.value, bits.value>;
+          auto [num_blocks, block_dims] =
+              get_launch_args(kernel, size, grid_shape, w.strides(), large);
+          enc.add_kernel_node(
+              kernel,
+              num_blocks,
+              block_dims,
+              w.data<DataType>(),
+              out.data<uint8_t>(),
+              outputs[1].data<DataType>(),
+              outputs[2].data<DataType>(),
+              w.size());
+        }
+      });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/reduce.cu

@@ -47,8 +47,7 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
     }
   }
   if (plan.type == GeneralReduce || broadcasted || !in.flags().contiguous) {
-    array in_copy(in.shape(), in.dtype(), nullptr, {});
-    copy_gpu(in, in_copy, CopyType::General, s);
+    array in_copy = contiguous_copy_gpu(in, s);
     encoder.add_temporary(in_copy);
     in = in_copy;
     plan = get_reduction_plan(in, axes_);

mlx/backend/cuda/reduce/all_reduce.cu

@@ -37,15 +37,15 @@ __global__ void all_reduce(T* in, U* out, size_t block_step, size_t size) {
   for (; i + block.size() * N <= check; i += block.size() * N) {
     cub::LoadDirectBlockedVectorized<T, N>(block.thread_rank(), in + i, vals);
     for (int j = 0; j < N; j++) {
-      accs[0] = op(accs[0], __cast<U, T>(vals[j]));
+      accs[0] = op(accs[0], cast_to<U>(vals[j]));
     }
   }
 
   if (i < check) {
     cub::LoadDirectBlocked(
-        block.thread_rank(), in + i, vals, check - i, __cast<T, U>(init));
+        block.thread_rank(), in + i, vals, check - i, cast_to<T>(init));
     for (int i = 0; i < N; i++) {
-      accs[0] = op(accs[0], __cast<U, T>(vals[i]));
+      accs[0] = op(accs[0], cast_to<U>(vals[i]));
     }
   }
 

mlx/backend/cuda/reduce/col_reduce.cu

@@ -3,7 +3,6 @@
 #include <numeric>
 
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cast_op.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 
 #include <cooperative_groups.h>
@@ -128,7 +127,7 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
         T vals[N_READS];
         cub::LoadDirectBlockedVectorized(thread_x, in + loop.location(), vals);
         for (int i = 0; i < N_READS; i++) {
-          totals[i] = op(totals[i], __cast<U, T>(vals[i]));
+          totals[i] = op(totals[i], cast_to<U>(vals[i]));
         }
         loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
       }
@@ -137,7 +136,7 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
         T vals[N_READS];
         cub::LoadDirectBlocked(thread_x, in + loop.location(), vals);
         for (int i = 0; i < N_READS; i++) {
-          totals[i] = op(totals[i], __cast<U, T>(vals[i]));
+          totals[i] = op(totals[i], cast_to<U>(vals[i]));
         }
         loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
       }
@@ -150,9 +149,9 @@ col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
           in + loop.location(),
           vals,
           args.reduction_stride - tile_x * BN,
-          __cast<T, U>(ReduceInit<Op, T>::value()));
+          cast_to<T>(ReduceInit<Op, T>::value()));
       for (int i = 0; i < N_READS; i++) {
-        totals[i] = op(totals[i], __cast<U, T>(vals[i]));
+        totals[i] = op(totals[i], cast_to<U>(vals[i]));
       }
       loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
     }

mlx/backend/cuda/reduce/reduce.cuh

@@ -3,7 +3,6 @@
 #include <type_traits>
 
 #include "mlx/backend/common/reduce.h"
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/cuda/reduce/reduce_ops.cuh"
 #include "mlx/dtype_utils.h"

mlx/backend/cuda/reduce/reduce_ops.cuh

@@ -2,6 +2,8 @@
 
 #pragma once
 
+#include "mlx/backend/cuda/device/atomic_ops.cuh"
+#include "mlx/backend/cuda/device/cast_op.cuh"
 #include "mlx/backend/cuda/device/utils.cuh"
 #include "mlx/backend/cuda/reduce/reduce_utils.cuh"
 
@@ -40,15 +42,15 @@ struct Sum {
   }
 
   __device__ void atomic_update(__nv_bfloat16* x, __nv_bfloat16 y) {
-    atomicAdd(x, y);
+    atomic_add(x, y);
   }
 
   __device__ void atomic_update(int* x, int y) {
-    atomicAdd(x, y);
+    atomic_add(x, y);
   }
 
   __device__ void atomic_update(float* x, float y) {
-    atomicAdd(x, y);
+    atomic_add(x, y);
   }
 };
 
@@ -67,6 +69,18 @@ struct Prod {
 struct Min {
   template <typename T>
   __device__ __forceinline__ T operator()(T a, T b) {
+    if constexpr (is_complex_v<T>) {
+      if (isnan(a.real()) || isnan(a.imag())) {
+        return a;
+      }
+      if (isnan(b.real()) || isnan(b.imag())) {
+        return b;
+      }
+    } else if constexpr (!cuda::std::is_integral_v<T>) {
+      if (isnan(a) || isnan(b)) {
+        return cuda::std::numeric_limits<float>::quiet_NaN();
+      }
+    }
     return a < b ? a : b;
   }
 
@@ -79,6 +93,18 @@ struct Min {
 struct Max {
   template <typename T>
   __device__ __forceinline__ T operator()(T a, T b) {
+    if constexpr (is_complex_v<T>) {
+      if (isnan(a.real()) || isnan(a.imag())) {
+        return a;
+      }
+      if (isnan(b.real()) || isnan(b.imag())) {
+        return b;
+      }
+    } else if constexpr (!cuda::std::is_integral_v<T>) {
+      if (isnan(a) || isnan(b)) {
+        return cuda::std::numeric_limits<float>::quiet_NaN();
+      }
+    }
     return a > b ? a : b;
   }
 
@@ -149,10 +175,10 @@ struct ReduceInit<Or, T> {
 template <typename T>
 struct ReduceInit<Sum, T> {
   static constexpr __host__ __device__ auto value() {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+    if constexpr (is_complex_v<T>) {
       return T{0, 0};
     } else {
-      return typename ReduceResult<Sum, T>::type{0};
+      return cast_to<typename ReduceResult<Sum, T>::type>(0);
     }
   }
 };
@@ -160,10 +186,10 @@ struct ReduceInit<Sum, T> {
 template <typename T>
 struct ReduceInit<Prod, T> {
   static constexpr __host__ __device__ auto value() {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+    if constexpr (is_complex_v<T>) {
       return T{1, 0};
     } else {
-      return typename ReduceResult<Prod, T>::type{1};
+      return cast_to<typename ReduceResult<Prod, T>::type>(1);
     }
   }
 };

mlx/backend/cuda/reduce/reduce_utils.cuh

@@ -4,6 +4,7 @@
 
 #include <numeric>
 
+#include "mlx/backend/common/utils.h"
 #include "mlx/backend/cuda/device/utils.cuh"
 
 #include <cooperative_groups.h>
@@ -55,22 +56,6 @@ __device__ void atomic_reduce(T* x, T y) {
   }
 }
 
-// TODO: Should make a custom complex type
-template <typename U, typename T>
-inline __device__ U __cast(T x) {
-  return static_cast<U>(x);
-}
-
-template <>
-inline __device__ bool __cast<bool, cuComplex>(cuComplex x) {
-  return x.x != 0 && x.y != 0;
-}
-
-template <>
-inline __device__ cuComplex __cast<cuComplex, bool>(bool x) {
-  return x ? make_cuFloatComplex(1, 1) : make_cuFloatComplex(0, 0);
-}
-
 template <typename T, int N, typename Block, typename Warp, typename Op>
 inline __device__ void
 block_reduce(Block block, Warp warp, T (&vals)[N], T* smem, Op op, T init) {

mlx/backend/cuda/reduce/row_reduce.cu

@@ -3,7 +3,6 @@
 #include <numeric>
 
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cast_op.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 
 #include <cooperative_groups.h>
@@ -113,7 +112,7 @@ __global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
             in + k * size + r * (block.size() * N),
             vals[k]);
         for (int j = 0; j < N; j++) {
-          accs[k] = op(accs[k], __cast<U, T>(vals[k][j]));
+          accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
         }
       }
     }
@@ -125,7 +124,7 @@ __global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
             in + k * size + r * (block.size() * N),
             vals[k]);
         for (int j = 0; j < N; j++) {
-          accs[k] = op(accs[k], __cast<U, T>(vals[k][j]));
+          accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
         }
       }
     }
@@ -138,9 +137,9 @@ __global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
           in + k * size + final_offset,
           vals[k],
           size,
-          __cast<T, U>(init));
+          cast_to<T>(init));
       for (int j = 0; j < N; j++) {
-        accs[k] = op(accs[k], __cast<U, T>(vals[k][j]));
+        accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
       }
     }
   }
@@ -199,7 +198,7 @@ __global__ void row_reduce_looped(
           in + loop.location() + r * BLOCK_DIM * N_READS,
           vals);
       for (int i = 0; i < N_READS; i++) {
-        total[0] = op(total[0], __cast<U, T>(vals[i]));
+        total[0] = op(total[0], cast_to<U>(vals[i]));
       }
     }
     if (final_offset < args.row_size) {
@@ -209,9 +208,9 @@ __global__ void row_reduce_looped(
           in + loop.location() + final_offset,
           vals,
           args.row_size - final_offset,
-          __cast<T, U>(init));
+          cast_to<T>(init));
       for (int i = 0; i < N_READS; i++) {
-        total[0] = op(total[0], __cast<U, T>(vals[i]));
+        total[0] = op(total[0], cast_to<U>(vals[i]));
       }
     }
     // TODO: Maybe block.sync() here?

mlx/backend/cuda/rms_norm.cu

@@ -1,7 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 #include "mlx/backend/gpu/copy.h"
@@ -74,7 +73,7 @@ __global__ void rms_norm(
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
     T xn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size, 0);
+    cub::LoadDirectBlocked(index, x, xn, axis_size, cast_to<T>(0));
     for (int i = 0; i < N_READS; ++i) {
       float t = static_cast<float>(xn[i]);
       normalizer += t * t;
@@ -89,7 +88,7 @@ __global__ void rms_norm(
     T xn[N_READS];
     T wn[N_READS];
     cub::LoadDirectBlocked(index, x, xn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    cub::LoadDirectBlocked(index, StridedIterator(w, w_stride), wn, axis_size);
     for (int i = 0; i < N_READS; ++i) {
       float norm = static_cast<float>(xn[i]) * normalizer;
       xn[i] = wn[i] * static_cast<T>(norm);
@@ -130,9 +129,9 @@ __global__ void rms_norm_vjp(
     T wn[N_READS] = {};
     T gn[N_READS] = {};
     auto index = r * BLOCK_DIM + block.thread_rank();
-    cub::LoadDirectBlocked(index, x, xn, axis_size, 0);
+    cub::LoadDirectBlocked(index, x, xn, axis_size, cast_to<T>(0));
     cub::LoadDirectBlocked(index, g, gn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    cub::LoadDirectBlocked(index, StridedIterator(w, w_stride), wn, axis_size);
     for (int i = 0; i < N_READS; i++) {
       float t = static_cast<float>(xn[i]);
       float wi = wn[i];
@@ -154,7 +153,7 @@ __global__ void rms_norm_vjp(
     T gn[N_READS];
     cub::LoadDirectBlocked(index, x, xn, axis_size);
     cub::LoadDirectBlocked(index, g, gn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    cub::LoadDirectBlocked(index, StridedIterator(w, w_stride), wn, axis_size);
     for (int i = 0; i < N_READS; i++) {
       float xi = xn[i];
       float wi = wn[i];
@@ -206,8 +205,7 @@ void RMSNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -259,9 +257,7 @@ void RMSNormVJP::eval_gpu(
       return x;
     }
     copied = true;
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
-    copy_gpu(x, x_copy, CopyType::General, s);
-    return x_copy;
+    return contiguous_copy_gpu(x, s);
   };
   bool donate_x = inputs[0].is_donatable();
   bool donate_g = inputs[2].is_donatable();

mlx/backend/cuda/scan.cu

@@ -0,0 +1,465 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/reduce/reduce_ops.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/scan.h>
+#include <nvtx3/nvtx3.hpp>
+
+#include <cassert>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename T>
+struct ScanResult {
+  using type = T;
+};
+
+template <>
+struct ScanResult<Sum, bool> {
+  using type = int32_t;
+};
+
+template <typename T>
+struct ReduceInit<LogAddExp, T> {
+  static constexpr __host__ __device__ T value() {
+    return Limits<T>::min();
+  }
+};
+
+template <bool reverse, typename T, typename U, int N_READS>
+inline __device__ void
+load_values(int index, const T* in, U (&values)[N_READS], int size, U init) {
+  int remaining = size - index * N_READS;
+  if constexpr (reverse) {
+    in += remaining - N_READS;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        values[N_READS - i - 1] =
+            (N_READS - i - 1 < remaining) ? cast_to<U>(in[i]) : init;
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        values[N_READS - i - 1] = cast_to<U>(in[i]);
+      }
+    }
+  } else {
+    in += index * N_READS;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        values[i] = (i < remaining) ? cast_to<U>(in[i]) : init;
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        values[i] = cast_to<U>(in[i]);
+      }
+    }
+  }
+}
+
+template <bool reverse, int offset, typename T, int N_READS>
+inline __device__ void
+store_values(int index, T* out, T (&values)[N_READS], int size) {
+  int start = index * N_READS + offset;
+  int remaining = size - start;
+  if constexpr (reverse) {
+    out += remaining - N_READS;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        if (N_READS - i - 1 < remaining) {
+          out[i] = values[N_READS - i - 1];
+        }
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        out[i] = values[N_READS - i - 1];
+      }
+    }
+  } else {
+    out += start;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        if (i < remaining) {
+          out[i] = values[i];
+        }
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        out[i] = values[i];
+      }
+    }
+  }
+}
+
+template <
+    typename T,
+    typename U,
+    typename Op,
+    int N_READS,
+    bool inclusive,
+    bool reverse>
+__global__ void contiguous_scan(const T* in, U* out, int32_t axis_size) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  in += grid.block_rank() * axis_size;
+  out += grid.block_rank() * axis_size;
+
+  __shared__ U warp_sums[WARP_SIZE];
+
+  Op op;
+  U init = ReduceInit<Op, T>::value();
+  U prefix = init;
+
+  // Scan per block.
+  for (int r = 0; r < cuda::ceil_div(axis_size, block.size() * N_READS); ++r) {
+    int32_t index = r * block.size() + block.thread_rank();
+    U values[N_READS];
+    load_values<reverse>(index, in, values, axis_size, init);
+
+    // Compute an inclusive scan per thread.
+    for (int i = 1; i < N_READS; ++i) {
+      values[i] = op(values[i], values[i - 1]);
+    }
+
+    // Compute exclusive scan of thread sums.
+    U prev_thread_sum = cg::exclusive_scan(warp, values[N_READS - 1], op);
+    if (warp.thread_rank() == 0) {
+      prev_thread_sum = init;
+    }
+
+    // Write wrap's sum to shared memory.
+    if (warp.thread_rank() == WARP_SIZE - 1) {
+      warp_sums[warp.meta_group_rank()] =
+          op(prev_thread_sum, values[N_READS - 1]);
+    }
+    block.sync();
+
+    // Compute exclusive scan of warp sums.
+    if (warp.meta_group_rank() == 0) {
+      U prev_warp_sum =
+          cg::exclusive_scan(warp, warp_sums[warp.thread_rank()], op);
+      if (warp.thread_rank() == 0) {
+        prev_warp_sum = init;
+      }
+      warp_sums[warp.thread_rank()] = prev_warp_sum;
+    }
+    block.sync();
+
+    // Compute the output.
+    for (int i = 0; i < N_READS; ++i) {
+      values[i] = op(values[i], prefix);
+      values[i] = op(values[i], warp_sums[warp.meta_group_rank()]);
+      values[i] = op(values[i], prev_thread_sum);
+    }
+
+    // Write the values.
+    if (inclusive) {
+      store_values<reverse, 0>(index, out, values, axis_size);
+    } else {
+      store_values<reverse, 1>(index, out, values, axis_size);
+      if (reverse) {
+        if (block.thread_rank() == 0 && index == 0) {
+          out[axis_size - 1] = init;
+        }
+      } else {
+        if (block.thread_rank() == 0 && index == 0) {
+          out[0] = init;
+        }
+      }
+    }
+    block.sync();
+
+    // Share the prefix.
+    if ((warp.meta_group_rank() == warp.meta_group_size() - 1) &&
+        (warp.thread_rank() == WARP_SIZE - 1)) {
+      warp_sums[0] = values[N_READS - 1];
+    }
+    block.sync();
+    prefix = warp_sums[0];
+  }
+}
+
+template <
+    typename T,
+    typename U,
+    typename Op,
+    int N_READS,
+    int BM,
+    int BN,
+    bool inclusive,
+    bool reverse>
+__global__ void strided_scan(
+    const T* in,
+    U* out,
+    int32_t axis_size,
+    int64_t stride,
+    int64_t stride_blocks) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  constexpr int BN_pad = WARP_SIZE + 16 / sizeof(U);
+  constexpr int n_warps = BN / N_READS;
+  constexpr int n_scans = BN / n_warps;
+
+  __shared__ U read_buffer[BM * BN_pad];
+
+  Op op;
+  U init = ReduceInit<Op, T>::value();
+  U values[n_scans];
+  U prefix[n_scans];
+  for (int i = 0; i < n_scans; ++i) {
+    prefix[i] = init;
+  }
+
+  // Compute offsets.
+  int64_t offset = (grid.block_rank() / stride_blocks) * axis_size * stride;
+  int64_t global_index_x = (grid.block_rank() % stride_blocks) * BN;
+  uint read_offset_y = (block.thread_rank() * N_READS) / BN;
+  uint read_offset_x = (block.thread_rank() * N_READS) % BN;
+  uint scan_offset_y = warp.thread_rank();
+  uint scan_offset_x = warp.meta_group_rank() * n_scans;
+
+  uint stride_limit = stride - global_index_x;
+  in += offset + global_index_x + read_offset_x;
+  out += offset + global_index_x + read_offset_x;
+  U* read_into = read_buffer + read_offset_y * BN_pad + read_offset_x;
+  U* read_from = read_buffer + scan_offset_y * BN_pad + scan_offset_x;
+
+  for (uint j = 0; j < axis_size; j += BM) {
+    // Calculate the indices for the current thread.
+    uint index_y = j + read_offset_y;
+    uint check_index_y = index_y;
+    if (reverse) {
+      index_y = axis_size - 1 - index_y;
+    }
+
+    // Read in SM.
+    if (check_index_y < axis_size && (read_offset_x + N_READS) < stride_limit) {
+      for (int i = 0; i < N_READS; ++i) {
+        read_into[i] = in[index_y * stride + i];
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        if (check_index_y < axis_size && (read_offset_x + i) < stride_limit) {
+          read_into[i] = in[index_y * stride + i];
+        } else {
+          read_into[i] = init;
+        }
+      }
+    }
+    block.sync();
+
+    // Read strided into registers.
+    for (int i = 0; i < n_scans; ++i) {
+      values[i] = read_from[i];
+    }
+
+    // Perform the scan.
+    for (int i = 0; i < n_scans; ++i) {
+      values[i] = cg::inclusive_scan(warp, values[i], op);
+      values[i] = op(values[i], prefix[i]);
+      prefix[i] = warp.shfl(values[i], WARP_SIZE - 1);
+    }
+
+    // Write to SM.
+    for (int i = 0; i < n_scans; ++i) {
+      read_from[i] = values[i];
+    }
+    block.sync();
+
+    // Write to device memory.
+    if (!inclusive) {
+      if (check_index_y == 0) {
+        if ((read_offset_x + N_READS) < stride_limit) {
+          for (int i = 0; i < N_READS; ++i) {
+            out[index_y * stride + i] = init;
+          }
+        } else {
+          for (int i = 0; i < N_READS; ++i) {
+            if ((read_offset_x + i) < stride_limit) {
+              out[index_y * stride + i] = init;
+            }
+          }
+        }
+      }
+      if (reverse) {
+        index_y -= 1;
+        check_index_y += 1;
+      } else {
+        index_y += 1;
+        check_index_y += 1;
+      }
+    }
+    if (check_index_y < axis_size && (read_offset_x + N_READS) < stride_limit) {
+      for (int i = 0; i < N_READS; ++i) {
+        out[index_y * stride + i] = read_into[i];
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        if (check_index_y < axis_size && (read_offset_x + i) < stride_limit) {
+          out[index_y * stride + i] = read_into[i];
+        }
+      }
+    }
+  }
+}
+
+} // namespace cu
+
+template <typename F>
+void dispatch_scan_ops(Scan::ReduceType scan_op, F&& f) {
+  if (scan_op == Scan::ReduceType::Max) {
+    f(type_identity<cu::Max>{});
+  } else if (scan_op == Scan::ReduceType::Min) {
+    f(type_identity<cu::Min>{});
+  } else if (scan_op == Scan::ReduceType::Sum) {
+    f(type_identity<cu::Sum>{});
+  } else if (scan_op == Scan::ReduceType::Prod) {
+    f(type_identity<cu::Prod>{});
+  } else if (scan_op == Scan::ReduceType::LogAddExp) {
+    f(type_identity<cu::LogAddExp>{});
+  } else {
+    throw std::invalid_argument("Unknown reduce type.");
+  }
+}
+
+template <typename Op>
+const char* op_to_string() {
+  if (cuda::std::is_same_v<Op, cu::Max>) {
+    return "Max";
+  } else if (cuda::std::is_same_v<Op, cu::Min>) {
+    return "Min";
+  } else if (cuda::std::is_same_v<Op, cu::Sum>) {
+    return "Sum";
+  } else if (cuda::std::is_same_v<Op, cu::Prod>) {
+    return "Prod";
+  } else if (cuda::std::is_same_v<Op, cu::LogAddExp>) {
+    return "LogAddExp";
+  } else {
+    throw std::invalid_argument("Unknown op.");
+  }
+}
+
+template <typename Op, typename T>
+constexpr bool supports_scan_op() {
+  if constexpr (cuda::std::is_same_v<Op, LogAddExp>) {
+    return is_inexact_v<T>;
+  } else {
+    return true;
+  }
+}
+
+void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Scan::eval_gpu");
+  assert(inputs.size() == 1);
+  auto in = inputs[0];
+  auto& s = stream();
+
+  if (in.flags().contiguous && in.strides()[axis_] != 0) {
+    if (in.is_donatable() && in.itemsize() == out.itemsize()) {
+      out.copy_shared_buffer(in);
+    } else {
+      out.set_data(
+          allocator::malloc(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+    }
+  } else {
+    in = contiguous_copy_gpu(in, s);
+    out.copy_shared_buffer(in);
+  }
+
+  constexpr int N_READS = 4;
+  int32_t axis_size = in.shape(axis_);
+  bool contiguous = in.strides()[axis_] == 1;
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    dispatch_scan_ops(reduce_type_, [&](auto scan_op_tag) {
+      using Op = MLX_GET_TYPE(scan_op_tag);
+      if constexpr (supports_scan_op<Op, T>) {
+        using U = typename cu::ScanResult<Op, T>::type;
+        dispatch_bool(inclusive_, [&](auto inclusive) {
+          dispatch_bool(reverse_, [&](auto reverse) {
+            if (contiguous) {
+              auto kernel = cu::contiguous_scan<
+                  T,
+                  U,
+                  Op,
+                  N_READS,
+                  inclusive.value,
+                  reverse.value>;
+              int block_dim = cuda::ceil_div(axis_size, N_READS);
+              block_dim = cuda::ceil_div(block_dim, WARP_SIZE) * WARP_SIZE;
+              block_dim = std::min(block_dim, WARP_SIZE * WARP_SIZE);
+              encoder.add_kernel_node(
+                  kernel,
+                  in.data_size() / axis_size,
+                  block_dim,
+                  in.data<T>(),
+                  out.data<U>(),
+                  axis_size);
+            } else {
+              constexpr int BM = WARP_SIZE;
+              constexpr int BN = WARP_SIZE;
+              auto kernel = cu::strided_scan<
+                  T,
+                  U,
+                  Op,
+                  N_READS,
+                  BM,
+                  BN,
+                  inclusive.value,
+                  reverse.value>;
+              int64_t stride = in.strides()[axis_];
+              int64_t stride_blocks = cuda::ceil_div(stride, BN);
+              dim3 num_blocks = get_2d_grid_dims(
+                  in.shape(), in.strides(), axis_size * stride);
+              if (num_blocks.x * stride_blocks <= UINT32_MAX) {
+                num_blocks.x *= stride_blocks;
+              } else {
+                num_blocks.y *= stride_blocks;
+              }
+              int block_dim = (BN / N_READS) * WARP_SIZE;
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dim,
+                  in.data<T>(),
+                  out.data<U>(),
+                  axis_size,
+                  stride,
+                  stride_blocks);
+            }
+          });
+        });
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do scan op {} on inputs of {} with result of {}.",
+            op_to_string<Op>(),
+            dtype_to_string(in.dtype()),
+            dtype_to_string(out.dtype())));
+      }
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/softmax.cu

@@ -43,7 +43,7 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
   // Thread reduce.
   AccT prevmax;
   AccT maxval = Limits<AccT>::finite_min();
-  AccT normalizer = 0;
+  AccT normalizer = cast_to<AccT>(0);
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
     AccT vals[N_READS];
     cub::LoadDirectBlocked(
@@ -125,8 +125,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }

mlx/backend/cuda/sort.cu

@@ -72,8 +72,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
   bool is_segmented_sort = in.flags().contiguous && in.strides()[axis] == 1;
   if (!is_segmented_sort) {
     array trans = swapaxes_in_eval(in, axis, last_dim);
-    in = array(trans.shape(), trans.dtype(), nullptr, {});
-    copy_gpu(trans, in, CopyType::General, s);
+    in = contiguous_copy_gpu(trans, s);
     encoder.add_temporary(in);
     out = array(allocator::malloc(out.nbytes()), in.shape(), out.dtype());
     encoder.add_temporary(out);

mlx/backend/cuda/ternary.cu

@@ -15,12 +15,27 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename Op, typename T, typename IdxT>
+template <typename Op, typename T, typename IdxT, int N_READS>
 __global__ void
 ternary_v(const bool* a, const T* b, const T* c, T* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[index], c[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[i], c[i]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+    auto c_vec = load_vector<N_READS>(c, index);
+
+    AlignedVector<T, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b_vec.val[i], c_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -61,7 +76,7 @@ __global__ void ternary_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [a_idx, b_idx, c_idx] = elem_to_loc_4d(
+    auto [a_idx, b_idx, c_idx] = elem_to_loc(
         index,
         shape.data(),
         a_strides.data(),
@@ -149,11 +164,18 @@ void ternary_op_gpu_inplace(
             }
           });
     } else {
-      dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
+      dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
         using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-        auto kernel = cu::ternary_v<Op, DType, IdxT>;
+        // TODO: Choose optimized value based on type size.
+        constexpr int N_READS = 4;
+        auto kernel = cu::ternary_v<Op, DType, IdxT, N_READS>;
         auto [num_blocks, block_dims] = get_launch_args(
-            kernel, out.data_size(), out.shape(), out.strides(), large());
+            kernel,
+            out.data_size(),
+            out.shape(),
+            out.strides(),
+            large(),
+            N_READS);
         encoder.add_kernel_node(
             kernel,
             num_blocks,

mlx/backend/cuda/unary.cu

@@ -2,9 +2,7 @@
 
 #include "mlx/backend/common/unary.h"
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/device/unary_ops.cuh"
-#include "mlx/backend/cuda/iterators/general_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -18,11 +16,24 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void unary_v(const In* in, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(in[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(in[i]);
+    }
+  } else {
+    auto in_vec = load_vector<N_READS>(in, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(in_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -36,7 +47,7 @@ __global__ void unary_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto idx = elem_to_loc_4d(index, shape.data(), strides.data(), ndim);
+    auto idx = elem_to_loc(index, shape.data(), strides.data(), ndim);
     out[index] = Op{}(in[idx]);
   }
 }
@@ -58,10 +69,10 @@ constexpr bool supports_unary_op() {
         !std::is_same_v<In, bool>;
   }
   if (std::is_same_v<Op, Ceil> || std::is_same_v<Op, Floor>) {
-    return std::is_same_v<In, Out> && !std::is_same_v<In, complex64_t>;
+    return std::is_same_v<In, Out> && !mlx::core::is_complex_v<In>;
   }
   if (std::is_same_v<Op, Conjugate>) {
-    return std::is_same_v<In, Out> && std::is_same_v<In, complex64_t>;
+    return std::is_same_v<In, Out> && mlx::core::is_complex_v<In>;
   }
   if (std::is_same_v<Op, ArcCos> || std::is_same_v<Op, ArcSin> ||
       std::is_same_v<Op, ArcTan> || std::is_same_v<Op, Cos> ||
@@ -75,7 +86,7 @@ constexpr bool supports_unary_op() {
     return std::is_same_v<In, Out> && is_inexact_v<In>;
   }
   if (std::is_same_v<Op, Imag> || std::is_same_v<Op, Real>) {
-    return std::is_same_v<In, complex64_t> && std::is_same_v<Out, float>;
+    return mlx::core::is_complex_v<In> && std::is_same_v<Out, float>;
   }
   if (std::is_same_v<Op, LogicalNot>) {
     return std::is_same_v<In, Out> && std::is_same_v<In, bool>;
@@ -89,7 +100,7 @@ template <typename Op>
 void unary_op_gpu_inplace(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   auto& in = inputs[0];
   if (in.size() == 0) {
@@ -112,14 +123,20 @@ void unary_op_gpu_inplace(
       using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
       if constexpr (cu::supports_unary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
         dispatch_bool(large, [&](auto large) {
-          using IdxT = std::conditional_t<large(), int64_t, int32_t>;
           using InType = cuda_type_t<CTYPE_IN>;
           using OutType = cuda_type_t<CTYPE_OUT>;
-          using IdxT = std::conditional_t<large(), int64_t, int32_t>;
           if (contig) {
-            auto kernel = cu::unary_v<Op, InType, OutType, IdxT>;
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::unary_v<Op, InType, OutType, IdxT, N_READS>;
             auto [num_blocks, block_dims] = get_launch_args(
-                kernel, out.data_size(), out.shape(), out.strides(), large);
+                kernel,
+                out.data_size(),
+                out.shape(),
+                out.strides(),
+                large,
+                N_READS);
             encoder.add_kernel_node(
                 kernel,
                 num_blocks,
@@ -128,6 +145,7 @@ void unary_op_gpu_inplace(
                 out.data<OutType>(),
                 out.data_size());
           } else {
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
             auto [shape, strides] = collapse_contiguous_dims(in);
             auto kernel = cu::unary_g<Op, InType, OutType, IdxT>;
             auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
@@ -158,17 +176,17 @@ template <typename Op>
 void unary_op_gpu(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   set_unary_output_data(inputs[0], out);
   unary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 
-#define UNARY_GPU(func)                                                 \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) {   \
-    nvtx3::scoped_range r(#func "::eval_gpu");                          \
-    auto& s = out.primitive().stream();                                 \
-    unary_op_gpu<cu::func>(inputs, out, get_primitive_string(this), s); \
+#define UNARY_GPU(func)                                               \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    nvtx3::scoped_range r(#func "::eval_gpu");                        \
+    auto& s = out.primitive().stream();                               \
+    unary_op_gpu<cu::func>(inputs, out, name(), s);                   \
   }
 
 UNARY_GPU(Abs)
@@ -204,16 +222,15 @@ UNARY_GPU(Tanh)
 void Log::eval_gpu(const std::vector<array>& inputs, array& out) {
   nvtx3::scoped_range r("Log::eval_gpu");
   auto& s = out.primitive().stream();
-  auto op = get_primitive_string(this);
   switch (base_) {
     case Base::e:
-      unary_op_gpu<cu::Log>(inputs, out, op, s);
+      unary_op_gpu<cu::Log>(inputs, out, name(), s);
       break;
     case Base::two:
-      unary_op_gpu<cu::Log2>(inputs, out, op, s);
+      unary_op_gpu<cu::Log2>(inputs, out, name(), s);
       break;
     case Base::ten:
-      unary_op_gpu<cu::Log10>(inputs, out, op, s);
+      unary_op_gpu<cu::Log10>(inputs, out, name(), s);
       break;
   }
 }
@@ -224,7 +241,7 @@ void Round::eval_gpu(const std::vector<array>& inputs, array& out) {
   const auto& in = inputs[0];
   auto& s = out.primitive().stream();
   if (issubdtype(in.dtype(), inexact)) {
-    unary_op_gpu<cu::Round>(inputs, out, get_primitive_string(this), s);
+    unary_op_gpu<cu::Round>(inputs, out, name(), s);
   } else {
     // No-op integer types
     out.copy_shared_buffer(in);

mlx/backend/cuda/utils.cpp

@@ -17,6 +17,14 @@ CudaStream::~CudaStream() {
   CHECK_CUDA_ERROR(cudaStreamDestroy(stream_));
 }
 
+void check_cublas_error(const char* name, cublasStatus_t err) {
+  if (err != CUBLAS_STATUS_SUCCESS) {
+    // TODO: Use cublasGetStatusString when it is widely available.
+    throw std::runtime_error(
+        fmt::format("{} failed with code: {}.", name, static_cast<int>(err)));
+  }
+}
+
 void check_cuda_error(const char* name, cudaError_t err) {
   if (err != cudaSuccess) {
     throw std::runtime_error(
@@ -61,7 +69,7 @@ const char* dtype_to_cuda_type(const Dtype& dtype) {
     case float64:
       return "double";
     case complex64:
-      return "cuComplex";
+      return "complex64_t";
     default:
       return "unknown";
   }

mlx/backend/cuda/utils.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <cublasLt.h>
 #include <cuda.h>
 #include <cuda_runtime.h>
 
@@ -33,10 +34,12 @@ class CudaStream {
 };
 
 // Throw exception if the cuda API does not succeed.
+void check_cublas_error(const char* name, cublasStatus_t err);
 void check_cuda_error(const char* name, cudaError_t err);
 void check_cuda_error(const char* name, CUresult err);
 
 // The macro version that prints the command that failed.
+#define CHECK_CUBLAS_ERROR(cmd) check_cublas_error(#cmd, (cmd))
 #define CHECK_CUDA_ERROR(cmd) check_cuda_error(#cmd, (cmd))
 
 // Convert Dtype to CUDA C++ types.

mlx/backend/cuda/worker.cpp

@@ -1,7 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/worker.h"
-#include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/device.h"
 
 namespace mlx::core::cu {
@@ -12,10 +11,10 @@ Worker::Worker()
 
 Worker::~Worker() {
   {
-    std::lock_guard lock(worker_mutex_);
+    std::lock_guard lock(mtx_);
     stop_ = true;
   }
-  worker_event_.signal(batch_ + 1);
+  cond_.notify_one();
   worker_.join();
 }
 
@@ -23,53 +22,41 @@ void Worker::add_task(std::function<void()> task) {
   pending_tasks_.push_back(std::move(task));
 }
 
-void Worker::consume_in_this_thread() {
-  for (auto& task : pending_tasks_) {
-    task();
-  }
-  pending_tasks_.clear();
-}
-
-void Worker::end_batch() {
-  batch_++;
+void Worker::signal(void* data) {
+  auto w = static_cast<Worker*>(data);
   {
-    std::lock_guard lock(worker_mutex_);
-    worker_tasks_[batch_] = std::move(pending_tasks_);
+    std::lock_guard lock(w->mtx_);
+    w->signaled_batch_++;
   }
-  uncommited_batches_++;
-}
-
-void Worker::commit() {
-  if (uncommited_batches_ == 0) {
-    return;
-  }
-  uncommited_batches_ = 0;
-  worker_event_.signal(batch_);
+  w->cond_.notify_one();
 }
 
 void Worker::commit(cudaStream_t stream) {
-  if (uncommited_batches_ == 0) {
+  // Move pending tasks into tasks
+  if (pending_tasks_.empty()) {
     return;
   }
-  uncommited_batches_ = 0;
-  // Signal the |worker_event_| in |signal_stream_| after the kernels in
-  // |stream_| finish running.
+  {
+    std::lock_guard lock(mtx_);
+    // Move pending tasks into ready tasks
+    worker_tasks_[++committed_batch_] = std::move(pending_tasks_);
+  }
   signal_event_.record(stream);
   signal_event_.wait(signal_stream_);
-  worker_event_.signal(signal_stream_, batch_);
+  cudaLaunchHostFunc(signal_stream_, signal, this);
 }
 
 void Worker::thread_fn() {
-  // The worker thread is safe to free buffers.
-  allocator().register_this_thread();
-
   while (!stop_) {
-    uint64_t batch = worker_event_.value();
+    uint64_t current_batch = 0;
     Tasks tasks;
     {
-      std::lock_guard lock(worker_mutex_);
-      // Move tasks in signaled batches.
-      auto end = worker_tasks_.upper_bound(batch);
+      std::unique_lock<std::mutex> lk(mtx_);
+      cond_.wait(lk, [this, &current_batch] {
+        return this->signaled_batch_ > current_batch || this->stop_;
+      });
+      current_batch = signaled_batch_;
+      auto end = worker_tasks_.upper_bound(current_batch);
       for (auto it = worker_tasks_.begin(); it != end; ++it) {
         if (tasks.empty()) {
           tasks = std::move(it->second);
@@ -85,7 +72,6 @@ void Worker::thread_fn() {
       auto task = std::move(tasks[i]);
       task();
     }
-    worker_event_.wait(batch + 1);
   }
 }
 

mlx/backend/cuda/worker.h

@@ -5,6 +5,7 @@
 #include "mlx/backend/cuda/event.h"
 #include "mlx/backend/cuda/utils.h"
 
+#include <condition_variable>
 #include <functional>
 #include <map>
 #include <mutex>
@@ -24,38 +25,24 @@ class Worker {
   // Add a pending |task| that will run when consumed or commited.
   void add_task(std::function<void()> task);
 
-  // Run pending tasks immediately in current thread.
-  void consume_in_this_thread();
-
-  // Put pending tasks in a batch.
-  void end_batch();
-
-  // Inform worker thread to run current batches now.
-  void commit();
-
   // Inform worker thread to run current batches after kernels in |stream|
   // finish running.
   void commit(cudaStream_t stream);
 
-  // Return how many batches have been added but not committed yet.
-  size_t uncommited_batches() const {
-    return uncommited_batches_;
-  }
-
  private:
-  void thread_fn();
+  static void signal(void*);
 
-  uint64_t batch_{0};
-  size_t uncommited_batches_{0};
+  void thread_fn();
+  std::mutex mtx_;
+  std::condition_variable cond_;
+
+  uint64_t committed_batch_{0};
+  uint64_t signaled_batch_{0};
 
   // Cuda stream and event for signaling kernel completion.
   CudaStream signal_stream_;
   CudaEvent signal_event_;
 
-  // Worker thread.
-  SharedEvent worker_event_;
-  std::thread worker_;
-  std::mutex worker_mutex_;
   bool stop_{false};
 
   // Tasks are put in |pending_tasks_| first, and then moved to
@@ -63,6 +50,7 @@ class Worker {
   using Tasks = std::vector<std::function<void()>>;
   Tasks pending_tasks_;
   std::map<uint64_t, Tasks> worker_tasks_;
+  std::thread worker_;
 };
 
 } // namespace mlx::core::cu

mlx/backend/gpu/copy.cpp

@@ -46,4 +46,10 @@ void copy_gpu_inplace(
       in, out, in.shape(), i_strides, out.strides(), i_offset, 0, ctype, s);
 }
 
+array contiguous_copy_gpu(const array& arr, const Stream& s) {
+  array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+  copy_gpu(arr, arr_copy, CopyType::General, s);
+  return arr_copy;
+}
+
 } // namespace mlx::core

mlx/backend/gpu/copy.h

@@ -43,4 +43,7 @@ void copy_gpu_inplace(
 // Fill the output with the scalar val
 void fill_gpu(const array& val, array& out, const Stream& s);
 
+// Return a contiguous array with same shape that copies the data of |arr|.
+array contiguous_copy_gpu(const array& arr, const Stream& s);
+
 } // namespace mlx::core

mlx/backend/metal/CMakeLists.txt

@@ -63,6 +63,7 @@ if(MLX_METAL_JIT)
   make_jit_source(steel/gemm/kernels/steel_gemm_masked kernels/steel/defines.h)
   make_jit_source(steel/gemm/kernels/steel_gemm_gather)
   make_jit_source(steel/gemm/kernels/steel_gemm_splitk)
+  make_jit_source(steel/gemm/kernels/steel_gemm_segmented)
   make_jit_source(
     steel/conv/conv
     kernels/steel/utils.h

mlx/backend/metal/allocator.cpp

@@ -128,8 +128,7 @@ Buffer MetalAllocator::malloc(size_t size) {
 
     auto pool = metal::new_scoped_memory_pool();
 
-    // If we have a lot of memory pressure or are over the maximum cache size,
-    // try to reclaim memory from the cache
+    // If we have a lot of memory pressure try to reclaim memory from the cache
     if (mem_required >= gc_limit_ || num_resources_ >= resource_limit_) {
       num_resources_ -=
           buffer_cache_.release_cached_buffers(mem_required - gc_limit_);

mlx/backend/metal/binary.cpp

@@ -7,20 +7,20 @@
 
 #define BINARY_GPU(func)                                              \
   void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
-    binary_op_gpu(inputs, out, get_primitive_string(this));           \
+    binary_op_gpu(inputs, out, name());                               \
   }
 
 #define BINARY_GPU_MULTI(func)                                         \
   void func::eval_gpu(                                                 \
       const std::vector<array>& inputs, std::vector<array>& outputs) { \
-    binary_op_gpu(inputs, outputs, get_primitive_string(this));        \
+    binary_op_gpu(inputs, outputs, name());                            \
   }
 
 namespace mlx::core {
 
 std::string get_kernel_name(
     BinaryOpType bopt,
-    const std::string& op,
+    const char* op,
     const array& a,
     bool large,
     int ndim,
@@ -65,7 +65,7 @@ std::string get_kernel_name(
 void binary_op_gpu_inplace(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   auto& a = inputs[0];
   auto& b = inputs[1];
@@ -165,7 +165,7 @@ void binary_op_gpu_inplace(
 void binary_op_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
@@ -179,7 +179,7 @@ void binary_op_gpu(
 void binary_op_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::string& op) {
+    const char* op) {
   auto& s = outputs[0].primitive().stream();
   binary_op_gpu(inputs, outputs, op, s);
 }
@@ -187,7 +187,7 @@ void binary_op_gpu(
 void binary_op_gpu_inplace(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   std::vector<array> outputs = {out};
   binary_op_gpu_inplace(inputs, outputs, op, s);
@@ -196,7 +196,7 @@ void binary_op_gpu_inplace(
 void binary_op_gpu(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   assert(inputs.size() == 2);
   auto& a = inputs[0];
@@ -209,7 +209,7 @@ void binary_op_gpu(
 void binary_op_gpu(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op) {
+    const char* op) {
   auto& s = out.primitive().stream();
   binary_op_gpu(inputs, out, op, s);
 }
@@ -237,19 +237,19 @@ BINARY_GPU(Subtract)
 void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
   switch (op_) {
     case BitwiseBinary::And:
-      binary_op_gpu(inputs, out, get_primitive_string(this));
+      binary_op_gpu(inputs, out, name());
       break;
     case BitwiseBinary::Or:
-      binary_op_gpu(inputs, out, get_primitive_string(this));
+      binary_op_gpu(inputs, out, name());
       break;
     case BitwiseBinary::Xor:
-      binary_op_gpu(inputs, out, get_primitive_string(this));
+      binary_op_gpu(inputs, out, name());
       break;
     case BitwiseBinary::LeftShift:
-      binary_op_gpu(inputs, out, get_primitive_string(this));
+      binary_op_gpu(inputs, out, name());
       break;
     case BitwiseBinary::RightShift:
-      binary_op_gpu(inputs, out, get_primitive_string(this));
+      binary_op_gpu(inputs, out, name());
       break;
   }
 }

mlx/backend/metal/binary.h

@@ -9,25 +9,25 @@ namespace mlx::core {
 void binary_op_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::string& op,
+    const char* op,
     const Stream& s);
 
 void binary_op_gpu(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s);
 
 void binary_op_gpu_inplace(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
-    const std::string& op,
+    const char* op,
     const Stream& s);
 
 void binary_op_gpu_inplace(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s);
 
 } // namespace mlx::core

mlx/backend/metal/compiled.cpp

@@ -212,9 +212,7 @@ inline void build_kernel(
           get_type_string(x.dtype()),
           namer.get_name(x.inputs()[0]));
     } else {
-      std::ostringstream ss;
-      x.primitive().print(ss);
-      os += ss.str();
+      os += x.primitive().name();
       os += "()(";
       for (int i = 0; i < x.inputs().size() - 1; i++) {
         os += fmt::format("tmp_{0}, ", namer.get_name(x.inputs()[i]));

mlx/backend/metal/conv.cpp

@@ -149,8 +149,7 @@ void explicit_gemm_conv_group_ND_gpu(
       wt, {wt.strides(0), 1, C_per_group}, wt.flags(), wt.size());
 
   // Materialize
-  auto wt_transpose = array(wt_view.shape(), wt_view.dtype(), nullptr, {});
-  copy_gpu(wt_view, wt_transpose, CopyType::General, s);
+  array wt_transpose = contiguous_copy_gpu(wt_view, s);
 
   // Perform gemm
   std::vector<array> copies = {in_unfolded, wt_transpose};
@@ -961,16 +960,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto in = inputs[0];
   auto wt = inputs[1];
   if (!in.flags().row_contiguous) {
-    array arr_copy(in.shape(), in.dtype(), nullptr, {});
-    copy_gpu(in, arr_copy, CopyType::General, s);
-    copies.push_back(arr_copy);
-    in = arr_copy;
+    in = contiguous_copy_gpu(in, s);
+    copies.push_back(in);
   }
   if (!wt.flags().row_contiguous) {
-    array arr_copy(wt.shape(), wt.dtype(), nullptr, {});
-    copy_gpu(wt, arr_copy, CopyType::General, s);
-    copies.push_back(arr_copy);
-    wt = arr_copy;
+    wt = contiguous_copy_gpu(wt, s);
+    copies.push_back(wt);
   }
 
   // 3D conv