Improve the ring backend initialization

2025-12-16 01:49:05 +08:00 · 2025-07-11 15:31:28 -07:00
193 changed files with 2873 additions and 7515 deletions
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -7,9 +7,18 @@ 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:
@@ -64,9 +73,9 @@ jobs:
                 git push -f origin gh-pages

  linux_build_and_test:
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
+    docker:
+      - image: cimg/python:3.9
+
    steps:
      - checkout
      - run:
@@ -78,36 +87,35 @@ jobs:
      - run:
          name: Install dependencies
          command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
+            pip install --upgrade cmake
+            pip install nanobind==2.4.0
+            pip install numpy
            sudo apt-get update
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-            curl -LsSf https://astral.sh/uv/install.sh | sh
      - run:
          name: Install Python package
          command: |
-            uv venv
-            uv pip install cmake
-            uv pip install -e ".[dev]" -v
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+              python3 setup.py build_ext --inplace
+            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
+              python3 setup.py develop
      - run:
          name: Generate package stubs
          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
+            echo "stubs"
+            pip install typing_extensions
+            python setup.py generate_stubs
      - run:
          name: Run Python tests
          command: |
-            source .venv/bin/activate
-            python -m unittest discover python/tests -v
+            python3 -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 -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
      - run:
          name: Build CPP only
          command: |
-            source .venv/bin/activate
-            mkdir -p build && cd build
+            mkdir -p build && cd build 
            cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
            make -j `nproc`
      - run:
@@ -132,49 +140,50 @@ jobs:
      - run:
          name: Install dependencies
          command: |
-            HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
-              brew install openmpi uv
+            brew install python@3.9
+            brew install openmpi
+            python3.9 -m venv env
+            source env/bin/activate
+            pip install --upgrade pip
+            pip install --upgrade cmake
+            pip install nanobind==2.4.0
+            pip install numpy
+            pip install torch
+            pip install tensorflow
+            pip install unittest-xml-reporting
      - run:
          name: Install Python package
          command: |
-            uv venv --python 3.9
-            uv pip install \
-              nanobind==2.4.0 \
-              cmake \
-              numpy \
-              torch \
-              tensorflow \
-              unittest-xml-reporting
+            source env/bin/activate
            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e . -v
+              pip install -e . -v
      - run:
          name: Generate package stubs
          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
+            source env/bin/activate
+            pip install typing_extensions
+            python setup.py generate_stubs
      - run:
          name: Run Python tests
          command: |
-            source .venv/bin/activate
+            source env/bin/activate
            LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
            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 -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
      - run:
          name: Build example extension
          command: |
-            source .venv/bin/activate
+            source env/bin/activate
            cd examples/extensions
-            uv pip install -r requirements.txt
-            uv run --no-project setup.py build_ext --inplace
-            uv run --no-project python test.py
+            pip install -r requirements.txt
+            python setup.py build_ext -j8
      - store_test_results:
          path: test-results
      - run:
          name: Build CPP only
          command: |
-            source .venv/bin/activate
+            source env/bin/activate
            mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
      - run:
          name: Run CPP tests
@@ -183,7 +192,7 @@ jobs:
      - run:
          name: Build small binary
          command: |
-            source .venv/bin/activate
+            source env/bin/activate
            cd build/
            cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
              -DBUILD_SHARED_LIBS=ON \
@@ -195,60 +204,34 @@ jobs:
      - run:
          name: Run Python tests with JIT
          command: |
+            source env/bin/activate
            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e .
+              pip install -e . -v
            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
              METAL_DEBUG_ERROR_MODE=0 \
-              uv run --no-project python -m xmlrunner discover \
-                -v python/tests \
-                -o test-results/gpu_jit
+              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:<< parameters.image_date >>"
+      image: linux-cuda-12:default
      resource_class: gpu.nvidia.small.gen2
    steps:
      - checkout
-      - restore_cache:
-          keys:
-            - cuda-<< parameters.image_date >>-{{ arch }}-
-      - run:
-          name: Install dependencies
-          command: |
-            sudo apt-get update
-            sudo apt-get install libcudnn9-dev-cuda-12
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
-            sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
-            rm -rf ccache-4.11.3-linux-x86_64
-            curl -LsSf https://astral.sh/uv/install.sh | sh
      - run:
          name: Install Python package
          command: |
-            uv venv
+            sudo apt-get update
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            python -m venv env
+            source env/bin/activate
            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              uv pip install -e ".[dev]" -v
+              pip install -e ".[dev]"
      - run:
          name: Run Python tests
          command: |
-            source .venv/bin/activate
+            source env/bin/activate
            LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
            LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
-      - run:
-          name: CCache report
-          command: |
-            ccache --show-stats
-            ccache --zero-stats
-            ccache --max-size 400MB
-            ccache --cleanup
-      - save_cache:
-          key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
-          paths:
-            - /home/circleci/.cache/ccache

  build_release:
    parameters:
@@ -301,18 +284,7 @@ jobs:
          name: Build Python package
          command: |
            source env/bin/activate
-            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
+            << parameters.build_env >> python -m build -w
      - when:
          condition: << parameters.build_env >>
          steps:
@@ -329,100 +301,88 @@ jobs:
      python_version:
        type: string
        default: "3.9"
-      build_env:
+      extra_env:
        type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
+        default: "DEV_RELEASE=1"
+    docker:
+      - image: ubuntu:20.04
    steps:
      - checkout
      - run:
          name: Build wheel
          command: |
            PYTHON=python<< parameters.python_version >>
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            TZ=Etc/UTC sudo apt-get -y install tzdata
-            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
+            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
            $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.build_env >> pip install ".[dev]" -v
+            << parameters.extra_env >> pip install . -v
            pip install typing_extensions
            python setup.py generate_stubs
-            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
+            << parameters.extra_env >> 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/*
      - store_artifacts:
          path: wheelhouse/

  build_cuda_release:
    parameters:
-      build_env:
+      python_version:
        type: string
-        default: ""
+        default: "3.9"
+      extra_env:
+        type: string
+        default: "DEV_RELEASE=1"
    machine:
-      image: ubuntu-2204:current
-      resource_class: large
+      image: linux-cuda-12:default
+      resource_class: gpu.nvidia.small.gen2
    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
-            sudo apt-get install zip
+            python -m venv env
+            source env/bin/activate
            pip install auditwheel
            pip install patchelf
            pip install build
            pip install twine
-            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 \
+            << parameters.extra_env >> \
              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              python -m build -w
+              pip install ".[dev]" -v
+            python setup.py generate_stubs
+            << parameters.extra_env >> \
+              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              python -m build --wheel
            bash python/scripts/repair_cuda.sh
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  twine upload wheelhouse/*.whl
+      - run:
+          name: Upload package
+          command: |
+            source env/bin/activate
+            twine upload wheelhouse/*.whl
      - store_artifacts:
          path: wheelhouse/

@@ -434,6 +394,7 @@ 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:
@@ -441,16 +402,14 @@ workflows:
            parameters:
              macosx_deployment_target: ["13.5", "14.0"]
      - linux_build_and_test
-      - cuda_build_and_test:
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
+      - cuda_build_and_test 
      - 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:
@@ -542,16 +501,7 @@ workflows:
          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"]
+              extra_env: ["PYPI_RELEASE=1"]

  prb:
    when:
@@ -572,9 +522,6 @@ workflows:
          requires: [ hold ]
      - cuda_build_and_test:
          requires: [ hold ]
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
  nightly_build:
    when:
      and:
@@ -633,17 +580,11 @@ workflows:
              - macosx_deployment_target: "15.0"
                xcode_version: "15.0.0"
                python_version: "3.13"
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-      - build_cuda_release
-
-  build_dev_release:
+  weekly_build:
    when:
      and:
        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.test_release >>
+        - << pipeline.parameters.weekly_build >>
    jobs:
      - build_release:
          matrix:
@@ -713,12 +654,25 @@ 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"]
-              build_env: ["DEV_RELEASE=1"]
+              extra_env: ["PYPI_RELEASE=1"]
+  cuda_test_release:
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.cuda_release >>
+    jobs:
      - build_cuda_release:
          matrix:
            parameters:
-              build_env: ["DEV_RELEASE=1"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              extra_env: ["PYPI_RELEASE=1"]
--- a/ACKNOWLEDGMENTS.md
+++ b/ACKNOWLEDGMENTS.md
@@ -19,7 +19,6 @@ 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" />
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -41,9 +41,7 @@ option(MLX_BUILD_GGUF "Include support for GGUF format" ON)
 option(MLX_BUILD_SAFETENSORS "Include support for safetensors format" ON)
 option(MLX_BUILD_BLAS_FROM_SOURCE "Build OpenBLAS from source code" OFF)
 option(MLX_METAL_JIT "Use JIT compilation for Metal kernels" OFF)
-option(MLX_USE_CCACHE "Use CCache for compilation cache when available" ON)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
-option(USE_SYSTEM_FMT "Use system's provided fmt library" OFF)

 # --------------------- Processor tests -------------------------
 message(
@@ -66,17 +64,10 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
      message(WARNING "Building for x86_64 arch is not officially supported.")
    endif()
  endif()
+
 else()
  set(MLX_BUILD_METAL OFF)
-endif()
-
-if(MLX_USE_CCACHE)
-  find_program(CCACHE_PROGRAM ccache)
-  if(CCACHE_PROGRAM)
-    set(CMAKE_C_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
-    set(CMAKE_CXX_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
-    set(CMAKE_CUDA_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
-  endif()
+  message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
 endif()

 # ----------------------------- Lib -----------------------------
@@ -243,16 +234,12 @@ target_include_directories(
 # Do not add mlx_EXPORTS define for shared library.
 set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")

-if(USE_SYSTEM_FMT)
-  find_package(fmt REQUIRED)
-else()
-  FetchContent_Declare(
-    fmt
-    GIT_REPOSITORY https://github.com/fmtlib/fmt.git
-    GIT_TAG 10.2.1
-    EXCLUDE_FROM_ALL)
-  FetchContent_MakeAvailable(fmt)
-endif()
+FetchContent_Declare(
+  fmt
+  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+  GIT_TAG 10.2.1
+  EXCLUDE_FROM_ALL)
+FetchContent_MakeAvailable(fmt)
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:fmt::fmt-header-only>)

 if(MLX_BUILD_PYTHON_BINDINGS)
--- a/README.md
+++ b/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.

@@ -68,23 +68,18 @@ in the documentation.

 ## Installation

-MLX is available on [PyPI](https://pypi.org/project/mlx/). To install MLX on
-macOS, run:
+MLX is available on [PyPI](https://pypi.org/project/mlx/). To install the Python API, run:

-```bash
+**With `pip`**:
+
+```
 pip install mlx
 ```

-To install the CUDA backend on Linux, run:
+**With `conda`**:

-```bash
-pip install mlx[cuda]
 ```
-
-To install a CPU-only Linux package, run:
-
-```bash
-pip install mlx[cpu]
+conda install -c conda-forge mlx
 ```

 Checkout the
--- a/docs/src/dev/extensions.rst
+++ b/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.
        */
-        std::pair<std::vector<array>, std::vector<int>> vmap(
+        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
            const std::vector<array>& inputs,
            const std::vector<int>& axes) override;

-        /** The name of primitive. */
-        const char* name() const override {
-          return "Axpby";
+        /** Print the primitive. */
+        void print(std::ostream& os) override {
+            os << "Axpby";
        }

        /** Equivalence check **/
@@ -394,14 +394,14 @@ below.
        out.set_data(allocator::malloc(out.nbytes()));

        // Resolve name of kernel
-        std::stream kname;
-        kname = "axpby_general_" + type_to_name(out);
+        std::ostringstream kname;
+        kname << "axpby_" << "general_" << type_to_name(out);

        // Load the metal library
-        auto lib = d.get_library("mlx_ext", current_binary_dir());
+        auto lib = d.get_library("mlx_ext");

        // Make a kernel from this metal library
-        auto kernel = d.get_kernel(kname, lib);
+        auto kernel = d.get_kernel(kname.str(), lib);

        // Prepare to encode kernel
        auto& compute_encoder = d.get_command_encoder(s.index);
--- a/docs/src/install.rst
+++ b/docs/src/install.rst
@@ -13,7 +13,7 @@ silicon computer is

    pip install mlx

-To install from PyPI your system must meet the following requirements:
+To install from PyPI you must meet the following requirements:

 - Using an M series chip (Apple silicon)
 - Using a native Python >= 3.9
@@ -23,39 +23,22 @@ To install from PyPI your system 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 install with:
+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:

 .. code-block:: shell

-    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
+    pip install mlx-cuda


 Troubleshooting
--- a/docs/src/python/optimizers/common_optimizers.rst
+++ b/docs/src/python/optimizers/common_optimizers.rst
@@ -19,4 +19,3 @@ Common Optimizers
   Adamax
   Lion
   MultiOptimizer
-   Muon
--- a/examples/extensions/axpby/axpby.cpp
+++ b/examples/extensions/axpby/axpby.cpp
@@ -1,6 +1,5 @@
 // Copyright © 2023-2025 Apple Inc.

-#include <dlfcn.h>
 #include <iostream>
 #include <sstream>

@@ -17,19 +16,6 @@

 namespace my_ext {

-// A helper function to find the location of the current binary on disk.
-// The Metal library ("mlx_ext.mtllib"), should be in the same directory.
-std::string current_binary_dir() {
-  static std::string 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().string();
-  }();
-  return binary_dir;
-}
-
 ///////////////////////////////////////////////////////////////////////////////
 // Operation Implementation
 ///////////////////////////////////////////////////////////////////////////////
@@ -181,15 +167,16 @@ void Axpby::eval_gpu(
  }

  // Resolve name of kernel (corresponds to axpby.metal)
-  std::string kname = "axpby_";
-  kname += (contiguous_kernel ? "contiguous_" : "general_");
-  kname += type_to_name(out);
+  std::ostringstream kname;
+  kname << "axpby_";
+  kname << (contiguous_kernel ? "contiguous_" : "general_");
+  kname << type_to_name(out);

  // Load the metal library
-  auto lib = d.get_library("mlx_ext", current_binary_dir());
+  auto lib = d.get_library("mlx_ext");

  // Make a kernel from this metal library
-  auto kernel = d.get_kernel(kname, lib);
+  auto kernel = d.get_kernel(kname.str(), lib);

  // Prepare to encode kernel
  auto& compute_encoder = d.get_command_encoder(s.index);
--- a/examples/extensions/axpby/axpby.h
+++ b/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;

-  /** The name of primitive. */
-  const char* name() const override {
-    return "Axpby";
+  /** Print the primitive. */
+  void print(std::ostream& os) override {
+    os << "Axpby";
  }

  /** Equivalence check **/
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,4 +1,4 @@
 setuptools>=42
 cmake>=3.25
 mlx>=0.21.0
-nanobind==2.4.0
+nanobind==2.2.0
--- a/examples/extensions/test.py
+++ b/examples/extensions/test.py
@@ -3,10 +3,8 @@ from mlx_sample_extensions import axpby

 a = mx.ones((3, 4))
 b = mx.ones((3, 4))
-c_cpu = axpby(a, b, 4.0, 2.0, stream=mx.cpu)
-c_gpu = axpby(a, b, 4.0, 2.0, stream=mx.gpu)
+c = axpby(a, b, 4.0, 2.0, stream=mx.cpu)

-print(f"c shape: {c_cpu.shape}")
-print(f"c dtype: {c_cpu.dtype}")
-print(f"c_cpu correct: {mx.all(c_cpu == 6.0).item()}")
-print(f"c_gpu correct: {mx.all(c_gpu == 6.0).item()}")
+print(f"c shape: {c.shape}")
+print(f"c dtype: {c.dtype}")
+print(f"c correct: {mx.all(c == 6.0).item()}")
--- a/mlx/array.h
+++ b/mlx/array.h
@@ -10,7 +10,6 @@
 #include "mlx/allocator.h"
 #include "mlx/dtype.h"
 #include "mlx/event.h"
-#include "mlx/small_vector.h"

 namespace mlx::core {

@@ -19,8 +18,8 @@ class Primitive;

 using Deleter = std::function<void(allocator::Buffer)>;
 using ShapeElem = int32_t;
-using Shape = SmallVector<ShapeElem>;
-using Strides = SmallVector<int64_t>;
+using Shape = std::vector<ShapeElem>;
+using Strides = std::vector<int64_t>;

 class array {
  /* An array is really a node in a graph. It contains a shared ArrayDesc
--- a/mlx/backend/common/utils.cpp
+++ b/mlx/backend/common/utils.cpp
@@ -1,20 +1,14 @@
 // Copyright © 2023-2024 Apple Inc.

-#include <dlfcn.h>
-
 #include "mlx/backend/common/utils.h"
+#include "mlx/primitives.h"

 namespace mlx::core {

-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::string get_primitive_string(Primitive* primitive) {
+  std::ostringstream op_t;
+  primitive->print(op_t);
+  return op_t.str();
 }

 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
@@ -228,31 +222,4 @@ std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
      std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
 }

-array swapaxes_in_eval(const array& x, int axis1, int axis2) {
-  int ndim = x.ndim();
-  if (axis1 < 0) {
-    axis1 += ndim;
-  }
-  if (axis2 < 0) {
-    axis2 += ndim;
-  }
-
-  auto shape = x.shape();
-  std::swap(shape[axis1], shape[axis2]);
-  auto strides = x.strides();
-  std::swap(strides[axis1], strides[axis2]);
-
-  auto [data_size, row_contiguous, col_contiguous] =
-      check_contiguity(shape, strides);
-  bool contiguous = data_size == x.data_size();
-
-  array out(std::move(shape), x.dtype(), nullptr, {});
-  out.copy_shared_buffer(
-      x,
-      std::move(strides),
-      {contiguous, row_contiguous, col_contiguous},
-      x.data_size());
-  return out;
-}
-
 } // namespace mlx::core
--- a/mlx/backend/common/utils.h
+++ b/mlx/backend/common/utils.h
@@ -2,7 +2,6 @@

 #pragma once

-#include <filesystem>
 #include <tuple>
 #include <vector>

@@ -10,8 +9,7 @@

 namespace mlx::core {

-// Return the directory that contains current shared library.
-std::filesystem::path current_binary_dir();
+std::string get_primitive_string(Primitive* primitive);

 inline int64_t
 elem_to_loc(int elem, const Shape& shape, const Strides& strides) {
@@ -196,11 +194,8 @@ void shared_buffer_reshape(
    const Strides& out_strides,
    array& out);

-// Like the swapaxes op but safe to call in eval_gpu.
-array swapaxes_in_eval(const array& x, int axis1, int axis2);
-
 template <typename T>
-inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
+inline std::vector<T> remove_index(std::vector<T> vec, size_t index) {
  vec.erase(std::next(vec.begin(), index));
  return vec;
 }
--- a/mlx/backend/cpu/cholesky.cpp
+++ b/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_cpu(
+  copy(
      a,
      factor,
      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
--- a/mlx/backend/cpu/compiled.cpp
+++ b/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 {
-      os << x.primitive().name();
+      x.primitive().print(os);
      os << "()(";
      for (int i = 0; i < x.inputs().size() - 1; i++) {
        os << "tmp_" << namer.get_name(x.inputs()[i]) << ", ";
@@ -288,14 +288,6 @@ void Compiled::eval_cpu(
  auto [contiguous, shape, strides] =
      compiled_collapse_contiguous_dims(inputs, outputs[0], is_constant_);

-  // Force allocating shape/strides on heap so we can take their data() first
-  // and then std::move them.
-  // TODO: Refactor code to avoid heap allocation.
-  shape.grow();
-  for (auto& s : strides) {
-    s.grow();
-  }
-
  // Collect function input arguments.
  std::vector<void*> args;
  int strides_index = 1;
--- a/mlx/backend/cpu/conv.cpp
+++ b/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_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy(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_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_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_cpu(in_strided_view, in_strided, CopyType::General, stream);
+  copy(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_cpu(wt_transpose, gemm_wt, CopyType::General, stream);
+    copy(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_cpu(wt, gemm_wt, ctype, stream);
+    copy(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_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_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_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy(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_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_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_cpu(in_strided_view, in_strided, CopyType::General, stream);
+  copy(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_cpu(wt, gemm_wt, ctype, stream);
+    copy(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_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_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_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy(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_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_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_cpu(in_strided_view, in_strided, CopyType::General, stream);
+  copy(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_cpu(wt, gemm_wt, ctype, stream);
+    copy(wt, gemm_wt, ctype, stream);
    temps.push_back(gemm_wt);
  }

  if (flip) {
    auto gemm_wt_ = array(gemm_wt.shape(), float32, nullptr, {});
-    copy_cpu(gemm_wt, gemm_wt_, CopyType::Vector, stream);
+    copy(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_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_inplace(gemm_out, out, CopyType::Vector, stream);
  }
  encoder.add_temporaries(std::move(temps));
 }
--- a/mlx/backend/cpu/copy.cpp
+++ b/mlx/backend/cpu/copy.cpp
@@ -295,11 +295,7 @@ inline void copy_inplace_dispatch(

 } // namespace

-void copy_cpu_inplace(
-    const array& src,
-    array& dst,
-    CopyType ctype,
-    Stream stream) {
+void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
  auto& encoder = cpu::get_command_encoder(stream);
  encoder.set_input_array(src);
  encoder.set_output_array(dst);
@@ -309,7 +305,7 @@ void copy_cpu_inplace(
       ctype]() mutable { copy_inplace_dispatch(src, dst, ctype); });
 }

-void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
  bool donated = set_copy_output_data(src, dst, ctype);
  if (donated && src.dtype() == dst.dtype()) {
    // If the output has the same type as the input then there is nothing to
@@ -319,10 +315,10 @@ void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream) {
  if (ctype == CopyType::GeneralGeneral) {
    ctype = CopyType::General;
  }
-  copy_cpu_inplace(src, dst, ctype, stream);
+  copy_inplace(src, dst, ctype, stream);
 }

-void copy_cpu_inplace(
+void copy_inplace(
    const array& src,
    array& dst,
    const Shape& data_shape,
@@ -377,10 +373,4 @@ void copy_cpu_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
--- a/mlx/backend/cpu/copy.h
+++ b/mlx/backend/cpu/copy.h
@@ -10,14 +10,10 @@

 namespace mlx::core {

-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(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream);

-void copy_cpu_inplace(
+void copy_inplace(
    const array& src,
    array& dst,
    const Shape& data_shape,
@@ -30,7 +26,4 @@ void copy_cpu_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
--- a/mlx/backend/cpu/distributed.cpp
+++ b/mlx/backend/cpu/distributed.cpp
@@ -13,7 +13,9 @@ std::pair<array, bool> ensure_row_contiguous(const array& arr, Stream stream) {
  if (arr.flags().row_contiguous) {
    return {arr, false};
  } else {
-    return {contiguous_copy_cpu(arr, stream), true};
+    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+    copy(arr, arr_copy, CopyType::General, stream);
+    return {arr_copy, true};
  }
 };

@@ -32,7 +34,8 @@ void AllReduce::eval_cpu(
      }
      return in;
    } else {
-      array arr_copy = contiguous_copy_cpu(in, s);
+      array arr_copy(in.shape(), in.dtype(), nullptr, {});
+      copy(in, arr_copy, CopyType::General, s);
      out.copy_shared_buffer(arr_copy);
      return arr_copy;
    }
--- a/mlx/backend/cpu/eig.cpp
+++ b/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_cpu(
+  copy(
      a,
      a_copy,
      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
--- a/mlx/backend/cpu/eigh.cpp
+++ b/mlx/backend/cpu/eigh.cpp
@@ -196,7 +196,7 @@ void Eigh::eval_cpu(

  values.set_data(allocator::malloc(values.nbytes()));

-  copy_cpu(
+  copy(
      a,
      vectors,
      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
--- a/mlx/backend/cpu/hadamard.cpp
+++ b/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_cpu(
+    copy(
        in,
        out,
        in.flags().row_contiguous ? CopyType::Vector : CopyType::General,
--- a/mlx/backend/cpu/indexing.cpp
+++ b/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_cpu(src, out, ctype, stream());
+  copy(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_cpu(src, out, ctype, stream());
+  copy(src, out, ctype, stream());

  auto& encoder = cpu::get_command_encoder(stream());
  encoder.set_input_array(idx);
--- a/mlx/backend/cpu/inverse.cpp
+++ b/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_cpu(
+  copy(
      a,
      inv,
      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
--- a/mlx/backend/cpu/jit_compiler.cpp
+++ b/mlx/backend/cpu/jit_compiler.cpp
@@ -2,7 +2,6 @@

 #include "mlx/backend/cpu/jit_compiler.h"

-#include <algorithm>
 #include <sstream>
 #include <vector>

--- a/mlx/backend/cpu/logsumexp.cpp
+++ b/mlx/backend/cpu/logsumexp.cpp
@@ -87,7 +87,8 @@ void LogSumExp::eval_cpu(const std::vector<array>& inputs, array& out) {
    if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
      return x;
    } else {
-      array x_copy = contiguous_copy_cpu(x, s);
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy(x, x_copy, CopyType::General, s);
      encoder.add_temporary(x_copy);
      return x_copy;
    }
--- a/mlx/backend/cpu/luf.cpp
+++ b/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_cpu_inplace(
+  copy_inplace(
      a,
      lu,
      a.shape(),
--- a/mlx/backend/cpu/masked_mm.cpp
+++ b/mlx/backend/cpu/masked_mm.cpp
@@ -124,20 +124,21 @@ 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_cpu(arr, arr_copy, CopyType::Vector, s);
+            copy(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_cpu(arr, arr_copy, CopyType::Vector, s);
+            copy(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);
        }
      };
@@ -385,7 +386,7 @@ 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_cpu(arr, temps.back(), CopyType::General, s);
+      copy(arr, temps.back(), CopyType::General, s);
      int64_t stx = arr.shape(-1);
      return std::make_tuple(false, stx, temps.back());
    }
@@ -503,7 +504,7 @@ void SegmentedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
      return std::make_tuple(true, sty, x);
    } else {
      array xc(x.shape(), x.dtype(), nullptr, {});
-      copy_cpu(x, xc, CopyType::General, s);
+      copy(x, xc, CopyType::General, s);
      encoder.add_temporary(xc);
      int64_t stx = x.shape(-1);
      return std::make_tuple(false, stx, xc);
--- a/mlx/backend/cpu/matmul.cpp
+++ b/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_cpu(arr, temps.back(), CopyType::General, stream);
+      copy(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_cpu(c, out, ctype, stream());
+  copy(c, out, ctype, stream());
  if (inputs[0].shape(-1) == 0) {
    return;
  }
--- a/mlx/backend/cpu/primitives.cpp
+++ b/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_cpu_inplace(in, out, CopyType::General, out.primitive().stream());
+    copy_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_cpu(in, out, ctype, stream());
+  copy(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_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
+    copy_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_cpu(in, out, CopyType::General, stream());
+    copy(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_cpu(in, out, ctype, stream());
+  copy(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_cpu(val, out, CopyType::Scalar, stream());
+  copy(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_cpu_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
+  copy_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_cpu_inplace(
+  copy_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_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());

  auto [out_offset, donated] =
      compute_dynamic_offset(inputs[2], out.strides(), axes_, stream());
-  copy_cpu_inplace(
+  copy_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_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());

  // Calculate out strides, initial offset and if copy needs to be made
  auto [data_offset, out_strides] =
      prepare_slice(out, start_indices_, strides_);

  // Do copy
-  copy_cpu_inplace(
+  copy_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_cpu_inplace(in_tmp, tmp, CopyType::General, stream());
+      copy_inplace(in_tmp, tmp, CopyType::General, stream());
    } else {
-      copy_cpu_inplace(in, tmp, CopyType::General, stream());
+      copy_inplace(in, tmp, CopyType::General, stream());
    }

    auto flags = out.flags();
--- a/mlx/backend/cpu/qrf.cpp
+++ b/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_cpu_inplace(a, in, CopyType::GeneralGeneral, stream);
+  copy_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()));
--- a/mlx/backend/cpu/quantized.cpp
+++ b/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_cpu(arr, temps.back(), CopyType::General, s);
+      copy(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_cpu(arr, temps.back(), CopyType::General, s);
+      copy(arr, temps.back(), CopyType::General, s);
      return temps.back();
    }
  };
@@ -712,7 +712,9 @@ void fast::AffineQuantize::eval_cpu(
    if (arr.flags().row_contiguous) {
      return std::make_pair(arr, false);
    } else {
-      return std::make_pair(contiguous_copy_cpu(arr, s), true);
+      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+      copy(arr, arr_copy, CopyType::General, s);
+      return std::make_pair(arr_copy, true);
    }
  };

--- a/mlx/backend/cpu/scan.cpp
+++ b/mlx/backend/cpu/scan.cpp
@@ -250,8 +250,10 @@ void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
  // Ensure contiguity
  auto in = inputs[0];
  if (!in.flags().row_contiguous) {
-    in = contiguous_copy_cpu(in, stream());
-    encoder.add_temporary(in);
+    array arr_copy(in.shape(), in.dtype(), nullptr, {});
+    copy(in, arr_copy, CopyType::General, stream());
+    in = arr_copy;
+    encoder.add_temporary(arr_copy);
  }
  out.set_data(allocator::malloc(out.nbytes()));

--- a/mlx/backend/cpu/softmax.cpp
+++ b/mlx/backend/cpu/softmax.cpp
@@ -131,7 +131,8 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
      }
      return x;
    } else {
-      array x_copy = contiguous_copy_cpu(x, s);
+      array x_copy(x.shape(), x.dtype(), nullptr, {});
+      copy(x, x_copy, CopyType::General, s);
      out.copy_shared_buffer(x_copy);
      return x_copy;
    }
--- a/mlx/backend/cpu/sort.cpp
+++ b/mlx/backend/cpu/sort.cpp
@@ -8,7 +8,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
-#include "mlx/dtype_utils.h"
+
 #include "mlx/primitives.h"

 namespace mlx::core {
@@ -333,24 +333,45 @@ void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 1);
  auto& in = inputs[0];

-  int axis = axis_;
-  if (axis < 0) {
-    axis += in.ndim();
-  }
-
  // Copy input to output
-  CopyType ctype = (in.flags().contiguous && in.strides()[axis] != 0)
-      ? CopyType::Vector
-      : CopyType::General;
-  copy_cpu(in, out, ctype, stream());
+  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
+  copy(in, out, ctype, stream());

  auto& encoder = cpu::get_command_encoder(stream());
  encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out), axis]() mutable {
-    dispatch_all_types(out.dtype(), [&](auto type_tag) {
-      sort<MLX_GET_TYPE(type_tag)>(out, axis);
-    });
-  });
+  encoder.dispatch(
+      [out = array::unsafe_weak_copy(out), axis_ = axis_]() mutable {
+        switch (out.dtype()) {
+          case bool_:
+            return sort<bool>(out, axis_);
+          case uint8:
+            return sort<uint8_t>(out, axis_);
+          case uint16:
+            return sort<uint16_t>(out, axis_);
+          case uint32:
+            return sort<uint32_t>(out, axis_);
+          case uint64:
+            return sort<uint64_t>(out, axis_);
+          case int8:
+            return sort<int8_t>(out, axis_);
+          case int16:
+            return sort<int16_t>(out, axis_);
+          case int32:
+            return sort<int32_t>(out, axis_);
+          case int64:
+            return sort<int64_t>(out, axis_);
+          case float32:
+            return sort<float>(out, axis_);
+          case float64:
+            return sort<double>(out, axis_);
+          case float16:
+            return sort<float16_t>(out, axis_);
+          case bfloat16:
+            return sort<bfloat16_t>(out, axis_);
+          case complex64:
+            return sort<complex64_t>(out, axis_);
+        }
+      });
 }

 void ArgPartition::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -405,10 +426,8 @@ void Partition::eval_cpu(const std::vector<array>& inputs, array& out) {
  auto& in = inputs[0];

  // Copy input to output
-  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
-      ? CopyType::Vector
-      : CopyType::General;
-  copy_cpu(in, out, ctype, stream());
+  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
+  copy(in, out, ctype, stream());

  auto& encoder = cpu::get_command_encoder(stream());
  encoder.set_output_array(out);
--- a/mlx/backend/cpu/svd.cpp
+++ b/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_cpu(
+  copy(
      a,
      in,
      a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
--- a/mlx/backend/cuda/CMakeLists.txt
+++ b/mlx/backend/cuda/CMakeLists.txt
@@ -6,7 +6,6 @@
 target_sources(
  mlx
  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/arange.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
@@ -16,21 +15,18 @@ 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
          ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/layer_norm.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/logsumexp.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/random.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/all_reduce.cu
@@ -39,7 +35,6 @@ 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}/scaled_dot_product_attention.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
@@ -47,18 +42,8 @@ target_sources(
          ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cpp
          ${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.
@@ -101,18 +86,11 @@ 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 >= 7.0 is required for synchronization between CPU/GPU with
-# managed memory.
-if(NOT DEFINED MLX_CUDA_ARCHITECTURES)
-  set(MLX_CUDA_ARCHITECTURES "native")
-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
+    "70;80"
+    CACHE STRING "CUDA architectures")
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
                                     "${MLX_CUDA_ARCHITECTURES}")
@@ -144,27 +122,6 @@ 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)
--- a/mlx/backend/cuda/allocator.cpp
+++ b/mlx/backend/cuda/allocator.cpp
@@ -2,6 +2,7 @@

 #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>
@@ -16,66 +17,14 @@ 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); }) {
+          [this](CudaBuffer* buf) {
+            cuda_free(buf->data);
+            delete buf;
+          }) {
  // TODO: Set memory limit for multi-device.
  size_t free, total;
  CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
@@ -87,9 +36,7 @@ Buffer CudaAllocator::malloc(size_t size) {
  // Find available buffer from cache.
  auto orig_size = size;
  std::unique_lock lock(mutex_);
-  if (size <= small_block_size) {
-    size = 8;
-  } else if (size < page_size) {
+  if (size < page_size) {
    size = next_power_of_2(size);
  } else {
    size = page_size * ((size + page_size - 1) / page_size);
@@ -97,25 +44,19 @@ Buffer CudaAllocator::malloc(size_t size) {

  CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
  if (!buf) {
-    // 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);
+    // 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_);
    }

-    // Try the scalar pool first
-    if (size <= small_block_size) {
-      buf = scalar_pool_.malloc();
-    }
    lock.unlock();
-    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)));
-      }
+    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();
  }
@@ -126,6 +67,7 @@ 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};
 }

@@ -140,7 +82,9 @@ void CudaAllocator::free(Buffer buffer) {
  if (get_cache_memory() < max_pool_size_) {
    buffer_cache_.recycle_to_cache(buf);
  } else {
-    cuda_free(buf);
+    lock.unlock();
+    cuda_free(buf->data);
+    delete buf;
  }
 }

@@ -152,14 +96,27 @@ size_t CudaAllocator::size(Buffer buffer) const {
  return buf->size;
 }

-// 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;
+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;
+    }
  }
+  cudaFree(buf);
 }

 size_t CudaAllocator::get_active_memory() const {
--- a/mlx/backend/cuda/allocator.h
+++ b/mlx/backend/cuda/allocator.h
@@ -7,10 +7,13 @@

 #include <mutex>
 #include <set>
+#include <thread>
 #include <utility>

 namespace mlx::core::cu {

+class Worker;
+
 using allocator::Buffer;

 // Stores cuda-managed unified memory.
@@ -19,35 +22,21 @@ 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();
@@ -58,18 +47,19 @@ 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();
--- a/mlx/backend/cuda/arange.cu
+++ b/mlx/backend/cuda/arange.cu
@@ -1,55 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/fp16_math.cuh"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-#include "mlx/primitives.h"
-
-#include <nvtx3/nvtx3.hpp>
-#include <thrust/device_ptr.h>
-#include <thrust/transform.h>
-
-namespace mlx::core {
-
-namespace cu {
-
-template <typename T>
-struct Arange {
-  const T start;
-  const T step;
-
-  __device__ T operator()(uint32_t i) const {
-    return start + i * step;
-  }
-};
-
-} // namespace cu
-
-void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Arange::eval_gpu");
-  if (out.size() == 0) {
-    return;
-  }
-  out.set_data(allocator::malloc(out.nbytes()));
-
-  auto& encoder = cu::get_command_encoder(stream());
-  encoder.set_output_array(out);
-
-  auto capture = encoder.capture_context();
-  dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
-    using CTYPE = MLX_GET_TYPE(type_tag);
-    using OutType = cuda_type_t<CTYPE>;
-    CTYPE step =
-        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
-    thrust::transform(
-        cu::thrust_policy(encoder.stream()),
-        thrust::counting_iterator<uint32_t>(0),
-        thrust::counting_iterator<uint32_t>(out.data_size()),
-        thrust::device_pointer_cast(out.data<OutType>()),
-        cu::Arange<OutType>{
-            static_cast<OutType>(start_), static_cast<OutType>(step)});
-  });
-}
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/arg_reduce.cu
+++ b/mlx/backend/cuda/arg_reduce.cu
@@ -1,8 +1,8 @@
 // Copyright © 2025 Apple Inc.
-
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/device/fp16_math.cuh"
+#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -44,11 +44,8 @@ struct ArgMin {
  }

  template <int N>
-  __device__ IndexValPair<T> reduce_many(
-      IndexValPair<T> best,
-      const AlignedVector<T, N>& vals,
-      uint32_t offset) {
-#pragma unroll
+  __device__ IndexValPair<T>
+  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
    for (int i = 0; i < N; i++) {
      if (vals[i] < best.val) {
        best.val = vals[i];
@@ -77,11 +74,8 @@ struct ArgMax {
  }

  template <int N>
-  __device__ IndexValPair<T> reduce_many(
-      IndexValPair<T> best,
-      const AlignedVector<T, N>& vals,
-      uint32_t offset) {
-#pragma unroll
+  __device__ IndexValPair<T>
+  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
    for (int i = 0; i < N; i++) {
      if (vals[i] > best.val) {
        best.val = vals[i];
@@ -112,15 +106,16 @@ __global__ void arg_reduce_general(

  int64_t in_idx = elem_to_loc(index, shape.data(), in_strides.data(), ndim);
  int64_t out_idx = elem_to_loc(index, shape.data(), out_strides.data(), ndim);
-  in += in_idx;

  Op op;
  T init = op.init();
  IndexValPair<T> best{0, init};

  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    T vals[N_READS];
    auto tid = r * BLOCK_DIM + block.thread_index().x;
-    auto vals = load_vector<N_READS>(in, tid, axis_size, axis_stride, init);
+    cub::LoadDirectBlocked(
+        tid, strided_iterator(in + in_idx, axis_stride), vals, axis_size, init);
    best = op.reduce_many(best, vals, tid * N_READS);
  }

@@ -171,7 +166,6 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
          kernel,
          num_blocks,
          block_dim(),
-          0,
          in.data<T>(),
          out.data<uint32_t>(),
          out.size(),
--- a/mlx/backend/cuda/binary.cu
+++ b/mlx/backend/cuda/binary.cu
@@ -3,6 +3,7 @@
 #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"
@@ -28,7 +29,7 @@ __global__ void binary_ss(const In* a, const In* b, Out* out, IdxT size) {
    AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      out_vec[i] = Op{}(a[0], b[0]);
+      out_vec.val[i] = Op{}(a[0], b[0]);
    }

    store_vector<N_READS>(out, index, out_vec);
@@ -49,7 +50,7 @@ __global__ void binary_sv(const In* a, const In* b, Out* out, IdxT size) {
    AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      out_vec[i] = Op{}(a[0], b_vec[i]);
+      out_vec.val[i] = Op{}(a[0], b_vec.val[i]);
    }

    store_vector<N_READS>(out, index, out_vec);
@@ -70,7 +71,7 @@ __global__ void binary_vs(const In* a, const In* b, Out* out, IdxT size) {
    AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      out_vec[i] = Op{}(a_vec[i], b[0]);
+      out_vec.val[i] = Op{}(a_vec.val[i], b[0]);
    }

    store_vector<N_READS>(out, index, out_vec);
@@ -92,7 +93,7 @@ __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
    AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      out_vec[i] = Op{}(a_vec[i], b_vec[i]);
+      out_vec.val[i] = Op{}(a_vec.val[i], b_vec.val[i]);
    }

    store_vector<N_READS>(out, index, out_vec);
@@ -128,7 +129,7 @@ __global__ void binary_g(
    int ndim) {
  IdxT index = cg::this_grid().thread_rank();
  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc(
+    auto [a_idx, b_idx] = elem_to_loc_4d(
        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
    out[index] = Op{}(a[a_idx], b[b_idx]);
  }
@@ -176,7 +177,7 @@ template <typename Op>
 void binary_op_gpu_inplace(
    const std::vector<array>& inputs,
    array& out,
-    const char* op,
+    std::string_view op,
    const Stream& s) {
  assert(inputs.size() > 1);
  const auto& a = inputs[0];
@@ -211,18 +212,14 @@ 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 [num_blocks, block_dims] =
-                        get_launch_args(out, large());
+                        get_launch_args(kernel, out, large());
                    encoder.add_kernel_node(
-                        cu::binary_g_nd<
-                            Op,
-                            InType,
-                            OutType,
-                            IdxT,
-                            dims_constant()>,
+                        kernel,
                        num_blocks,
                        block_dims,
-                        0,
                        a.data<InType>(),
                        b.data<InType>(),
                        out.data<OutType>(),
@@ -232,12 +229,13 @@ void binary_op_gpu_inplace(
                        const_param<dims_constant()>(b_strides));
                  });
                } else {
-                  auto [num_blocks, block_dims] = get_launch_args(out, large());
+                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
+                  auto [num_blocks, block_dims] =
+                      get_launch_args(kernel, out, large());
                  encoder.add_kernel_node(
-                      cu::binary_g<Op, InType, OutType, IdxT>,
+                      kernel,
                      num_blocks,
                      block_dims,
-                      0,
                      a.data<InType>(),
                      b.data<InType>(),
                      out.data<OutType>(),
@@ -251,7 +249,8 @@ void binary_op_gpu_inplace(
        } else {
          dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            constexpr int N_READS = 16 / sizeof(InType);
+            // 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, N_READS>;
@@ -261,12 +260,16 @@ void binary_op_gpu_inplace(
              kernel = cu::binary_vv<Op, InType, OutType, IdxT, N_READS>;
            }
            auto [num_blocks, block_dims] = get_launch_args(
-                out.data_size(), out.shape(), out.strides(), large(), N_READS);
+                kernel,
+                out.data_size(),
+                out.shape(),
+                out.strides(),
+                large(),
+                N_READS);
            encoder.add_kernel_node(
                kernel,
                num_blocks,
                block_dims,
-                0,
                a.data<InType>(),
                b.data<InType>(),
                out.data<OutType>(),
@@ -288,7 +291,7 @@ template <typename Op>
 void binary_op_gpu(
    const std::vector<array>& inputs,
    array& out,
-    const char* op,
+    std::string_view op,
    const Stream& s) {
  auto& a = inputs[0];
  auto& b = inputs[1];
@@ -297,11 +300,11 @@ 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, name(), 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); \
  }

 BINARY_GPU(Add)
@@ -325,31 +328,33 @@ 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, name(), s);
+    binary_op_gpu<cu::NaNEqual>(inputs, out, op, s);
  } else {
-    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
+    binary_op_gpu<cu::Equal>(inputs, out, op, 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, name(), s);
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, op, s);
      break;
    case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, op, s);
      break;
    case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, op, s);
      break;
    case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
+      binary_op_gpu<cu::LeftShift>(inputs, out, op, s);
      break;
    case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
+      binary_op_gpu<cu::RightShift>(inputs, out, op, s);
      break;
  }
 }
--- a/mlx/backend/cuda/binary_two.cu
+++ b/mlx/backend/cuda/binary_two.cu
@@ -3,6 +3,7 @@
 #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"
@@ -33,8 +34,8 @@ binary_two_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
      auto out = Op{}(a[0], b[0]);
-      out_a_vec[i] = out[0];
-      out_b_vec[i] = out[1];
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
    }

    store_vector<N_READS>(out_a, index, out_a_vec);
@@ -60,9 +61,9 @@ binary_two_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
    AlignedVector<Out, N_READS> out_b_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      auto out = Op{}(a[0], b_vec[i]);
-      out_a_vec[i] = out[0];
-      out_b_vec[i] = out[1];
+      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);
@@ -88,9 +89,9 @@ binary_two_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
    AlignedVector<Out, N_READS> out_b_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      auto out = Op{}(a_vec[i], b[0]);
-      out_a_vec[i] = out[0];
-      out_b_vec[i] = out[1];
+      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);
@@ -117,9 +118,9 @@ binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
    AlignedVector<Out, N_READS> out_b_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      auto out = Op{}(a_vec[i], b_vec[i]);
-      out_a_vec[i] = out[0];
-      out_b_vec[i] = out[1];
+      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);
@@ -160,7 +161,7 @@ __global__ void binary_two_g(
    int ndim) {
  IdxT index = cg::this_grid().thread_rank();
  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc(
+    auto [a_idx, b_idx] = elem_to_loc_4d(
        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
    auto out = Op{}(a[a_idx], b[b_idx]);
    out_a[index] = out[0];
@@ -183,7 +184,7 @@ template <typename Op>
 void binary_two_op_gpu_inplace(
    const std::vector<array>& inputs,
    std::vector<array>& outputs,
-    const char* op,
+    std::string_view op,
    const Stream& s) {
  assert(inputs.size() > 1);
  const auto& a = inputs[0];
@@ -227,18 +228,18 @@ void binary_two_op_gpu_inplace(
                int ndim = shape.size();
                if (ndim <= 3) {
                  dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                    auto kernel = cu::binary_two_g_nd<
+                        Op,
+                        InType,
+                        OutType,
+                        IdxT,
+                        dims_constant()>;
                    auto [num_blocks, block_dims] =
-                        get_launch_args(out_a, large());
+                        get_launch_args(kernel, out_a, large());
                    encoder.add_kernel_node(
-                        cu::binary_two_g_nd<
-                            Op,
-                            InType,
-                            OutType,
-                            IdxT,
-                            dims_constant()>,
+                        kernel,
                        num_blocks,
                        block_dims,
-                        0,
                        a.data<InType>(),
                        b.data<InType>(),
                        out_a.data<OutType>(),
@@ -249,13 +250,13 @@ void binary_two_op_gpu_inplace(
                        const_param<dims_constant()>(b_strides));
                  });
                } else {
+                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT>;
                  auto [num_blocks, block_dims] =
-                      get_launch_args(out_a, large());
+                      get_launch_args(kernel, out_a, large());
                  encoder.add_kernel_node(
-                      cu::binary_two_g<Op, InType, OutType, IdxT>,
+                      kernel,
                      num_blocks,
                      block_dims,
-                      0,
                      a.data<InType>(),
                      b.data<InType>(),
                      out_a.data<OutType>(),
@@ -270,7 +271,8 @@ void binary_two_op_gpu_inplace(
        } else {
          dispatch_bool(out_a.data_size() > UINT32_MAX, [&](auto large) {
            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            constexpr int N_READS = 16 / sizeof(InType);
+            // 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_two_sv<Op, InType, OutType, IdxT, N_READS>;
@@ -280,6 +282,7 @@ void binary_two_op_gpu_inplace(
              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(),
@@ -289,7 +292,6 @@ void binary_two_op_gpu_inplace(
                kernel,
                num_blocks,
                block_dims,
-                0,
                a.data<InType>(),
                b.data<InType>(),
                out_a.data<OutType>(),
@@ -312,7 +314,7 @@ template <typename Op>
 void binary_two_op_gpu(
    const std::vector<array>& inputs,
    std::vector<array>& outputs,
-    const char* op,
+    std::string_view op,
    const Stream& s) {
  auto& a = inputs[0];
  auto& b = inputs[1];
@@ -327,7 +329,7 @@ void DivMod::eval_gpu(
    std::vector<array>& outputs) {
  nvtx3::scoped_range r("DivMod::eval_gpu");
  auto& s = outputs[0].primitive().stream();
-  binary_two_op_gpu<cu::DivMod>(inputs, outputs, name(), s);
+  binary_two_op_gpu<cu::DivMod>(inputs, outputs, get_primitive_string(this), s);
 }

 } // namespace mlx::core
--- a/mlx/backend/cuda/compiled.cpp
+++ b/mlx/backend/cuda/compiled.cpp
@@ -53,10 +53,9 @@ struct FusedKernelBuilder {

    // Build function signature.
    if (contiguous) {
-      os += "template <typename IdxT = uint32_t, int work_per_thread = 1>\n";
+      os += "template <typename IdxT = uint32_t>\n";
    } else {
-      os +=
-          "template <int NDIM, typename IdxT = uint32_t, int work_per_thread = 1>\n";
+      os += "template <int NDIM, typename IdxT = uint32_t>\n";
    }
    os += fmt::format("__global__ void {}(\n", kernel_name + name);
    for (size_t i = 0; i < params.size(); ++i) {
@@ -68,77 +67,12 @@ struct FusedKernelBuilder {
    }
    os += ") {\n";

-    // Index. For non contiguous kernels we create a separate index
-    // variable per variable otherwise everyone uses `index`.
+    // Index.
    os +=
-        "  IdxT index = cg::this_grid().thread_rank() * work_per_thread;\n"
+        "  IdxT index = cg::this_grid().thread_rank();\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";
-    }
-
-    // Vectorized read loop
-    if (contiguous) {
-      for (size_t i = 0; i < inputs.size(); ++i) {
-        const auto& x = inputs[i];
-        if (is_scalar(x) || is_constant(i)) {
-          continue;
-        }
-        const std::string& xname = namer.get_name(x);
-        std::string type = dtype_to_cuda_type(x.dtype());
-        os += fmt::format(
-            "  auto vec_{0} = load_vector<work_per_thread, {1}>({0} + index, 0, size - index, 0);\n",
-            xname,
-            type);
-      }
-    }
-
-    // Create some space for the outputs
-    for (const auto& x : outputs) {
-      const std::string& xname = namer.get_name(x);
-      std::string type = dtype_to_cuda_type(x.dtype());
-      os += fmt::format(
-          "  AlignedVector<{}, work_per_thread> vec_{};\n", type, xname);
-    }
-
-    // Work loop
-    if (!contiguous) {
-      os +=
-          "\n"
-          "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
-    } else {
-      os +=
-          "\n"
-          "  #pragma unroll\n"
-          "  for (int i = 0; i < work_per_thread; i++) {\n";
-    }

    // Read inputs.
    for (size_t i = 0; i < inputs.size(); ++i) {
@@ -153,11 +87,14 @@ struct FusedKernelBuilder {
      } else if (is_scalar(x)) {
        value = fmt::format("{}[0]", xname);
      } else if (contiguous) {
-        value = fmt::format("vec_{}[i]", xname);
+        value = fmt::format("{}[index]", xname);
      } else {
-        value = fmt::format("{}[{}_idx]", xname, xname);
+        std::string index = fmt::format(
+            "elem_to_loc_nd<NDIM>(index, shape.data(), {}_strides.data())",
+            xname);
+        value = fmt::format("{}[{}]", xname, index);
      }
-      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
    }

    // Write tape.
@@ -169,40 +106,21 @@ struct FusedKernelBuilder {
        value = fmt::format(
            "static_cast<{}>(tmp_{})", type, namer.get_name(x.inputs()[0]));
      } else {
-        value = x.primitive().name();
+        std::ostringstream ss;
+        x.primitive().print(ss);
+        value = ss.str();
        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("    vec_{0}[i] = tmp_{0};\n", namer.get_name(x));
-    }
-
-    // End of work loop
-    if (!contiguous) {
-      os += "\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 += fmt::format("    {0}_idx += {0}_strides[NDIM - 1];\n", xname);
-      }
-    }
-    os += "  }\n";
-
-    // Store the output to global memory
-    for (const auto& x : outputs) {
-      os += fmt::format(
-          "  store_vector({0} + index, 0, vec_{0}, size - index);\n",
-          namer.get_name(x));
+      os += fmt::format("  {0}[index] = tmp_{0};\n", namer.get_name(x));
    }

    os += "}\n";
@@ -228,15 +146,6 @@ void Compiled::eval_gpu(
  nvtx3::scoped_range r("Compiled::eval_gpu");
  auto& s = stream();

-  // Determine the work per thread for the vectorized reads/writes. We take it
-  // as 16 over the max itemsize for the outputs. Another heuristic could be
-  // over the max itemsize of all arrays.
-  int max_size = 1;
-  for (const auto& x : outputs) {
-    max_size = (max_size > x.itemsize()) ? max_size : x.itemsize();
-  }
-  int work_per_thread = 16 / max_size;
-
  cu::JitModule& mod = cu::get_jit_module(s.device, lib_name(), [&]() {
    // Build source code.
    cu::FusedKernelBuilder builder{
@@ -249,24 +158,16 @@ 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;
-    kernel_names.push_back(fmt::format(
-        "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 (auto wpt : std::array<int, 2>{1, 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, wpt));
-        kernel_names.push_back(fmt::format(
-            "mlx::core::cu::{}_strided<{}, int64_t, {}>", lib_name(), i, wpt));
-      }
+    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) {
+      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));
    }
-
    return std::make_pair(std::move(builder.os), std::move(kernel_names));
  });

@@ -308,20 +209,13 @@ void Compiled::eval_gpu(
    args.append<uint32_t>(outputs[0].data_size());
  }

-  // Choose work per thread
-  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, work_per_thread);
+    kernel_name += fmt::format("_contiguous<{}>", index_type);
  } else {
-    kernel_name += fmt::format(
-        "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
+    kernel_name += fmt::format("_strided<{}, {}>", shape.size(), index_type);
  }
  auto& encoder = cu::get_command_encoder(s);
  for (const auto& in : inputs) {
@@ -332,9 +226,8 @@ void Compiled::eval_gpu(
  }

  auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] =
-      get_launch_args(outputs[0], large, work_per_thread);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(kernel, outputs[0], large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
 }

 } // namespace mlx::core
--- a/mlx/backend/cuda/conv.cpp
+++ b/mlx/backend/cuda/conv.cpp
@@ -1,546 +0,0 @@
-// 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>
-
-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
-
-// Alias for better readability.
-#define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
-#define CONV_BACKWARD_INPUT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DATA_DESCRIPTOR
-#define CONV_BACKWARD_WEIGHT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
-
-struct ConvCacheKey {
-  int device_id;
-  cudnnDataType_t cudnn_dtype;
-  std::array<int, MAX_NDIM> input_shape;
-  std::array<int, MAX_NDIM> weight_shape;
-  std::array<int, MAX_NDIM> stride;
-  std::array<int, MAX_NDIM> padding_lo;
-  std::array<int, MAX_NDIM> padding_hi;
-  std::array<int, MAX_NDIM> dilation;
-  int groups;
-  bool flip;
-  uint8_t input_alignment;
-  uint8_t weight_alignment;
-  uint8_t output_alignment;
-};
-
-auto& conv_cache() {
-  static LRUBytesKeyCache<
-      ConvCacheKey,
-      std::pair<cudnnBackendDescriptorType_t, cudnn_frontend::ExecutionPlan>>
-      cache(/* capacity */ 128);
-  return cache;
-}
-
-template <typename T, typename Vec>
-inline SmallVector<T> convert_vector(const Vec& vec) {
-  return SmallVector<T>(vec.begin(), vec.end());
-}
-
-template <typename T, template <typename U> class Vec>
-inline std::array<T, MAX_NDIM> fixed_vector(const Vec<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(std::move(shape), std::move(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,
-    array& x,
-    array& w,
-    array& y) {
-  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] = {
-      x.data<void>(),
-      w.data<void>(),
-      y.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,
-    const ConvCacheKey& cache_key,
-    cudnnBackendDescriptorType_t backend_type,
-    cudnn_frontend::EngineConfigList& configs,
-    const std::string& op_graph_tag,
-    array& x,
-    array& w,
-    array& y) {
-  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, x, w, y)) {
-        conv_cache().emplace(
-            cache_key, std::make_pair(backend_type, std::move(plan)));
-        return true;
-      }
-    } catch (cudnn_frontend::cudnnException& error) {
-      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
-        throw;
-      }
-    }
-  }
-  return false;
-}
-
-auto get_conv_op_settings(
-    cudnnBackendDescriptorType_t backend_type,
-    array& x,
-    array& w,
-    array& y,
-    const std::vector<int>& kernel_strides,
-    const std::vector<int>& padding_lo_,
-    const std::vector<int>& padding_hi_,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation) {
-  auto padding_lo = convert_vector<int64_t>(padding_lo_);
-  auto padding_hi = convert_vector<int64_t>(padding_hi_);
-
-  if (backend_type == CONV_BACKWARD_INPUT) {
-    for (int i = 0; i < padding_lo.size(); ++i) {
-      int wt_size = 1 + kernel_dilation[i] * (w.shape(1 + i) - 1);
-      padding_lo[i] = wt_size - padding_lo[i] - 1;
-      int in_size = 1 + kernel_strides[i] * (x.shape(1 + i) - 1);
-      int out_size = 1 + input_dilation[i] * (y.shape(1 + i) - 1);
-      padding_hi[i] = out_size - in_size + padding_hi[i];
-    }
-    return std::make_tuple(
-        convert_vector<int64_t>(input_dilation),
-        std::move(padding_lo),
-        std::move(padding_hi),
-        convert_vector<int64_t>(kernel_dilation));
-
-  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    padding_hi = padding_lo;
-    return std::make_tuple(
-        convert_vector<int64_t>(kernel_dilation),
-        std::move(padding_lo),
-        std::move(padding_hi),
-        convert_vector<int64_t>(kernel_strides));
-
-  } else {
-    return std::make_tuple(
-        convert_vector<int64_t>(kernel_strides),
-        std::move(padding_lo),
-        std::move(padding_hi),
-        convert_vector<int64_t>(kernel_dilation));
-  }
-}
-
-std::optional<cudnn_frontend::OperationGraph> build_op_graph(
-    cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    array& x,
-    array& w,
-    array& y,
-    const SmallVector<int64_t>& stride,
-    const SmallVector<int64_t>& padding_lo,
-    const SmallVector<int64_t>& padding_hi,
-    const SmallVector<int64_t>& dilation) {
-  try {
-    auto compute_dtype = (dtype == float16 || dtype == bfloat16)
-        ? CUDNN_DATA_FLOAT
-        : dtype_to_cudnn_type(dtype);
-    auto conv_desc = cudnn_frontend::ConvDescBuilder()
-                         .setDataType(compute_dtype)
-                         .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', x))
-                  .setwDesc(build_tensor('w', w))
-                  .setyDesc(build_tensor('y', y))
-                  .setcDesc(conv_desc)
-                  .build();
-
-    std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
-    return cudnn_frontend::OperationGraphBuilder()
-        .setHandle(encoder.device().cudnn_handle())
-        .setOperationGraph(ops.size(), ops.data())
-        .build();
-  } catch (cudnn_frontend::cudnnException& error) {
-    if (error.getCudnnStatus() != CUDNN_STATUS_BAD_PARAM) {
-      throw;
-    }
-    return std::nullopt;
-  }
-}
-
-// Do necessary transposes and copies to prepare the inputs and outputs for
-// building the cuDNN conv op. It is safe to be called multiple times in one
-// eval_gpu, with cost of possible redundant copies.
-std::tuple<array, array, array> prepare_args(
-    cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
-    array in,
-    array wt,
-    array out,
-    Stream s) {
-  // Transpose the args depending on the backend type.
-  // TODO: Handle groups.
-  if (backend_type == CONV_BACKWARD_INPUT) {
-    wt = swapaxes_in_eval(wt, 0, -1);
-  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    in = swapaxes_in_eval(in, 0, -1);
-    wt = swapaxes_in_eval(wt, 0, -1);
-    // Create a contiguous array that shares the data with |out|, but with dim
-    // C_in and C_out swapped.
-    Shape shape(out.shape());
-    std::swap(shape.front(), shape.back());
-    Strides strides(shape.size(), 1);
-    for (int i = shape.size() - 2; i >= 0; --i) {
-      strides[i] = shape[i + 1] * strides[i + 1];
-    }
-    array intermediate(std::move(shape), out.dtype(), nullptr, {});
-    intermediate.copy_shared_buffer(
-        out, std::move(strides), {true, true, false}, out.data_size());
-    out = intermediate;
-  }
-
-  // cuDNN requires contiguous input.
-  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);
-  }
-
-  return {std::move(in), std::move(wt), std::move(out)};
-}
-
-// Get the x/w/y args from the in/wt/out args depending on backend type.
-inline std::tuple<array&, array&, array&> dispatch_args(
-    cudnnBackendDescriptorType_t backend_type,
-    array& in,
-    array& wt,
-    array& out) {
-  switch (backend_type) {
-    case CONV_BACKWARD_INPUT:
-      return {out, wt, in};
-    case CONV_BACKWARD_WEIGHT:
-      return {in, out, wt};
-    default:
-      return {in, wt, out};
-  }
-}
-
-// Register inputs and outputs before actually running conv op. Can only be
-// called once per eval_gpu.
-void register_args(
-    cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
-    array& in,
-    array& wt,
-    array& intermediate_out,
-    array& final_out) {
-  encoder.set_input_array(in);
-  encoder.set_input_array(wt);
-  encoder.set_output_array(final_out);
-
-  if (backend_type == CONV_BACKWARD_WEIGHT) {
-    // Turn |out| into a strided array, which will have C_in and C_out swapped
-    // in vjp and the final |grad_weight| will then be contiguous.
-    Strides strides = intermediate_out.strides();
-    std::swap(strides.front(), strides.back());
-    final_out.copy_shared_buffer(
-        intermediate_out,
-        std::move(strides),
-        {false, false, false},
-        intermediate_out.data_size());
-  }
-}
-
-} // 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];
-  array out = out_;
-  out.set_data(allocator::malloc(out.nbytes()));
-  Dtype dtype = out.dtype();
-
-  auto& s = stream();
-  auto& encoder = cu::get_command_encoder(s);
-
-  // Search cache.
-  ConvCacheKey cache_key{
-      encoder.device().cuda_device(),
-      dtype_to_cudnn_type(dtype),
-      fixed_vector(in.shape()),
-      fixed_vector(wt.shape()),
-      fixed_vector(kernel_strides_),
-      fixed_vector(padding_lo_),
-      fixed_vector(padding_hi_),
-      fixed_vector(kernel_dilation_),
-      groups_,
-      flip_,
-      get_alignment(in),
-      get_alignment(wt),
-      get_alignment(out)};
-  if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
-    auto& [backend_type, plan] = it->second;
-    std::tie(in, wt, out) = prepare_args(encoder, backend_type, in, wt, out, s);
-    register_args(encoder, backend_type, in, wt, out, out_);
-    auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-    if (!execute_plan(encoder, plan, x, w, y)) {
-      throw std::runtime_error("[conv] Cached plan failed to execute.");
-    }
-    return;
-  }
-
-  // There is no reliable way to deduce the proper cuDNN backend for the
-  // convolution, so we make a best guess and then try.
-  SmallVector<cudnnBackendDescriptorType_t, 2> try_backends;
-  if (flip_) {
-    // When weight is flipped, we assume it is backward input convolution.
-    try_backends.push_back(CONV_BACKWARD_INPUT);
-  } else {
-    // Otherwise it could be backward weight convolution or forward convolution,
-    // mathematically there is no difference so we have to use heuristics.
-    // Empirically backward convolutions have large kernel dimensions, and
-    // usually have |in| and |wt| transposed.
-    if (!in.flags().row_contiguous && !wt.flags().row_contiguous &&
-        wt.shape(2) > out.shape(2)) {
-      try_backends = {CONV_BACKWARD_WEIGHT, CONV_FORWARD};
-    } else {
-      try_backends = {CONV_FORWARD, CONV_BACKWARD_WEIGHT};
-    }
-  }
-
-  // Try to build op graph.
-  cudnnBackendDescriptorType_t backend_type;
-  std::optional<cudnn_frontend::OperationGraph> op_graph;
-  for (auto try_backend : try_backends) {
-    auto [in_copy, wt_copy, out_copy] =
-        prepare_args(encoder, try_backend, in, wt, out, s);
-    auto [x, w, y] = dispatch_args(try_backend, in_copy, wt_copy, out_copy);
-    auto [stride, padding_lo, padding_hi, dilation] = get_conv_op_settings(
-        try_backend,
-        x,
-        w,
-        y,
-        kernel_strides_,
-        padding_lo_,
-        padding_hi_,
-        kernel_dilation_,
-        input_dilation_);
-    op_graph = build_op_graph(
-        encoder,
-        try_backend,
-        dtype,
-        x,
-        w,
-        y,
-        stride,
-        padding_lo,
-        padding_hi,
-        dilation);
-    if (op_graph) {
-      backend_type = try_backend;
-      in = std::move(in_copy);
-      wt = std::move(wt_copy);
-      out = std::move(out_copy);
-      break;
-    }
-  }
-  if (!op_graph) {
-    throw std::runtime_error("[conv] Can not build op graph.");
-  }
-
-  // Get ready to execute the graph.
-  register_args(encoder, backend_type, in, wt, out, out_);
-
-  // Try to run plans based on heuristics.
-  auto configs = get_engine_configs(backend_type, dtype, *op_graph);
-  auto tag = op_graph->getTag();
-  auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-    return;
-  }
-  // Then try fallback plans.
-  configs = get_engine_configs(backend_type, dtype, *op_graph);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-    return;
-  }
-  throw std::runtime_error("[conv] Unable to find a working engine.");
-}
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/copy/copy_contiguous.cu
+++ b/mlx/backend/cuda/copy/copy_contiguous.cu
@@ -22,7 +22,7 @@ __global__ void copy_s(const In* in, Out* out, IdxT size) {
    AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      out_vec[i] = cast_to<Out>(in[0]);
+      out_vec.val[i] = cast_to<Out>(in[0]);
    }

    store_vector<N_READS>(out, index, out_vec);
@@ -43,7 +43,7 @@ __global__ void copy_v(const In* in, Out* out, IdxT size) {
    AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
    for (int i = 0; i < N_READS; ++i) {
-      out_vec[i] = cast_to<Out>(in_vec[i]);
+      out_vec.val[i] = cast_to<Out>(in_vec.val[i]);
    }

    store_vector<N_READS>(out, index, out_vec);
@@ -65,18 +65,23 @@ 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>;
-        constexpr int N_READS = 16 / sizeof(InType);
+        // 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, N_READS>;
        }
        auto [num_blocks, block_dims] = get_launch_args(
-            out.data_size(), out.shape(), out.strides(), large(), N_READS);
+            kernel,
+            out.data_size(),
+            out.shape(),
+            out.strides(),
+            large(),
+            N_READS);
        encoder.add_kernel_node(
            kernel,
            num_blocks,
            block_dims,
-            0,
            in.data<InType>() + in_offset,
            out.data<OutType>() + out_offset,
            out.data_size());
--- a/mlx/backend/cuda/copy/copy_general.cu
+++ b/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(
+    auto [idx_in, idx_out] = elem_to_loc_4d(
        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
  }
@@ -71,13 +71,14 @@ void copy_general(
              data_size *= s;
            if (ndim <= 3) {
              dispatch_1_2_3(ndim, [&](auto ndim_constant) {
-                auto [num_blocks, block_dims] =
-                    get_launch_args(data_size, shape, out.strides(), large());
+                auto kernel =
+                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>;
+                auto [num_blocks, block_dims] = get_launch_args(
+                    kernel, data_size, shape, out.strides(), large());
                encoder.add_kernel_node(
-                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>,
+                    kernel,
                    num_blocks,
                    block_dims,
-                    0,
                    in_ptr,
                    out_ptr,
                    data_size,
@@ -86,13 +87,13 @@ void copy_general(
                    const_param<ndim_constant()>(strides_out));
              });
            } else { // ndim >= 4
-              auto [num_blocks, block_dims] =
-                  get_launch_args(data_size, shape, out.strides(), large());
+              auto kernel = cu::copy_gg<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] = get_launch_args(
+                  kernel, data_size, shape, out.strides(), large());
              encoder.add_kernel_node(
-                  cu::copy_gg<InType, OutType, IdxT>,
+                  kernel,
                  num_blocks,
                  block_dims,
-                  0,
                  in_ptr,
                  out_ptr,
                  data_size,
--- a/mlx/backend/cuda/copy/copy_general_dynamic.cu
+++ b/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(
+    auto [idx_in, idx_out] = elem_to_loc_4d(
        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
    out[idx_out + *offset_out] = CastOp<In, Out>{}(in[idx_in + *offset_in]);
  }
@@ -74,16 +74,14 @@ void copy_general_dynamic(
            int ndim = shape.size();
            if (ndim <= 3) {
              dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                auto [num_blocks, block_dims] = get_launch_args(out, large());
+                auto kernel = cu::
+                    copy_gg_dynamic_nd<InType, OutType, IdxT, dims_constant()>;
+                auto [num_blocks, block_dims] =
+                    get_launch_args(kernel, out, large());
                encoder.add_kernel_node(
-                    cu::copy_gg_dynamic_nd<
-                        InType,
-                        OutType,
-                        IdxT,
-                        dims_constant()>,
+                    kernel,
                    num_blocks,
                    block_dims,
-                    0,
                    in_ptr,
                    out_ptr,
                    out.size(),
@@ -94,12 +92,13 @@ void copy_general_dynamic(
                    dynamic_offset_out.data<int64_t>());
              });
            } else { // ndim >= 4
-              auto [num_blocks, block_dims] = get_launch_args(out, large());
+              auto kernel = cu::copy_gg_dynamic<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] =
+                  get_launch_args(kernel, out, large());
              encoder.add_kernel_node(
-                  cu::copy_gg_dynamic<InType, OutType, IdxT>,
+                  kernel,
                  num_blocks,
                  block_dims,
-                  0,
                  in_ptr,
                  out_ptr,
                  out.size(),
--- a/mlx/backend/cuda/copy/copy_general_input.cu
+++ b/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(index, shape.data(), strides_in.data(), ndim);
+    IdxT idx_in = elem_to_loc_4d(index, shape.data(), strides_in.data(), ndim);
    out[index] = CastOp<In, Out>{}(in[idx_in]);
  }
 }
@@ -63,12 +63,14 @@ void copy_general_input(
            int ndim = shape.size();
            if (ndim <= 3) {
              dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                auto [num_blocks, block_dims] = get_launch_args(out, large());
+                auto kernel =
+                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>;
+                auto [num_blocks, block_dims] =
+                    get_launch_args(kernel, out, large());
                encoder.add_kernel_node(
-                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>,
+                    kernel,
                    num_blocks,
                    block_dims,
-                    0,
                    in_ptr,
                    out_ptr,
                    out.size(),
@@ -76,12 +78,13 @@ void copy_general_input(
                    const_param<dims_constant()>(strides_in));
              });
            } else { // ndim >= 4
-              auto [num_blocks, block_dims] = get_launch_args(out, large());
+              auto kernel = cu::copy_g<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] =
+                  get_launch_args(kernel, out, large());
              encoder.add_kernel_node(
-                  cu::copy_g<InType, OutType, IdxT>,
+                  kernel,
                  num_blocks,
                  block_dims,
-                  0,
                  in_ptr,
                  out_ptr,
                  out.size(),
--- a/mlx/backend/cuda/device.cpp
+++ b/mlx/backend/cuda/device.cpp
@@ -1,7 +1,6 @@
 // Copyright © 2025 Apple Inc.

 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/jit_module.h"
 #include "mlx/backend/cuda/worker.h"
 #include "mlx/utils.h"

@@ -10,23 +9,12 @@
 #include <future>
 #include <unordered_set>

-namespace mlx::core::cu {
-
-namespace {
+namespace mlx::core {

 // 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);
@@ -34,7 +22,7 @@ int cuda_graph_cache_size() {
  return cache_size;
 }

-} // namespace
+namespace cu {

 Device::Device(int device) : device_(device) {
  CHECK_CUDA_ERROR(cudaDeviceGetAttribute(
@@ -52,18 +40,11 @@ Device::Device(int device) : device_(device) {
  }
  // The cublasLt handle is used by matmul.
  make_current();
-  CHECK_CUBLAS_ERROR(cublasLtCreate(&lt_));
-  // The cudnn handle is used by Convolution.
-  CHECK_CUDNN_ERROR(cudnnCreate(&cudnn_));
-
-  // Initialize the jit module cache here ensures it is not
-  // unloaded before any evaluation is done
-  get_jit_module_cache();
+  cublasLtCreate(&lt_);
 }

 Device::~Device() {
-  CHECK_CUDNN_ERROR(cudnnDestroy(cudnn_));
-  CHECK_CUBLAS_ERROR(cublasLtDestroy(lt_));
+  cublasLtDestroy(lt_);
 }

 void Device::make_current() {
@@ -85,19 +66,30 @@ CommandEncoder& Device::get_command_encoder(Stream s) {
 }

 CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
-  enc.device().make_current();
+  CHECK_CUDA_ERROR(cudaGraphCreate(&graph, 0));
  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;
+  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'});
  }
-  enc.add_graph_node(graph);
+  CHECK_CUDA_ERROR(cudaGraphDestroy(graph));
 }

 CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
@@ -184,11 +176,21 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
  }
 }

-CommandEncoder::CommandEncoder(Device& d)
-    : device_(d), stream_(d), graph_cache_(cuda_graph_cache_size()) {
+CommandEncoder::CommandEncoder(Device& d) : device_(d), stream_(d) {
  CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
 }

+void clear_graphs(std::unordered_map<std::string, cudaGraphExec_t>& graphs) {
+  for (auto& [_, graph_exec] : graphs) {
+    CHECK_CUDA_ERROR(cudaGraphExecDestroy(graph_exec));
+  }
+  graphs.clear();
+}
+
+CommandEncoder::~CommandEncoder() {
+  clear_graphs(graph_cache_);
+}
+
 void CommandEncoder::add_completed_handler(std::function<void()> task) {
  worker_.add_task(std::move(task));
 }
@@ -214,22 +216,22 @@ void CommandEncoder::add_kernel_node(
    void* func,
    dim3 grid_dim,
    dim3 block_dim,
-    uint32_t smem_bytes,
    void** params) {
  cudaKernelNodeParams kernel_params = {0};
  kernel_params.func = func;
  kernel_params.gridDim = grid_dim;
  kernel_params.blockDim = block_dim;
  kernel_params.kernelParams = params;
-  kernel_params.sharedMemBytes = smem_bytes;
-  add_kernel_node(kernel_params);
+  cudaGraphNode_t node;
+  CHECK_CUDA_ERROR(
+      cudaGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
+  insert_graph_dependencies(GraphNode{node, 'K'});
 }

 void CommandEncoder::add_kernel_node(
    CUfunction func,
    dim3 grid_dim,
    dim3 block_dim,
-    uint32_t smem_bytes,
    void** params) {
  CUDA_KERNEL_NODE_PARAMS kernel_params = {0};
  kernel_params.func = func;
@@ -240,30 +242,13 @@ void CommandEncoder::add_kernel_node(
  kernel_params.blockDimY = block_dim.y;
  kernel_params.blockDimZ = block_dim.z;
  kernel_params.kernelParams = params;
-  kernel_params.sharedMemBytes = smem_bytes;
-  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));
+  CHECK_CUDA_ERROR(
+      cuGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_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() {
-  nvtx3::scoped_range r("CommandEncoder::commit");
  if (!temporaries_.empty()) {
    add_completed_handler([temporaries = std::move(temporaries_)]() {});
  }
@@ -280,7 +265,7 @@ void CommandEncoder::commit() {
    graph_key_ += ".";
    graph_key_ += std::to_string(empty_node_count_);

-    CudaGraphExec& graph_exec = graph_cache_[graph_key_];
+    cudaGraphExec_t& graph_exec = graph_cache_[graph_key_];

    if (graph_exec != nullptr) {
      cudaGraphExecUpdateResult update_result;
@@ -294,19 +279,25 @@ void CommandEncoder::commit() {
 #endif // CUDART_VERSION >= 12000
      if (update_result != cudaGraphExecUpdateSuccess) {
        cudaGetLastError(); // reset error
-        graph_exec.reset();
+        CHECK_CUDA_ERROR(cudaGraphExecDestroy(graph_exec));
+        graph_exec = nullptr;
      }
    }
    if (graph_exec == nullptr) {
-      graph_exec.instantiate(graph_);
+      CHECK_CUDA_ERROR(
+          cudaGraphInstantiate(&graph_exec, graph_, NULL, NULL, 0));
    }
    device_.make_current();
    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));

+    // TODO smarter cache policy
+    if (graph_cache_.size() > cuda_graph_cache_size()) {
+      clear_graphs(graph_cache_);
+    }
+
    // Reset state
    node_count_ = 0;
    graph_node_count_ = 0;
-    empty_node_count_ = 0;
    from_nodes_.clear();
    to_nodes_.clear();
    graph_key_.clear();
@@ -316,6 +307,7 @@ void CommandEncoder::commit() {
  }

  // Put completion handlers in a batch.
+  worker_.end_batch();
  worker_.commit(stream_);
 }

@@ -324,6 +316,7 @@ 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();
 }
@@ -341,4 +334,6 @@ CommandEncoder& get_command_encoder(Stream s) {
  return device(s.device).get_command_encoder(s);
 }

-} // namespace mlx::core::cu
+} // namespace cu
+
+} // namespace mlx::core
--- a/mlx/backend/cuda/device.h
+++ b/mlx/backend/cuda/device.h
@@ -3,13 +3,11 @@
 #pragma once

 #include "mlx/array.h"
-#include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/cuda/worker.h"
 #include "mlx/stream.h"

 #include <cublasLt.h>
 #include <cuda.h>
-#include <cudnn.h>
 #include <thrust/execution_policy.h>

 #include <unordered_map>
@@ -23,7 +21,6 @@ class CommandEncoder {
    ~CaptureContext();
    cudaGraph_t graph;
    CommandEncoder& enc;
-    bool discard{false};
  };
  struct ConcurrentContext {
    ConcurrentContext(CommandEncoder& enc);
@@ -32,6 +29,7 @@ class CommandEncoder {
  };

  explicit CommandEncoder(Device& d);
+  ~CommandEncoder();

  CommandEncoder(const CommandEncoder&) = delete;
  CommandEncoder& operator=(const CommandEncoder&) = delete;
@@ -47,39 +45,25 @@ class CommandEncoder {
  void set_output_array(const array& arr);

  template <typename F, typename... Params>
-  void add_kernel_node(
-      F* func,
-      dim3 grid_dim,
-      dim3 block_dim,
-      uint32_t smem_bytes,
-      Params&&... params) {
+  void
+  add_kernel_node(F* func, dim3 grid_dim, dim3 block_dim, Params&&... params) {
    constexpr size_t num = sizeof...(Params);
    void* ptrs[num];
    size_t i = 0;
    ([&](auto&& p) { ptrs[i++] = static_cast<void*>(&p); }(
         std::forward<Params>(params)),
     ...);
-    add_kernel_node((void*)func, grid_dim, block_dim, smem_bytes, ptrs);
+    add_kernel_node((void*)func, grid_dim, block_dim, ptrs);
  }

  void add_kernel_node(
      CUfunction func,
      dim3 grid_dim,
      dim3 block_dim,
-      uint32_t smem_bytes,
      void** params);

-  void add_kernel_node(
-      void* func,
-      dim3 grid_dim,
-      dim3 block_dim,
-      uint32_t smem_bytes,
-      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_kernel_node(void* func, dim3 grid_dim, dim3 block_dim, void** params);

  void add_temporary(const array& arr) {
    temporaries_.push_back(arr.data_shared_ptr());
@@ -89,10 +73,6 @@ class CommandEncoder {
  void maybe_commit();
  void commit();

-  Device& device() {
-    return device_;
-  }
-
  CudaStream& stream() {
    return stream_;
  }
@@ -126,7 +106,7 @@ class CommandEncoder {
  std::string graph_key_;
  std::vector<GraphNode> concurrent_nodes_;
  std::vector<std::shared_ptr<array::Data>> temporaries_;
-  LRUCache<std::string, CudaGraphExec> graph_cache_;
+  std::unordered_map<std::string, cudaGraphExec_t> graph_cache_;
  std::vector<std::uintptr_t> active_deps_;
  std::vector<std::uintptr_t> active_outputs_;
  std::unordered_map<std::uintptr_t, GraphNode> node_map_;
@@ -157,16 +137,12 @@ 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_;
 };

--- a/mlx/backend/cuda/device/arange.cuh
+++ b/mlx/backend/cuda/device/arange.cuh
@@ -0,0 +1,15 @@
+// Copyright © 2025 Apple Inc.
+
+namespace mlx::core::cu {
+
+template <typename T>
+struct Arange {
+  const T start;
+  const T step;
+
+  __device__ T operator()(uint32_t i) const {
+    return start + i * step;
+  }
+};
+
+} // namespace mlx::core::cu
--- a/mlx/backend/cuda/device/atomic_ops.cuh
+++ b/mlx/backend/cuda/device/atomic_ops.cuh
@@ -2,7 +2,7 @@

 #pragma once

-#include "mlx/backend/cuda/device/complex.cuh"
+#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/device/fp16_math.cuh"

 #include <cuda/atomic>
@@ -48,16 +48,25 @@ inline __device__ void atomic_add(__half* out, __half val) {
  atomicAdd(out, val);
 }

-inline __device__ void atomic_add(complex64_t* out, complex64_t val) {
-  atomic_add_general(out, val);
-}
-
-inline __device__ void atomic_add(__nv_bfloat16* out, __nv_bfloat16 val) {
-#if __CUDA_ARCH__ < 800
+inline __device__ void atomic_add(cuComplex* out, cuComplex val) {
+#if __CUDA_ARCH__ < 900
  atomic_add_general(out, val);
 #else
  atomicAdd(out, val);
 #endif
 }

+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
+}
+
 } // namespace mlx::core::cu
--- a/mlx/backend/cuda/device/binary_ops.cuh
+++ b/mlx/backend/cuda/device/binary_ops.cuh
@@ -44,7 +44,7 @@ struct Remainder {
      } else {
        return x % y;
      }
-    } else if constexpr (is_complex_v<T>) {
+    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
      return x % y;
    } else {
      T r = fmod(x, y);
@@ -66,12 +66,14 @@ struct Equal {
 struct NaNEqual {
  template <typename T>
  __device__ bool operator()(T x, T y) {
-    if constexpr (is_complex_v<T>) {
+    if constexpr (std::is_same_v<T, cuComplex>) {
      return x == 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());
+          (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));
    } else {
      return x == y || (isnan(x) && isnan(y));
    }
@@ -109,17 +111,17 @@ struct LessEqual {
 struct LogAddExp {
  template <typename T>
  __device__ T operator()(T x, T y) {
-    if constexpr (is_complex_v<T>) {
-      if (isnan(x.real()) || isnan(x.imag()) || isnan(y.real()) ||
-          isnan(y.imag())) {
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      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()};
      }
-      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();
+      auto max = cuCrealf(x) > cuCrealf(y) ? x : y;
+      auto min = cuCrealf(x) < cuCrealf(y) ? x : y;
+      auto min_real = cuCrealf(min);
+      auto max_real = cuCrealf(max);
      if (!isfinite(min_real) && (min_real == max_real)) {
        if (min_real < 0) {
          return min;
@@ -148,8 +150,8 @@ struct Maximum {
  __device__ T operator()(T x, T y) {
    if constexpr (cuda::std::is_integral_v<T>) {
      return max(x, y);
-    } else if constexpr (is_complex_v<T>) {
-      if (isnan(x.real()) || isnan(x.imag())) {
+    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      if (isnan(cuCrealf(x)) || isnan(cuCimagf(x))) {
        return x;
      }
      return x > y ? x : y;
@@ -167,8 +169,8 @@ struct Minimum {
  __device__ T operator()(T x, T y) {
    if constexpr (cuda::std::is_integral_v<T>) {
      return min(x, y);
-    } else if constexpr (is_complex_v<T>) {
-      if (isnan(x.real()) || isnan(x.imag())) {
+    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      if (isnan(cuCrealf(x)) || isnan(cuCimagf(x))) {
        return x;
      }
      return x < y ? x : y;
@@ -191,8 +193,8 @@ struct Multiply {
 struct NotEqual {
  template <typename T>
  __device__ bool operator()(T x, T y) {
-    if constexpr (is_complex_v<T>) {
-      return x.real() != y.real() || x.imag() != y.imag();
+    if constexpr (std::is_same_v<T, cuComplex>) {
+      return cuCrealf(x) != cuCrealf(y) || cuCimagf(x) != cuCimagf(y);
    } else {
      return x != y;
    }
@@ -212,8 +214,19 @@ struct Power {
        base *= base;
      }
      return res;
-    } else if constexpr (is_complex_v<T>) {
-      return pow(base, exp);
+    } 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 {
      return powf(base, exp);
    }
--- a/mlx/backend/cuda/device/cast_op.cuh
+++ b/mlx/backend/cuda/device/cast_op.cuh
@@ -2,8 +2,7 @@

 #pragma once

-#include "mlx/backend/cuda/device/complex.cuh"
-
+#include <cuComplex.h>
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 #include <thrust/iterator/transform_iterator.h>
@@ -21,43 +20,50 @@ struct CastOp {
 };

 // Castings between complex and boolean.
-template <typename T>
-struct CastOp<complex_t<T>, bool> {
+// TODO: Should make a custom complex type.
+template <>
+struct CastOp<cuComplex, bool> {
  static constexpr bool is_castable = true;

-  __device__ bool operator()(complex_t<T> x) {
-    return x.real() != 0 && x.imag() != 0;
+  __device__ bool operator()(cuComplex x) {
+    return x.x != 0 && x.y != 0;
  }
 };

-template <typename T>
-struct CastOp<bool, complex_t<T>> {
+template <>
+struct CastOp<bool, cuComplex> {
  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};
+  __device__ cuComplex operator()(bool x) {
+    return x ? make_cuFloatComplex(1, 1) : make_cuFloatComplex(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>;
+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>;

-  __device__ DstT operator()(complex_t<T> x) {
-    static_assert(!is_complex_v<DstT>);
-    return static_cast<DstT>(x.real());
+  __device__ DstT operator()(cuComplex x) {
+    static_assert(!cuda::std::is_same_v<cuComplex, DstT>);
+    return static_cast<DstT>(cuCrealf(x));
  }
 };

 // Allow converting a real number to complex number.
-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>;
+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>;

-  __device__ complex_t<T> operator()(SrcT x) {
-    static_assert(!is_complex_v<SrcT>);
-    return complex_t<T>{static_cast<T>(x), 0};
+  __device__ cuComplex operator()(SrcT x) {
+    static_assert(!cuda::std::is_same_v<SrcT, cuComplex>);
+    return cuComplex{static_cast<float>(x), 0};
  }
 };

@@ -82,7 +88,8 @@ struct CastOp<
    SrcT,
    DstT,
    cuda::std::enable_if_t<
-        !cuda::std::is_convertible_v<SrcT, DstT> && !is_complex_v<SrcT> &&
+        !cuda::std::is_convertible_v<SrcT, DstT> &&
+        !cuda::std::is_same_v<SrcT, cuComplex> &&
        (cuda::std::is_same_v<DstT, __half> ||
         cuda::std::is_same_v<DstT, __nv_bfloat16>)>> {
  static constexpr bool is_castable = true;
@@ -97,7 +104,8 @@ struct CastOp<
    SrcT,
    DstT,
    cuda::std::enable_if_t<
-        !cuda::std::is_convertible_v<SrcT, DstT> && !is_complex_v<SrcT> &&
+        !cuda::std::is_convertible_v<SrcT, DstT> &&
+        !cuda::std::is_same_v<DstT, cuComplex> &&
        !cuda::std::is_same_v<DstT, __half> &&
        !cuda::std::is_same_v<DstT, __nv_bfloat16> &&
        (cuda::std::is_same_v<SrcT, __half> ||
--- a/mlx/backend/cuda/device/cexpf.cuh
+++ b/mlx/backend/cuda/device/cexpf.cuh
@@ -0,0 +1,138 @@
+// Copyright © 2025 Apple Inc.
+// Copyright © 2008-2013 NVIDIA Corporation
+// Copyright © 2013 Filipe RNC Maia
+//
+// 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/NVIDIA/cccl/blob/main/thrust/thrust/detail/complex/cexpf.h
+
+// TODO: We should use thrust::exp but the thrust header in old CUDA versions
+// can not be used in JIT.
+
+#pragma once
+
+#include <cuComplex.h>
+#include <cuda/std/cstdint>
+
+namespace mlx::core::cu::detail {
+
+using ieee_float_shape_type = union {
+  float value;
+  uint32_t word;
+};
+
+inline __device__ void get_float_word(uint32_t& i, float d) {
+  ieee_float_shape_type gf_u;
+  gf_u.value = (d);
+  (i) = gf_u.word;
+}
+
+inline __device__ void get_float_word(int32_t& i, float d) {
+  ieee_float_shape_type gf_u;
+  gf_u.value = (d);
+  (i) = gf_u.word;
+}
+
+inline __device__ void set_float_word(float& d, uint32_t i) {
+  ieee_float_shape_type sf_u;
+  sf_u.word = (i);
+  (d) = sf_u.value;
+}
+
+inline __device__ float frexp_expf(float x, int* expt) {
+  const uint32_t k = 235;
+  const float kln2 = 162.88958740F;
+
+  float exp_x;
+  uint32_t hx;
+
+  exp_x = expf(x - kln2);
+  get_float_word(hx, exp_x);
+  *expt = (hx >> 23) - (0x7f + 127) + k;
+  set_float_word(exp_x, (hx & 0x7fffff) | ((0x7f + 127) << 23));
+  return exp_x;
+}
+
+inline __device__ cuComplex ldexp_cexpf(cuComplex z, int expt) {
+  float x, y, exp_x, scale1, scale2;
+  int ex_expt, half_expt;
+
+  x = cuCrealf(z);
+  y = cuCimagf(z);
+  exp_x = frexp_expf(x, &ex_expt);
+  expt += ex_expt;
+
+  half_expt = expt / 2;
+  set_float_word(scale1, (0x7f + half_expt) << 23);
+  half_expt = expt - half_expt;
+  set_float_word(scale2, (0x7f + half_expt) << 23);
+
+  return cuComplex{
+      cosf(y) * exp_x * scale1 * scale2, sinf(y) * exp_x * scale1 * scale2};
+}
+
+inline __device__ cuComplex cexpf(const cuComplex& z) {
+  float x, y, exp_x;
+  uint32_t hx, hy;
+
+  const uint32_t exp_ovfl = 0x42b17218, cexp_ovfl = 0x43400074;
+
+  x = cuCrealf(z);
+  y = cuCimagf(z);
+
+  get_float_word(hy, y);
+  hy &= 0x7fffffff;
+
+  /* cexp(x + I 0) = exp(x) + I 0 */
+  if (hy == 0) {
+    return cuComplex{expf(x), y};
+  }
+  get_float_word(hx, x);
+  /* cexp(0 + I y) = cos(y) + I sin(y) */
+  if ((hx & 0x7fffffff) == 0) {
+    return cuComplex{cosf(y), sinf(y)};
+  }
+  if (hy >= 0x7f800000) {
+    if ((hx & 0x7fffffff) != 0x7f800000) {
+      /* cexp(finite|NaN +- I Inf|NaN) = NaN + I NaN */
+      return cuComplex{y - y, y - y};
+    } else if (hx & 0x80000000) {
+      /* cexp(-Inf +- I Inf|NaN) = 0 + I 0 */
+      return cuComplex{0.0, 0.0};
+    } else {
+      /* cexp(+Inf +- I Inf|NaN) = Inf + I NaN */
+      return cuComplex{x, y - y};
+    }
+  }
+
+  if (hx >= exp_ovfl && hx <= cexp_ovfl) {
+    /*
+     * x is between 88.7 and 192, so we must scale to avoid
+     * overflow in expf(x).
+     */
+    return ldexp_cexpf(z, 0);
+  } else {
+    /*
+     * Cases covered here:
+     *  -  x < exp_ovfl and exp(x) won't overflow (common case)
+     *  -  x > cexp_ovfl, so exp(x) * s overflows for all s > 0
+     *  -  x = +-Inf (generated by exp())
+     *  -  x = NaN (spurious inexact exception from y)
+     */
+    exp_x = expf(x);
+    return cuComplex{exp_x * cosf(y), exp_x * sinf(y)};
+  }
+}
+
+} // namespace mlx::core::cu::detail
--- a/mlx/backend/cuda/device/complex.cuh
+++ b/mlx/backend/cuda/device/complex.cuh
@@ -1,60 +0,0 @@
-// 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
--- a/mlx/backend/cuda/device/cucomplex_math.cuh
+++ b/mlx/backend/cuda/device/cucomplex_math.cuh
@@ -0,0 +1,240 @@
+// 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);
+}
--- a/mlx/backend/cuda/device/unary_ops.cuh
+++ b/mlx/backend/cuda/device/unary_ops.cuh
@@ -2,6 +2,8 @@

 #pragma once

+#include "mlx/backend/cuda/device/cexpf.cuh"
+#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/device/fp16_math.cuh"
 #include "mlx/backend/cuda/device/utils.cuh"

@@ -14,6 +16,8 @@ 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);
    }
@@ -25,6 +29,8 @@ struct ArcCos {
  __device__ T operator()(T x) {
    return acos(x);
  }
+
+  __device__ cuComplex operator()(cuComplex x);
 };

 struct ArcCosh {
@@ -39,6 +45,8 @@ struct ArcSin {
  __device__ T operator()(T x) {
    return asin(x);
  }
+
+  __device__ cuComplex operator()(cuComplex x);
 };

 struct ArcSinh {
@@ -53,6 +61,8 @@ struct ArcTan {
  __device__ T operator()(T x) {
    return atan(x);
  }
+
+  __device__ cuComplex operator()(cuComplex x);
 };

 struct ArcTanh {
@@ -74,8 +84,6 @@ 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);
    }
@@ -83,23 +91,34 @@ struct Ceil {
 };

 struct Conjugate {
-  template <typename T>
-  __device__ complex_t<T> operator()(complex_t<T> x) {
-    return conj(x);
+  __device__ cuComplex operator()(cuComplex x) {
+    return {cuCrealf(x), -cuCimagf(x)};
  }
 };

 struct Cos {
  template <typename T>
  __device__ T operator()(T x) {
-    return cos(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);
+    }
  }
 };

 struct Cosh {
  template <typename T>
  __device__ T operator()(T x) {
-    return cosh(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);
+    }
  }
 };

@@ -132,7 +151,11 @@ struct ErfInv {
 struct Exp {
  template <typename T>
  __device__ T operator()(T x) {
-    return exp(x);
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      return detail::cexpf(x);
+    } else {
+      return exp(x);
+    }
  }
 };

@@ -154,8 +177,6 @@ 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);
    }
@@ -163,25 +184,30 @@ struct Floor {
 };

 struct Imag {
-  template <typename T>
-  __device__ auto operator()(complex_t<T> x) {
-    return x.imag();
+  __device__ float operator()(cuComplex x) {
+    return cuCimagf(x);
  }
 };

 struct Log {
  template <typename T>
  __device__ T operator()(T x) {
-    return log(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);
+    }
  }
 };

 struct Log2 {
  template <typename T>
  __device__ T operator()(T x) {
-    if constexpr (is_complex_v<T>) {
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
      auto y = Log{}(x);
-      return {y.real() / CUDART_LN2_F, y.imag() / CUDART_LN2_F};
+      return {cuCrealf(y) / CUDART_LN2_F, cuCimagf(y) / CUDART_LN2_F};
    } else {
      return log2(x);
    }
@@ -191,17 +217,23 @@ struct Log2 {
 struct Log10 {
  template <typename T>
  __device__ T operator()(T x) {
-    return log10(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);
+    }
  }
 };

 struct Log1p {
  template <typename T>
  __device__ T operator()(T z) {
-    if constexpr (is_complex_v<T>) {
-      float x = z.real();
-      float y = z.imag();
-      float zabs = Abs{}(z).real();
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      float x = cuCrealf(z);
+      float y = cuCimagf(z);
+      float zabs = cuCrealf(Abs{}(z));
      float theta = atan2f(y, x + 1);
      if (zabs < 0.5f) {
        float r = x * (2 + x) + y * y;
@@ -228,8 +260,8 @@ struct LogicalNot {
 struct Negative {
  template <typename T>
  __device__ T operator()(T x) {
-    if constexpr (is_complex_v<T>) {
-      return T{0, 0} - x;
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      return 0 - x;
    } else {
      return -x;
    }
@@ -237,17 +269,16 @@ struct Negative {
 };

 struct Real {
-  template <typename T>
-  __device__ auto operator()(complex_t<T> x) {
-    return x.real();
+  __device__ float operator()(cuComplex x) {
+    return cuCrealf(x);
  }
 };

 struct Round {
  template <typename T>
  __device__ T operator()(T x) {
-    if constexpr (is_complex_v<T>) {
-      return {rint(x.real()), rint(x.imag())};
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      return {rint(cuCrealf(x)), rint(cuCimagf(x))};
    } else {
      return rint(x);
    }
@@ -267,8 +298,8 @@ struct Sign {
  __device__ T operator()(T x) {
    if constexpr (cuda::std::is_unsigned_v<T>) {
      return x != 0;
-    } else if constexpr (is_complex_v<T>) {
-      if (x.real() == 0 && x.imag() == 0) {
+    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+      if (cuCrealf(x) == 0 && cuCimagf(x) == 0) {
        return x;
      } else {
        return x / Abs()(x);
@@ -284,14 +315,26 @@ struct Sign {
 struct Sin {
  template <typename T>
  __device__ T operator()(T x) {
-    return sin(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);
+    }
  }
 };

 struct Sinh {
  template <typename T>
  __device__ T operator()(T x) {
-    return sinh(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);
+    }
  }
 };

@@ -307,31 +350,77 @@ 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) {
-    if constexpr (is_complex_v<T>) {
-      return 1.0f / Sqrt{}(x);
-    } else {
-      return rsqrt(x);
-    }
+    return rsqrt(x);
+  }
+  __device__ cuComplex operator()(cuComplex x) {
+    return 1.0f / Sqrt{}(x);
  }
 };

 struct Tan {
  template <typename T>
  __device__ T operator()(T x) {
-    return tan(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);
+    }
  }
 };

 struct Tanh {
  template <typename T>
  __device__ T operator()(T x) {
-    return tanh(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);
+    }
  }
 };

+inline __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)};
+};
+
+inline __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)};
+};
+
+inline __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
--- a/mlx/backend/cuda/device/utils.cuh
+++ b/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>
@@ -32,118 +32,21 @@ using Strides = cuda::std::array<int64_t, MAX_NDIM>;
 template <typename T, int N>
 struct alignas(sizeof(T) * N) AlignedVector {
  T val[N];
-
-  __device__ T& operator[](int i) {
-    return val[i];
-  }
-
-  __device__ T operator[](int i) const {
-    return val[i];
-  }
 };

 template <int N, typename T>
-inline __host__ __device__ bool is_aligned(T* x) {
-  return (reinterpret_cast<uintptr_t>(x) % (N * sizeof(T))) == 0;
-}
-
-template <int N, typename T>
-inline __device__ AlignedVector<T, N> unsafe_load_vector(
+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__ AlignedVector<T, N> load_vector(
-    const T* ptr,
-    uint32_t offset) {
-  if (is_aligned<N>(ptr)) {
-    auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
-    return from[offset];
-  } else {
-    AlignedVector<T, N> v;
-#pragma unroll
-    for (int i = 0; i < N; ++i) {
-      v[i] = ptr[offset * N + i];
-    }
-    return v;
-  }
-}
-
-template <int N, typename T, typename SizeT>
-inline __device__ AlignedVector<T, N>
-load_vector(const T* ptr, uint32_t offset, SizeT size, T fallback) {
-  if (is_aligned<N>(ptr) && (offset + 1) * N <= size) {
-    auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
-    return from[offset];
-  } else {
-    AlignedVector<T, N> v;
-#pragma unroll
-    for (int i = 0; i < N; ++i) {
-      v[i] = (N * offset + i) < size ? ptr[offset * N + i] : fallback;
-    }
-    return v;
-  }
-}
-
-template <int N, typename T, typename SizeT>
-inline __device__ AlignedVector<T, N> load_vector(
-    const T* ptr,
-    uint32_t offset,
-    SizeT size,
-    int64_t stride,
-    T fallback) {
-  if (is_aligned<N>(ptr) && stride == 1 && (offset + 1) * N <= size) {
-    auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
-    return from[offset];
-  } else {
-    AlignedVector<T, N> v;
-#pragma unroll
-    for (int i = 0; i < N; ++i) {
-      v[i] =
-          (N * offset + i) < size ? ptr[stride * (offset * N + i)] : fallback;
-    }
-    return v;
-  }
-}
-
-template <int N, typename T>
-inline __device__ void
-unsafe_store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
-  auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
-  to[offset] = vec;
-}
-
 template <int N, typename T>
 inline __device__ void
 store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
-  if (is_aligned<N>(ptr)) {
-    auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
-    to[offset] = vec;
-  } else {
-#pragma unroll
-    for (int i = 0; i < N; ++i) {
-      ptr[offset * N + i] = vec[i];
-    }
-  }
-}
-
-template <int N, typename T, typename SizeT>
-inline __device__ void store_vector(
-    T* ptr,
-    uint32_t offset,
-    const AlignedVector<T, N>& vec,
-    SizeT size) {
-  if (is_aligned<N>(ptr) && (offset + 1) * N <= size) {
-    auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
-    to[offset] = vec;
-  } else {
-    for (int i = 0; (offset * N + i) < size && i < N; ++i) {
-      ptr[offset * N + i] = vec[i];
-    }
-  }
+  auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
+  to[offset] = vec;
 }

 ///////////////////////////////////////////////////////////////////////////////
@@ -224,13 +127,13 @@ struct Limits<bool> {
  }
 };

-template <typename T>
-struct Limits<complex_t<T>> {
-  static constexpr __host__ __device__ complex_t<T> max() {
-    return {Limits<T>::max(), Limits<T>::max()};
+template <>
+struct Limits<cuComplex> {
+  static constexpr __host__ __device__ cuComplex max() {
+    return {Limits<float>::max(), Limits<float>::max()};
  }
-  static constexpr __host__ __device__ complex_t<T> min() {
-    return {Limits<T>::min(), Limits<T>::min()};
+  static constexpr __host__ __device__ cuComplex min() {
+    return {Limits<float>::min(), Limits<float>::min()};
  }
 };

@@ -301,8 +204,20 @@ 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__ cuda::std::tuple<IdxT, IdxT> elem_to_loc(
+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(
    IdxT elem,
    const int* shape,
    const int64_t* a_strides,
@@ -320,7 +235,7 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc(
 }

 template <typename IdxT = int64_t>
-inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc(
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_4d(
    IdxT elem,
    const int* shape,
    const int64_t* a_strides,
--- a/mlx/backend/cuda/eval.cpp
+++ b/mlx/backend/cuda/eval.cpp
@@ -19,6 +19,8 @@ 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) {
@@ -36,15 +38,18 @@ void eval(array& arr) {

  auto& encoder = cu::get_command_encoder(arr.primitive().stream());
  // Keep used buffers alive until kernel finishes running.
+  std::unordered_set<std::shared_ptr<array::Data>> buffers;
  for (auto& in : arr.inputs()) {
-    // Except for the donated one.
-    if (in.data_shared_ptr() != arr.data_shared_ptr()) {
-      encoder.add_temporary(in);
-    }
+    buffers.insert(in.data_shared_ptr());
  }
  for (auto& s : arr.siblings()) {
-    encoder.add_temporary(s);
+    buffers.insert(s.data_shared_ptr());
  }
+  // Remove the output if it was donated to by an input.
+  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+    buffers.erase(it);
+  }
+  encoder.add_completed_handler([buffers = std::move(buffers)]() {});
  encoder.maybe_commit();
 }

--- a/mlx/backend/cuda/event.cu
+++ b/mlx/backend/cuda/event.cu
@@ -90,6 +90,8 @@ 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) {
@@ -110,26 +112,26 @@ __global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
  event_signal(ac, value);
 }

-SharedEvent::Atomic* to_atomic(std::shared_ptr<Buffer> buf) {
-  return static_cast<SharedEvent::Atomic*>(buf->raw_ptr());
-}
+} // namespace

 SharedEvent::SharedEvent() {
-  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;
+  // 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);
+  });
 }

 void SharedEvent::wait(uint64_t value) {
  nvtx3::scoped_range r("cu::SharedEvent::wait");
-  event_wait(to_atomic(buf_), value);
+  event_wait(ac_.get(), value);
 }

 void SharedEvent::wait(cudaStream_t stream, uint64_t value) {
-  event_wait_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
+  event_wait_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
 }

 void SharedEvent::wait(Stream s, uint64_t value) {
@@ -140,17 +142,17 @@ void SharedEvent::wait(Stream s, uint64_t value) {
    auto& encoder = get_command_encoder(s);
    encoder.commit();
    wait(encoder.stream(), value);
-    encoder.add_completed_handler([buf = buf_]() {});
+    encoder.add_completed_handler([ac = ac_]() {});
  }
 }

 void SharedEvent::signal(uint64_t value) {
  nvtx3::scoped_range r("cu::SharedEvent::signal");
-  event_signal(to_atomic(buf_), value);
+  event_signal(ac_.get(), value);
 }

 void SharedEvent::signal(cudaStream_t stream, uint64_t value) {
-  event_signal_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
+  event_signal_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
 }

 void SharedEvent::signal(Stream s, uint64_t value) {
@@ -164,18 +166,18 @@ void SharedEvent::signal(Stream s, uint64_t value) {
    auto& encoder = get_command_encoder(s);
    encoder.commit();
    signal(encoder.stream(), value);
-    encoder.add_completed_handler([buf = buf_]() {});
+    encoder.add_completed_handler([ac = ac_]() {});
  }
 }

 bool SharedEvent::is_signaled(uint64_t value) const {
  nvtx3::scoped_range r("cu::SharedEvent::is_signaled");
-  return to_atomic(buf_)->load() >= value;
+  return ac_->load() >= value;
 }

 uint64_t SharedEvent::value() const {
  nvtx3::scoped_range r("cu::SharedEvent::value");
-  return to_atomic(buf_)->load();
+  return ac_->load();
 }

 } // namespace cu
--- a/mlx/backend/cuda/event.h
+++ b/mlx/backend/cuda/event.h
@@ -2,7 +2,6 @@

 #pragma once

-#include "mlx/allocator.h"
 #include "mlx/stream.h"

 #include <cuda_runtime.h>
@@ -56,8 +55,12 @@ 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<mlx::core::allocator::Buffer> buf_;
+  std::shared_ptr<Atomic> ac_;
 };

 } // namespace mlx::core::cu
--- a/mlx/backend/cuda/gemms/cublas_batched_gemm_12_0.cpp
+++ b/mlx/backend/cuda/gemms/cublas_batched_gemm_12_0.cpp
@@ -1,73 +0,0 @@
-// 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
--- a/mlx/backend/cuda/gemms/cublas_batched_gemm_12_9.cu
+++ b/mlx/backend/cuda/gemms/cublas_batched_gemm_12_9.cu
@@ -1,208 +0,0 @@
-// 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,
-      0,
-      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,
-      0,
-      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
--- a/mlx/backend/cuda/gemms/cublas_gemm.cpp
+++ b/mlx/backend/cuda/gemms/cublas_gemm.cpp
@@ -1,284 +0,0 @@
-// 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_,
-        c ? c_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) {
-    // Ensure workspace is 256-byte aligned
-    int nbytes = cuda::ceil_div(heuristic_.workspaceSize, 256) * 256;
-    array workspace(
-        allocator::malloc(nbytes),
-        {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
--- a/mlx/backend/cuda/gemms/cublas_gemm.h
+++ b/mlx/backend/cuda/gemms/cublas_gemm.h
@@ -1,100 +0,0 @@
-// 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
--- a/mlx/backend/cuda/gemms/gemv.cu
+++ b/mlx/backend/cuda/gemms/gemv.cu
@@ -1,173 +0,0 @@
-// 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 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 =
-          unsafe_load_vector<n_per_thread>(mat + row * cols + col, 0);
-      auto local_vec = unsafe_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[j]) * static_cast<float>(local_vec[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 % 32 == 0 && ((M == 1 && b_transposed) || (N == 1 && !a_transposed));
-}
-
-template <typename F>
-void dispatch_n_per_thread(int n_per_thread, F&& f) {
-  switch (n_per_thread) {
-    case 1:
-      f(std::integral_constant<int, 1>{});
-      break;
-    case 2:
-      f(std::integral_constant<int, 2>{});
-      break;
-    case 4:
-      f(std::integral_constant<int, 4>{});
-      break;
-  }
-}
-
-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;
-    int n_per_t;
-    if (K % 128 == 0 && is_aligned<4>(mat) && is_aligned<4>(vec)) {
-      n_per_t = 4;
-    } else if (K % 64 == 0 && is_aligned<2>(mat) && is_aligned<2>(vec)) {
-      n_per_t = 2;
-    } else {
-      n_per_t = 1;
-    }
-    dispatch_n_per_thread(n_per_t, [&](auto n_per_thread) {
-      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,
-            0,
-            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,
-            0,
-            mat,
-            vec,
-            out.data<DataType>(),
-            rows,
-            cols,
-            const_param(batch_shape),
-            mat_strides,
-            vec_strides,
-            batch_shape.size());
-      }
-    });
-  });
-}
-
-} // namespace mlx::core::cu
--- a/mlx/backend/cuda/gemms/gemv.h
+++ b/mlx/backend/cuda/gemms/gemv.h
@@ -1,24 +0,0 @@
-// 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
--- a/mlx/backend/cuda/indexing.cpp
+++ b/mlx/backend/cuda/indexing.cpp
@@ -29,12 +29,12 @@ void append_indices_arg(
    const std::vector<array>& inputs,
    int nidx,
    int idx_ndim) {
-  SmallVector<const void*> indices(nidx);
+  std::vector<const void*> indices(nidx);
  for (int i = 0; i < nidx; ++i) {
    indices[i] = inputs[i + 1].data<void>();
  }
  args.append(std::move(indices));
-  SmallVector<int32_t> indices_shape(nidx * idx_ndim);
+  std::vector<int32_t> indices_shape(nidx * idx_ndim);
  for (int i = 0; i < nidx; ++i) {
    std::copy_n(
        inputs[i + 1].shape().begin(),
@@ -42,7 +42,7 @@ void append_indices_arg(
        indices_shape.data() + i * idx_ndim);
  }
  args.append(std::move(indices_shape));
-  SmallVector<int64_t> indices_strides(nidx * idx_ndim);
+  std::vector<int64_t> indices_strides(nidx * idx_ndim);
  for (int i = 0; i < nidx; ++i) {
    std::copy_n(
        inputs[i + 1].strides().begin(),
@@ -110,7 +110,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
  args.append<int32_t>(src.ndim());
  args.append_ndim(slice_sizes_);
  args.append(slice_size);
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
  append_indices_arg(args, inputs, nidx, idx_ndim);

  std::string kernel_name = fmt::format(
@@ -128,8 +128,8 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
  encoder.set_output_array(out);

  auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(out, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
 }

 void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -211,7 +211,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
  args.append_ndim(out.shape());
  args.append_ndim(out.strides());
  args.append<int32_t>(out.ndim());
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
  append_indices_arg(args, inputs, nidx, idx_ndim);

  std::string kernel_name = fmt::format(
@@ -229,8 +229,8 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
  }
  encoder.set_output_array(out);
  auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(upd, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(kernel, upd, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
 }

 void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -317,8 +317,8 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
  }
  encoder.set_output_array(out);
  auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(idx, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
 }

 void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -421,8 +421,8 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
  }
  encoder.set_output_array(out);
  auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(idx, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
 }

 } // namespace mlx::core
--- a/mlx/backend/cuda/iterators/general_iterator.cuh
+++ b/mlx/backend/cuda/iterators/general_iterator.cuh
@@ -0,0 +1,121 @@
+// 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
--- a/mlx/backend/cuda/iterators/strided_iterator.cuh
+++ b/mlx/backend/cuda/iterators/strided_iterator.cuh
@@ -0,0 +1,60 @@
+// 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
--- a/mlx/backend/cuda/jit_module.cpp
+++ b/mlx/backend/cuda/jit_module.cpp
@@ -9,10 +9,10 @@
 #include <cstdlib>
 #include <filesystem>
 #include <fstream>
+#include <unordered_map>

 #include <fmt/format.h>
 #include <nvrtc.h>
-#include <unistd.h>

 namespace mlx::core::cu {

@@ -50,32 +50,6 @@ 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 {
@@ -136,8 +110,7 @@ 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::string& source_code) {
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
  if (cache_dir.empty()) {
    return;
  }
@@ -150,9 +123,6 @@ 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.
@@ -191,8 +161,9 @@ constexpr const char* g_include_names[] = {
    INCLUDE_PREFIX "atomic_ops.cuh",
    INCLUDE_PREFIX "binary_ops.cuh",
    INCLUDE_PREFIX "cast_op.cuh",
+    INCLUDE_PREFIX "cexpf.cuh",
    INCLUDE_PREFIX "config.h",
-    INCLUDE_PREFIX "complex.cuh",
+    INCLUDE_PREFIX "cucomplex_math.cuh",
    INCLUDE_PREFIX "fp16_math.cuh",
    INCLUDE_PREFIX "indexing.cuh",
    INCLUDE_PREFIX "scatter_ops.cuh",
@@ -207,8 +178,9 @@ constexpr const char* g_headers[] = {
    jit_source_atomic_ops,
    jit_source_binary_ops,
    jit_source_cast_op,
+    jit_source_cexpf,
    jit_source_config,
-    jit_source_complex,
+    jit_source_cucomplex_math,
    jit_source_fp16_math,
    jit_source_indexing,
    jit_source_scatter_ops,
@@ -245,24 +217,16 @@ 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());
-    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());
+    std::string include = fmt::format("--include-path={}/include", cuda_home());
+    const char* args[] = {compute.c_str(), include.c_str()};
    nvrtcResult compile_result =
-        nvrtcCompileProgram(prog, args.size(), args.data());
+        nvrtcCompileProgram(prog, std::size(args), args);
    if (compile_result != NVRTC_SUCCESS) {
      size_t log_size;
      CHECK_NVRTC_ERROR(nvrtcGetProgramLogSize(prog, &log_size));
@@ -292,8 +256,7 @@ JitModule::JitModule(
    } else {
      CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
    }
-    write_cached_ptx(
-        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
+    write_cached_ptx(ptx_cache_dir(), module_name, ptx, ptx_kernels);
  }

  // Load module.
@@ -329,16 +292,11 @@ CUfunction JitModule::get_kernel(const std::string& kernel_name) {
  return it->second;
 }

-std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
-  static std::unordered_map<std::string, JitModule> map;
-  return map;
-}
-
 JitModule& get_jit_module(
    const mlx::core::Device& device,
    const std::string& name,
    const KernelBuilder& builder) {
-  auto& map = get_jit_module_cache();
+  static std::unordered_map<std::string, JitModule> map;
  auto it = map.find(name);
  if (it == map.end()) {
    it = map.try_emplace(name, cu::device(device), name, builder).first;
--- a/mlx/backend/cuda/jit_module.h
+++ b/mlx/backend/cuda/jit_module.h
@@ -40,14 +40,19 @@ struct KernelArgs {
  }

  template <typename T>
-  void append(SmallVector<T> vec) {
-    storage_.emplace_back(std::move(vec));
-    append_ptr(std::get<SmallVector<T>>(storage_.back()).data());
+  void append(std::vector<T> vec) {
+    if (vec.empty()) {
+      // The nullptr can not be used as arg, pass something not null.
+      append(std::monostate{});
+    } else {
+      append_ptr(vec.data());
+      storage_.emplace_back(std::move(vec));
+    }
  }

  // Make sure the arg is copied to an array with size of NDIM.
  template <size_t NDIM = MAX_NDIM, typename T>
-  void append_ndim(SmallVector<T> vec) {
+  void append_ndim(std::vector<T> vec) {
    if (vec.size() > NDIM) {
      throw std::runtime_error(
          fmt::format("ndim can not be larger than {}.", NDIM));
@@ -71,9 +76,9 @@ struct KernelArgs {
      int32_t,
      uint32_t,
      int64_t,
-      SmallVector<const void*>,
-      SmallVector<int32_t>,
-      SmallVector<int64_t>>;
+      std::vector<const void*>,
+      std::vector<int32_t>,
+      std::vector<int64_t>>;
  std::deque<Arg> storage_;
 };

@@ -94,8 +99,6 @@ class JitModule {
  std::unordered_map<std::string, CUfunction> kernels_;
 };

-std::unordered_map<std::string, JitModule>& get_jit_module_cache();
-
 JitModule& get_jit_module(
    const mlx::core::Device& device,
    const std::string& name,
--- a/mlx/backend/cuda/kernel_utils.cu
+++ b/mlx/backend/cuda/kernel_utils.cu
@@ -30,25 +30,4 @@ std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2) {
  return std::make_pair(dim3(gx, gy, gz), dim3(bx, by, bz));
 }

-std::tuple<dim3, uint> get_launch_args(
-    size_t size,
-    const Shape& shape,
-    const Strides& strides,
-    bool large,
-    int work_per_thread) {
-  size_t nthreads = cuda::ceil_div(size, work_per_thread);
-  uint block_dim = 1024;
-  if (block_dim > nthreads) {
-    block_dim = nthreads;
-  }
-  dim3 num_blocks;
-  if (large) {
-    num_blocks = get_2d_grid_dims(shape, strides, work_per_thread);
-    num_blocks.x = cuda::ceil_div(num_blocks.x, block_dim);
-  } else {
-    num_blocks.x = cuda::ceil_div(nthreads, block_dim);
-  }
-  return std::make_tuple(num_blocks, block_dim);
-}
-
 } // namespace mlx::core
--- a/mlx/backend/cuda/kernel_utils.cuh
+++ b/mlx/backend/cuda/kernel_utils.cuh
@@ -11,6 +11,7 @@
 #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>
@@ -78,7 +79,7 @@ struct CTypeToCudaType<bfloat16_t> {

 template <>
 struct CTypeToCudaType<complex64_t> {
-  using type = cu::complex64_t;
+  using type = cuComplex;
 };

 template <typename T>
@@ -90,18 +91,14 @@ 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> || is_complex_v<T>;
+inline constexpr bool is_inexact_v =
+    is_floating_v<T> || cuda::std::is_same_v<T, complex64_t>;

 // Utility to copy data from vector to array in host.
 template <int NDIM = MAX_NDIM, typename T = int32_t>
-inline cuda::std::array<T, NDIM> const_param(const SmallVector<T>& vec) {
+inline cuda::std::array<T, NDIM> const_param(const std::vector<T>& vec) {
  if (vec.size() > NDIM) {
    throw std::runtime_error(
        fmt::format("ndim can not be larger than {}.", NDIM));
@@ -120,19 +117,53 @@ dim3 get_2d_grid_dims(
    size_t divisor);
 std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2);

+// Return a block size that achieves maximum potential occupancy for kernel.
+template <typename T>
+inline uint max_occupancy_block_dim(T kernel) {
+  int _, block_dim;
+  if constexpr (std::is_same_v<T, CUfunction>) {
+    CHECK_CUDA_ERROR(
+        cuOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel, 0, 0, 0));
+  } else {
+    CHECK_CUDA_ERROR(
+        cudaOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel));
+  }
+  return block_dim;
+}
+
 // Get the num_blocks and block_dims that maximize occupancy for |kernel|,
 // assuming each thread handles |work_per_thread| elements of |arr|.
-std::tuple<dim3, uint> get_launch_args(
+template <typename T>
+inline std::tuple<dim3, uint> get_launch_args(
+    T kernel,
    size_t size,
    const Shape& shape,
    const Strides& strides,
    bool large,
-    int work_per_thread = 1);
+    int work_per_thread = 1) {
+  size_t nthreads = cuda::ceil_div(size, work_per_thread);
+  uint block_dim = max_occupancy_block_dim(kernel);
+  if (block_dim > nthreads) {
+    block_dim = nthreads;
+  }
+  dim3 num_blocks;
+  if (large) {
+    num_blocks = get_2d_grid_dims(shape, strides, work_per_thread);
+    num_blocks.x = cuda::ceil_div(num_blocks.x, block_dim);
+  } else {
+    num_blocks.x = cuda::ceil_div(nthreads, block_dim);
+  }
+  return std::make_tuple(num_blocks, block_dim);
+}

-inline std::tuple<dim3, uint>
-get_launch_args(const array& arr, bool large, int work_per_thread = 1) {
+template <typename T>
+inline std::tuple<dim3, uint> get_launch_args(
+    T kernel,
+    const array& arr,
+    bool large,
+    int work_per_thread = 1) {
  return get_launch_args(
-      arr.size(), arr.shape(), arr.strides(), large, work_per_thread);
+      kernel, arr.size(), arr.shape(), arr.strides(), large, work_per_thread);
 }

 } // namespace mlx::core
--- a/mlx/backend/cuda/layer_norm.cu
+++ b/mlx/backend/cuda/layer_norm.cu
@@ -1,6 +1,7 @@
 // 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"
@@ -10,6 +11,8 @@
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 #include <nvtx3/nvtx3.hpp>
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_reduce.cuh>

 namespace mlx::core {

@@ -72,11 +75,9 @@ __global__ void layer_norm(
  float sum = 0;
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-#pragma unroll
-    for (int i = 0; i < N_READS; ++i) {
-      sum += static_cast<float>(xn[i]);
-    }
+    T xn[N_READS] = {};
+    cub::LoadDirectBlocked(index, x, xn, axis_size);
+    sum += static_cast<float>(cub::ThreadReduce(xn, cuda::std::plus<>{}));
  }
  sum = BlockReduceT{block, temp}.Sum(sum);

@@ -87,18 +88,11 @@ __global__ void layer_norm(
  float normalizer = 0;
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    if ((index + 1) * N_READS <= axis_size) {
-      auto xn = load_vector<N_READS>(x, index);
-#pragma unroll
-      for (int i = 0; i < N_READS; ++i) {
-        float t = static_cast<float>(xn[i]) - mean;
-        normalizer += t * t;
-      }
-    } else {
-      for (int i = index * N_READS; i < axis_size; ++i) {
-        float t = static_cast<float>(x[i]) - mean;
-        normalizer += t * t;
-      }
+    T xn[N_READS];
+    cub::LoadDirectBlocked(index, x, xn, axis_size, mean);
+    for (int i = 0; i < N_READS; ++i) {
+      float t = static_cast<float>(xn[i]) - mean;
+      normalizer += t * t;
    }
  }
  normalizer = BlockReduceT{block, temp}.Sum(normalizer);
@@ -107,15 +101,17 @@ __global__ void layer_norm(
  // Outputs.
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
-    auto bn = load_vector<N_READS>(b, index, axis_size, b_stride, T(0));
-#pragma unroll
+    T xn[N_READS];
+    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);
    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];
    }
-    store_vector<N_READS>(out, index, xn, axis_size);
+    cub::StoreDirectBlocked(index, out, xn, axis_size);
  }
 }

@@ -148,11 +144,9 @@ __global__ void layer_norm_vjp(
  float sum = 0;
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-#pragma unroll
-    for (int i = 0; i < N_READS; ++i) {
-      sum += static_cast<float>(xn[i]);
-    }
+    T xn[N_READS] = {};
+    cub::LoadDirectBlocked(index, x, xn, axis_size);
+    sum += static_cast<float>(cub::ThreadReduce(xn, cuda::std::plus<>{}));
  }
  sum = BlockReduceF{block, temp.f}.Sum(sum);

@@ -162,28 +156,19 @@ __global__ void layer_norm_vjp(
  // Normalizer.
  float3 factors = {};
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    T xn[N_READS];
+    T wn[N_READS] = {};
+    T gn[N_READS] = {};
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
-    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
-
-    if ((index + 1) * N_READS <= axis_size) {
-      auto xn = load_vector<N_READS>(x, index);
-#pragma unroll
-      for (int i = 0; i < N_READS; ++i) {
-        float t = static_cast<float>(xn[i]) - mean;
-        float wi = wn[i];
-        float gi = gn[i];
-        float wg = wi * gi;
-        factors = plus_f3(factors, {wg, wg * t, t * t});
-      }
-    } else {
-      for (int i = index * N_READS; i < axis_size; ++i) {
-        float t = static_cast<float>(x[i]) - mean;
-        float wi = wn[i];
-        float gi = gn[i];
-        float wg = wi * gi;
-        factors = plus_f3(factors, {wg, wg * t, t * t});
-      }
+    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);
+    for (int i = 0; i < N_READS; i++) {
+      float t = static_cast<float>(xn[i]) - mean;
+      float wi = wn[i];
+      float gi = gn[i];
+      float wg = wi * gi;
+      factors = plus_f3(factors, {wg, wg * t, t * t});
    }
  }
  factors = BlockReduceF3{block, temp.f3}.Reduce(factors, plus_f3, {});
@@ -195,10 +180,12 @@ __global__ void layer_norm_vjp(
  // Outputs.
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
-    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
-
+    T xn[N_READS];
+    T wn[N_READS];
+    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);
    for (int i = 0; i < N_READS; i++) {
      float xi = (static_cast<float>(xn[i]) - mean) * normalizer;
      float wi = wn[i];
@@ -208,9 +195,9 @@ __global__ void layer_norm_vjp(
        wn[i] = gi * xi;
      }
    }
-    store_vector<N_READS>(gx, index, xn, axis_size);
+    cub::StoreDirectBlocked(index, gx, xn, axis_size);
    if constexpr (HAS_W) {
-      store_vector<N_READS>(gw, index, wn, axis_size);
+      cub::StoreDirectBlocked(index, gw, wn, axis_size);
    }
  }
 }
@@ -250,7 +237,8 @@ void LayerNorm::eval_gpu(
      }
      return x;
    } else {
-      array x_copy = contiguous_copy_gpu(x, s);
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy_gpu(x, x_copy, CopyType::General, s);
      out.copy_shared_buffer(x_copy);
      return x_copy;
    }
@@ -271,15 +259,14 @@ void LayerNorm::eval_gpu(
  encoder.set_input_array(b);
  encoder.set_output_array(out);
  dispatch_float_types(out.dtype(), "layernorm", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    constexpr int N_READS = 16 / sizeof(DataType);
+    constexpr uint32_t N_READS = 4;
    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
      auto kernel = cu::layer_norm<DataType, block_dim(), N_READS>;
      encoder.add_kernel_node(
          kernel,
          n_rows,
          block_dim(),
-          0,
          x.data<DataType>(),
          w.data<DataType>(),
          b.data<DataType>(),
@@ -308,7 +295,9 @@ void LayerNormVJP::eval_gpu(
      return x;
    }
    copied = true;
-    return contiguous_copy_gpu(x, s);
+    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    copy_gpu(x, x_copy, CopyType::General, s);
+    return x_copy;
  };
  bool donate_x = inputs[0].is_donatable();
  bool donate_g = inputs[3].is_donatable();
@@ -379,10 +368,10 @@ void LayerNormVJP::eval_gpu(
  encoder.set_output_array(gw_temp);
  dispatch_float_types(gx.dtype(), "layernorm_vjp", [&](auto type_tag) {
    dispatch_bool(has_w, [&](auto has_w_constant) {
-      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-      constexpr int N_READS = 16 / sizeof(DataType);
+      constexpr int N_READS = 4;
      dispatch_block_dim(
          cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+            using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
            auto kernel = cu::layer_norm_vjp<
                DataType,
                has_w_constant.value,
@@ -392,7 +381,6 @@ void LayerNormVJP::eval_gpu(
                kernel,
                n_rows,
                block_dim(),
-                0,
                x.data<DataType>(),
                w.data<DataType>(),
                g.data<DataType>(),
--- a/mlx/backend/cuda/logsumexp.cu
+++ b/mlx/backend/cuda/logsumexp.cu
@@ -43,19 +43,20 @@ __global__ void logsumexp(const T* in, T* out, int axis_size) {
  AccT maxval = Limits<AccT>::finite_min();
  AccT normalizer = 0;
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
-    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto vals = load_vector<N_READS>(in, index, axis_size, Limits<T>::min());
+    AccT vals[N_READS];
+    cub::LoadDirectBlocked(
+        r * BLOCK_DIM + block.thread_rank(),
+        make_cast_iterator<AccT>(in),
+        vals,
+        axis_size,
+        Limits<AccT>::min());
    prevmax = maxval;
-#pragma unroll
-    for (int i = 0; i < N_READS; ++i) {
-      maxval = max_op(maxval, static_cast<AccT>(vals[i]));
-    }
+    maxval = max_op(maxval, cub::ThreadReduce(vals, max_op));
    // Online normalizer calculation for softmax:
    // https://github.com/NVIDIA/online-softmax
    normalizer = normalizer * softmax_exp(prevmax - maxval);
    for (int i = 0; i < N_READS; i++) {
-      normalizer =
-          normalizer + softmax_exp(static_cast<AccT>(vals[i]) - maxval);
+      normalizer = normalizer + softmax_exp(vals[i] - maxval);
    }
  }

@@ -107,7 +108,8 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
    if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
      return x;
    } else {
-      array x_copy = contiguous_copy_gpu(x, s);
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy_gpu(x, x_copy, CopyType::General, s);
      encoder.add_temporary(x_copy);
      return x_copy;
    }
@@ -142,15 +144,14 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
  encoder.set_input_array(in);
  encoder.set_output_array(out);
  dispatch_float_types(out.dtype(), "logsumexp", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    constexpr int N_READS = 16 / sizeof(DataType);
+    constexpr int N_READS = 4;
    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
      auto kernel = cu::logsumexp<DataType, float, block_dim(), N_READS>;
      encoder.add_kernel_node(
          kernel,
          n_rows,
          block_dim(),
-          0,
          in.data<DataType>(),
          out.data<DataType>(),
          axis_size);
--- a/mlx/backend/cuda/lru_cache.h
+++ b/mlx/backend/cuda/lru_cache.h
@@ -1,159 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include <cstring>
-#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) {
-    if (capacity == 0) {
-      throw std::runtime_error("LRUCache requires capacity > 0.");
-    }
-  }
-
-  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);
-  }
-
-  V& operator[](const K& key) {
-    auto it = find(key);
-    if (it == end()) {
-      it = emplace(key, V{}).first;
-    }
-    return it->second;
-  }
-
- 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
--- a/mlx/backend/cuda/matmul.cpp
+++ b/mlx/backend/cuda/matmul.cpp
@@ -2,15 +2,275 @@

 #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>
@@ -22,7 +282,8 @@ 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 = contiguous_copy_gpu(arr, s);
+    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+    copy_gpu(arr, arr_copy, CopyType::General, s);
    enc.add_temporary(arr_copy);
    return std::make_tuple(false, arr.shape(-1), arr_copy);
  }
@@ -79,25 +340,10 @@ 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,
@@ -112,13 +358,27 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
      a_batch_strides.back(),
      b_batch_strides.back());

-  if ((batch_count / batch_shape.back()) == 1) {
-    matmul.run(encoder, out, a, b);
+  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>());
    return;
  }

-  matmul.run_batched(
-      encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides);
+  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();
+  }
 }

 void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -129,7 +389,9 @@ 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 = inputs[2];
+  auto& c_pre = inputs[2];
+
+  out.set_data(allocator::malloc(out.nbytes()));

  /////////////////////////////////////////////////////////////////////////////
  // Init checks and prep
@@ -142,24 +404,7 @@ 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);
-
-  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;
-    }
-  }
+  auto [c_transposed, ldc, c] = check_transpose(encoder, s, c_pre);

  /////////////////////////////////////////////////////////////////////////////
  // Check and collapse batch dimensions
@@ -186,7 +431,7 @@ 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,
@@ -197,28 +442,48 @@ 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());

-  if ((batch_count / batch_shape.back()) == 1) {
-    matmul.run(encoder, out, a, b, c, alpha_, beta_);
+  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_);
    return;
  }
-  matmul.run_batched(
-      encoder,
-      out,
-      a,
-      b,
-      c,
-      batch_shape,
-      a_batch_strides,
-      b_batch_strides,
-      c_batch_strides,
-      alpha_,
-      beta_);
+
+  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();
+  }
 }

 } // namespace mlx::core
--- a/mlx/backend/cuda/primitives.cpp
+++ b/mlx/backend/cuda/primitives.cpp
@@ -1,55 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/distributed/primitives.h"
-#include "mlx/fast_primitives.h"
-#include "mlx/primitives.h"
-
-namespace mlx::core {
-
-#define NO_GPU_MULTI(func)                                             \
-  void func::eval_gpu(                                                 \
-      const std::vector<array>& inputs, std::vector<array>& outputs) { \
-    throw std::runtime_error(#func " has no CUDA implementation.");    \
-  }
-
-#define NO_GPU_USE_FALLBACK(func)     \
-  bool func::use_fallback(Stream s) { \
-    return true;                      \
-  }                                   \
-  NO_GPU_MULTI(func)
-
-#define NO_GPU(func)                                                  \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
-    throw std::runtime_error(#func " has no CUDA implementation.");   \
-  }
-
-NO_GPU(BlockMaskedMM)
-NO_GPU(DynamicSlice)
-NO_GPU(DynamicSliceUpdate)
-NO_GPU(FFT)
-NO_GPU(GatherMM)
-NO_GPU(GatherQMM)
-NO_GPU(Hadamard)
-NO_GPU(Load)
-NO_GPU_MULTI(LUF)
-NO_GPU_MULTI(QRF)
-NO_GPU(QuantizedMatmul)
-NO_GPU(SegmentedMM)
-NO_GPU_MULTI(SVD)
-NO_GPU(Inverse)
-NO_GPU(Cholesky)
-NO_GPU_MULTI(Eig)
-NO_GPU_MULTI(Eigh)
-
-namespace fast {
-NO_GPU_MULTI(CustomKernel)
-} // namespace fast
-
-namespace distributed {
-NO_GPU_MULTI(AllReduce)
-NO_GPU_MULTI(AllGather)
-NO_GPU_MULTI(Send)
-NO_GPU_MULTI(Recv)
-} // namespace distributed
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/primitives.cu
+++ b/mlx/backend/cuda/primitives.cu
@@ -0,0 +1,105 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/arange.cuh"
+#include "mlx/backend/cuda/device/fp16_math.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/distributed/primitives.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/fast_primitives.h"
+#include "mlx/primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+#include <thrust/device_ptr.h>
+#include <thrust/transform.h>
+
+#include <cassert>
+
+namespace mlx::core {
+
+void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Arange::eval_gpu");
+  assert(inputs.size() == 0);
+  out.set_data(allocator::malloc(out.nbytes()));
+  if (out.size() == 0) {
+    return;
+  }
+  auto& encoder = cu::get_command_encoder(stream());
+  encoder.set_output_array(out);
+  auto capture = encoder.capture_context();
+  dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
+    using CTYPE = MLX_GET_TYPE(type_tag);
+    using OutType = cuda_type_t<CTYPE>;
+    CTYPE step =
+        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
+    thrust::transform(
+        cu::thrust_policy(encoder.stream()),
+        thrust::counting_iterator<uint32_t>(0),
+        thrust::counting_iterator<uint32_t>(out.data_size()),
+        thrust::device_pointer_cast(out.data<OutType>()),
+        cu::Arange<OutType>{
+            static_cast<OutType>(start_), static_cast<OutType>(step)});
+  });
+}
+
+bool fast::ScaledDotProductAttention::use_fallback(
+    const array& q,
+    const array& k,
+    const array& v,
+    bool has_mask,
+    bool has_arr_mask,
+    bool do_causal,
+    Stream s) {
+  return true;
+}
+
+#define NO_GPU_MULTI(func)                                             \
+  void func::eval_gpu(                                                 \
+      const std::vector<array>& inputs, std::vector<array>& outputs) { \
+    throw std::runtime_error(#func " has no CUDA implementation.");    \
+  }
+
+#define NO_GPU_USE_FALLBACK(func)     \
+  bool func::use_fallback(Stream s) { \
+    return true;                      \
+  }                                   \
+  NO_GPU_MULTI(func)
+
+#define NO_GPU(func)                                                  \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    throw std::runtime_error(#func " has no CUDA implementation.");   \
+  }
+
+NO_GPU(BlockMaskedMM)
+NO_GPU(Convolution)
+NO_GPU(DynamicSlice)
+NO_GPU(DynamicSliceUpdate)
+NO_GPU(FFT)
+NO_GPU(GatherMM)
+NO_GPU(GatherQMM)
+NO_GPU(Hadamard)
+NO_GPU(Load)
+NO_GPU_MULTI(LUF)
+NO_GPU_MULTI(QRF)
+NO_GPU(QuantizedMatmul)
+NO_GPU(SegmentedMM)
+NO_GPU_MULTI(SVD)
+NO_GPU(Inverse)
+NO_GPU(Cholesky)
+NO_GPU_MULTI(Eig)
+NO_GPU_MULTI(Eigh)
+
+namespace fast {
+NO_GPU(ScaledDotProductAttention)
+NO_GPU_MULTI(AffineQuantize)
+NO_GPU_MULTI(CustomKernel)
+} // namespace fast
+
+namespace distributed {
+NO_GPU_MULTI(AllReduce)
+NO_GPU_MULTI(AllGather)
+NO_GPU_MULTI(Send)
+NO_GPU_MULTI(Recv)
+} // namespace distributed
+
+} // namespace mlx::core
--- a/mlx/backend/cuda/quantized/affine_quantize.cu
+++ b/mlx/backend/cuda/quantized/affine_quantize.cu
@@ -1,331 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/backend/cuda/quantized/quantized_utils.cuh"
-#include "mlx/dtype_utils.h"
-
-#include <cooperative_groups.h>
-#include <cooperative_groups/reduce.h>
-
-namespace mlx::core {
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-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
-
-void affine_quantize(
-    const array& w,
-    array& wq,
-    array& scales,
-    array& biases,
-    int group_size_,
-    int bits_,
-    cu::CommandEncoder& enc,
-    const Stream& s) {
-  // Calculate the number of elements per thread
-  int per_thread = group_size_ / WARP_SIZE;
-  size_t size = w.size() / per_thread;
-
-  // Calculate the thread grid that we need to launch
-  bool large = size > UINT_MAX;
-  auto grid_shape = w.shape();
-  grid_shape.back() /= per_thread;
-
-  enc.set_input_array(w);
-  enc.set_output_array(wq);
-  enc.set_output_array(scales);
-  enc.set_output_array(biases);
-  dispatch_float_types(w.dtype(), "affine_quantize", [&](auto type_tag) {
-    dispatch_groups(group_size_, [&](auto group_size) {
-      dispatch_bits(bits_, [&](auto bits) {
-        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-        auto kernel = cu::affine_quantize<T, group_size.value, bits.value>;
-        auto [num_blocks, block_dims] =
-            get_launch_args(size, grid_shape, w.strides(), large);
-        enc.add_kernel_node(
-            kernel,
-            num_blocks,
-            block_dims,
-            0,
-            w.data<T>(),
-            wq.data<uint8_t>(),
-            scales.data<T>(),
-            biases.data<T>(),
-            w.size());
-      });
-    });
-  });
-}
-
-void affine_dequantize(
-    const array& wq,
-    const array& scales,
-    const array& biases,
-    array& w,
-    int group_size_,
-    int bits_,
-    cu::CommandEncoder& enc,
-    const Stream& s) {
-  // Calculate how many numbers we pack together. For 2, 4, 8 bits we pack in
-  // one uint8, for 3, 6 in 3 uint8 and for 5 in 5 uint8.
-  constexpr int uint8_per_uint32 = 4;
-  int packs_per_int;
-  switch (bits_) {
-    case 3:
-    case 5:
-      packs_per_int = 8;
-      break;
-    case 6:
-      packs_per_int = 4;
-      break;
-    default:
-      packs_per_int = 8 / bits_;
-  }
-
-  size_t size = w.size() / packs_per_int;
-  bool large = size > UINT_MAX;
-  auto grid_shape = w.shape();
-  grid_shape.back() *= uint8_per_uint32;
-
-  enc.set_input_array(wq);
-  enc.set_input_array(scales);
-  enc.set_input_array(biases);
-  enc.set_output_array(w);
-  dispatch_float_types(w.dtype(), "affine_quantize", [&](auto type_tag) {
-    dispatch_groups(group_size_, [&](auto group_size) {
-      dispatch_bits(bits_, [&](auto bits) {
-        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-        auto kernel = cu::affine_dequantize<T, group_size.value, bits.value>;
-        auto [num_blocks, block_dims] =
-            get_launch_args(size, grid_shape, w.strides(), large);
-        enc.add_kernel_node(
-            kernel,
-            num_blocks,
-            block_dims,
-            0,
-            wq.data<uint8_t>(),
-            scales.data<T>(),
-            biases.data<T>(),
-            w.data<T>(),
-            w.size());
-      });
-    });
-  });
-}
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/quantized/quantized.cpp
+++ b/mlx/backend/cuda/quantized/quantized.cpp
@@ -1,80 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/quantized/quantized.h"
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/gpu/copy.h"
-#include "mlx/fast_primitives.h"
-
-#include <nvtx3/nvtx3.hpp>
-
-namespace mlx::core {
-
-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;
-  }
-}
-
-inline array ensure_row_contiguous_matrix(
-    const array& x,
-    cu::CommandEncoder& enc,
-    const Stream& s) {
-  if (x.ndim() < 2) {
-    if (x.strides()[0] == 1) {
-      return x;
-    }
-  } else {
-    auto stride_0 = x.strides()[x.ndim() - 2];
-    auto stride_1 = x.strides()[x.ndim() - 1];
-    if (stride_0 == x.shape(-1) && stride_1 == 1) {
-      return x;
-    }
-  }
-  array x_copy = contiguous_copy_gpu(x, s);
-  enc.add_temporary(x_copy);
-  return x_copy;
-}
-
-} // namespace
-
-void fast::AffineQuantize::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  nvtx3::scoped_range r("AffineQuantize::eval_gpu");
-  auto& s = stream();
-  auto& d = cu::device(s.device);
-  auto& enc = d.get_command_encoder(s);
-
-  if (dequantize_) {
-    auto wq = ensure_row_contiguous(inputs[0], enc, s);
-    auto scales = ensure_row_contiguous(inputs[1], enc, s);
-    auto biases = ensure_row_contiguous(inputs[2], enc, s);
-    auto& w = outputs[0];
-
-    w.set_data(allocator::malloc(w.nbytes()));
-
-    affine_dequantize(wq, scales, biases, w, group_size_, bits_, enc, s);
-  } else {
-    auto w = ensure_row_contiguous(inputs[0], enc, s);
-    auto& wq = outputs[0];
-    auto& scales = outputs[1];
-    auto& biases = outputs[2];
-
-    wq.set_data(allocator::malloc(wq.nbytes()));
-    scales.set_data(allocator::malloc(scales.nbytes()));
-    biases.set_data(allocator::malloc(biases.nbytes()));
-
-    affine_quantize(w, wq, scales, biases, group_size_, bits_, enc, s);
-  }
-}
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/quantized/quantized.h
+++ b/mlx/backend/cuda/quantized/quantized.h
@@ -1,27 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-
-namespace mlx::core {
-
-void affine_quantize(
-    const array& w,
-    array& wq,
-    array& scales,
-    array& biases,
-    int group_size_,
-    int bits_,
-    cu::CommandEncoder& enc,
-    const Stream& s);
-
-void affine_dequantize(
-    const array& wq,
-    const array& scales,
-    const array& biases,
-    array& w,
-    int group_size_,
-    int bits_,
-    cu::CommandEncoder& enc,
-    const Stream& s);
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/quantized/quantized_utils.cuh
+++ b/mlx/backend/cuda/quantized/quantized_utils.cuh
@@ -1,59 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-namespace mlx::core {
-
-namespace cu {
-
-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);
-}
-
-} // namespace cu
-
-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;
-  }
-}
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/random.cu
+++ b/mlx/backend/cuda/random.cu
@@ -170,7 +170,6 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
        cu::rbitsc,
        grid,
        block,
-        0,
        keys.data<uint32_t>(),
        out.data<uint8_t>(),
        grid_dims,
@@ -181,7 +180,6 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
        cu::rbits,
        grid,
        block,
-        0,
        keys.data<uint32_t>(),
        out.data<uint8_t>(),
        grid_dims,
--- a/mlx/backend/cuda/reduce.cu
+++ b/mlx/backend/cuda/reduce.cu
@@ -5,6 +5,8 @@
 #include "mlx/backend/gpu/copy.h"

 #include <nvtx3/nvtx3.hpp>
+#include <thrust/device_ptr.h>
+#include <thrust/fill.h>

 #include <cassert>

@@ -45,7 +47,8 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
    }
  }
  if (plan.type == GeneralReduce || broadcasted || !in.flags().contiguous) {
-    array in_copy = contiguous_copy_gpu(in, s);
+    array in_copy(in.shape(), in.dtype(), nullptr, {});
+    copy_gpu(in, in_copy, CopyType::General, s);
    encoder.add_temporary(in_copy);
    in = in_copy;
    plan = get_reduction_plan(in, axes_);
--- a/mlx/backend/cuda/reduce/all_reduce.cu
+++ b/mlx/backend/cuda/reduce/all_reduce.cu
@@ -120,7 +120,6 @@ void all_reduce(
            kernel,
            blocks,
            threads,
-            0,
            static_cast<T*>(indata),
            intermediate.data<U>(),
            block_step,
@@ -147,7 +146,6 @@ void all_reduce(
          kernel,
          blocks,
          threads,
-          0,
          static_cast<T*>(indata),
          out.data<U>(),
          block_step,
--- a/mlx/backend/cuda/reduce/col_reduce.cu
+++ b/mlx/backend/cuda/reduce/col_reduce.cu
@@ -230,7 +230,7 @@ void col_reduce_looped(
        auto kernel =
            cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
        encoder.add_kernel_node(
-            kernel, grid, blocks, 0, indata, out.data<U>(), args);
+            kernel, grid, blocks, indata, out.data<U>(), args);
      });
    });
  });
--- a/mlx/backend/cuda/reduce/init_reduce.cu
+++ b/mlx/backend/cuda/reduce/init_reduce.cu
@@ -41,8 +41,7 @@ void init_reduce(
      dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
      dim3 block(grid.x < 1024 ? grid.x : 1024, 1, 1);
      grid.x = (grid.x + 1023) / 1024;
-      encoder.add_kernel_node(
-          kernel, grid, block, 0, out.data<U>(), out.size());
+      encoder.add_kernel_node(kernel, grid, block, out.data<U>(), out.size());
    });
  });
 }
--- a/mlx/backend/cuda/reduce/reduce.cuh
+++ b/mlx/backend/cuda/reduce/reduce.cuh
@@ -3,6 +3,7 @@
 #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"
--- a/mlx/backend/cuda/reduce/reduce_ops.cuh
+++ b/mlx/backend/cuda/reduce/reduce_ops.cuh
@@ -69,18 +69,6 @@ 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;
  }

@@ -93,18 +81,6 @@ 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;
  }

@@ -175,7 +151,7 @@ struct ReduceInit<Or, T> {
 template <typename T>
 struct ReduceInit<Sum, T> {
  static constexpr __host__ __device__ auto value() {
-    if constexpr (is_complex_v<T>) {
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
      return T{0, 0};
    } else {
      return cast_to<typename ReduceResult<Sum, T>::type>(0);
@@ -186,7 +162,7 @@ struct ReduceInit<Sum, T> {
 template <typename T>
 struct ReduceInit<Prod, T> {
  static constexpr __host__ __device__ auto value() {
-    if constexpr (is_complex_v<T>) {
+    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
      return T{1, 0};
    } else {
      return cast_to<typename ReduceResult<Prod, T>::type>(1);
--- a/mlx/backend/cuda/reduce/row_reduce.cu
+++ b/mlx/backend/cuda/reduce/row_reduce.cu
@@ -269,7 +269,7 @@ void row_reduce_simple(

      int size = plan.shape.back();
      encoder.add_kernel_node(
-          kernel, grid, block, 0, indata, out.data<U>(), out.size(), size);
+          kernel, grid, block, indata, out.data<U>(), out.size(), size);
    });
  });
 }
@@ -322,7 +322,7 @@ void row_reduce_looped(
      });

      encoder.add_kernel_node(
-          kernel, grid, block, 0, indata, out.data<U>(), out.size(), args);
+          kernel, grid, block, indata, out.data<U>(), out.size(), args);
    });
  });
 }
--- a/mlx/backend/cuda/rms_norm.cu
+++ b/mlx/backend/cuda/rms_norm.cu
@@ -1,6 +1,7 @@
 // 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"
@@ -10,6 +11,8 @@
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 #include <nvtx3/nvtx3.hpp>
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_reduce.cuh>

 namespace mlx::core {

@@ -55,7 +58,7 @@ __global__ void rms_norm(
    const T* w,
    T* out,
    float eps,
-    uint32_t axis_size,
+    int32_t axis_size,
    int64_t w_stride) {
  auto grid = cg::this_grid();
  auto block = cg::this_thread_block();
@@ -70,8 +73,8 @@ __global__ void rms_norm(
  float normalizer = 0;
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-#pragma unroll
+    T xn[N_READS];
+    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;
@@ -83,14 +86,15 @@ __global__ void rms_norm(
  // Outputs.
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
-#pragma unroll
+    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);
    for (int i = 0; i < N_READS; ++i) {
-      float y = static_cast<float>(xn[i]) * normalizer;
-      xn[i] = wn[i] * static_cast<T>(y);
+      float norm = static_cast<float>(xn[i]) * normalizer;
+      xn[i] = wn[i] * static_cast<T>(norm);
    }
-    store_vector<N_READS>(out, index, xn, axis_size);
+    cub::StoreDirectBlocked(index, out, xn, axis_size);
  }
 }

@@ -122,10 +126,13 @@ __global__ void rms_norm_vjp(
  // Normalizer.
  float2 factors = {};
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
+    T xn[N_READS];
+    T wn[N_READS] = {};
+    T gn[N_READS] = {};
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
-    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(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);
    for (int i = 0; i < N_READS; i++) {
      float t = static_cast<float>(xn[i]);
      float wi = wn[i];
@@ -142,9 +149,12 @@ __global__ void rms_norm_vjp(
  // Outputs.
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
-    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
-    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+    T xn[N_READS];
+    T wn[N_READS];
+    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);
    for (int i = 0; i < N_READS; i++) {
      float xi = xn[i];
      float wi = wn[i];
@@ -154,9 +164,9 @@ __global__ void rms_norm_vjp(
        wn[i] = static_cast<T>(gi * xi * normalizer);
      }
    }
-    store_vector<N_READS>(gx, index, xn, axis_size);
+    cub::StoreDirectBlocked(index, gx, xn, axis_size);
    if constexpr (HAS_W) {
-      store_vector<N_READS>(gw, index, wn, axis_size);
+      cub::StoreDirectBlocked(index, gw, wn, axis_size);
    }
  }
 }
@@ -196,7 +206,8 @@ void RMSNorm::eval_gpu(
      }
      return x;
    } else {
-      array x_copy = contiguous_copy_gpu(x, s);
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy_gpu(x, x_copy, CopyType::General, s);
      out.copy_shared_buffer(x_copy);
      return x_copy;
    }
@@ -214,15 +225,14 @@ void RMSNorm::eval_gpu(
  encoder.set_input_array(w);
  encoder.set_output_array(out);
  dispatch_float_types(out.dtype(), "rms_norm", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    constexpr int N_READS = 16 / sizeof(DataType);
+    constexpr uint32_t N_READS = 4;
    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
      auto kernel = cu::rms_norm<DataType, block_dim(), N_READS>;
      encoder.add_kernel_node(
          kernel,
          n_rows,
          block_dim(),
-          0,
          x.data<DataType>(),
          w.data<DataType>(),
          out.data<DataType>(),
@@ -249,7 +259,9 @@ void RMSNormVJP::eval_gpu(
      return x;
    }
    copied = true;
-    return contiguous_copy_gpu(x, s);
+    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    copy_gpu(x, x_copy, CopyType::General, s);
+    return x_copy;
  };
  bool donate_x = inputs[0].is_donatable();
  bool donate_g = inputs[2].is_donatable();
@@ -304,10 +316,11 @@ void RMSNormVJP::eval_gpu(
  encoder.set_output_array(gw_temp);
  dispatch_float_types(gx.dtype(), "rms_norm_vjp", [&](auto type_tag) {
    dispatch_bool(has_w, [&](auto has_w_constant) {
-      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-      constexpr int N_READS = 16 / sizeof(DataType);
+      constexpr int N_READS = 4;
      dispatch_block_dim(
          cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+            using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+            constexpr int N_READS = 4;
            auto kernel = cu::rms_norm_vjp<
                DataType,
                has_w_constant.value,
@@ -317,7 +330,6 @@ void RMSNormVJP::eval_gpu(
                kernel,
                n_rows,
                block_dim(),
-                0,
                x.data<DataType>(),
                w.data<DataType>(),
                g.data<DataType>(),
--- a/mlx/backend/cuda/rope.cu
+++ b/mlx/backend/cuda/rope.cu
@@ -325,7 +325,6 @@ void RoPE::eval_gpu(
              kernel,
              grid,
              block,
-              0,
              (donated ? out : in).data<DataType>(),
              out.data<DataType>(),
              offset.data<int32_t>(),
@@ -342,7 +341,6 @@ void RoPE::eval_gpu(
              kernel,
              grid,
              block,
-              0,
              (donated ? out : in).data<DataType>(),
              out.data<DataType>(),
              offset.data<int32_t>(),
@@ -362,7 +360,6 @@ void RoPE::eval_gpu(
              kernel,
              grid,
              block,
-              0,
              (donated ? out : in).data<DataType>(),
              out.data<DataType>(),
              offset.data<int32_t>(),
@@ -384,7 +381,6 @@ void RoPE::eval_gpu(
              kernel,
              grid,
              block,
-              0,
              (donated ? out : in).data<DataType>(),
              out.data<DataType>(),
              offset.data<int32_t>(),
--- a/mlx/backend/cuda/scaled_dot_product_attention.cu
+++ b/mlx/backend/cuda/scaled_dot_product_attention.cu
--- a/mlx/backend/cuda/scan.cu
+++ b/mlx/backend/cuda/scan.cu
@@ -379,7 +379,9 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
          in.flags());
    }
  } else {
-    in = contiguous_copy_gpu(in, s);
+    array arr_copy(in.shape(), in.dtype(), nullptr, {});
+    copy_gpu(in, arr_copy, CopyType::General, s);
+    in = std::move(arr_copy);
    out.copy_shared_buffer(in);
  }

@@ -414,7 +416,6 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
                  kernel,
                  in.data_size() / axis_size,
                  block_dim,
-                  0,
                  in.data<T>(),
                  out.data<U>(),
                  axis_size);
@@ -444,7 +445,6 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
                  kernel,
                  num_blocks,
                  block_dim,
-                  0,
                  in.data<T>(),
                  out.data<U>(),
                  axis_size,
--- a/mlx/backend/cuda/softmax.cu
+++ b/mlx/backend/cuda/softmax.cu
@@ -11,6 +11,7 @@
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 #include <nvtx3/nvtx3.hpp>
+#include <cub/block/block_load.cuh>

 #include <cassert>

@@ -44,21 +45,20 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
  AccT maxval = Limits<AccT>::finite_min();
  AccT normalizer = cast_to<AccT>(0);
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
-    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto vals = load_vector<N_READS>(in, index, axis_size, Limits<T>::min());
+    AccT vals[N_READS];
+    cub::LoadDirectBlocked(
+        r * BLOCK_DIM + block.thread_rank(),
+        make_cast_iterator<AccT>(in),
+        vals,
+        axis_size,
+        Limits<AccT>::min());
    prevmax = maxval;
-#pragma unroll
-    for (int i = 0; i < N_READS; ++i) {
-      maxval = max_op(maxval, static_cast<AccT>(vals[i]));
-    }
-
+    maxval = max_op(maxval, cub::ThreadReduce(vals, max_op));
    // Online normalizer calculation for softmax:
    // https://github.com/NVIDIA/online-softmax
    normalizer = normalizer * softmax_exp(prevmax - maxval);
-#pragma unroll
    for (int i = 0; i < N_READS; i++) {
-      normalizer =
-          normalizer + softmax_exp(static_cast<AccT>(vals[i]) - maxval);
+      normalizer = normalizer + softmax_exp(vals[i] - maxval);
    }
  }

@@ -95,11 +95,12 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
  // Write output.
  for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
    auto index = r * BLOCK_DIM + block.thread_rank();
-    auto vals = load_vector<N_READS>(in, index, axis_size, T(0));
+    T vals[N_READS];
+    cub::LoadDirectBlocked(index, in, vals, axis_size);
    for (int i = 0; i < N_READS; i++) {
      vals[i] = softmax_exp(static_cast<AccT>(vals[i]) - maxval) * normalizer;
    }
-    store_vector<N_READS>(out, index, vals, axis_size);
+    cub::StoreDirectBlocked(index, out, vals, axis_size);
  }
 }

@@ -124,7 +125,8 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
      }
      return x;
    } else {
-      array x_copy = contiguous_copy_gpu(x, s);
+      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      copy_gpu(x, x_copy, CopyType::General, s);
      out.copy_shared_buffer(x_copy);
      return x_copy;
    }
@@ -140,9 +142,9 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
  encoder.set_input_array(in);
  encoder.set_output_array(out);
  dispatch_float_types(out.dtype(), "softmax", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    constexpr int N_READS = 16 / sizeof(DataType);
+    constexpr int N_READS = 4;
    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
      auto kernel = cu::softmax<DataType, DataType, block_dim(), N_READS>;
      if (precise) {
        kernel = cu::softmax<DataType, float, block_dim(), N_READS>;
@@ -151,7 +153,6 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
          kernel,
          n_rows,
          block_dim(),
-          0,
          in.data<DataType>(),
          out.data<DataType>(),
          axis_size);
--- a/Show More
+++ b/Show More