Support unaligned M

* speedup scalars in cuda

* comment

.circleci/config.yml

@@ -7,15 +7,6 @@ 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
 
 jobs:
   build_documentation:
@@ -38,7 +29,7 @@ jobs:
             pip install --upgrade pip
             pip install --upgrade cmake
             pip install -r docs/requirements.txt
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` pip install . -v
+            pip install . -v
       - when:
           condition:
             not: << parameters.upload-docs >>
@@ -70,9 +61,9 @@ jobs:
                  git push -f origin gh-pages
 
   linux_build_and_test:
-    docker:
-      - image: cimg/python:3.9
-
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
@@ -84,37 +75,34 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
             sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get upgrade -y
+            pip install --upgrade cmake
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
             sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
       - run:
           name: Install Python package
           command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py develop
+            pip install -e ".[dev]"
       - run:
           name: Generate package stubs
           command: |
             echo "stubs"
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
-            python3 -m unittest discover python/tests -v
+            python -m unittest discover python/tests -v
             mpirun --bind-to none -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build CPP only
           command: |
-            mkdir -p build && cd build 
+            mkdir -p build && cd build
             cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
             make -j `nproc`
       - run:
@@ -154,15 +142,14 @@ jobs:
           name: Install Python package
           command: |
             source env/bin/activate
-            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-            CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
               pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
@@ -170,7 +157,8 @@ jobs:
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
             mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py
+            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
       - run:
           name: Build example extension
           command: |
@@ -205,8 +193,7 @@ jobs:
           name: Run Python tests with JIT
           command: |
             source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
+            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
               pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
@@ -214,7 +201,7 @@ jobs:
 
   cuda_build_and_test:
     machine:
-      image: linux-cuda-12:default
+      image: linux-cuda-12:2023.11.1
       resource_class: gpu.nvidia.small.gen2
     steps:
       - checkout
@@ -223,11 +210,9 @@ jobs:
           command: |
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-            python -m venv env
+            python3 -m venv env
             source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               pip install -e ".[dev]"
       - run:
           name: Run Python tests
@@ -276,21 +261,29 @@ jobs:
           command: |
             source env/bin/activate
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
             source env/bin/activate
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
             source env/bin/activate
-            << parameters.build_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              python -m build -w
+            python setup.py clean --all
+            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
+      - when:
+          condition:
+            equal: ["3.9", << parameters.python_version >>]
+          steps:
+            - run:
+                name: Build common package
+                command: |
+                  source env/bin/activate
+                  python setup.py clean --all
+                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
           condition: << parameters.build_env >>
           steps:
@@ -307,52 +300,100 @@ jobs:
       python_version:
         type: string
         default: "3.9"
-      extra_env:
+      build_env:
         type: string
-        default: "DEV_RELEASE=1"
-    docker:
-      - image: ubuntu:20.04
+        default: ""
+    machine:
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
           name: Build wheel
           command: |
             PYTHON=python<< parameters.python_version >>
-            apt-get update
-            apt-get upgrade -y
-            DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata
-            apt-get install -y apt-utils
-            apt-get install -y software-properties-common
-            add-apt-repository -y ppa:deadsnakes/ppa
-            apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            apt-get install -y build-essential git
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
+            sudo apt-get update
+            sudo apt-get upgrade -y
+            TZ=Etc/UTC sudo apt-get -y install tzdata
+            sudo apt-get install -y apt-utils
+            sudo apt-get install -y software-properties-common
+            sudo add-apt-repository -y ppa:deadsnakes/ppa
+            sudo apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
+            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install -y build-essential git
             $PYTHON -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              pip install . -v
+            << parameters.build_env >> pip install ".[dev]" -v
             pip install typing_extensions
-            python setup.py generate_stubs 
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python -m build --wheel
-            auditwheel show dist/*
-            auditwheel repair dist/* --plat manylinux_2_31_x86_64
+            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
+      - store_artifacts:
+          path: wheelhouse/
+
+  build_cuda_release:
+    parameters:
+      build_env:
+        type: string
+        default: ""
+    machine:
+      image: linux-cuda-12:2024.11.1
+      resource_class: gpu.nvidia.small.gen2
+    steps:
+      - checkout
       - run:
-          name: Upload package
+          name: Build wheel
           command: |
+            sudo apt-get update
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            sudo apt-get install zip
+            python -m venv env
             source env/bin/activate
-            twine upload wheelhouse/*
+            pip install auditwheel
+            pip install patchelf
+            pip install build
+            pip install twine
+            << parameters.build_env >> MLX_BUILD_STAGE=2 \
+              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              python -m build -w
+            bash python/scripts/repair_cuda.sh
+      - when:
+          condition: << parameters.build_env >>
+          steps:
+            - run:
+                name: Upload package
+                command: |
+                  source env/bin/activate
+                  twine upload wheelhouse/*.whl
       - store_artifacts:
           path: wheelhouse/
 
@@ -364,8 +405,6 @@ workflows:
             pattern: "^(?!pull/)[-\\w]+$"
             value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
     jobs:
       - mac_build_and_test:
           matrix:
@@ -379,8 +418,6 @@ workflows:
     when:
       and:
         - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
     jobs:
       - build_release:
           filters:
@@ -462,6 +499,25 @@ workflows:
             branches:
               ignore: /.*/
           upload-docs: true
+      - build_linux_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              build_env: ["PYPI_RELEASE=1"]
+      - build_cuda_release:
+          filters:
+            tags:
+              only: /^v.*/
+            branches:
+              ignore: /.*/
+          matrix:
+            parameters:
+              build_env: ["PYPI_RELEASE=1"]
 
   prb:
     when:
@@ -540,88 +596,8 @@ workflows:
               - macosx_deployment_target: "15.0"
                 xcode_version: "15.0.0"
                 python_version: "3.13"
-  weekly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.weekly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-  linux_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.linux_release >>
-    jobs:
       - build_linux_release:
           matrix:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]
+      - build_cuda_release

ACKNOWLEDGMENTS.md

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

CMakeLists.txt

@@ -22,7 +22,7 @@ project(
 
 # ----------------------------- Setup -----------------------------
 set(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)
@@ -64,10 +64,8 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
       message(WARNING "Building for x86_64 arch is not officially supported.")
     endif()
   endif()
-
 else()
   set(MLX_BUILD_METAL OFF)
-  message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
 endif()
 
 # ----------------------------- Lib -----------------------------

benchmarks/cpp/single_ops.cpp

@@ -192,6 +192,22 @@ void time_reductions() {
 
   auto argmin_along_1 = [&a]() { return mx::argmin(a, 1, false); };
   TIME(argmin_along_1);
+
+  auto indices = mx::array({1});
+  auto updates = mx::reshape(mx::array({NAN}), {1, 1, 1});
+  std::vector<int> axes{0};
+  auto b = scatter(a, {indices}, updates, axes);
+  mx::eval(b);
+
+  auto max_along_0 = [&b]() { return mx::max(b, 0, false); };
+  TIME(max_along_0);
+  auto max_along_1 = [&b]() { return mx::max(b, 1, false); };
+  TIME(max_along_1);
+
+  auto min_along_0 = [&b]() { return mx::min(b, 0, false); };
+  TIME(min_along_0);
+  auto min_along_1 = [&b]() { return mx::min(b, 1, false); };
+  TIME(min_along_1);
 }
 
 void time_gather_scatter() {

benchmarks/python/comparative/bench_torch.py

@@ -5,6 +5,7 @@ import os
 import time
 
 import torch
+import torch.cuda
 import torch.mps
 
 
@@ -44,8 +45,10 @@ def bench(f, *args):
 
 
 def sync_if_needed(x):
-    if x.device != torch.device("cpu"):
+    if x.device == torch.device("mps"):
         torch.mps.synchronize()
+    elif x.device == torch.device("cuda"):
+        torch.cuda.synchronize()
 
 
 @torch.no_grad()
@@ -99,6 +102,14 @@ def reduction(op, axis, x):
     sync_if_needed(x)
 
 
+@torch.no_grad()
+def sum_and_add(axis, x, y):
+    z = x.sum(axis=axis, keepdims=True)
+    for i in range(50):
+        z = (z + y).sum(axis=axis, keepdims=True)
+    sync_if_needed(x)
+
+
 @torch.no_grad()
 def softmax(axis, x):
     ys = []
@@ -340,7 +351,11 @@ if __name__ == "__main__":
         args.axis.pop(0)
 
     torch.set_num_threads(1)
-    device = "cpu" if args.cpu else "mps"
+    device = "mps"
+    if torch.cuda.is_available():
+        device = "cuda"
+    if args.cpu:
+        device = "cpu"
 
     types = args.dtype
     if not types:
@@ -460,5 +475,8 @@ if __name__ == "__main__":
     elif args.benchmark == "selu":
         print(bench(selu, x))
 
+    elif args.benchmark == "sum_and_add":
+        print(bench(sum_and_add, axis, *xs))
+
     else:
         raise ValueError(f"Unknown benchmark `{args.benchmark}`.")

benchmarks/python/single_ops.py

@@ -51,6 +51,20 @@ def time_maximum():
     time_fn(mx.maximum, a, b)
 
 
+def time_max():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    a[1, 1] = mx.nan
+    mx.eval(a)
+    time_fn(mx.max, a, 0)
+
+
+def time_min():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    a[1, 1] = mx.nan
+    mx.eval(a)
+    time_fn(mx.min, a, 0)
+
+
 def time_negative():
     a = mx.random.uniform(shape=(10000, 1000))
     mx.eval(a)
@@ -108,6 +122,8 @@ if __name__ == "__main__":
 
     time_add()
     time_matmul()
+    time_min()
+    time_max()
     time_maximum()
     time_exp()
     time_negative()

docs/src/dev/extensions.rst

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

docs/src/install.rst

@@ -23,13 +23,24 @@ To install from PyPI you must meet the following requirements:
     MLX is only available on devices running macOS >= 13.5
     It is highly recommended to use macOS 14 (Sonoma)
 
+CUDA
+^^^^
 
-MLX is also available on conda-forge. To install MLX with conda do:
+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
 
-   conda install conda-forge::mlx
+    pip install "mlx[cuda]"
 
+CPU-only (Linux)
+^^^^^^^^^^^^^^^^
+
+For a CPU-only version of MLX that runs on Linux use:
+
+.. code-block:: shell
+
+    pip install "mlx[cpu]"
 
 Troubleshooting
 ^^^^^^^^^^^^^^^
@@ -65,6 +76,8 @@ Build Requirements
 Python API
 ^^^^^^^^^^
 
+.. _python install:
+
 To build and install the MLX python library from source, first, clone MLX from
 `its GitHub repo <https://github.com/ml-explore/mlx>`_:
 
@@ -76,20 +89,20 @@ Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
+  pip install .
 
 For developing, install the package with development dependencies, and use an
 editable install:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
+  pip install -e ".[dev]"
 
 Once the development dependencies are installed, you can build faster with:
 
 .. code-block:: shell
 
- CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
+ python setup.py build_ext --inplace
 
 Run the tests with:
 
@@ -107,6 +120,8 @@ IDE:
 C++ API
 ^^^^^^^
 
+.. _cpp install:
+
 Currently, MLX must be built and installed from source.
 
 Similarly to the python library, to build and install the MLX C++ library start
@@ -185,6 +200,7 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
+
 Binary Size Minimization
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -213,6 +229,50 @@ be anwywhere from a few hundred millisecond to a few seconds depending on the
 application. Once a kernel is compiled, it will be cached by the system. The
 Metal kernel cache persists across reboots.
 
+Linux
+^^^^^
+
+To build from source on Linux (CPU only), install the BLAS and LAPACK headers.
+For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   apt-get update -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+
+From here follow the instructions to install either the :ref:`Python <python
+install>` or :ref:`C++ <cpp install>` APIs.
+
+CUDA
+^^^^
+
+To build from source on Linux with CUDA, install the BLAS and LAPACK headers
+and the CUDA toolkit. For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+   dpkg -i cuda-keyring_1.1-1_all.deb
+   apt-get update -y
+   apt-get -y install cuda-toolkit-12-9
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+
+
+When building either the Python or C++ APIs make sure to pass the cmake flag
+``MLX_BUILD_CUDA=ON``. For example, to build the Python API run:
+
+.. code-block:: shell
+
+  CMAKE_ARGS="-DMLX_BUILD_CUDA=ON" pip install -e ".[dev]"
+
+To build the C++ package run:
+
+.. code-block:: shell
+
+   mkdir -p build && cd build
+   cmake .. -DMLX_BUILD_CUDA=ON && make -j
+
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 

docs/src/python/optimizers/common_optimizers.rst

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

examples/extensions/axpby/axpby.h

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

mlx/backend/common/compiled.cpp

@@ -14,6 +14,8 @@ void print_constant(std::ostream& os, const array& x) {
       return print_float_constant<float16_t>(os, x);
     case bfloat16:
       return print_float_constant<bfloat16_t>(os, x);
+    case float64:
+      return print_float_constant<double>(os, x);
     case complex64:
       return print_complex_constant<complex64_t>(os, x);
     case int8:
@@ -50,6 +52,8 @@ std::string get_type_string(Dtype d) {
       return "float16_t";
     case bfloat16:
       return "bfloat16_t";
+    case float64:
+      return "double";
     case complex64:
       return "complex64_t";
     case bool_:

mlx/backend/common/compiled.h

@@ -18,8 +18,12 @@ std::string get_type_string(Dtype d);
 template <typename T>
 void print_float_constant(std::ostream& os, const array& x) {
   auto old_precision = os.precision();
-  os << std::setprecision(std::numeric_limits<float>::digits10 + 1)
-     << x.item<T>() << std::setprecision(old_precision);
+  if constexpr (std::is_same_v<T, double>) {
+    os << std::setprecision(std::numeric_limits<double>::digits10 + 1);
+  } else {
+    os << std::setprecision(std::numeric_limits<float>::digits10 + 1);
+  }
+  os << x.item<T>() << std::setprecision(old_precision);
 }
 
 template <typename T>

mlx/backend/common/matmul.h

@@ -12,16 +12,11 @@ namespace mlx::core {
 inline std::tuple<Shape, Strides, Strides> collapse_batches(
     const array& a,
     const array& b) {
-  // Get and check the shape for the batched dims
-  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
-  Shape B_bshape{b.shape().begin(), b.shape().end() - 2};
-  if (A_bshape != B_bshape) {
-    std::ostringstream msg;
-    msg << "[matmul] Got matrices with incorrectly broadcasted shapes: " << "A "
-        << a.shape() << ", B " << b.shape() << ".";
-    throw std::runtime_error(msg.str());
+  if (a.ndim() == 2) {
+    return {{1}, {0}, {0}};
   }
 
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
   Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
   Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
 
@@ -42,17 +37,11 @@ inline std::tuple<Shape, Strides, Strides> collapse_batches(
 
 inline std::tuple<Shape, Strides, Strides, Strides>
 collapse_batches(const array& a, const array& b, const array& c) {
-  // Get and check the shape for the batched dims
-  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
-  Shape B_bshape{b.shape().begin(), b.shape().end() - 2};
-  Shape C_bshape{c.shape().begin(), c.shape().end() - 2};
-  if (A_bshape != B_bshape || A_bshape != C_bshape) {
-    std::ostringstream msg;
-    msg << "[addmm] Got matrices with incorrectly broadcasted shapes: " << "A "
-        << a.shape() << ", B " << b.shape() << ", B " << c.shape() << ".";
-    throw std::runtime_error(msg.str());
+  if (a.ndim() == 2) {
+    return {{1}, {0}, {0}, {0}};
   }
 
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
   Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
   Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
   Strides C_bstride{c.strides().begin(), c.strides().end() - 2};

mlx/backend/common/reduce.cpp

@@ -5,11 +5,9 @@
 namespace mlx::core {
 
 std::pair<Shape, Strides> shapes_without_reduction_axes(
-    const array& x,
+    Shape shape,
+    Strides strides,
     const std::vector<int>& axes) {
-  auto shape = x.shape();
-  auto strides = x.strides();
-
   for (int i = axes.size() - 1; i >= 0; i--) {
     int a = axes[i];
     shape.erase(shape.begin() + a);
@@ -19,6 +17,15 @@ std::pair<Shape, Strides> shapes_without_reduction_axes(
   return std::make_pair(shape, strides);
 }
 
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes) {
+  auto shape = x.shape();
+  auto strides = x.strides();
+  return shapes_without_reduction_axes(
+      std::move(shape), std::move(strides), axes);
+}
+
 ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
   // The data is all there and we are reducing over everything
   if (x.size() == x.data_size() && axes.size() == x.ndim() &&

mlx/backend/common/reduce.h

@@ -51,5 +51,9 @@ ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes);
 std::pair<Shape, Strides> shapes_without_reduction_axes(
     const array& x,
     const std::vector<int>& axes);
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    Shape shape,
+    Strides strides,
+    const std::vector<int>& axes);
 
 } // namespace mlx::core

mlx/backend/common/utils.cpp

@@ -1,14 +1,20 @@
 // Copyright © 2023-2024 Apple Inc.
 
+#include <dlfcn.h>
+
 #include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive) {
-  std::ostringstream op_t;
-  primitive->print(op_t);
-  return op_t.str();
+std::filesystem::path current_binary_dir() {
+  static std::filesystem::path binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path();
+  }();
+  return binary_dir;
 }
 
 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
@@ -199,12 +205,15 @@ Dims get_2d_grid_dims_common(
       }
     }
   }
-  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX || divisor > 1) {
+  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX) {
     throw std::runtime_error("Unable to safely factor shape.");
   }
   if (grid_y > grid_x) {
     std::swap(grid_x, grid_y);
   }
+  if (divisor > 1) {
+    grid_x = ((grid_x + divisor - 1) / divisor) * divisor;
+  }
   return std::make_tuple(
       static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
 }

mlx/backend/common/utils.h

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

mlx/backend/cpu/cholesky.cpp

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

mlx/backend/cpu/compiled.cpp

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

mlx/backend/cpu/conv.cpp

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

mlx/backend/cpu/copy.cpp

@@ -295,7 +295,11 @@ inline void copy_inplace_dispatch(
 
 } // namespace
 
-void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu_inplace(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Stream stream) {
   auto& encoder = cpu::get_command_encoder(stream);
   encoder.set_input_array(src);
   encoder.set_output_array(dst);
@@ -305,7 +309,7 @@ void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
        ctype]() mutable { copy_inplace_dispatch(src, dst, ctype); });
 }
 
-void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream) {
   bool donated = set_copy_output_data(src, dst, ctype);
   if (donated && src.dtype() == dst.dtype()) {
     // If the output has the same type as the input then there is nothing to
@@ -315,10 +319,10 @@ void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;
   }
-  copy_inplace(src, dst, ctype, stream);
+  copy_cpu_inplace(src, dst, ctype, stream);
 }
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,

mlx/backend/cpu/copy.h

@@ -10,10 +10,14 @@
 
 namespace mlx::core {
 
-void copy(const array& src, array& dst, CopyType ctype, Stream stream);
-void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_cpu_inplace(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Stream stream);
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,

mlx/backend/cpu/distributed.cpp

@@ -14,7 +14,7 @@ std::pair<array, bool> ensure_row_contiguous(const array& arr, Stream stream) {
     return {arr, false};
   } else {
     array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy(arr, arr_copy, CopyType::General, stream);
+    copy_cpu(arr, arr_copy, CopyType::General, stream);
     return {arr_copy, true};
   }
 };
@@ -35,7 +35,7 @@ void AllReduce::eval_cpu(
       return in;
     } else {
       array arr_copy(in.shape(), in.dtype(), nullptr, {});
-      copy(in, arr_copy, CopyType::General, s);
+      copy_cpu(in, arr_copy, CopyType::General, s);
       out.copy_shared_buffer(arr_copy);
       return arr_copy;
     }

mlx/backend/cpu/eig.cpp

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

mlx/backend/cpu/eigh.cpp

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

mlx/backend/cpu/hadamard.cpp

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

mlx/backend/cpu/indexing.cpp

@@ -517,7 +517,7 @@ void Scatter::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   std::vector<array> inds;
@@ -686,7 +686,7 @@ void ScatterAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(idx);

mlx/backend/cpu/inverse.cpp

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

mlx/backend/cpu/logsumexp.cpp

@@ -88,7 +88,7 @@ void LogSumExp::eval_cpu(const std::vector<array>& inputs, array& out) {
       return x;
     } else {
       auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      copy_cpu(x, x_copy, CopyType::General, s);
       encoder.add_temporary(x_copy);
       return x_copy;
     }

mlx/backend/cpu/luf.cpp

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

mlx/backend/cpu/masked_mm.cpp

@@ -6,6 +6,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/gemm.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
@@ -52,6 +53,58 @@ inline void mask_matrix(
   }
 }
 
+template <typename T>
+inline void segmented_mm(
+    const T* a,
+    const T* b,
+    const uint32_t* segments,
+    T* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides,
+    size_t num_segments,
+    const Shape& segments_shape,
+    const Strides& segments_strides) {
+  int ndim = a_shape.size();
+  Shape a_copy = a_shape;
+  Shape b_copy = b_shape;
+  int32_t M = a_copy[ndim - 2];
+  int32_t N = b_copy[ndim - 1];
+  for (int i = 0; i < num_segments; i++) {
+    uint32_t k_start =
+        segments[elem_to_loc(2 * i, segments_shape, segments_strides)];
+    uint32_t k_end =
+        segments[elem_to_loc(2 * i + 1, segments_shape, segments_strides)];
+    if (k_end <= k_start) {
+      std::fill_n(out + i * M * N, M * N, T(0));
+      continue;
+    }
+    a_copy[ndim - 1] = k_end - k_start;
+    b_copy[ndim - 2] = k_end - k_start;
+    matmul<T>(
+        a + k_start * a_strides[ndim - 1],
+        b + k_start * b_strides[ndim - 2],
+        out + i * M * N,
+        a_transposed,
+        b_transposed,
+        lda,
+        ldb,
+        N,
+        1.0,
+        0.0,
+        1,
+        a_copy,
+        a_strides,
+        b_copy,
+        b_strides);
+  }
+}
+
 } // namespace
 
 void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -71,20 +124,20 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
         if (!expand_all && stx == arr.shape(-1) && sty == 1) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(false, stx, arr_copy, true);
           }
           return std::make_tuple(false, stx, arr, false);
         } else if (!expand_all && stx == 1 && sty == arr.shape(-2)) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(true, sty, arr_copy, true);
           }
           return std::make_tuple(true, sty, arr, false);
         } else {
           array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy(arr, arr_copy, CopyType::General, s);
+          copy_cpu(arr, arr_copy, CopyType::General, s);
           int64_t stx = arr.shape(-1);
           return std::make_tuple(false, stx, arr_copy, true);
         }
@@ -333,7 +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(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       int64_t stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -437,4 +490,121 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   encoder.add_temporaries(std::move(temps));
 }
 
+void SegmentedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto check_transpose = [&s, &encoder](const array& x) {
+    auto stx = x.strides()[x.ndim() - 2];
+    auto sty = x.strides()[x.ndim() - 1];
+    if (stx == x.shape(-1) && sty == 1) {
+      return std::make_tuple(false, stx, x);
+    } else if (stx == 1 && sty == x.shape(-2)) {
+      return std::make_tuple(true, sty, x);
+    } else {
+      array xc(x.shape(), x.dtype(), nullptr, {});
+      copy_cpu(x, xc, CopyType::General, s);
+      encoder.add_temporary(xc);
+      int64_t stx = x.shape(-1);
+      return std::make_tuple(false, stx, xc);
+    }
+  };
+
+  auto [a_transposed, lda, a] = check_transpose(inputs[0]);
+  auto [b_transposed, ldb, b] = check_transpose(inputs[1]);
+  auto& segments = inputs[2];
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(segments);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    segments = array::unsafe_weak_copy(segments),
+                    out_ptr = out.data<void>(),
+                    a_transposed = a_transposed,
+                    b_transposed = b_transposed,
+                    lda = lda,
+                    ldb = ldb]() {
+    switch (a.dtype()) {
+      case float64:
+        segmented_mm<double>(
+            a.data<double>(),
+            b.data<double>(),
+            segments.data<uint32_t>(),
+            static_cast<double*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float32:
+        segmented_mm<float>(
+            a.data<float>(),
+            b.data<float>(),
+            segments.data<uint32_t>(),
+            static_cast<float*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float16:
+        segmented_mm<float16_t>(
+            a.data<float16_t>(),
+            b.data<float16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<float16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case bfloat16:
+        segmented_mm<bfloat16_t>(
+            a.data<bfloat16_t>(),
+            b.data<bfloat16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<bfloat16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      default:
+        throw std::invalid_argument(
+            "Segmented mm supports only real float types.");
+    }
+  });
+}
+
 } // namespace mlx::core

mlx/backend/cpu/matmul.cpp

@@ -81,7 +81,7 @@ void matmul_general(
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, stream);
+      copy_cpu(arr, temps.back(), CopyType::General, stream);
       stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -142,7 +142,7 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   CopyType ctype = c.data_size() == 1
       ? CopyType::Scalar
       : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
-  copy(c, out, ctype, stream());
+  copy_cpu(c, out, ctype, stream());
   if (inputs[0].shape(-1) == 0) {
     return;
   }

mlx/backend/cpu/primitives.cpp

@@ -22,7 +22,7 @@ void reshape(const array& in, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
     out.set_data(allocator::malloc(out.nbytes()));
-    copy_inplace(in, out, CopyType::General, out.primitive().stream());
+    copy_cpu_inplace(in, out, CopyType::General, out.primitive().stream());
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -175,7 +175,7 @@ void AsType::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -198,7 +198,7 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
     size_t data_offset = strides[axis_] * sizes[i];
     out_slice.copy_shared_buffer(
         out, strides, flags, out_slice.size(), data_offset);
-    copy_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
+    copy_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
   }
 }
 
@@ -211,7 +211,7 @@ void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
        (allow_col_major_ && in.flags().col_contiguous))) {
     out.copy_shared_buffer(in);
   } else {
-    copy(in, out, CopyType::General, stream());
+    copy_cpu(in, out, CopyType::General, stream());
   }
 }
 
@@ -235,7 +235,7 @@ void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   } else {
     ctype = CopyType::General;
   }
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -251,7 +251,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
 
   // Fill output with val
-  copy(val, out, CopyType::Scalar, stream());
+  copy_cpu(val, out, CopyType::Scalar, stream());
 
   // Find offset for start of input values
   size_t data_offset = 0;
@@ -266,7 +266,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
       out, out.strides(), out.flags(), out_slice.size(), data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
+  copy_cpu_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
 }
 
 void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -340,7 +340,7 @@ void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
   out.set_data(allocator::malloc(out.nbytes()));
   auto [in_offset, donated] =
       compute_dynamic_offset(inputs[1], in.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ in,
       /* array& dst = */ out,
       /* const Shape& data_shape = */ out.shape(),
@@ -372,11 +372,11 @@ void DynamicSliceUpdate::eval_cpu(
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   auto [out_offset, donated] =
       compute_dynamic_offset(inputs[2], out.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -412,14 +412,14 @@ void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   // Calculate out strides, initial offset and if copy needs to be made
   auto [data_offset, out_strides] =
       prepare_slice(out, start_indices_, strides_);
 
   // Do copy
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -456,9 +456,9 @@ void View::eval_cpu(const std::vector<array>& inputs, array& out) {
     if (in.dtype() == bool_) {
       auto in_tmp = array(in.shape(), uint8, nullptr, {});
       in_tmp.copy_shared_buffer(in);
-      copy_inplace(in_tmp, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in_tmp, tmp, CopyType::General, stream());
     } else {
-      copy_inplace(in, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in, tmp, CopyType::General, stream());
     }
 
     auto flags = out.flags();

mlx/backend/cpu/qrf.cpp

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

mlx/backend/cpu/quantized.cpp

@@ -529,7 +529,7 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
       return arr;
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       return temps.back();
     }
   };
@@ -579,7 +579,7 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return arr;
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       return temps.back();
     }
   };
@@ -713,7 +713,7 @@ void fast::AffineQuantize::eval_cpu(
       return std::make_pair(arr, false);
     } else {
       array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General, s);
+      copy_cpu(arr, arr_copy, CopyType::General, s);
       return std::make_pair(arr_copy, true);
     }
   };

mlx/backend/cpu/reduce.cpp

@@ -325,7 +325,15 @@ struct MaxReduce {
   };
 
   template <int N, typename T>
-  T operator()(simd::Simd<T, N> x) {
+  std::enable_if_t<std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    return simd::max(x);
+  };
+
+  template <int N, typename T>
+  std::enable_if_t<!std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    if (simd::any(x != x)) {
+      return static_cast<T>(NAN);
+    }
     return simd::max(x);
   };
 };
@@ -342,7 +350,15 @@ struct MinReduce {
   };
 
   template <int N, typename T>
-  T operator()(simd::Simd<T, N> x) {
+  std::enable_if_t<std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    return simd::min(x);
+  };
+
+  template <int N, typename T>
+  std::enable_if_t<!std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    if (simd::any(x != x)) {
+      return static_cast<T>(NAN);
+    }
     return simd::min(x);
   };
 };
@@ -527,10 +543,10 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
             reduce_dispatch_min_max<uint64_t>(in, out, reduce_type_, axes_);
             break;
           case int8:
-            reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
+            reduce_dispatch_min_max<int8_t>(in, out, reduce_type_, axes_);
             break;
           case int16:
-            reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
+            reduce_dispatch_min_max<int16_t>(in, out, reduce_type_, axes_);
             break;
           case int32:
             reduce_dispatch_min_max<int32_t>(in, out, reduce_type_, axes_);

mlx/backend/cpu/scan.cpp

@@ -251,7 +251,7 @@ void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto in = inputs[0];
   if (!in.flags().row_contiguous) {
     array arr_copy(in.shape(), in.dtype(), nullptr, {});
-    copy(in, arr_copy, CopyType::General, stream());
+    copy_cpu(in, arr_copy, CopyType::General, stream());
     in = arr_copy;
     encoder.add_temporary(arr_copy);
   }

mlx/backend/cpu/softmax.cpp

@@ -132,7 +132,7 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
       return x;
     } else {
       array x_copy(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      copy_cpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }

mlx/backend/cpu/sort.cpp

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

mlx/backend/cpu/svd.cpp

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

mlx/backend/cuda/CMakeLists.txt

@@ -8,6 +8,7 @@ target_sources(
   PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_contiguous.cu
@@ -28,17 +29,22 @@ target_sources(
           ${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
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/col_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/init_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/segmented_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/sort.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/qmm.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
@@ -65,6 +71,11 @@ target_include_directories(mlx PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/gen")
 target_compile_options(mlx
                        PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>")
 
+# Enable calling host constexpr functions from device. This is needed because
+# the constexpr version of isnan is host only.
+target_compile_options(
+  mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>")
+
 # CUDA 12.8 emits warning #20280-D for copy kernels which is a false positive.
 # Explicitly pass this flag to suppress the warning, it is safe to set it to
 # true but the warning wouldn't be suppressed.
@@ -81,7 +92,7 @@ target_compile_options(
 # 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"
+    "80"
     CACHE STRING "CUDA architectures")
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
@@ -117,3 +128,16 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 # 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 ${CMAKE_INSTALL_INCLUDEDIR}/cccl)
+
+# Make Thunderkittens available
+FetchContent_Declare(
+  kittens
+  GIT_REPOSITORY https://github.com/HazyResearch/ThunderKittens.git
+  GIT_TAG aaab847f430ed313ed466e64b25b9177babd1db8
+  GIT_SHALLOW TRUE)
+FetchContent_MakeAvailable(kittens)
+target_include_directories(mlx BEFORE PRIVATE "${kittens_SOURCE_DIR}/include")

mlx/backend/cuda/allocator.cpp

@@ -3,6 +3,7 @@
 #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>
 #include <fmt/format.h>
@@ -14,9 +15,57 @@ namespace mlx::core {
 
 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() {
+  CHECK_CUDA_ERROR(cudaMallocManaged(&buffer_, small_pool_size));
+  end_ = reinterpret_cast<void*>(
+      reinterpret_cast<char*>(buffer_) + small_pool_size);
+  next_free_ = reinterpret_cast<Block*>(buffer_);
+
+  auto num_blocks = small_pool_size / small_block_size;
+  auto curr = next_free_;
+  for (size_t i = 0; i < num_blocks - 1; ++i) {
+    curr->next = reinterpret_cast<Block*>(
+        reinterpret_cast<char*>(buffer_) + (i + 1) * small_block_size);
+    curr = curr->next;
+  }
+  curr->next = nullptr;
+}
+
+SmallSizePool::~SmallSizePool() {
+  CHECK_CUDA_ERROR(cudaFree(buffer_));
+}
+
+void* SmallSizePool::malloc() {
+  if (next_free_ == nullptr) {
+    return nullptr;
+  }
+  Block* b = next_free_;
+  next_free_ = next_free_->next;
+  return static_cast<void*>(b);
+}
+
+void SmallSizePool::free(void* p) {
+  auto b = static_cast<Block*>(p);
+  b->next = next_free_;
+  next_free_ = b;
+}
+
+bool SmallSizePool::in_pool(void* p) {
+  return (p >= buffer_) && (p < end_);
+}
+
 CudaAllocator::CudaAllocator()
     : buffer_cache_(
-          getpagesize(),
+          page_size,
           [](CudaBuffer* buf) { return buf->size; },
           [this](CudaBuffer* buf) {
             cuda_free(buf->data);
@@ -31,7 +80,16 @@ CudaAllocator::CudaAllocator()
 
 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) {
+    size = next_power_of_2(size);
+  } else {
+    size = page_size * ((size + page_size - 1) / page_size);
+  }
+
   CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
   if (!buf) {
     // If we have a lot of memory pressure or are over the maximum cache size,
@@ -43,11 +101,19 @@ Buffer CudaAllocator::malloc(size_t size) {
 
     lock.unlock();
     buf = new CudaBuffer{nullptr, size};
-    cudaError_t err = cudaMallocManaged(&buf->data, size);
-    if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
-      throw std::runtime_error(fmt::format(
-          "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+
+    // Try the scalar pool first
+    if (size <= small_block_size) {
+      buf->data = scalar_pool_.malloc();
     }
+    if (!buf->data) {
+      cudaError_t err = cudaMallocManaged(&buf->data, size);
+      if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
+        throw std::runtime_error(fmt::format(
+            "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+      }
+    }
+
     lock.lock();
   }
   active_memory_ += size;
@@ -106,8 +172,11 @@ void CudaAllocator::cuda_free(void* buf) {
       return;
     }
   }
-
-  cudaFree(buf);
+  if (scalar_pool_.in_pool(buf)) {
+    scalar_pool_.free(buf);
+  } else {
+    cudaFree(buf);
+  }
 }
 
 size_t CudaAllocator::get_active_memory() const {

mlx/backend/cuda/allocator.h

@@ -22,6 +22,28 @@ struct CudaBuffer {
   size_t size;
 };
 
+class SmallSizePool {
+ private:
+  struct Block {
+    Block* next;
+  };
+
+  void* buffer_{nullptr};
+  Block* next_free_{nullptr};
+  void* end_{nullptr};
+
+ public:
+  SmallSizePool();
+  ~SmallSizePool();
+
+  SmallSizePool(const SmallSizePool&) = delete;
+  SmallSizePool& operator=(const SmallSizePool&) = delete;
+
+  void* malloc();
+  void free(void* p);
+  bool in_pool(void* p);
+};
+
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
@@ -60,6 +82,7 @@ class CudaAllocator : public allocator::Allocator {
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
+  SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();

mlx/backend/cuda/arg_reduce.cu

@@ -1,6 +1,7 @@
 // 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"
@@ -151,36 +152,30 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_REAL_TYPES_CHECKED(in.dtype(), "ArgReduce", CTYPE, {
-      using InType = cuda_type_t<CTYPE>;
-      constexpr uint32_t N_READS = 4;
-      MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-        dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
-        dim3 block_dims{BLOCK_DIM, 1, 1};
-        auto kernel = &cu::arg_reduce_general<
-            InType,
-            cu::ArgMax<InType>,
-            BLOCK_DIM,
-            N_READS>;
-        if (reduce_type_ == ArgReduce::ArgMin) {
-          kernel = &cu::arg_reduce_general<
-              InType,
-              cu::ArgMin<InType>,
-              BLOCK_DIM,
-              N_READS>;
-        }
-        kernel<<<num_blocks, block_dims, 0, stream>>>(
-            in.data<InType>(),
-            out.data<uint32_t>(),
-            out.size(),
-            const_param(shape),
-            const_param(in_strides),
-            const_param(out_strides),
-            ndim,
-            axis_stride,
-            axis_size);
-      });
+  dispatch_real_types(in.dtype(), "ArgReduce", [&](auto type_tag) {
+    using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr uint32_t N_READS = 4;
+    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
+      dim3 num_blocks = get_2d_grid_dims(out.shape(), out.strides());
+      auto kernel =
+          cu::arg_reduce_general<T, cu::ArgMax<T>, block_dim(), N_READS>;
+      if (reduce_type_ == ArgReduce::ArgMin) {
+        kernel = cu::arg_reduce_general<T, cu::ArgMin<T>, block_dim(), N_READS>;
+      }
+      encoder.add_kernel_node(
+          kernel,
+          num_blocks,
+          block_dim(),
+          0,
+          in.data<T>(),
+          out.data<uint32_t>(),
+          out.size(),
+          const_param(shape),
+          const_param(in_strides),
+          const_param(out_strides),
+          ndim,
+          axis_stride,
+          axis_size);
     });
   });
 }

mlx/backend/cuda/binary.cu

@@ -3,7 +3,6 @@
 #include "mlx/backend/common/binary.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/device/binary_ops.cuh"
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -17,35 +16,86 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_ss(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[0], b[0]);
+
+  if ((index + 1) * N_READS > size) {
+    for (int i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a[0], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_sv(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[0], b[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[0], b[i]);
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a[0], b_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_vs(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[0]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[0]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[i]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -125,13 +175,12 @@ constexpr bool supports_binary_op() {
 template <typename Op>
 void binary_op_gpu_inplace(
     const std::vector<array>& inputs,
-    std::vector<array>& outputs,
-    std::string_view op,
+    array& out,
+    const char* op,
     const Stream& s) {
   assert(inputs.size() > 1);
   const auto& a = inputs[0];
   const auto& b = inputs[1];
-  auto& out = outputs[0];
   if (out.size() == 0) {
     return;
   }
@@ -140,126 +189,124 @@ void binary_op_gpu_inplace(
   encoder.set_input_array(a);
   encoder.set_input_array(b);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(a.dtype(), CTYPE_IN, {
-      MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, {
-        if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
-          using InType = cuda_type_t<CTYPE_IN>;
-          using OutType = cuda_type_t<CTYPE_OUT>;
-
-          auto bopt = get_binary_op_type(a, b);
-          if (bopt == BinaryOpType::General) {
-            auto [shape, strides] = collapse_contiguous_dims(a, b, out);
-            auto& a_strides = strides[0];
-            auto& b_strides = strides[1];
-            bool large = a.data_size() > INT32_MAX ||
-                b.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-            MLX_SWITCH_BOOL(large, LARGE, {
-              using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-              int ndim = shape.size();
-              if (ndim <= 3) {
-                MLX_SWITCH_1_2_3(ndim, NDIM, {
-                  auto kernel =
-                      &cu::binary_g_nd<Op, InType, OutType, IdxT, NDIM>;
+  dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
+      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
+      if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+        using InType = cuda_type_t<CTYPE_IN>;
+        using OutType = cuda_type_t<CTYPE_OUT>;
+        auto bopt = get_binary_op_type(a, b);
+        if (bopt == BinaryOpType::General) {
+          dispatch_bool(
+              a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
+                  out.data_size() > INT32_MAX,
+              [&](auto large) {
+                using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+                Shape shape;
+                std::vector<Strides> strides;
+                std::tie(shape, strides) = collapse_contiguous_dims(a, b, out);
+                auto& a_strides = strides[0];
+                auto& b_strides = strides[1];
+                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(kernel, out, large());
+                    encoder.add_kernel_node(
+                        kernel,
+                        num_blocks,
+                        block_dims,
+                        0,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out.data<OutType>(),
+                        out.size(),
+                        const_param<dims_constant()>(shape),
+                        const_param<dims_constant()>(a_strides),
+                        const_param<dims_constant()>(b_strides));
+                  });
+                } else {
+                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
                   auto [num_blocks, block_dims] =
-                      get_launch_args(kernel, out, large);
-                  kernel<<<num_blocks, block_dims, 0, stream>>>(
+                      get_launch_args(kernel, out, large());
+                  encoder.add_kernel_node(
+                      kernel,
+                      num_blocks,
+                      block_dims,
+                      0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out.data<OutType>(),
                       out.size(),
-                      const_param<NDIM>(shape),
-                      const_param<NDIM>(a_strides),
-                      const_param<NDIM>(b_strides));
-                });
-              } else {
-                auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
-                auto [num_blocks, block_dims] =
-                    get_launch_args(kernel, out, large);
-                kernel<<<num_blocks, block_dims, 0, stream>>>(
-                    a.data<InType>(),
-                    b.data<InType>(),
-                    out.data<OutType>(),
-                    out.size(),
-                    const_param(shape),
-                    const_param(a_strides),
-                    const_param(b_strides),
-                    ndim);
-              }
-            });
-          } else {
-            MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
-              using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
-              auto kernel = cu::binary_ss<Op, InType, OutType, IdxT>;
-              if (bopt == BinaryOpType::ScalarVector) {
-                kernel = cu::binary_sv<Op, InType, OutType, IdxT>;
-              } else if (bopt == BinaryOpType::VectorScalar) {
-                kernel = cu::binary_vs<Op, InType, OutType, IdxT>;
-              } else if (bopt == BinaryOpType::VectorVector) {
-                kernel = cu::binary_vv<Op, InType, OutType, IdxT>;
-              }
-              auto [num_blocks, block_dims] = get_launch_args(
-                  kernel, out.data_size(), out.shape(), out.strides(), LARGE);
-              kernel<<<num_blocks, block_dims, 0, stream>>>(
-                  a.data<InType>(),
-                  b.data<InType>(),
-                  out.data<OutType>(),
-                  out.data_size());
-            });
-          }
+                      const_param(shape),
+                      const_param(a_strides),
+                      const_param(b_strides),
+                      ndim);
+                }
+              });
         } else {
-          throw std::runtime_error(fmt::format(
-              "Can not do binary op {} on inputs of {} with result of {}.",
-              op,
-              dtype_to_string(a.dtype()),
-              dtype_to_string(out.dtype())));
+          dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::binary_ss<Op, InType, OutType, IdxT, N_READS>;
+            if (bopt == BinaryOpType::ScalarVector) {
+              kernel = cu::binary_sv<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorScalar) {
+              kernel = cu::binary_vs<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorVector) {
+              kernel = cu::binary_vv<Op, InType, OutType, IdxT, N_READS>;
+            }
+            auto [num_blocks, block_dims] = get_launch_args(
+                kernel,
+                out.data_size(),
+                out.shape(),
+                out.strides(),
+                large(),
+                N_READS);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                a.data<InType>(),
+                b.data<InType>(),
+                out.data<OutType>(),
+                out.data_size());
+          });
         }
-      });
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do binary op {} on inputs of {} with result of {}.",
+            op,
+            dtype_to_string(a.dtype()),
+            dtype_to_string(out.dtype())));
+      }
     });
   });
 }
 
-template <typename Op>
-void binary_op_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs,
-    std::string_view op,
-    const Stream& s) {
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, outputs[0], bopt);
-  set_binary_op_output_data(a, b, outputs[1], bopt);
-  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
-}
-
 template <typename Op>
 void binary_op_gpu(
     const std::vector<array>& inputs,
     array& out,
-    std::string_view op,
+    const char* op,
     const Stream& s) {
   auto& a = inputs[0];
   auto& b = inputs[1];
   auto bopt = get_binary_op_type(a, b);
   set_binary_op_output_data(a, b, out, bopt);
-  std::vector<array> outputs{out};
-  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
+  binary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 
-#define BINARY_GPU(func)                                                 \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) {    \
-    nvtx3::scoped_range r(#func "::eval_gpu");                           \
-    auto& s = out.primitive().stream();                                  \
-    binary_op_gpu<cu::func>(inputs, out, get_primitive_string(this), s); \
-  }
-
-#define BINARY_GPU_MULTI(func)                                               \
-  void func::eval_gpu(                                                       \
-      const std::vector<array>& inputs, std::vector<array>& outputs) {       \
-    nvtx3::scoped_range r(#func "::eval_gpu");                               \
-    auto& s = outputs[0].primitive().stream();                               \
-    binary_op_gpu<cu::func>(inputs, outputs, get_primitive_string(this), s); \
+#define BINARY_GPU(func)                                              \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    nvtx3::scoped_range r(#func "::eval_gpu");                        \
+    auto& s = out.primitive().stream();                               \
+    binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
   }
 
 BINARY_GPU(Add)
@@ -283,33 +330,31 @@ BINARY_GPU(Subtract)
 void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
   nvtx3::scoped_range r("Equal::eval_gpu");
   auto& s = out.primitive().stream();
-  auto op = get_primitive_string(this);
   if (equal_nan_) {
-    binary_op_gpu<cu::NaNEqual>(inputs, out, op, s);
+    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
   } else {
-    binary_op_gpu<cu::Equal>(inputs, out, op, s);
+    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
   }
 }
 
 void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
   nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
   auto& s = out.primitive().stream();
-  auto op = get_primitive_string(this);
   switch (op_) {
     case BitwiseBinary::And:
-      binary_op_gpu<cu::BitwiseAnd>(inputs, out, op, s);
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
       break;
     case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, op, s);
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
       break;
     case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, op, s);
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
       break;
     case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, op, s);
+      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
       break;
     case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, op, s);
+      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
       break;
   }
 }

mlx/backend/cuda/binary_two.cu

@@ -0,0 +1,337 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b[0]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[0], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a[0], b_vec.val[i]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[0]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec.val[i], b[0]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void
+binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      auto out = Op{}(a[i], b[i]);
+      out_a[i] = out[0];
+      out_b[i] = out[1];
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+
+    AlignedVector<Out, N_READS> out_a_vec;
+    AlignedVector<Out, N_READS> out_b_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      auto out = Op{}(a_vec.val[i], b_vec.val[i]);
+      out_a_vec.val[i] = out[0];
+      out_b_vec.val[i] = out[1];
+    }
+
+    store_vector<N_READS>(out_a, index, out_a_vec);
+    store_vector<N_READS>(out_b, index, out_b_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+__global__ void binary_two_g_nd(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+        index, shape.data(), a_strides.data(), b_strides.data());
+    auto out = Op{}(a[a_idx], b[b_idx]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void binary_two_g(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides a_strides,
+    const __grid_constant__ Strides b_strides,
+    int ndim) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    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];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out>
+constexpr bool supports_binary_two_op() {
+  if (std::is_same_v<Op, DivMod>) {
+    return std::is_same_v<In, Out> &&
+        (std::is_integral_v<Out> || is_floating_v<Out>);
+  }
+  return false;
+}
+
+} // namespace cu
+
+template <typename Op>
+void binary_two_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const char* op,
+    const Stream& s) {
+  assert(inputs.size() > 1);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  auto& out_a = outputs[0];
+  auto& out_b = outputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out_a, bopt);
+  set_binary_op_output_data(a, b, out_b, bopt);
+
+  if (out_a.size() == 0) {
+    return;
+  }
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out_a);
+  encoder.set_output_array(out_b);
+  dispatch_all_types(a.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out_a.dtype(), [&](auto out_type_tag) {
+      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
+      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
+      if constexpr (cu::supports_binary_two_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+        using InType = cuda_type_t<CTYPE_IN>;
+        using OutType = cuda_type_t<CTYPE_OUT>;
+
+        auto bopt = get_binary_op_type(a, b);
+        if (bopt == BinaryOpType::General) {
+          dispatch_bool(
+              a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
+                  out_a.data_size() > INT32_MAX,
+              [&](auto large) {
+                using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+                Shape shape;
+                std::vector<Strides> strides;
+                std::tie(shape, strides) =
+                    collapse_contiguous_dims(a, b, out_a);
+                auto& a_strides = strides[0];
+                auto& b_strides = strides[1];
+                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(kernel, out_a, large());
+                    encoder.add_kernel_node(
+                        kernel,
+                        num_blocks,
+                        block_dims,
+                        0,
+                        a.data<InType>(),
+                        b.data<InType>(),
+                        out_a.data<OutType>(),
+                        out_b.data<OutType>(),
+                        out_a.size(),
+                        const_param<dims_constant()>(shape),
+                        const_param<dims_constant()>(a_strides),
+                        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(kernel, out_a, large());
+                  encoder.add_kernel_node(
+                      kernel,
+                      num_blocks,
+                      block_dims,
+                      0,
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out_a.data<OutType>(),
+                      out_b.data<OutType>(),
+                      out_a.size(),
+                      const_param(shape),
+                      const_param(a_strides),
+                      const_param(b_strides),
+                      ndim);
+                }
+              });
+        } else {
+          dispatch_bool(out_a.data_size() > UINT32_MAX, [&](auto large) {
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::binary_two_ss<Op, InType, OutType, IdxT, N_READS>;
+            if (bopt == BinaryOpType::ScalarVector) {
+              kernel = cu::binary_two_sv<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorScalar) {
+              kernel = cu::binary_two_vs<Op, InType, OutType, IdxT, N_READS>;
+            } else if (bopt == BinaryOpType::VectorVector) {
+              kernel = cu::binary_two_vv<Op, InType, OutType, IdxT, N_READS>;
+            }
+            auto [num_blocks, block_dims] = get_launch_args(
+                kernel,
+                out_a.data_size(),
+                out_a.shape(),
+                out_a.strides(),
+                large(),
+                N_READS);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                a.data<InType>(),
+                b.data<InType>(),
+                out_a.data<OutType>(),
+                out_b.data<OutType>(),
+                out_a.data_size());
+          });
+        }
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do binary op {} on inputs of {} with result of {}.",
+            op,
+            dtype_to_string(a.dtype()),
+            dtype_to_string(out_a.dtype())));
+      }
+    });
+  });
+}
+
+template <typename Op>
+void binary_two_op_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    const char* op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, outputs[0], bopt);
+  set_binary_op_output_data(a, b, outputs[1], bopt);
+  binary_two_op_gpu_inplace<Op>(inputs, outputs, op, s);
+}
+
+void DivMod::eval_gpu(
+    const std::vector<array>& inputs,
+    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);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/compiled.cpp

@@ -3,6 +3,7 @@
 #include "mlx/backend/common/compiled.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/jit_module.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/graph_utils.h"
 #include "mlx/primitives.h"
 
@@ -52,9 +53,10 @@ struct FusedKernelBuilder {
 
     // Build function signature.
     if (contiguous) {
-      os += "template <typename IdxT = uint32_t>\n";
+      os += "template <typename IdxT = uint32_t, int work_per_thread = 1>\n";
     } else {
-      os += "template <int NDIM, typename IdxT = uint32_t>\n";
+      os +=
+          "template <int NDIM, typename IdxT = uint32_t, int work_per_thread = 1>\n";
     }
     os += fmt::format("__global__ void {}(\n", kernel_name + name);
     for (size_t i = 0; i < params.size(); ++i) {
@@ -66,12 +68,46 @@ struct FusedKernelBuilder {
     }
     os += ") {\n";
 
-    // Index.
+    // Index. For non contiguous kernels we create a separate index
+    // variable per variable otherwise everyone uses `index`.
     os +=
-        "  IdxT index = cg::this_grid().thread_rank();\n"
+        "  IdxT index = cg::this_grid().thread_rank() * work_per_thread;\n"
         "  if (index >= size) {\n"
         "    return;\n"
         "  }\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "  IdxT " + xname + "_idx = 0;\n";
+      }
+      os += "  {\n";
+      os += "    IdxT loc = index;\n";
+      os +=
+          "    #pragma unroll\n"
+          "    for (int i = NDIM - 1; i >= 0; i--) {\n";
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "      " + xname + "_idx += (loc \% shape[i]) * IdxT(" + xname +
+            "_strides[i]);\n";
+      }
+      os +=
+          "      loc /= shape[i];\n"
+          "    }\n"
+          "  }\n";
+    }
+
+    // Work loop
+    os +=
+        "\n"
+        "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
 
     // Read inputs.
     for (size_t i = 0; i < inputs.size(); ++i) {
@@ -88,12 +124,9 @@ struct FusedKernelBuilder {
       } else if (contiguous) {
         value = fmt::format("{}[index]", xname);
       } else {
-        std::string index = fmt::format(
-            "elem_to_loc_nd<NDIM>(index, shape.data(), {}_strides.data())",
-            xname);
-        value = fmt::format("{}[{}]", xname, index);
+        value = fmt::format("{}[{}_idx]", xname, xname);
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write tape.
@@ -105,23 +138,37 @@ struct FusedKernelBuilder {
         value = fmt::format(
             "static_cast<{}>(tmp_{})", type, namer.get_name(x.inputs()[0]));
       } else {
-        std::ostringstream ss;
-        x.primitive().print(ss);
-        value = ss.str();
+        value = x.primitive().name();
         value += "{}(";
         for (size_t i = 0; i < x.inputs().size() - 1; ++i) {
           value += fmt::format("tmp_{}, ", namer.get_name(x.inputs()[i]));
         }
         value += fmt::format("tmp_{})", namer.get_name(x.inputs().back()));
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write output.
     for (const auto& x : outputs) {
-      os += fmt::format("  {0}[index] = tmp_{0};\n", namer.get_name(x));
+      os += fmt::format("    {0}[index] = tmp_{0};\n", namer.get_name(x));
     }
 
+    // End of work loop
+    os +=
+        "\n"
+        "    index++;\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "    " + xname + "_idx += " + xname + "_strides[NDIM - 1];\n";
+      }
+    }
+    os += "  }\n";
+
     os += "}\n";
   }
 };
@@ -157,15 +204,28 @@ void Compiled::eval_gpu(
     builder.build("_strided", false);
     builder.os += "\n} // namespace mlx::core::cu\n";
     // Build kernel names.
-    std::vector<std::string> kernel_names = {
-        fmt::format("mlx::core::cu::{}_contiguous<uint32_t>", lib_name()),
-        fmt::format("mlx::core::cu::{}_contiguous<int64_t>", lib_name()),
-    };
-    for (int i = 1; i <= MAX_NDIM; ++i) {
+    std::vector<std::string> kernel_names;
+    for (auto work_per_thread : std::array<int, 2>{1, 4}) {
       kernel_names.push_back(fmt::format(
-          "mlx::core::cu::{}_strided<{}, uint32_t>", lib_name(), i));
-      kernel_names.push_back(
-          fmt::format("mlx::core::cu::{}_strided<{}, int64_t>", lib_name(), i));
+          "mlx::core::cu::{}_contiguous<uint32_t, {}>",
+          lib_name(),
+          work_per_thread));
+      kernel_names.push_back(fmt::format(
+          "mlx::core::cu::{}_contiguous<int64_t, {}>",
+          lib_name(),
+          work_per_thread));
+      for (int i = 1; i <= MAX_NDIM; ++i) {
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, uint32_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, int64_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+      }
     }
     return std::make_pair(std::move(builder.os), std::move(kernel_names));
   });
@@ -178,6 +238,7 @@ void Compiled::eval_gpu(
   // Whether to use large index.
   bool large = compiled_use_large_index(inputs, outputs, contiguous);
 
+  cu::KernelArgs args;
   // Put inputs.
   int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
@@ -185,35 +246,43 @@ void Compiled::eval_gpu(
       continue;
     }
     const auto& x = inputs[i];
-    mod.append_arg(x);
+    args.append(x);
     if (!contiguous && !is_scalar(x)) {
-      mod.append_arg(strides_vec[strides_index++]);
+      args.append_ptr(strides_vec[strides_index++].data());
     }
   }
 
   // Put outputs.
   compiled_allocate_outputs(inputs, outputs, is_constant_, contiguous);
   for (auto& x : outputs) {
-    mod.append_arg(x);
+    args.append(x);
   }
 
   // Put shape and size.
   if (!contiguous) {
-    mod.append_arg(shape);
+    args.append_ptr(shape.data());
   }
   if (large) {
-    mod.append_arg<int64_t>(outputs[0].data_size());
+    args.append<int64_t>(outputs[0].data_size());
   } else {
-    mod.append_arg<uint32_t>(outputs[0].data_size());
+    args.append<uint32_t>(outputs[0].data_size());
+  }
+
+  // Choose work per thread
+  int work_per_thread = 4;
+  if (!contiguous && shape.back() % work_per_thread != 0) {
+    work_per_thread = 1;
   }
 
   // Launch kernel.
   const char* index_type = large ? "int64_t" : "uint32_t";
   std::string kernel_name = fmt::format("mlx::core::cu::{}", lib_name());
   if (contiguous) {
-    kernel_name += fmt::format("_contiguous<{}>", index_type);
+    kernel_name +=
+        fmt::format("_contiguous<{}, {}>", index_type, work_per_thread);
   } else {
-    kernel_name += fmt::format("_strided<{}, {}>", shape.size(), index_type);
+    kernel_name += fmt::format(
+        "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
   }
   auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
@@ -222,9 +291,11 @@ void Compiled::eval_gpu(
   for (const auto& out : outputs) {
     encoder.set_output_array(out);
   }
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    mod.launch_kernel(stream, kernel_name, outputs[0], large);
-  });
+
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] =
+      get_launch_args(kernel, outputs[0], large, work_per_thread);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 } // namespace mlx::core

mlx/backend/cuda/copy.cu

@@ -24,7 +24,6 @@ void copy_gpu_inplace(
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-
   if (ctype == CopyType::Scalar || ctype == CopyType::Vector) {
     copy_contiguous(encoder, ctype, in, out, offset_in, offset_out);
     return;

mlx/backend/cuda/copy/copy.cuh

@@ -10,15 +10,6 @@
 
 namespace mlx::core {
 
-#define MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, ...) \
-  MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE_IN, {               \
-    MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, {           \
-      using InType = cuda_type_t<CTYPE_IN>;                  \
-      using OutType = cuda_type_t<CTYPE_OUT>;                \
-      __VA_ARGS__;                                           \
-    });                                                      \
-  })
-
 void copy_contiguous(
     cu::CommandEncoder& encoder,
     CopyType ctype,

mlx/backend/cuda/copy/copy_contiguous.cu

@@ -10,19 +10,43 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_s(const In* in, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = CastOp<In, Out>{}(in[0]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[0]);
+    }
+  } else {
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = cast_to<Out>(in[0]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
-template <typename In, typename Out, typename IdxT>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_v(const In* in, Out* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = CastOp<In, Out>{}(in[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = cast_to<Out>(in[i]);
+    }
+  } else {
+    auto in_vec = load_vector<N_READS>(in, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = cast_to<Out>(in_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -35,17 +59,30 @@ void copy_contiguous(
     array& out,
     int64_t in_offset,
     int64_t out_offset) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
-        auto kernel = cu::copy_s<InType, OutType, IdxT>;
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+        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>;
+        // TODO: Choose optimized value based on type size.
+        constexpr int N_READS = 4;
+        auto kernel = cu::copy_s<InType, OutType, IdxT, N_READS>;
         if (ctype == CopyType::Vector) {
-          kernel = cu::copy_v<InType, OutType, IdxT>;
+          kernel = cu::copy_v<InType, OutType, IdxT, N_READS>;
         }
         auto [num_blocks, block_dims] = get_launch_args(
-            kernel, out.data_size(), out.shape(), out.strides(), LARGE);
-        kernel<<<num_blocks, block_dims, 0, stream>>>(
+            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());

mlx/backend/cuda/copy/copy_general.cu

@@ -55,39 +55,56 @@ void copy_general(
     const Shape& shape,
     const Strides& strides_in,
     const Strides& strides_out) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      const InType* in_ptr = in.data<InType>() + offset_in;
-      OutType* out_ptr = out.data<OutType>() + offset_out;
-      bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-      MLX_SWITCH_BOOL(large, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-        int ndim = shape.size();
-        if (ndim <= 3) {
-          MLX_SWITCH_1_2_3(ndim, NDIM, {
-            auto kernel = cu::copy_gg_nd<InType, OutType, IdxT, NDIM>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-            kernel<<<num_blocks, block_dims, 0, stream>>>(
-                in_ptr,
-                out_ptr,
-                out.size(),
-                const_param<NDIM>(shape),
-                const_param<NDIM>(strides_in),
-                const_param<NDIM>(strides_out));
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            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, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            int ndim = shape.size();
+            size_t data_size = 1;
+            for (auto& s : shape)
+              data_size *= s;
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto ndim_constant) {
+                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(
+                    kernel,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    data_size,
+                    const_param<ndim_constant()>(shape),
+                    const_param<ndim_constant()>(strides_in),
+                    const_param<ndim_constant()>(strides_out));
+              });
+            } else { // ndim >= 4
+              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(
+                  kernel,
+                  num_blocks,
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  data_size,
+                  const_param(shape),
+                  const_param(strides_in),
+                  const_param(strides_out),
+                  ndim);
+            }
           });
-        } else { // ndim >= 4
-          auto kernel = cu::copy_gg<InType, OutType, IdxT>;
-          auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in_ptr,
-              out_ptr,
-              out.size(),
-              const_param(shape),
-              const_param(strides_in),
-              const_param(strides_out),
-              ndim);
-        }
-      });
     });
   });
 }

mlx/backend/cuda/copy/copy_general_dynamic.cu

@@ -61,43 +61,57 @@ void copy_general_dynamic(
     const Strides& strides_out,
     const array& dynamic_offset_in,
     const array& dynamic_offset_out) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      const InType* in_ptr = in.data<InType>() + offset_in;
-      OutType* out_ptr = out.data<OutType>() + offset_out;
-      bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-      MLX_SWITCH_BOOL(large, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-        int ndim = shape.size();
-        if (ndim <= 3) {
-          MLX_SWITCH_1_2_3(ndim, NDIM, {
-            auto kernel = cu::copy_gg_dynamic_nd<InType, OutType, IdxT, NDIM>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-            kernel<<<num_blocks, block_dims, 0, stream>>>(
-                in_ptr,
-                out_ptr,
-                out.size(),
-                const_param<NDIM>(shape),
-                const_param<NDIM>(strides_in),
-                const_param<NDIM>(strides_out),
-                dynamic_offset_in.data<int64_t>(),
-                dynamic_offset_out.data<int64_t>());
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            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, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            int ndim = shape.size();
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                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(
+                    kernel,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    out.size(),
+                    const_param<dims_constant()>(shape),
+                    const_param<dims_constant()>(strides_in),
+                    const_param<dims_constant()>(strides_out),
+                    dynamic_offset_in.data<int64_t>(),
+                    dynamic_offset_out.data<int64_t>());
+              });
+            } else { // ndim >= 4
+              auto kernel = cu::copy_gg_dynamic<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] =
+                  get_launch_args(kernel, out, large());
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  out.size(),
+                  const_param(shape),
+                  const_param(strides_in),
+                  const_param(strides_out),
+                  ndim,
+                  dynamic_offset_in.data<int64_t>(),
+                  dynamic_offset_out.data<int64_t>());
+            }
           });
-        } else { // ndim >= 4
-          auto kernel = cu::copy_gg_dynamic<InType, OutType, IdxT>;
-          auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in_ptr,
-              out_ptr,
-              out.size(),
-              const_param(shape),
-              const_param(strides_in),
-              const_param(strides_out),
-              ndim,
-              dynamic_offset_in.data<int64_t>(),
-              dynamic_offset_out.data<int64_t>());
-        }
-      });
     });
   });
 }

mlx/backend/cuda/copy/copy_general_input.cu

@@ -50,37 +50,51 @@ void copy_general_input(
     int64_t offset_out,
     const Shape& shape,
     const Strides& strides_in) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_COPY_TYPES(in, out, InType, OutType, {
-      const InType* in_ptr = in.data<InType>() + offset_in;
-      OutType* out_ptr = out.data<OutType>() + offset_out;
-      bool large = in.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-      MLX_SWITCH_BOOL(large, LARGE, {
-        using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-        int ndim = shape.size();
-        if (ndim <= 3) {
-          MLX_SWITCH_1_2_3(ndim, NDIM, {
-            auto kernel = cu::copy_g_nd<InType, OutType, IdxT, NDIM>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-            kernel<<<num_blocks, block_dims, 0, stream>>>(
-                in_ptr,
-                out_ptr,
-                out.size(),
-                const_param<NDIM>(shape),
-                const_param<NDIM>(strides_in));
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      dispatch_bool(
+          in.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            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, int32_t>;
+            const InType* in_ptr = in.data<InType>() + offset_in;
+            OutType* out_ptr = out.data<OutType>() + offset_out;
+            int ndim = shape.size();
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                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(
+                    kernel,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    in_ptr,
+                    out_ptr,
+                    out.size(),
+                    const_param<dims_constant()>(shape),
+                    const_param<dims_constant()>(strides_in));
+              });
+            } else { // ndim >= 4
+              auto kernel = cu::copy_g<InType, OutType, IdxT>;
+              auto [num_blocks, block_dims] =
+                  get_launch_args(kernel, out, large());
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dims,
+                  0,
+                  in_ptr,
+                  out_ptr,
+                  out.size(),
+                  const_param(shape),
+                  const_param(strides_in),
+                  ndim);
+            }
           });
-        } else { // ndim >= 4
-          auto kernel = cu::copy_g<InType, OutType, IdxT>;
-          auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in_ptr,
-              out_ptr,
-              out.size(),
-              const_param(shape),
-              const_param(strides_in),
-              ndim);
-        }
-      });
     });
   });
 }

mlx/backend/cuda/device.cpp

@@ -2,37 +2,28 @@
 
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/worker.h"
-#include "mlx/backend/metal/metal.h"
+#include "mlx/utils.h"
 
 #include <fmt/format.h>
 #include <nvtx3/nvtx3.hpp>
+#include <future>
+#include <unordered_set>
 
 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;
+
+int cuda_graph_cache_size() {
+  static int cache_size = []() {
+    return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 100);
+  }();
+  return cache_size;
+}
+
 namespace cu {
 
-DeviceStream::DeviceStream(Device& device) : device_(device), stream_(device) {}
-
-void DeviceStream::synchronize() {
-  cudaStreamSynchronize(stream_);
-}
-
-cudaStream_t DeviceStream::schedule_cuda_stream() {
-  // TODO: Return a stream that maximizes parallelism.
-  return stream_;
-}
-
-cudaStream_t DeviceStream::last_cuda_stream() {
-  return stream_;
-}
-
-CommandEncoder& DeviceStream::get_encoder() {
-  if (!encoder_) {
-    encoder_ = std::make_unique<CommandEncoder>(*this);
-  }
-  return *encoder_;
-}
-
 Device::Device(int device) : device_(device) {
   CHECK_CUDA_ERROR(cudaDeviceGetAttribute(
       &compute_capability_major_, cudaDevAttrComputeCapabilityMajor, device_));
@@ -66,45 +57,271 @@ void Device::make_current() {
   }
 }
 
-DeviceStream& Device::get_stream(Stream s) {
-  auto it = streams_.find(s.index);
-  if (it == streams_.end()) {
-    it = streams_.try_emplace(s.index, *this).first;
+CommandEncoder& Device::get_command_encoder(Stream s) {
+  auto it = encoders_.find(s.index);
+  if (it == encoders_.end()) {
+    it = encoders_.try_emplace(s.index, *this).first;
   }
   return it->second;
 }
 
-CommandEncoder::CommandEncoder(DeviceStream& s)
-    : device_(s.device()), stream_(s) {}
+CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
+  CHECK_CUDA_ERROR(
+      cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
+}
+
+CommandEncoder::CaptureContext::~CaptureContext() {
+  CHECK_CUDA_ERROR(cudaStreamEndCapture(enc.stream(), &graph));
+  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'});
+  }
+  CHECK_CUDA_ERROR(cudaGraphDestroy(graph));
+}
+
+CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
+    : enc(enc) {
+  enc.in_concurrent_ = true;
+}
+
+CommandEncoder::ConcurrentContext::~ConcurrentContext() {
+  enc.in_concurrent_ = false;
+
+  // Use an empty graph node for synchronization
+  CommandEncoder::GraphNode empty{NULL, 'E', std::to_string(enc.node_count_++)};
+  enc.empty_node_count_++;
+  CHECK_CUDA_ERROR(cudaGraphAddEmptyNode(&empty.node, enc.graph_, NULL, 0));
+
+  // Insert the concurrent -> empty node dependencies
+  for (auto& from : enc.concurrent_nodes_) {
+    enc.from_nodes_.push_back(from.node);
+    enc.to_nodes_.push_back(empty.node);
+    enc.graph_key_ += from.id;
+    enc.graph_key_ += from.node_type;
+    enc.graph_key_ += empty.id;
+    enc.graph_key_ += empty.node_type;
+  }
+
+  // Insert the input -> concurrent node dependencies without updating output
+  // nodes
+  auto outputs = std::move(enc.active_outputs_);
+  enc.insert_graph_dependencies(std::move(enc.concurrent_nodes_));
+
+  // Update output node to be the empty node
+  for (auto o : outputs) {
+    enc.node_map_.emplace(o, empty).first->second = empty;
+  }
+}
+
+void CommandEncoder::insert_graph_dependencies(GraphNode node) {
+  if (node.node_type == 'G') {
+    graph_node_count_++;
+  }
+  node.id = std::to_string(node_count_++);
+  if (in_concurrent_) {
+    concurrent_nodes_.push_back(std::move(node));
+  } else {
+    std::vector<GraphNode> nodes;
+    nodes.push_back(std::move(node));
+    insert_graph_dependencies(std::move(nodes));
+  }
+}
+
+void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
+  std::vector<GraphNode> deps;
+  {
+    // Dependencies must be added in the same order to produce a consistent
+    // topology
+    std::unordered_set<cudaGraphNode_t> set_deps;
+    for (auto d : active_deps_) {
+      if (auto it = node_map_.find(d); it != node_map_.end()) {
+        auto [_, inserted] = set_deps.insert(it->second.node);
+        if (inserted) {
+          deps.push_back(it->second);
+        }
+      }
+    }
+  }
+  active_deps_.clear();
+
+  for (auto o : active_outputs_) {
+    for (auto& node : nodes) {
+      node_map_.emplace(o, node).first->second = node;
+    }
+  }
+  active_outputs_.clear();
+
+  for (auto& from : deps) {
+    for (auto& to : nodes) {
+      from_nodes_.push_back(from.node);
+      to_nodes_.push_back(to.node);
+      graph_key_ += from.id;
+      graph_key_ += from.node_type;
+      graph_key_ += to.id;
+      graph_key_ += to.node_type;
+    }
+  }
+}
+
+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));
 }
 
-void CommandEncoder::end_encoding() {
-  if (!temporaries_.empty()) {
-    add_completed_handler([temporaries = std::move(temporaries_)]() {});
-  }
+void CommandEncoder::set_input_array(const array& arr) {
+  auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
+  active_deps_.push_back(id);
+}
 
-  // There is no kernel running, run completion handlers immediately.
-  if (!has_gpu_work_) {
-    worker_.consume_in_this_thread();
-    return;
-  }
-  has_gpu_work_ = false;
+void CommandEncoder::set_output_array(const array& arr) {
+  auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
+  active_deps_.push_back(id);
+  active_outputs_.push_back(id);
+}
 
-  // Put completion handlers in a batch.
-  worker_.end_batch();
-
-  // Signaling kernel completion is expensive, delay until enough batches.
-  // TODO: This number is arbitrarily picked, profile for a better stragety.
-  if (worker_.uncommited_batches() > 8) {
+void CommandEncoder::maybe_commit() {
+  if (node_count_ >= env::max_ops_per_buffer(default_max_nodes_per_graph)) {
     commit();
   }
 }
 
+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;
+  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;
+  kernel_params.gridDimX = grid_dim.x;
+  kernel_params.gridDimY = grid_dim.y;
+  kernel_params.gridDimZ = grid_dim.z;
+  kernel_params.blockDimX = block_dim.x;
+  kernel_params.blockDimY = block_dim.y;
+  kernel_params.blockDimZ = block_dim.z;
+  kernel_params.kernelParams = params;
+  kernel_params.sharedMemBytes = smem_bytes;
+  CUgraphNode node;
+  CHECK_CUDA_ERROR(
+      cuGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
+  insert_graph_dependencies(GraphNode{node, 'K'});
+}
+
 void CommandEncoder::commit() {
-  worker_.commit(stream_.last_cuda_stream());
+  if (!temporaries_.empty()) {
+    add_completed_handler([temporaries = std::move(temporaries_)]() {});
+  }
+  if (node_count_ > 0) {
+    if (!from_nodes_.empty()) {
+      CHECK_CUDA_ERROR(cudaGraphAddDependencies(
+          graph_, from_nodes_.data(), to_nodes_.data(), from_nodes_.size()));
+    }
+
+    graph_key_ += ".";
+    graph_key_ += std::to_string(node_count_);
+    graph_key_ += ".";
+    graph_key_ += std::to_string(graph_node_count_);
+    graph_key_ += ".";
+    graph_key_ += std::to_string(empty_node_count_);
+
+    cudaGraphExec_t& graph_exec = graph_cache_[graph_key_];
+
+    if (graph_exec != nullptr) {
+      cudaGraphExecUpdateResult update_result;
+#if CUDART_VERSION >= 12000
+      cudaGraphExecUpdateResultInfo info;
+      cudaGraphExecUpdate(graph_exec, graph_, &info);
+      update_result = info.result;
+#else
+      cudaGraphNode_t error_node;
+      cudaGraphExecUpdate(graph_exec, graph_, &error_node, &update_result);
+#endif // CUDART_VERSION >= 12000
+      if (update_result != cudaGraphExecUpdateSuccess) {
+        cudaGetLastError(); // reset error
+        CHECK_CUDA_ERROR(cudaGraphExecDestroy(graph_exec));
+        graph_exec = nullptr;
+      }
+    }
+    if (graph_exec == nullptr) {
+      CHECK_CUDA_ERROR(
+          cudaGraphInstantiate(&graph_exec, graph_, NULL, NULL, 0));
+    }
+    device_.make_current();
+    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+
+    // TODO smarter cache policy
+    if (graph_cache_.size() > cuda_graph_cache_size()) {
+      clear_graphs(graph_cache_);
+    }
+
+    // Reset state
+    node_count_ = 0;
+    graph_node_count_ = 0;
+    from_nodes_.clear();
+    to_nodes_.clear();
+    graph_key_.clear();
+    node_map_.clear();
+    CHECK_CUDA_ERROR(cudaGraphDestroy(graph_));
+    CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
+  }
+
+  // Put completion handlers in a batch.
+  worker_.end_batch();
+  worker_.commit(stream_);
+}
+
+void CommandEncoder::synchronize() {
+  cudaStreamSynchronize(stream_);
+  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();
 }
 
 Device& device(mlx::core::Device device) {
@@ -116,12 +333,8 @@ Device& device(mlx::core::Device device) {
   return it->second;
 }
 
-DeviceStream& get_stream(Stream s) {
-  return device(s.device).get_stream(s);
-}
-
 CommandEncoder& get_command_encoder(Stream s) {
-  return get_stream(s).get_encoder();
+  return device(s.device).get_command_encoder(s);
 }
 
 } // namespace cu

mlx/backend/cuda/device.h

@@ -7,41 +7,118 @@
 #include "mlx/stream.h"
 
 #include <cublasLt.h>
+#include <cuda.h>
 #include <thrust/execution_policy.h>
 
 #include <unordered_map>
 
 namespace mlx::core::cu {
 
-class Device;
-class CommandEncoder;
-
-class DeviceStream {
+class CommandEncoder {
  public:
-  explicit DeviceStream(Device& device);
+  struct CaptureContext {
+    CaptureContext(CommandEncoder& enc);
+    ~CaptureContext();
+    cudaGraph_t graph;
+    CommandEncoder& enc;
+  };
+  struct ConcurrentContext {
+    ConcurrentContext(CommandEncoder& enc);
+    ~ConcurrentContext();
+    CommandEncoder& enc;
+  };
 
-  DeviceStream(const DeviceStream&) = delete;
-  DeviceStream& operator=(const DeviceStream&) = delete;
+  explicit CommandEncoder(Device& d);
+  ~CommandEncoder();
 
-  // Wait until kernels in the stream complete.
-  void synchronize();
+  CommandEncoder(const CommandEncoder&) = delete;
+  CommandEncoder& operator=(const CommandEncoder&) = delete;
 
-  // Return a cuda stream for launching kernels.
-  cudaStream_t schedule_cuda_stream();
-
-  // Return the last cuda stream used.
-  cudaStream_t last_cuda_stream();
-
-  CommandEncoder& get_encoder();
-
-  Device& device() {
-    return device_;
+  CaptureContext capture_context() {
+    return CaptureContext{*this};
+  }
+  ConcurrentContext concurrent_context() {
+    return ConcurrentContext{*this};
   }
 
+  void set_input_array(const array& arr);
+  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) {
+    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);
+  }
+
+  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);
+
+  void add_temporary(const array& arr) {
+    temporaries_.push_back(arr.data_shared_ptr());
+  }
+
+  void add_completed_handler(std::function<void()> task);
+  void maybe_commit();
+  void commit();
+
+  CudaStream& stream() {
+    return stream_;
+  }
+
+  // Wait until kernels and completion handlers are finished
+  void synchronize();
+
  private:
+  struct GraphNode {
+    cudaGraphNode_t node;
+    // K = kernel
+    // E = empty
+    // G = subgraph
+    char node_type;
+    std::string id;
+  };
+
+  void insert_graph_dependencies(GraphNode node);
+  void insert_graph_dependencies(std::vector<GraphNode> nodes);
+
   Device& device_;
   CudaStream stream_;
-  std::unique_ptr<CommandEncoder> encoder_;
+  cudaGraph_t graph_;
+  Worker worker_;
+  char node_count_{0};
+  char graph_node_count_{0};
+  char empty_node_count_{0};
+  bool in_concurrent_{false};
+  std::vector<cudaGraphNode_t> from_nodes_;
+  std::vector<cudaGraphNode_t> to_nodes_;
+  std::string graph_key_;
+  std::vector<GraphNode> concurrent_nodes_;
+  std::vector<std::shared_ptr<array::Data>> temporaries_;
+  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_;
 };
 
 class Device {
@@ -55,7 +132,7 @@ class Device {
   // Make this device the current cuda device, required by some cuda calls.
   void make_current();
 
-  DeviceStream& get_stream(Stream s);
+  CommandEncoder& get_command_encoder(Stream s);
 
   int cuda_device() const {
     return device_;
@@ -75,64 +152,10 @@ class Device {
   int compute_capability_major_;
   int compute_capability_minor_;
   cublasLtHandle_t lt_;
-  std::unordered_map<int, DeviceStream> streams_;
-};
-
-class CommandEncoder {
- public:
-  explicit CommandEncoder(DeviceStream& stream);
-
-  CommandEncoder(const CommandEncoder&) = delete;
-  CommandEncoder& operator=(const CommandEncoder&) = delete;
-
-  void set_input_array(const array& arr) {}
-  void set_output_array(const array& arr) {}
-
-  void add_temporary(const array& arr) {
-    temporaries_.push_back(arr.data_shared_ptr());
-  }
-
-  void add_completed_handler(std::function<void()> task);
-  void end_encoding();
-  void commit();
-
-  // Schedule a cuda stream for |fun| to launch kernels, and check error
-  // afterwards.
-  template <typename F>
-  void launch_kernel(F&& fun) {
-    launch_kernel(stream_.schedule_cuda_stream(), std::forward<F>(fun));
-  }
-
-  template <typename F>
-  void launch_kernel(cudaStream_t stream, F&& fun) {
-    device_.make_current();
-    fun(stream);
-    check_cuda_error("kernel launch", cudaGetLastError());
-    has_gpu_work_ = true;
-  }
-
-  Device& device() {
-    return device_;
-  }
-
-  DeviceStream& stream() {
-    return stream_;
-  }
-
-  bool has_gpu_work() const {
-    return has_gpu_work_;
-  }
-
- private:
-  Device& device_;
-  DeviceStream& stream_;
-  Worker worker_;
-  bool has_gpu_work_{false};
-  std::vector<std::shared_ptr<array::Data>> temporaries_;
+  std::unordered_map<int, CommandEncoder> encoders_;
 };
 
 Device& device(mlx::core::Device device);
-DeviceStream& get_stream(Stream s);
 CommandEncoder& get_command_encoder(Stream s);
 
 // Return an execution policy that does not sync for result.

mlx/backend/cuda/device/atomic_ops.cuh

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

mlx/backend/cuda/device/binary_ops.cuh

@@ -1,10 +1,7 @@
 // Copyright © 2025 Apple Inc.
 
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
-#include "mlx/backend/cuda/device/fp16_math.cuh"
-#include "mlx/backend/cuda/device/utils.cuh"
+#include "mlx/backend/cuda/device/unary_ops.cuh"
 
-#include <cuComplex.h>
 #include <cuda/std/array>
 
 namespace mlx::core::cu {
@@ -22,7 +19,7 @@ struct FloorDivide {
     if constexpr (cuda::std::is_integral_v<T>) {
       return x / y;
     } else {
-      return trunc(x / y);
+      return truncf(x / y);
     }
   }
 };
@@ -47,7 +44,7 @@ struct Remainder {
       } else {
         return x % y;
       }
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+    } else if constexpr (is_complex_v<T>) {
       return x % y;
     } else {
       T r = fmod(x, y);
@@ -69,14 +66,12 @@ struct Equal {
 struct NaNEqual {
   template <typename T>
   __device__ bool operator()(T x, T y) {
-    if constexpr (std::is_same_v<T, cuComplex>) {
+    if constexpr (is_complex_v<T>) {
       return x == y ||
-          (isnan(cuCrealf(x)) && isnan(cuCrealf(y)) && isnan(cuCimagf(x)) &&
-           isnan(cuCimagf(y))) ||
-          (cuCrealf(x) == cuCrealf(y) && isnan(cuCimagf(x)) &&
-           isnan(cuCimagf(y))) ||
-          (isnan(cuCrealf(x)) && isnan(cuCrealf(y)) &&
-           cuCimagf(x) == cuCimagf(y));
+          (isnan(x.real()) && isnan(y.real()) && isnan(x.imag()) &&
+           isnan(y.imag())) ||
+          (x.real() == y.real() && isnan(x.imag()) && isnan(y.imag())) ||
+          (isnan(x.real()) && isnan(y.real()) && x.imag() == y.imag());
     } else {
       return x == y || (isnan(x) && isnan(y));
     }
@@ -114,36 +109,38 @@ struct LessEqual {
 struct LogAddExp {
   template <typename T>
   __device__ T operator()(T x, T y) {
-    if (isnan(x) || isnan(y)) {
-      return cuda::std::numeric_limits<T>::quiet_NaN();
+    if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag()) || isnan(y.real()) ||
+          isnan(y.imag())) {
+        return {
+            cuda::std::numeric_limits<float>::quiet_NaN(),
+            cuda::std::numeric_limits<float>::quiet_NaN()};
+      }
+      auto max = x.real() > y.real() ? x : y;
+      auto min = x.real() < y.real() ? x : y;
+      auto min_real = min.real();
+      auto max_real = max.real();
+      if (!isfinite(min_real) && (min_real == max_real)) {
+        if (min_real < 0) {
+          return min;
+        } else {
+          return Log{}(Exp{}(min) + Exp{}(max));
+        }
+      } else {
+        return Log1p{}(Exp{}(min - max)) + max;
+      }
+    } else {
+      if (isnan(x) || isnan(y)) {
+        return cuda::std::numeric_limits<T>::quiet_NaN();
+      }
+      T maxval = max(x, y);
+      T minval = min(x, y);
+      return (minval == -cuda::std::numeric_limits<T>::infinity() ||
+              maxval == cuda::std::numeric_limits<T>::infinity())
+          ? maxval
+          : T(float(maxval) + log1p(expf(minval - maxval)));
     }
-    T maxval = max(x, y);
-    T minval = min(x, y);
-    return (minval == -cuda::std::numeric_limits<T>::infinity() ||
-            maxval == cuda::std::numeric_limits<T>::infinity())
-        ? maxval
-        : T(float(maxval) + log1p(expf(minval - maxval)));
   };
-
-  __device__ cuComplex operator()(cuComplex x, cuComplex y) {
-    if (isnan(cuCrealf(x)) || isnan(cuCimagf(x)) || isnan(cuCrealf(y)) ||
-        isnan(cuCimagf(y))) {
-      return {
-          cuda::std::numeric_limits<float>::quiet_NaN(),
-          cuda::std::numeric_limits<float>::quiet_NaN()};
-    }
-    constexpr float inf = cuda::std::numeric_limits<float>::infinity();
-    auto maxval = x > y ? x : y;
-    auto minval = x < y ? x : y;
-    if (cuCrealf(minval) == -inf || cuCrealf(maxval) == inf)
-      return maxval;
-    float m = exp(cuCrealf(minval) - cuCrealf(maxval));
-    cuComplex dexp{
-        m * cos(cuCimagf(minval) - cuCimagf(maxval)),
-        m * sin(cuCimagf(minval) - cuCimagf(maxval)),
-    };
-    return maxval + log1p(dexp);
-  }
 };
 
 struct Maximum {
@@ -151,8 +148,8 @@ struct Maximum {
   __device__ T operator()(T x, T y) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return max(x, y);
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (isnan(cuCrealf(x)) || isnan(cuCimagf(x))) {
+    } else if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag())) {
         return x;
       }
       return x > y ? x : y;
@@ -170,8 +167,8 @@ struct Minimum {
   __device__ T operator()(T x, T y) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return min(x, y);
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (isnan(cuCrealf(x)) || isnan(cuCimagf(x))) {
+    } else if constexpr (is_complex_v<T>) {
+      if (isnan(x.real()) || isnan(x.imag())) {
         return x;
       }
       return x < y ? x : y;
@@ -194,8 +191,8 @@ struct Multiply {
 struct NotEqual {
   template <typename T>
   __device__ bool operator()(T x, T y) {
-    if constexpr (std::is_same_v<T, cuComplex>) {
-      return cuCrealf(x) != cuCrealf(y) || cuCimagf(x) != cuCimagf(y);
+    if constexpr (is_complex_v<T>) {
+      return x.real() != y.real() || x.imag() != y.imag();
     } else {
       return x != y;
     }
@@ -215,19 +212,8 @@ struct Power {
         base *= base;
       }
       return res;
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (base.y == 0 && base.x == 0) {
-        if (isnan(exp.x) || isnan(exp.y)) {
-          auto nan = cuda::std::numeric_limits<float>::quiet_NaN();
-          return make_cuFloatComplex(nan, nan);
-        }
-        return make_cuFloatComplex(0.0, 0.0);
-      }
-      auto x_theta = atan2f(base.y, base.x);
-      auto x_ln_r = 0.5 * logf(base.x * base.x + base.y * base.y);
-      auto mag = expf(exp.x * x_ln_r - exp.y * x_theta);
-      auto phase = exp.y * x_ln_r + exp.x * x_theta;
-      return make_cuFloatComplex(mag * cosf(phase), mag * sinf(phase));
+    } else if constexpr (is_complex_v<T>) {
+      return pow(base, exp);
     } else {
       return powf(base, exp);
     }

mlx/backend/cuda/device/cast_op.cuh

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

mlx/backend/cuda/device/complex.cuh

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

mlx/backend/cuda/device/config.h

@@ -5,7 +5,7 @@
 #pragma once
 
 // The maximum dimensions of shape/strides passed as kernel parameters.
-#define MAX_NDIM 8
+#define MAX_NDIM 10
 
 // All existing NVIDIA hardware has a fixed 32 warp size. Though a built-in
 // warpSize variable exists, using it would prevent compile-time optimizations.

mlx/backend/cuda/device/cucomplex_math.cuh

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

mlx/backend/cuda/device/unary_ops.cuh

@@ -14,8 +14,6 @@ struct Abs {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_unsigned_v<T>) {
       return x;
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {sqrt(cuCrealf(x) * cuCrealf(x) + cuCimagf(x) * cuCimagf(x)), 0};
     } else {
       return abs(x);
     }
@@ -76,6 +74,8 @@ struct Ceil {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return x;
+    } else if constexpr (is_complex_v<T>) {
+      return T{ceil(x.real()), ceil(x.imag())};
     } else {
       return ceil(x);
     }
@@ -83,34 +83,23 @@ struct Ceil {
 };
 
 struct Conjugate {
-  __device__ cuComplex operator()(cuComplex x) {
-    return {cuCrealf(x), -cuCimagf(x)};
+  template <typename T>
+  __device__ complex_t<T> operator()(complex_t<T> x) {
+    return conj(x);
   }
 };
 
 struct Cos {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          cos(cuCrealf(x)) * cosh(cuCimagf(x)),
-          -sin(cuCrealf(x)) * sinh(cuCimagf(x))};
-    } else {
-      return cos(x);
-    }
+    return cos(x);
   }
 };
 
 struct Cosh {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          cosh(cuCrealf(x)) * cos(cuCimagf(x)),
-          sinh(cuCrealf(x)) * sin(cuCimagf(x))};
-    } else {
-      return cosh(x);
-    }
+    return cosh(x);
   }
 };
 
@@ -143,12 +132,7 @@ struct ErfInv {
 struct Exp {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      auto m = exp(cuCrealf(x));
-      return {m * cos(cuCimagf(x)), m * sinh(cuCimagf(x))};
-    } else {
-      return exp(x);
-    }
+    return exp(x);
   }
 };
 
@@ -170,6 +154,8 @@ struct Floor {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_integral_v<T>) {
       return x;
+    } else if constexpr (is_complex_v<T>) {
+      return T{floor(x.real()), floor(x.imag())};
     } else {
       return floor(x);
     }
@@ -177,30 +163,25 @@ struct Floor {
 };
 
 struct Imag {
-  __device__ float operator()(cuComplex x) {
-    return cuCimagf(x);
+  template <typename T>
+  __device__ auto operator()(complex_t<T> x) {
+    return x.imag();
   }
 };
 
 struct Log {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      auto r = log(cuCrealf(Abs{}(x)));
-      auto i = atan2f(cuCimagf(x), cuCrealf(x));
-      return {r, i};
-    } else {
-      return log(x);
-    }
+    return log(x);
   }
 };
 
 struct Log2 {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+    if constexpr (is_complex_v<T>) {
       auto y = Log{}(x);
-      return {cuCrealf(y) / CUDART_LN2_F, cuCimagf(y) / CUDART_LN2_F};
+      return {y.real() / CUDART_LN2_F, y.imag() / CUDART_LN2_F};
     } else {
       return log2(x);
     }
@@ -210,20 +191,31 @@ struct Log2 {
 struct Log10 {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      auto y = Log{}(x);
-      return {cuCrealf(y) / CUDART_LNT_F, cuCimagf(y) / CUDART_LNT_F};
-      return y;
-    } else {
-      return log10(x);
-    }
+    return log10(x);
   }
 };
 
 struct Log1p {
   template <typename T>
-  __device__ T operator()(T x) {
-    return log1p(x);
+  __device__ T operator()(T z) {
+    if constexpr (is_complex_v<T>) {
+      float x = z.real();
+      float y = z.imag();
+      float zabs = Abs{}(z).real();
+      float theta = atan2f(y, x + 1);
+      if (zabs < 0.5f) {
+        float r = x * (2 + x) + y * y;
+        if (r == 0) { // handle underflow
+          return {x, theta};
+        }
+        return {0.5f * log1pf(r), theta};
+      } else {
+        float z0 = hypotf(x + 1, y);
+        return {logf(z0), theta};
+      }
+    } else {
+      return log1p(z);
+    }
   }
 };
 
@@ -236,8 +228,8 @@ struct LogicalNot {
 struct Negative {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return 0 - x;
+    if constexpr (is_complex_v<T>) {
+      return T{0, 0} - x;
     } else {
       return -x;
     }
@@ -245,29 +237,23 @@ struct Negative {
 };
 
 struct Real {
-  __device__ float operator()(cuComplex x) {
-    return cuCrealf(x);
+  template <typename T>
+  __device__ auto operator()(complex_t<T> x) {
+    return x.real();
   }
 };
 
 struct Round {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {rint(cuCrealf(x)), rint(cuCimagf(x))};
+    if constexpr (is_complex_v<T>) {
+      return {rint(x.real()), rint(x.imag())};
     } else {
       return rint(x);
     }
   }
 };
 
-struct Rsqrt {
-  template <typename T>
-  __device__ T operator()(T x) {
-    return rsqrt(x);
-  }
-};
-
 struct Sigmoid {
   template <typename T>
   __device__ T operator()(T x) {
@@ -281,8 +267,8 @@ struct Sign {
   __device__ T operator()(T x) {
     if constexpr (cuda::std::is_unsigned_v<T>) {
       return x != 0;
-    } else if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      if (cuCrealf(x) == 0 && cuCimagf(x) == 0) {
+    } else if constexpr (is_complex_v<T>) {
+      if (x.real() == 0 && x.imag() == 0) {
         return x;
       } else {
         return x / Abs()(x);
@@ -298,26 +284,14 @@ struct Sign {
 struct Sin {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          sin(cuCrealf(x)) * cosh(cuCimagf(x)),
-          cos(cuCrealf(x)) * sinh(cuCimagf(x))};
-    } else {
-      return sin(x);
-    }
+    return sin(x);
   }
 };
 
 struct Sinh {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return {
-          sinh(cuCrealf(x)) * cos(cuCimagf(x)),
-          cosh(cuCrealf(x)) * sin(cuCimagf(x))};
-    } else {
-      return sinh(x);
-    }
+    return sinh(x);
   }
 };
 
@@ -335,33 +309,28 @@ struct Sqrt {
   }
 };
 
+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);
+    }
+  }
+};
+
 struct Tan {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      float tan_a = tan(cuCrealf(x));
-      float tanh_b = tanh(cuCimagf(x));
-      float t1 = tan_a * tanh_b;
-      float denom = 1. + t1 * t1;
-      return {(tan_a - tanh_b * t1) / denom, (tanh_b + tan_a * t1) / denom};
-    } else {
-      return tan(x);
-    }
+    return tan(x);
   }
 };
 
 struct Tanh {
   template <typename T>
   __device__ T operator()(T x) {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      float tanh_a = tanh(cuCrealf(x));
-      float tan_b = tan(cuCimagf(x));
-      float t1 = tanh_a * tan_b;
-      float denom = 1. + t1 * t1;
-      return {(tanh_a + tan_b * t1) / denom, (tan_b - tanh_a * t1) / denom};
-    } else {
-      return tanh(x);
-    }
+    return tanh(x);
   }
 };
 

mlx/backend/cuda/device/utils.cuh

@@ -8,9 +8,9 @@
 
 #pragma once
 
+#include "mlx/backend/cuda/device/complex.cuh"
 #include "mlx/backend/cuda/device/config.h"
 
-#include <cuComplex.h>
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 #include <cuda/std/array>
@@ -28,6 +28,27 @@ namespace mlx::core::cu {
 using Shape = cuda::std::array<int32_t, MAX_NDIM>;
 using Strides = cuda::std::array<int64_t, MAX_NDIM>;
 
+// Vectorized load/store.
+template <typename T, int N>
+struct alignas(sizeof(T) * N) AlignedVector {
+  T val[N];
+};
+
+template <int N, typename T>
+inline __device__ AlignedVector<T, N> load_vector(
+    const T* ptr,
+    uint32_t offset) {
+  auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
+  return from[offset];
+}
+
+template <int N, typename T>
+inline __device__ void
+store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
+  auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
+  to[offset] = vec;
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Type limits utils
 ///////////////////////////////////////////////////////////////////////////////
@@ -78,20 +99,20 @@ struct Limits<
     return cuda::std::numeric_limits<T>::infinity();
   }
   static constexpr __host__ __device__ T min() {
-#if defined(__CUDA_ARCH__) || CUDART_VERSION >= 12000
-    return -cuda::std::numeric_limits<T>::infinity();
-#else
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
     return -cuda::std::numeric_limits<float>::infinity();
+#else
+    return -cuda::std::numeric_limits<T>::infinity();
 #endif
   }
   static constexpr __host__ __device__ T finite_max() {
     return cuda::std::numeric_limits<T>::max();
   }
   static constexpr __host__ __device__ T finite_min() {
-#if defined(__CUDA_ARCH__) || CUDART_VERSION >= 12000
-    return cuda::std::numeric_limits<T>::lowest();
-#else
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
     return cuda::std::numeric_limits<float>::lowest();
+#else
+    return cuda::std::numeric_limits<T>::lowest();
 #endif
   }
 };
@@ -106,13 +127,13 @@ struct Limits<bool> {
   }
 };
 
-template <>
-struct Limits<cuComplex> {
-  static constexpr __host__ __device__ cuComplex max() {
-    return {Limits<float>::max(), Limits<float>::max()};
+template <typename T>
+struct Limits<complex_t<T>> {
+  static constexpr __host__ __device__ complex_t<T> max() {
+    return {Limits<T>::max(), Limits<T>::max()};
   }
-  static constexpr __host__ __device__ cuComplex min() {
-    return {Limits<float>::min(), Limits<float>::min()};
+  static constexpr __host__ __device__ complex_t<T> min() {
+    return {Limits<T>::min(), Limits<T>::min()};
   }
 };
 
@@ -155,8 +176,8 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_nd(
 #pragma unroll
   for (int i = NDIM - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
-    b_loc += dim_idx * b_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc);
@@ -175,9 +196,9 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_nd(
 #pragma unroll
   for (int i = NDIM - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
-    b_loc += dim_idx * b_strides[i];
-    c_loc += dim_idx * c_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
+    c_loc += dim_idx * IdxT(c_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc, c_loc);
@@ -206,8 +227,8 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_4d(
   IdxT b_loc = 0;
   for (int i = ndim - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
-    b_loc += dim_idx * b_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc);
@@ -226,9 +247,9 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_4d(
   IdxT c_loc = 0;
   for (int i = ndim - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
-    b_loc += dim_idx * b_strides[i];
-    c_loc += dim_idx * c_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
+    b_loc += dim_idx * IdxT(b_strides[i]);
+    c_loc += dim_idx * IdxT(c_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc, c_loc);
@@ -338,21 +359,4 @@ struct LoopedElemToLoc<1, false, OffsetT> {
   }
 };
 
-inline __device__ cuComplex log1p(cuComplex in) {
-  float x = cuCrealf(in);
-  float y = cuCimagf(in);
-  float zabs = sqrt(x * x + y * y);
-  float theta = atan2f(y, x + 1);
-  if (zabs < 0.5f) {
-    float r = x * (2 + x) + y * y;
-    if (r == 0) { // handle underflow
-      return {x, theta};
-    }
-    return {0.5f * log1pf(r), theta};
-  } else {
-    auto z0 = sqrt((x + 1) * (x + 1) + y * y);
-    return {log(z0), theta};
-  }
-}
-
 } // namespace mlx::core::cu

mlx/backend/cuda/eval.cpp

@@ -37,22 +37,20 @@ void eval(array& arr) {
   }
 
   auto& encoder = cu::get_command_encoder(arr.primitive().stream());
-  if (encoder.has_gpu_work()) {
-    // Keep used buffers alive until kernel finishes running.
-    std::unordered_set<std::shared_ptr<array::Data>> buffers;
-    for (auto& in : arr.inputs()) {
-      buffers.insert(in.data_shared_ptr());
-    }
-    for (auto& s : arr.siblings()) {
-      buffers.insert(s.data_shared_ptr());
-    }
-    // Remove the output if it was donated to by an input.
-    if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
-      buffers.erase(it);
-    }
-    encoder.add_completed_handler([buffers = std::move(buffers)]() {});
+  // Keep used buffers alive until kernel finishes running.
+  std::unordered_set<std::shared_ptr<array::Data>> buffers;
+  for (auto& in : arr.inputs()) {
+    buffers.insert(in.data_shared_ptr());
   }
-  encoder.end_encoding();
+  for (auto& s : arr.siblings()) {
+    buffers.insert(s.data_shared_ptr());
+  }
+  // Remove the output if it was donated to by an input.
+  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+    buffers.erase(it);
+  }
+  encoder.add_completed_handler([buffers = std::move(buffers)]() {});
+  encoder.maybe_commit();
 }
 
 void finalize(Stream s) {
@@ -62,7 +60,7 @@ void finalize(Stream s) {
 
 void synchronize(Stream s) {
   nvtx3::scoped_range r("gpu::synchronize");
-  cu::get_stream(s).synchronize();
+  cu::get_command_encoder(s).synchronize();
 }
 
 } // namespace mlx::core::gpu

mlx/backend/cuda/event.cu

@@ -61,7 +61,9 @@ void CudaEvent::wait(Stream s) {
   if (s.device == mlx::core::Device::cpu) {
     scheduler::enqueue(s, [*this]() mutable { wait(); });
   } else {
-    wait(cu::get_stream(s).last_cuda_stream());
+    auto& enc = cu::get_command_encoder(s);
+    enc.commit();
+    wait(enc.stream());
   }
 }
 
@@ -74,7 +76,9 @@ void CudaEvent::record(Stream s) {
   if (s.device == mlx::core::Device::cpu) {
     throw std::runtime_error("CudaEvent can not wait on cpu stream.");
   } else {
-    record(cu::get_stream(s).last_cuda_stream());
+    auto& enc = cu::get_command_encoder(s);
+    enc.commit();
+    record(enc.stream());
   }
 }
 
@@ -86,8 +90,6 @@ bool CudaEvent::completed() const {
 // SharedEvent implementations
 ///////////////////////////////////////////////////////////////////////////////
 
-namespace {
-
 __host__ __device__ void event_wait(SharedEvent::Atomic* ac, uint64_t value) {
   uint64_t current;
   while ((current = ac->load()) < value) {
@@ -108,8 +110,6 @@ __global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
   event_signal(ac, value);
 }
 
-} // namespace
-
 SharedEvent::SharedEvent() {
   // Allocate cuda::atomic on managed memory.
   Atomic* ac;
@@ -136,11 +136,9 @@ void SharedEvent::wait(Stream s, uint64_t value) {
     scheduler::enqueue(s, [*this, value]() mutable { wait(value); });
   } else {
     auto& encoder = get_command_encoder(s);
-    encoder.launch_kernel(
-        encoder.stream().last_cuda_stream(),
-        [this, value](cudaStream_t stream) { wait(stream, value); });
+    encoder.commit();
+    wait(encoder.stream(), value);
     encoder.add_completed_handler([ac = ac_]() {});
-    encoder.end_encoding();
   }
 }
 
@@ -162,11 +160,9 @@ void SharedEvent::signal(Stream s, uint64_t value) {
     scheduler::enqueue(s, [*this, value]() mutable { signal(stream, value); });
   } else {
     auto& encoder = get_command_encoder(s);
-    encoder.launch_kernel(
-        encoder.stream().last_cuda_stream(),
-        [this, value](cudaStream_t stream) { signal(stream, value); });
+    encoder.commit();
+    signal(encoder.stream(), value);
     encoder.add_completed_handler([ac = ac_]() {});
-    encoder.end_encoding();
   }
 }
 

mlx/backend/cuda/indexing.cpp

@@ -3,13 +3,16 @@
 #include "mlx/backend/common/compiled.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/jit_module.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
 #include "cuda_jit_sources.h"
 
+#include <cuda.h>
 #include <fmt/format.h>
+#include <nvrtc.h>
 #include <nvtx3/nvtx3.hpp>
 
 #include <cassert>
@@ -22,7 +25,7 @@ namespace {
 constexpr const char* g_scatter_ops[] = {"Max", "Min", "Sum", "Prod", "Assign"};
 
 void append_indices_arg(
-    cu::JitModule& mod,
+    cu::KernelArgs& args,
     const std::vector<array>& inputs,
     int nidx,
     int idx_ndim) {
@@ -30,7 +33,7 @@ void append_indices_arg(
   for (int i = 0; i < nidx; ++i) {
     indices[i] = inputs[i + 1].data<void>();
   }
-  mod.append_arg(std::move(indices));
+  args.append(std::move(indices));
   std::vector<int32_t> indices_shape(nidx * idx_ndim);
   for (int i = 0; i < nidx; ++i) {
     std::copy_n(
@@ -38,7 +41,7 @@ void append_indices_arg(
         idx_ndim,
         indices_shape.data() + i * idx_ndim);
   }
-  mod.append_arg(std::move(indices_shape));
+  args.append(std::move(indices_shape));
   std::vector<int64_t> indices_strides(nidx * idx_ndim);
   for (int i = 0; i < nidx; ++i) {
     std::copy_n(
@@ -46,7 +49,7 @@ void append_indices_arg(
         idx_ndim,
         indices_strides.data() + i * idx_ndim);
   }
-  mod.append_arg(std::move(indices_strides));
+  args.append(std::move(indices_strides));
 }
 
 } // namespace
@@ -94,20 +97,21 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
     return std::make_pair(jit_source_gather, std::move(kernel_names));
   });
 
-  mod.append_arg(src);
-  mod.append_arg(out);
+  cu::KernelArgs args;
+  args.append(src);
+  args.append(out);
   if (large) {
-    mod.append_arg<int64_t>(out.size());
+    args.append<int64_t>(out.size());
   } else {
-    mod.append_arg<int32_t>(out.size());
+    args.append<int32_t>(out.size());
   }
-  mod.append_ndim_arg(src.shape());
-  mod.append_ndim_arg(src.strides());
-  mod.append_arg<int32_t>(src.ndim());
-  mod.append_ndim_arg(slice_sizes_);
-  mod.append_arg(slice_size);
-  mod.append_arg(axes_);
-  append_indices_arg(mod, inputs, nidx, idx_ndim);
+  args.append_ndim(src.shape());
+  args.append_ndim(src.strides());
+  args.append<int32_t>(src.ndim());
+  args.append_ndim(slice_sizes_);
+  args.append(slice_size);
+  args.append(axes_);
+  append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
       "mlx::core::cu::gather<{}, {}, {}, {}, {}>",
@@ -122,9 +126,10 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
     encoder.set_input_array(in);
   }
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    mod.launch_kernel(stream, kernel_name, out, large);
-  });
+
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -187,26 +192,27 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
     return std::make_pair(jit_source_scatter, std::move(kernel_names));
   });
 
-  mod.append_arg(upd);
-  mod.append_arg(out);
+  cu::KernelArgs args;
+  args.append(upd);
+  args.append(out);
   if (large) {
-    mod.append_arg<int64_t>(upd.size());
+    args.append<int64_t>(upd.size());
   } else {
-    mod.append_arg<int32_t>(upd.size());
+    args.append<int32_t>(upd.size());
   }
-  mod.append_ndim_arg(upd.shape());
-  mod.append_ndim_arg(upd.strides());
-  mod.append_arg<int32_t>(upd.ndim());
+  args.append_ndim(upd.shape());
+  args.append_ndim(upd.strides());
+  args.append<int32_t>(upd.ndim());
   if (large) {
-    mod.append_arg<int64_t>(upd_post_idx_size);
+    args.append<int64_t>(upd_post_idx_size);
   } else {
-    mod.append_arg<int32_t>(upd_post_idx_size);
+    args.append<int32_t>(upd_post_idx_size);
   }
-  mod.append_ndim_arg(out.shape());
-  mod.append_ndim_arg(out.strides());
-  mod.append_arg<int32_t>(out.ndim());
-  mod.append_arg(axes_);
-  append_indices_arg(mod, inputs, nidx, idx_ndim);
+  args.append_ndim(out.shape());
+  args.append_ndim(out.strides());
+  args.append<int32_t>(out.ndim());
+  args.append(axes_);
+  append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
       "mlx::core::cu::scatter<{}, {}, mlx::core::cu::Scatter{}, {}, {}, {}>",
@@ -222,9 +228,9 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
     encoder.set_input_array(in);
   }
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    mod.launch_kernel(stream, kernel_name, upd, large);
-  });
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, upd, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -275,25 +281,26 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
   size_t idx_size_axis = idx.shape(axis_);
 
-  mod.append_arg(src);
-  mod.append_arg(idx);
-  mod.append_arg(out);
+  cu::KernelArgs args;
+  args.append(src);
+  args.append(idx);
+  args.append(out);
   if (large) {
-    mod.append_arg<int64_t>(idx_size_pre);
-    mod.append_arg<int64_t>(idx_size_axis);
-    mod.append_arg<int64_t>(idx_size_post);
+    args.append<int64_t>(idx_size_pre);
+    args.append<int64_t>(idx_size_axis);
+    args.append<int64_t>(idx_size_post);
   } else {
-    mod.append_arg<int32_t>(idx_size_pre);
-    mod.append_arg<int32_t>(idx_size_axis);
-    mod.append_arg<int32_t>(idx_size_post);
+    args.append<int32_t>(idx_size_pre);
+    args.append<int32_t>(idx_size_axis);
+    args.append<int32_t>(idx_size_post);
   }
-  mod.append_arg(remove_index(idx.shape(), axis_));
-  mod.append_arg(remove_index(src.strides(), axis_));
-  mod.append_arg(remove_index(idx.strides(), axis_));
-  mod.append_arg<int32_t>(axis_);
-  mod.append_arg(src.shape(axis_));
-  mod.append_arg(src.strides(axis_));
-  mod.append_arg(idx.strides(axis_));
+  args.append(remove_index(idx.shape(), axis_));
+  args.append(remove_index(src.strides(), axis_));
+  args.append(remove_index(idx.strides(), axis_));
+  args.append<int32_t>(axis_);
+  args.append(src.shape(axis_));
+  args.append(src.strides(axis_));
+  args.append(idx.strides(axis_));
 
   std::string kernel_name = fmt::format(
       "mlx::core::cu::gather_axis<{}, {}, {}, {}, {}, {}>",
@@ -309,9 +316,9 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
     encoder.set_input_array(in);
   }
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    mod.launch_kernel(stream, kernel_name, idx, large);
-  });
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -377,25 +384,26 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
   size_t idx_size_axis = idx.shape(axis_);
 
-  mod.append_arg(upd);
-  mod.append_arg(idx);
-  mod.append_arg(out);
+  cu::KernelArgs args;
+  args.append(upd);
+  args.append(idx);
+  args.append(out);
   if (large) {
-    mod.append_arg<int64_t>(idx_size_pre);
-    mod.append_arg<int64_t>(idx_size_axis);
-    mod.append_arg<int64_t>(idx_size_post);
+    args.append<int64_t>(idx_size_pre);
+    args.append<int64_t>(idx_size_axis);
+    args.append<int64_t>(idx_size_post);
   } else {
-    mod.append_arg<int32_t>(idx_size_pre);
-    mod.append_arg<int32_t>(idx_size_axis);
-    mod.append_arg<int32_t>(idx_size_post);
+    args.append<int32_t>(idx_size_pre);
+    args.append<int32_t>(idx_size_axis);
+    args.append<int32_t>(idx_size_post);
   }
-  mod.append_arg(remove_index(idx.shape(), axis_));
-  mod.append_arg(remove_index(upd.strides(), axis_));
-  mod.append_arg(remove_index(idx.strides(), axis_));
-  mod.append_arg<int32_t>(axis_);
-  mod.append_arg(out.shape(axis_));
-  mod.append_arg(upd.strides(axis_));
-  mod.append_arg(idx.strides(axis_));
+  args.append(remove_index(idx.shape(), axis_));
+  args.append(remove_index(upd.strides(), axis_));
+  args.append(remove_index(idx.strides(), axis_));
+  args.append<int32_t>(axis_);
+  args.append(out.shape(axis_));
+  args.append(upd.strides(axis_));
+  args.append(idx.strides(axis_));
 
   std::string kernel_name = fmt::format(
       "mlx::core::cu::scatter_axis<{}, {}, mlx::core::cu::Scatter{}, {}, {}, {}, {}>",
@@ -412,9 +420,9 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
     encoder.set_input_array(in);
   }
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    mod.launch_kernel(stream, kernel_name, idx, large);
-  });
+  auto kernel = mod.get_kernel(kernel_name);
+  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 } // namespace mlx::core

mlx/backend/cuda/jit_module.cpp

@@ -2,6 +2,7 @@
 
 #include "mlx/backend/cuda/jit_module.h"
 #include "mlx/backend/cuda/device.h"
+#include "mlx/version.h"
 
 #include "cuda_jit_sources.h"
 
@@ -12,6 +13,7 @@
 
 #include <fmt/format.h>
 #include <nvrtc.h>
+#include <unistd.h>
 
 namespace mlx::core::cu {
 
@@ -26,47 +28,74 @@ void check_nvrtc_error(const char* name, nvrtcResult err) {
   }
 }
 
-#define CHECK_CU_ERROR(cmd) check_cu_error(#cmd, (cmd))
-
-void check_cu_error(const char* name, CUresult err) {
-  if (err != CUDA_SUCCESS) {
-    const char* err_str = "Unknown error";
-    cuGetErrorString(err, &err_str);
-    throw std::runtime_error(fmt::format("{} failed: {}", name, err_str));
-  }
+// Return the location of the CUDA toolkit.
+const std::string& cuda_home() {
+  static std::string home = []() -> std::string {
+    const char* home = std::getenv("CUDA_HOME");
+    if (home) {
+      return home;
+    }
+    home = std::getenv("CUDA_PATH");
+    if (home) {
+      return home;
+    }
+#if defined(__linux__)
+    home = "/usr/local/cuda";
+    if (std::filesystem::exists(home)) {
+      return home;
+    }
+#endif
+    throw std::runtime_error(
+        "Environment variable CUDA_HOME or CUDA_PATH is not set.");
+  }();
+  return home;
 }
 
-// Return the location of the CUDA toolkit.
-const char* cuda_home() {
-  const char* home = std::getenv("CUDA_HOME");
-  if (home) {
-    return home;
-  }
-  home = std::getenv("CUDA_PATH");
-  if (home) {
-    return home;
-  }
-#if defined(__linux__)
-  home = "/usr/local/cuda";
-  if (std::filesystem::exists(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
-  throw std::runtime_error(
-      "Environment variable CUDA_HOME or CUDA_PATH is not set.");
+    // 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.
-bool get_ptx_cache_dir(std::filesystem::path* result) {
-  auto path = std::filesystem::temp_directory_path() / "mlx" / "ptx";
-  if (!std::filesystem::is_directory(path)) {
-    std::error_code error;
-    if (!std::filesystem::create_directories(path, error)) {
-      return false;
+const std::filesystem::path& ptx_cache_dir() {
+  static std::filesystem::path cache = []() -> std::filesystem::path {
+    std::filesystem::path cache;
+    if (auto c = std::getenv("MLX_PTX_CACHE_DIR"); c) {
+      cache = c;
+    } else {
+      cache =
+          std::filesystem::temp_directory_path() / "mlx" / version() / "ptx";
     }
-  }
-  *result = path;
-  return true;
+    if (!std::filesystem::exists(cache)) {
+      std::error_code error;
+      if (!std::filesystem::create_directories(cache, error)) {
+        return std::filesystem::path();
+      }
+    }
+    return cache;
+  }();
+  return cache;
 }
 
 // Try to read the cached |ptx| and |ptx_kernels| from |cache_dir|.
@@ -75,6 +104,10 @@ bool read_cached_ptx(
     const std::string& module_name,
     std::vector<char>* ptx,
     std::vector<std::pair<std::string, std::string>>* ptx_kernels) {
+  if (cache_dir.empty()) {
+    return false;
+  }
+
   auto ptx_path = cache_dir / (module_name + ".ptx");
   std::error_code error;
   auto ptx_size = std::filesystem::file_size(ptx_path, error);
@@ -104,7 +137,12 @@ void write_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
     const std::vector<char>& ptx,
-    const std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
+    const std::string& source_code) {
+  if (cache_dir.empty()) {
+    return;
+  }
+
   std::ofstream ptx_file(cache_dir / (module_name + ".ptx"), std::ios::binary);
   if (!ptx.empty()) {
     ptx_file.write(&ptx.front(), ptx.size());
@@ -113,6 +151,9 @@ void write_cached_ptx(
   for (const auto& [name, mangled] : ptx_kernels) {
     txt_file << name << "\t" << mangled << std::endl;
   }
+
+  std::ofstream source_file(cache_dir / (module_name + ".cu"));
+  source_file << source_code;
 }
 
 // Return if |device|'s version is not newer than |major|.|minor| version.
@@ -152,7 +193,7 @@ constexpr const char* g_include_names[] = {
     INCLUDE_PREFIX "binary_ops.cuh",
     INCLUDE_PREFIX "cast_op.cuh",
     INCLUDE_PREFIX "config.h",
-    INCLUDE_PREFIX "cucomplex_math.cuh",
+    INCLUDE_PREFIX "complex.cuh",
     INCLUDE_PREFIX "fp16_math.cuh",
     INCLUDE_PREFIX "indexing.cuh",
     INCLUDE_PREFIX "scatter_ops.cuh",
@@ -168,7 +209,7 @@ constexpr const char* g_headers[] = {
     jit_source_binary_ops,
     jit_source_cast_op,
     jit_source_config,
-    jit_source_cucomplex_math,
+    jit_source_complex,
     jit_source_fp16_math,
     jit_source_indexing,
     jit_source_scatter_ops,
@@ -184,11 +225,9 @@ JitModule::JitModule(
     const std::string& module_name,
     const KernelBuilder& builder) {
   // Check cache.
-  std::filesystem::path cache_dir;
   std::vector<char> ptx;
   std::vector<std::pair<std::string, std::string>> ptx_kernels;
-  if (!get_ptx_cache_dir(&cache_dir) ||
-      !read_cached_ptx(cache_dir, module_name, &ptx, &ptx_kernels)) {
+  if (!read_cached_ptx(ptx_cache_dir(), module_name, &ptx, &ptx_kernels)) {
     // Create program.
     auto [source_code, kernel_names] = builder();
     nvrtcProgram prog;
@@ -207,16 +246,24 @@ JitModule::JitModule(
     }
 
     // Compile program.
+    std::vector<const char*> args;
     bool use_sass = compiler_supports_device_sass(device);
     std::string compute = fmt::format(
         "--gpu-architecture={}_{}{}",
         use_sass ? "sm" : "compute",
         device.compute_capability_major(),
         device.compute_capability_minor());
-    std::string include = fmt::format("--include-path={}/include", cuda_home());
-    const char* args[] = {compute.c_str(), include.c_str()};
+    args.push_back(compute.c_str());
+    std::string cccl_include = cccl_dir();
+    if (!cccl_include.empty()) {
+      cccl_include = fmt::format("--include-path={}", cccl_include);
+      args.push_back(cccl_include.c_str());
+    }
+    std::string cuda_include =
+        fmt::format("--include-path={}/include", cuda_home());
+    args.push_back(cuda_include.c_str());
     nvrtcResult compile_result =
-        nvrtcCompileProgram(prog, std::size(args), args);
+        nvrtcCompileProgram(prog, args.size(), args.data());
     if (compile_result != NVRTC_SUCCESS) {
       size_t log_size;
       CHECK_NVRTC_ERROR(nvrtcGetProgramLogSize(prog, &log_size));
@@ -246,7 +293,8 @@ JitModule::JitModule(
     } else {
       CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
     }
-    write_cached_ptx(cache_dir, module_name, ptx, ptx_kernels);
+    write_cached_ptx(
+        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
   }
 
   // Load module.
@@ -264,60 +312,13 @@ JitModule::JitModule(
   // Load kernels.
   for (const auto& [name, mangled] : ptx_kernels) {
     CUfunction kernel;
-    CHECK_CU_ERROR(cuModuleGetFunction(&kernel, module_, mangled.c_str()));
+    CHECK_CUDA_ERROR(cuModuleGetFunction(&kernel, module_, mangled.c_str()));
     kernels_[name] = kernel;
   }
 }
 
 JitModule::~JitModule() {
-  CHECK_CU_ERROR(cuModuleUnload(module_));
-}
-
-void JitModule::launch_kernel(
-    CUstream stream,
-    const std::string& kernel_name,
-    const array& arr,
-    bool large,
-    int work_per_thread) {
-  CUfunction kernel = get_kernel(kernel_name);
-  size_t nthreads = cuda::ceil_div(arr.size(), work_per_thread);
-  int _, block_dim;
-  CHECK_CU_ERROR(
-      cuOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel, 0, 0, 0));
-  if (block_dim > nthreads) {
-    block_dim = nthreads;
-  }
-  Dims num_blocks{1, 1, 1};
-  if (large) {
-    num_blocks =
-        get_2d_grid_dims_common(arr.shape(), arr.strides(), work_per_thread);
-    std::get<0>(num_blocks) =
-        (std::get<0>(num_blocks) + block_dim - 1) / block_dim;
-  } else {
-    std::get<0>(num_blocks) = (nthreads + block_dim - 1) / block_dim;
-  }
-  launch_kernel(stream, kernel, num_blocks, Dims{block_dim, 1, 1});
-}
-
-void JitModule::launch_kernel(
-    CUstream stream,
-    CUfunction kernel,
-    Dims num_blocks,
-    Dims block_dims) {
-  CHECK_CU_ERROR(cuLaunchKernel(
-      kernel,
-      std::get<0>(num_blocks),
-      std::get<1>(num_blocks),
-      std::get<2>(num_blocks),
-      std::get<0>(block_dims),
-      std::get<1>(block_dims),
-      std::get<2>(block_dims),
-      0,
-      stream,
-      args_.data(),
-      nullptr));
-  args_.clear();
-  storage_.clear();
+  CHECK_CUDA_ERROR(cuModuleUnload(module_));
 }
 
 CUfunction JitModule::get_kernel(const std::string& kernel_name) {
@@ -329,10 +330,6 @@ CUfunction JitModule::get_kernel(const std::string& kernel_name) {
   return it->second;
 }
 
-void JitModule::append_ptr_arg(const void* v) {
-  args_.push_back(const_cast<void*>(v));
-}
-
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,

mlx/backend/cuda/jit_module.h

@@ -4,6 +4,7 @@
 
 #include "mlx/array.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/device/config.h"
 
 #include <deque>
@@ -23,72 +24,48 @@ using KernelBuilderResult = std::pair<
     /* kernel names */ std::vector<std::string>>;
 using KernelBuilder = std::function<KernelBuilderResult()>;
 
-class JitModule {
- public:
-  JitModule(
-      Device& device,
-      const std::string& module_name,
-      const KernelBuilder& builder);
-  ~JitModule();
+struct KernelArgs {
+  void** args() {
+    return args_.data();
+  }
 
-  JitModule(const JitModule&) = delete;
-  JitModule& operator=(const JitModule&) = delete;
-
-  void append_arg(const array& a) {
-    append_arg(reinterpret_cast<CUdeviceptr>(a.data<void>()));
+  void append(const array& a) {
+    append(reinterpret_cast<CUdeviceptr>(a.data<void>()));
   }
 
   template <typename T>
-  void append_arg(T val) {
+  void append(T val) {
     storage_.emplace_back(val);
-    append_ptr_arg(&storage_.back());
+    append_ptr(&storage_.back());
   }
 
   template <typename T>
-  void append_arg(std::vector<T> vec) {
+  void append(std::vector<T> vec) {
     if (vec.empty()) {
       // The nullptr can not be used as arg, pass something not null.
-      append_arg(std::monostate{});
+      append(std::monostate{});
     } else {
-      append_ptr_arg(vec.data());
+      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_arg(const std::vector<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));
     }
-    std::vector<T> copied(NDIM);
-    std::copy(vec.begin(), vec.end(), copied.data());
-    append_arg(std::move(copied));
+    vec.resize(NDIM);
+    append(std::move(vec));
   }
 
-  // Launch kernel with |kernel_name| that each thread works on
-  // |work_per_thread| elements of |arr|.
-  void launch_kernel(
-      CUstream stream,
-      const std::string& kernel_name,
-      const array& arr,
-      bool large,
-      int work_per_thread = 1);
-
-  void launch_kernel(
-      CUstream stream,
-      CUfunction kernel,
-      Dims num_blocks,
-      Dims block_dims);
-
-  CUfunction get_kernel(const std::string& kernel_name);
+  void append_ptr(const void* v) {
+    args_.push_back(const_cast<void*>(v));
+  }
 
  private:
-  void append_ptr_arg(const void* v);
-
-  CUmodule module_{nullptr};
-  std::unordered_map<std::string, CUfunction> kernels_;
   std::vector<void*> args_;
 
   // The cuLaunchKernel API requires passing pointers to arguments so store
@@ -105,6 +82,23 @@ class JitModule {
   std::deque<Arg> storage_;
 };
 
+class JitModule {
+ public:
+  JitModule(
+      Device& device,
+      const std::string& module_name,
+      const KernelBuilder& builder);
+  ~JitModule();
+
+  JitModule(const JitModule&) = delete;
+  JitModule& operator=(const JitModule&) = delete;
+  CUfunction get_kernel(const std::string& kernel_name);
+
+ private:
+  CUmodule module_{nullptr};
+  std::unordered_map<std::string, CUfunction> kernels_;
+};
+
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,

mlx/backend/cuda/kernel_utils.cuh

@@ -6,10 +6,12 @@
 
 #pragma once
 
+#include <type_traits>
+
 #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>
 #include <fmt/format.h>
@@ -17,60 +19,46 @@
 
 namespace mlx::core {
 
-// Convert a number between 1~3 to constexpr.
-#define MLX_SWITCH_1_2_3(N, NDIM, ...) \
-  switch (N) {                         \
-    case 1: {                          \
-      constexpr int NDIM = 1;          \
-      __VA_ARGS__;                     \
-      break;                           \
-    }                                  \
-    case 2: {                          \
-      constexpr int NDIM = 2;          \
-      __VA_ARGS__;                     \
-      break;                           \
-    }                                  \
-    case 3: {                          \
-      constexpr int NDIM = 3;          \
-      __VA_ARGS__;                     \
-      break;                           \
-    }                                  \
+template <typename F>
+void dispatch_1_2_3(int n, F&& f) {
+  switch (n) {
+    case 1:
+      f(std::integral_constant<int, 1>{});
+      break;
+    case 2:
+      f(std::integral_constant<int, 2>{});
+      break;
+    case 3:
+      f(std::integral_constant<int, 3>{});
+      break;
   }
+}
 
-// Like MLX_SWITCH_ALL_TYPES but for booleans.
-#define MLX_SWITCH_BOOL(BOOL, BOOL_ALIAS, ...) \
-  if (BOOL) {                                  \
-    constexpr bool BOOL_ALIAS = true;          \
-    __VA_ARGS__;                               \
-  } else {                                     \
-    constexpr bool BOOL_ALIAS = false;         \
-    __VA_ARGS__;                               \
+template <typename F>
+void dispatch_bool(bool v, F&& f) {
+  if (v) {
+    f(std::true_type{});
+  } else {
+    f(std::false_type{});
   }
+}
 
-// Convert a block_dim to constexpr between WARP_SIZE and WARP_SIZE ^ 2.
-#define MLX_SWITCH_BLOCK_DIM(NUM_THREADS, BLOCK_DIM, ...)   \
-  {                                                         \
-    uint32_t _num_threads = NUM_THREADS;                    \
-    if (_num_threads <= WARP_SIZE) {                        \
-      constexpr uint32_t BLOCK_DIM = WARP_SIZE;             \
-      __VA_ARGS__;                                          \
-    } else if (_num_threads <= WARP_SIZE * 2) {             \
-      constexpr uint32_t BLOCK_DIM = WARP_SIZE * 2;         \
-      __VA_ARGS__;                                          \
-    } else if (_num_threads <= WARP_SIZE * 4) {             \
-      constexpr uint32_t BLOCK_DIM = WARP_SIZE * 4;         \
-      __VA_ARGS__;                                          \
-    } else if (_num_threads <= WARP_SIZE * 8) {             \
-      constexpr uint32_t BLOCK_DIM = WARP_SIZE * 8;         \
-      __VA_ARGS__;                                          \
-    } else if (_num_threads <= WARP_SIZE * 16) {            \
-      constexpr uint32_t BLOCK_DIM = WARP_SIZE * 16;        \
-      __VA_ARGS__;                                          \
-    } else {                                                \
-      constexpr uint32_t BLOCK_DIM = WARP_SIZE * WARP_SIZE; \
-      __VA_ARGS__;                                          \
-    }                                                       \
+template <typename F>
+void dispatch_block_dim(int threads, F&& f) {
+  if (threads <= WARP_SIZE) {
+    f(std::integral_constant<int, WARP_SIZE>{});
+  } else if (threads <= WARP_SIZE * 2) {
+    f(std::integral_constant<int, WARP_SIZE * 2>{});
+  } else if (threads <= WARP_SIZE * 4) {
+    f(std::integral_constant<int, WARP_SIZE * 4>{});
+  } else if (threads <= WARP_SIZE * 8) {
+    f(std::integral_constant<int, WARP_SIZE * 8>{});
+  } else if (threads <= WARP_SIZE * 16) {
+    f(std::integral_constant<int, WARP_SIZE * 16>{});
+  } else {
+    f(std::integral_constant<int, WARP_SIZE * 32>{});
   }
+}
 
 // Maps CPU types to CUDA types.
 template <typename T>
@@ -90,7 +78,7 @@ struct CTypeToCudaType<bfloat16_t> {
 
 template <>
 struct CTypeToCudaType<complex64_t> {
-  using type = cuComplex;
+  using type = cu::complex64_t;
 };
 
 template <typename T>
@@ -102,10 +90,14 @@ inline constexpr bool is_floating_v =
     cuda::std::is_same_v<T, float> || cuda::std::is_same_v<T, double> ||
     cuda::std::is_same_v<T, float16_t> || cuda::std::is_same_v<T, bfloat16_t>;
 
+// Type traits for detecting complex numbers.
+template <typename T>
+inline constexpr bool is_complex_v = cuda::std::is_same_v<T, complex64_t> ||
+    cuda::std::is_same_v<T, complex128_t>;
+
 // Type traits for detecting complex or real floating point numbers.
 template <typename T>
-inline constexpr bool is_inexact_v =
-    is_floating_v<T> || cuda::std::is_same_v<T, complex64_t>;
+inline constexpr bool is_inexact_v = is_floating_v<T> || is_complex_v<T>;
 
 // Utility to copy data from vector to array in host.
 template <int NDIM = MAX_NDIM, typename T = int32_t>
@@ -132,7 +124,13 @@ std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2);
 template <typename T>
 inline uint max_occupancy_block_dim(T kernel) {
   int _, block_dim;
-  CHECK_CUDA_ERROR(cudaOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel));
+  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;
 }
 

mlx/backend/cuda/layer_norm.cu

@@ -237,8 +237,7 @@ void LayerNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -258,22 +257,24 @@ void LayerNorm::eval_gpu(
   encoder.set_input_array(w);
   encoder.set_input_array(b);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "layernorm", CTYPE, {
-      using DataType = cuda_type_t<CTYPE>;
-      constexpr uint32_t N_READS = 4;
-      MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-        auto kernel = cu::layer_norm<DataType, BLOCK_DIM, N_READS>;
-        kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
-            x.data<DataType>(),
-            w.data<DataType>(),
-            b.data<DataType>(),
-            out.data<DataType>(),
-            eps_,
-            axis_size,
-            w_stride,
-            b_stride);
-      });
+  dispatch_float_types(out.dtype(), "layernorm", [&](auto type_tag) {
+    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>(),
+          out.data<DataType>(),
+          eps_,
+          axis_size,
+          w_stride,
+          b_stride);
     });
   });
 }
@@ -288,21 +289,23 @@ void LayerNormVJP::eval_gpu(
   // Ensure row contiguity. We could relax this step by checking that the array
   // is contiguous (no broadcasts or holes) and that the input strides are the
   // same as the cotangent strides but for now this is simpler.
-  auto check_input = [&s](const array& x) -> std::pair<array, bool> {
+  auto check_input = [&s](const array& x, bool& copied) {
     if (x.flags().row_contiguous) {
-      return {x, false};
+      copied = false;
+      return x;
     }
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
-    copy_gpu(x, x_copy, CopyType::General, s);
-    return {x_copy, true};
+    copied = true;
+    return contiguous_copy_gpu(x, s);
   };
   bool donate_x = inputs[0].is_donatable();
   bool donate_g = inputs[3].is_donatable();
-  auto [x, copied] = check_input(inputs[0]);
+  bool copied;
+  auto x = check_input(inputs[0], copied);
   donate_x |= copied;
   const array& w = inputs[1];
   const array& b = inputs[2];
-  auto [g, g_copied] = check_input(inputs[3]);
+  bool g_copied;
+  auto g = check_input(inputs[3], g_copied);
   donate_g |= g_copied;
   array& gx = outputs[0];
   array& gw = outputs[1];
@@ -333,47 +336,59 @@ void LayerNormVJP::eval_gpu(
   // gradient accumulators.
   array gw_temp =
       (has_w) ? array({n_rows, x.shape().back()}, gw.dtype(), nullptr, {}) : w;
+  bool g_in_gw = false;
   if (has_w) {
     if (!g_in_gx && donate_g) {
+      g_in_gw = true;
       gw_temp.copy_shared_buffer(g);
     } else {
       gw_temp.set_data(allocator::malloc(gw_temp.nbytes()));
       encoder.add_temporary(gw_temp);
     }
   }
-  gw.set_data(allocator::malloc(gw.nbytes()));
-  gb.set_data(allocator::malloc(gb.nbytes()));
 
-  // Finish with the gradient for b in case we had a b.
-  if (gb.ndim() == 1 && gb.size() == axis_size) {
+  // The gradient for b in case we had a b.
+  bool has_gb = (gb.ndim() == 1 && gb.size() == axis_size);
+  if (has_gb) {
     ReductionPlan plan(
         ReductionOpType::ContiguousStridedReduce, {n_rows}, {axis_size});
     col_reduce(encoder, g, gb, Reduce::ReduceType::Sum, {0}, plan);
   }
 
+  // Insert dependency if `g` was donated
+  if ((g_in_gx || g_in_gw) && has_gb) {
+    encoder.set_input_array(gb);
+  }
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(g);
   encoder.set_output_array(gx);
   encoder.set_output_array(gw_temp);
-  encoder.launch_kernel([&, x = x, g = g](cudaStream_t stream) {
-    MLX_SWITCH_FLOAT_TYPES_CHECKED(gx.dtype(), "layernorm_vjp", CTYPE, {
-      using DataType = cuda_type_t<CTYPE>;
+  dispatch_float_types(gx.dtype(), "layernorm_vjp", [&](auto type_tag) {
+    dispatch_bool(has_w, [&](auto has_w_constant) {
       constexpr int N_READS = 4;
-      MLX_SWITCH_BOOL(has_w, HAS_W, {
-        MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-          auto kernel = cu::layer_norm_vjp<DataType, HAS_W, BLOCK_DIM, N_READS>;
-          kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
-              x.data<DataType>(),
-              w.data<DataType>(),
-              g.data<DataType>(),
-              gx.data<DataType>(),
-              gw_temp.data<DataType>(),
-              eps_,
-              axis_size,
-              w_stride);
-        });
-      });
+      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,
+                block_dim(),
+                N_READS>;
+            encoder.add_kernel_node(
+                kernel,
+                n_rows,
+                block_dim(),
+                0,
+                x.data<DataType>(),
+                w.data<DataType>(),
+                g.data<DataType>(),
+                gx.data<DataType>(),
+                gw_temp.data<DataType>(),
+                eps_,
+                axis_size,
+                w_stride);
+          });
     });
   });
 

mlx/backend/cuda/logsumexp.cu

@@ -108,8 +108,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
     if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       encoder.add_temporary(x_copy);
       return x_copy;
     }
@@ -143,15 +142,19 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "logsumexp", CTYPE, {
-      using DataType = cuda_type_t<CTYPE>;
-      constexpr int N_READS = 4;
-      MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-        auto kernel = cu::logsumexp<DataType, float, BLOCK_DIM, N_READS>;
-        kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
-            in.data<DataType>(), out.data<DataType>(), axis_size);
-      });
+  dispatch_float_types(out.dtype(), "logsumexp", [&](auto type_tag) {
+    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);
     });
   });
 }

mlx/backend/cuda/matmul.cpp

@@ -27,6 +27,35 @@ void check_cublas_error(const char* name, cublasStatus_t err) {
   }
 }
 
+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_;
+}
+
 class MatMul {
  public:
   MatMul(
@@ -42,7 +71,8 @@ class MatMul {
       int64_t ldb,
       int32_t batch_count,
       int64_t a_batch_stride,
-      int64_t b_batch_stride) {
+      int64_t b_batch_stride)
+      : handle_(device.lt_handle()), pref_(cublas_preference(device)) {
     heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
 
     auto scale_type = dtype_to_cuda_type(dtype);
@@ -76,20 +106,6 @@ class MatMul {
         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(
@@ -103,7 +119,6 @@ class MatMul {
       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,
@@ -125,15 +140,15 @@ class MatMul {
             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);
+        type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
   }
 
   ~MatMul() {
-    cublasLtMatrixLayoutDestroy(a_desc_);
-    cublasLtMatrixLayoutDestroy(b_desc_);
-    cublasLtMatrixLayoutDestroy(c_desc_);
-    cublasLtMatrixLayoutDestroy(out_desc_);
-    cublasLtMatmulDescDestroy(matmul_desc_);
+    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 run(
@@ -147,7 +162,7 @@ class MatMul {
     if (heuristic_.state != CUBLAS_STATUS_SUCCESS) {
       int ret = 0;
       CHECK_CUBLAS_ERROR(cublasLtMatmulAlgoGetHeuristic(
-          encoder.device().lt_handle(),
+          handle_,
           matmul_desc_,
           a_desc_,
           b_desc_,
@@ -162,31 +177,34 @@ class MatMul {
       }
     }
 
-    array workspace(
-        allocator::malloc(heuristic_.workspaceSize),
-        {static_cast<int>(heuristic_.workspaceSize)},
-        int8);
-    encoder.add_temporary(workspace);
+    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>();
+    }
 
-    encoder.launch_kernel([&](cudaStream_t stream) {
-      CHECK_CUBLAS_ERROR(cublasLtMatmul(
-          encoder.device().lt_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.data<void>(),
-          workspace.nbytes(),
-          stream));
-    });
+    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:
@@ -255,8 +273,9 @@ class MatMul {
     return desc;
   }
 
-  cublasLtMatmulDesc_t matmul_desc_{nullptr};
   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};
@@ -269,7 +288,7 @@ class MatMul {
 namespace {
 
 std::tuple<bool, int64_t, array>
-check_transpose(std::vector<array>& copies, const Stream& s, const array& arr) {
+check_transpose(cu::CommandEncoder& enc, const Stream& s, const array& arr) {
   auto stx = arr.strides()[arr.ndim() - 2];
   auto sty = arr.strides()[arr.ndim() - 1];
   if (sty == 1 && stx == arr.shape(-1)) {
@@ -277,9 +296,8 @@ check_transpose(std::vector<array>& copies, const Stream& s, const array& arr) {
   } else if (stx == 1 && sty == arr.shape(-2)) {
     return std::make_tuple(true, sty, arr);
   } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy_gpu(arr, arr_copy, CopyType::General, s);
-    copies.push_back(arr_copy);
+    array arr_copy = contiguous_copy_gpu(arr, s);
+    enc.add_temporary(arr_copy);
     return std::make_tuple(false, arr.shape(-1), arr_copy);
   }
 }
@@ -313,13 +331,8 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   // Keep a vector with copies to be cleared in the completed buffer to release
   // the arrays
-  std::vector<array> copies;
-  auto [a_transposed, lda, a] = check_transpose(copies, s, a_pre);
-  auto [b_transposed, ldb, b] = check_transpose(copies, s, b_pre);
-
-  for (auto& temp : copies) {
-    encoder.add_temporary(temp);
-  }
+  auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
+  auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
 
   /////////////////////////////////////////////////////////////////////////////
   // Check and collapse batch dimensions
@@ -344,7 +357,7 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   // Invoke cublasLt
 
   cu::MatMul matmul(
-      encoder.device(),
+      cu::device(s.device),
       a.dtype(),
       a_transposed,
       M,
@@ -358,9 +371,19 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
       a_batch_strides.back(),
       b_batch_strides.back());
 
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  auto nbatch = batch_count / batch_shape.back();
+  if (nbatch == 1) {
+    matmul.run(encoder, out.data<int8_t>(), a.data<int8_t>(), b.data<int8_t>());
+    return;
+  }
+
   ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
   ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
-  for (size_t i = 0; i < batch_count / batch_shape.back(); ++i) {
+  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,
@@ -379,9 +402,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 3);
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
-  auto& c_pre = inputs[2];
-
-  out.set_data(allocator::malloc(out.nbytes()));
+  auto c = inputs[2];
 
   /////////////////////////////////////////////////////////////////////////////
   // Init checks and prep
@@ -392,13 +413,25 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   // Keep a vector with copies to be cleared in the completed buffer to release
   // the arrays
-  std::vector<array> copies;
-  auto [a_transposed, lda, a] = check_transpose(copies, s, a_pre);
-  auto [b_transposed, ldb, b] = check_transpose(copies, s, b_pre);
-  auto [c_transposed, ldc, c] = check_transpose(copies, s, c_pre);
+  auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
+  auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
 
-  for (auto& temp : copies) {
-    encoder.add_temporary(temp);
+  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;
+    }
   }
 
   /////////////////////////////////////////////////////////////////////////////
@@ -427,7 +460,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   // Invoke cublasLt
 
   cu::MatMul matmul(
-      encoder.device(),
+      cu::device(s.device),
       a.dtype(),
       a_transposed,
       M,
@@ -437,17 +470,35 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       K,
       N,
       ldb,
-      c_transposed,
       ldc,
       batch_shape.back(),
       a_batch_strides.back(),
       b_batch_strides.back(),
       c_batch_strides.back());
 
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  auto nbatch = batch_count / batch_shape.back();
+  if (nbatch == 1) {
+    matmul.run(
+        encoder,
+        out.data<int8_t>(),
+        a.data<int8_t>(),
+        b.data<int8_t>(),
+        c.data<int8_t>(),
+        alpha_,
+        beta_);
+    return;
+  }
+
   ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
   ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
   ContiguousIterator c_it(batch_shape, c_batch_strides, batch_shape.size() - 1);
-  for (size_t i = 0; i < batch_count / batch_shape.back(); ++i) {
+  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,

mlx/backend/cuda/matmul/mma.cuh

@@ -0,0 +1,108 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/matmul/tiles.cuh"
+
+namespace mlx::core::cu {
+
+template <typename U, typename T>
+__device__ inline void
+mma_t(Tile16x16<U>& C, Tile16x16<T>& A, Tile16x16<T>& B) {}
+
+/**
+ * Multiply the 16x16 bfloat16 tiles and accumulate the result in one 16x16
+ * float tile.
+ *
+ * We actually perform C += A @ B.T
+ */
+__device__ inline void mma_t(
+    Tile16x16<float>& C,
+    Tile16x16<__nv_bfloat16>& A,
+    Tile16x16<__nv_bfloat16>& B) {
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0, %1, %2, %3}, "
+      "{%4, %5, %6, %7}, "
+      "{%8, %9}, "
+      "{%10, %11, %12, %13};"
+
+      // D matrix
+      : "+f"(C.values[0].x),
+        "+f"(C.values[0].y),
+        "+f"(C.values[1].x),
+        "+f"(C.values[1].y)
+
+      // A matrix
+      : "r"(*(uint32_t*)(&A.values[0])),
+        "r"(*(uint32_t*)(&A.values[1])),
+        "r"(*(uint32_t*)(&A.values[2])),
+        "r"(*(uint32_t*)(&A.values[3])),
+
+        // B matrix
+        "r"(*(uint32_t*)(&B.values[0])),
+        "r"(*(uint32_t*)(&B.values[2])),
+
+        // C matrix
+        "f"(C.values[0].x),
+        "f"(C.values[0].y),
+        "f"(C.values[1].x),
+        "f"(C.values[1].y));
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0, %1, %2, %3}, "
+      "{%4, %5, %6, %7}, "
+      "{%8, %9}, "
+      "{%10, %11, %12, %13};"
+
+      // D matrix
+      : "+f"(C.values[2].x),
+        "+f"(C.values[2].y),
+        "+f"(C.values[3].x),
+        "+f"(C.values[3].y)
+
+      // A matrix
+      : "r"(*(uint32_t*)(&A.values[0])),
+        "r"(*(uint32_t*)(&A.values[1])),
+        "r"(*(uint32_t*)(&A.values[2])),
+        "r"(*(uint32_t*)(&A.values[3])),
+
+        // B matrix
+        "r"(*(uint32_t*)(&B.values[1])),
+        "r"(*(uint32_t*)(&B.values[3])),
+
+        // C matrix
+        "f"(C.values[2].x),
+        "f"(C.values[2].y),
+        "f"(C.values[3].x),
+        "f"(C.values[3].y));
+}
+
+/**
+ * Multiply larger register tiles by delegating to mma_t.
+ */
+template <typename U, typename T, int M, int N, int K>
+__device__ inline void mma_t(
+    RegisterTile<U, M, N>& C,
+    RegisterTile<T, M, K>& A,
+    RegisterTile<T, N, K>& B) {
+  constexpr int TILES_M = RegisterTile<T, M, K>::TILES_Y;
+  constexpr int TILES_K = RegisterTile<T, M, K>::TILES_X;
+  constexpr int TILES_N = RegisterTile<T, N, K>::TILES_Y;
+
+  MLX_UNROLL
+  for (int k = 0; k < TILES_K; k++) {
+    MLX_UNROLL
+    for (int m = 0; m < TILES_M; m++) {
+      MLX_UNROLL
+      for (int n = 0; n < TILES_N; n++) {
+        mma_t(
+            C.data[m * TILES_N + n],
+            A.data[m * TILES_K + k],
+            B.data[n * TILES_K + k]);
+      }
+    }
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/matmul/tiles.cuh

@@ -0,0 +1,419 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#define MLX_UNROLL _Pragma("unroll")
+
+namespace mlx::core::cu {
+
+// Map types to their vector of 2 type float -> float2, double -> double2 etc
+template <typename T>
+struct Vector2;
+template <>
+struct Vector2<double> {
+  using type = double2;
+};
+template <>
+struct Vector2<float> {
+  using type = float2;
+};
+template <>
+struct Vector2<__half> {
+  using type = __half2;
+};
+template <>
+struct Vector2<__nv_bfloat16> {
+  using type = __nv_bfloat162;
+};
+template <typename T>
+using Vector2_t = typename Vector2<T>::type;
+
+/**
+ * The basic building block for Ampere mmas. A 16x16 tile distributed across
+ * the warp.
+ *
+ * Each thread holds 8 values. They are distributed according to
+ * https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-fragment-mma-16816-float
+ *
+ * For use instructions see the individual methods eg load().
+ */
+template <typename T>
+struct Tile16x16 {
+  using T2 = Vector2_t<T>;
+
+  T2 values[4];
+
+  __device__ inline void fill(T v) {
+    T2 v2 = {v, v};
+    for (int i = 0; i < 4; i++) {
+      values[i] = v2;
+    }
+  }
+
+  /**
+   * Load a 16x16 tile from shared memory.
+   *
+   * The instruction is a bit weird in the sense that the address provided by
+   * each thread and the elements loaded are not the same.
+   *
+   * We load 4 8x8 tiles. The tile rows are stored contiguously in memory. As a
+   * result the warp provides 4*8 = 32 addresses one per row.
+   *
+   * Threads 0-7 provide the addresses for the first tile, 8-15 for the second
+   * and so on. For instance to load a non swizzled tile we would do
+   *
+   *    base_addr + (laneid % 16) * BK + (laneid / 2) * 8
+   *
+   * See
+   * https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-instructions-ldmatrix
+   */
+  __device__ inline void load(uint32_t row_address) {
+    if constexpr (
+        std::is_same_v<T2, __nv_bfloat162> || std::is_same_v<T2, __half2>) {
+      asm volatile(
+          "ldmatrix.sync.aligned.m8n8.x4.shared::cta.b16 {%0, %1, %2, %3}, [%4];\n"
+          : "=r"(*(uint32_t*)&(values[0])),
+            "=r"(*(uint32_t*)&(values[1])),
+            "=r"(*(uint32_t*)&(values[2])),
+            "=r"(*(uint32_t*)&(values[3]))
+          : "r"(row_address));
+    }
+  }
+
+  /**
+   * Store the tile to the address pointed to by `x`.
+   *
+   * The provided pointer is a generic pointer but this is meant to be used to
+   * store to global memory. For storing to shared memory we should use
+   * `stmatrix`.
+   *
+   * This also showcases the format of the tile quite nicely. Each register is
+   * holding to adjacent values. The indices are
+   *
+   *    row + 0, col + 0
+   *    row + 8, col + 0
+   *    row + 0, col + 8
+   *    row + 8, col + 8
+   *
+   * Given that we are dealing with Vector2_t<U> the column offsets are 4
+   * instead of 8.
+   */
+  template <typename U>
+  __device__ inline void store_global(U* x, int N) {
+    using U2 = Vector2_t<U>;
+    U2* x2 = reinterpret_cast<U2*>(x);
+    const int laneid = threadIdx.x % 32;
+    const int row = laneid / 4;
+    const int col = laneid % 4;
+    if constexpr (std::is_same_v<U2, T2>) {
+      x2[(row + 0) * (N / 2) + col + 0] = values[0];
+      x2[(row + 0) * (N / 2) + col + 4] = values[2];
+      x2[(row + 8) * (N / 2) + col + 0] = values[1];
+      x2[(row + 8) * (N / 2) + col + 4] = values[3];
+    } else if constexpr (
+        std::is_same_v<T2, float2> && std::is_same_v<U, __nv_bfloat16>) {
+      x2[(row + 0) * (N / 2) + col + 0] =
+          __floats2bfloat162_rn(values[0].x, values[0].y);
+      x2[(row + 0) * (N / 2) + col + 4] =
+          __floats2bfloat162_rn(values[2].x, values[2].y);
+      x2[(row + 8) * (N / 2) + col + 0] =
+          __floats2bfloat162_rn(values[1].x, values[1].y);
+      x2[(row + 8) * (N / 2) + col + 4] =
+          __floats2bfloat162_rn(values[3].x, values[3].y);
+    }
+  }
+
+  template <typename U>
+  __device__ inline void store_global_safe(U* x, int N, int max_rows) {
+    const int laneid = threadIdx.x % 32;
+    const int row = laneid / 4;
+    const int col = laneid % 4;
+    if (row < max_rows) {
+      x[(row + 0) * N + 2 * col + 0] = static_cast<U>(values[0].x);
+      x[(row + 0) * N + 2 * col + 1] = static_cast<U>(values[0].y);
+      x[(row + 0) * N + 2 * col + 8] = static_cast<U>(values[2].x);
+      x[(row + 0) * N + 2 * col + 9] = static_cast<U>(values[2].y);
+    }
+    if (row + 8 < max_rows) {
+      x[(row + 8) * N + 2 * col + 0] = static_cast<U>(values[1].x);
+      x[(row + 8) * N + 2 * col + 1] = static_cast<U>(values[1].y);
+      x[(row + 8) * N + 2 * col + 8] = static_cast<U>(values[3].x);
+      x[(row + 8) * N + 2 * col + 9] = static_cast<U>(values[3].y);
+    }
+  }
+};
+
+/**
+ * A simple container of multiple Tile16x16.
+ *
+ * Provides utility functions for loading and manipulating collections of basic
+ * tiles.
+ */
+template <typename T, int ROWS_, int COLS_>
+struct RegisterTile {
+  static constexpr int ROWS = ROWS_;
+  static constexpr int COLS = COLS_;
+  static constexpr int TILES_X = COLS / 16;
+  static constexpr int TILES_Y = ROWS / 16;
+
+  Tile16x16<T> data[TILES_X * TILES_Y];
+
+  __device__ inline void fill(T v) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].fill(v);
+      }
+    }
+  }
+
+  template <typename Tile>
+  __device__ inline void
+  load(Tile& tile, uint32_t base_address, int row, int col) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].load(
+            tile.loc(base_address, row + i * 16, col + j * 16));
+      }
+    }
+  }
+
+  template <typename U>
+  __device__ inline void store_global(U* x, int N, int row, int col) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].store_global(
+            x + (row + i * 16) * N + col + j * 16, N);
+      }
+    }
+  }
+
+  template <typename U>
+  __device__ inline void
+  store_global_safe(U* x, int N, int row, int col, int max_rows) {
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        data[i * TILES_X + j].store_global_safe(
+            x + (row + i * 16) * N + col + j * 16, N, max_rows - row - i * 16);
+      }
+    }
+  }
+};
+
+template <typename T, int ROWS_, int COLS_>
+struct SharedTile {
+  static constexpr int ROWS = ROWS_;
+  static constexpr int COLS = COLS_;
+  static constexpr int TILES_X = COLS / 16;
+  static constexpr int TILES_Y = ROWS / 16;
+  static constexpr int NUMEL = ROWS * COLS;
+
+  // Swizzle taken from ThunderKittens.
+  //
+  // See inludes/types/shared/st.cuh
+  //
+  // I do feel that it is too math heavy and can be improved. Also the math is
+  // done every time although the addresses don't change from load to load. I
+  // guess we are expecting the compiler to figure that out.
+  static constexpr int swizzle_bytes =
+      (sizeof(T) == 2 ? (TILES_X % 4 == 0 ? 128 : (TILES_X % 2 == 0 ? 64 : 32))
+                      : (sizeof(T) == 4 ? (TILES_X % 2 == 0 ? 128 : 64) : 0));
+
+  T data[ROWS * COLS];
+
+  // Return a pointer to the element at (row, col) using the swizzle.
+  __device__ static inline T* ptr(T* ptr, int row, int col) {
+    if constexpr (swizzle_bytes > 0) {
+      static constexpr int swizzle_repeat = swizzle_bytes * 8;
+      static constexpr int subtile_cols = swizzle_bytes / sizeof(T);
+      const int outer_idx = col / subtile_cols;
+      const uint64_t addr =
+          (uint64_t)(&ptr
+                         [outer_idx * ROWS * subtile_cols + row * subtile_cols +
+                          col % subtile_cols]);
+      const int swizzle = ((addr % swizzle_repeat) >> 7) << 4;
+      return (T*)(addr ^ swizzle);
+    } else {
+      return ptr + row * COLS + col;
+    }
+  }
+
+  // Return the location of the element at (row, col) using the swizzle.
+  __device__ static inline uint32_t loc(uint32_t ptr, int row, int col) {
+    if constexpr (swizzle_bytes > 0) {
+      static constexpr int swizzle_repeat = swizzle_bytes * 8;
+      static constexpr int subtile_cols = swizzle_bytes / sizeof(T);
+      const int outer_idx = col / subtile_cols;
+      const uint32_t addr = ptr +
+          sizeof(T) *
+              (outer_idx * ROWS * subtile_cols + row * subtile_cols +
+               col % subtile_cols);
+      const int swizzle = ((addr % swizzle_repeat) >> 7) << 4;
+      return (addr ^ swizzle);
+    } else {
+      return ptr + sizeof(T) * (row * COLS + col);
+    }
+  }
+
+  // Convenience functions to edit elements going through the swizzle.
+  __device__ inline T& operator()(int row, int col) {
+    return *ptr(data, row, col);
+  }
+  __device__ inline void store(float4& v, int row, int col) {
+    *(reinterpret_cast<float4*>(ptr(data, row, col))) = v;
+  }
+  __device__ inline void store(float2& v, int row, int col) {
+    *(reinterpret_cast<float2*>(ptr(data, row, col))) = v;
+  }
+  __device__ inline void store(float& v, int row, int col) {
+    *(reinterpret_cast<float*>(ptr(data, row, col))) = v;
+  }
+  template <int N>
+  __device__ inline void store(T (&v)[N], int row, int col) {
+    if constexpr (sizeof(T) * N == 4) {
+      store(*(reinterpret_cast<float*>(&v[0])), row, col);
+    } else if constexpr (sizeof(T) * N == 8) {
+      store(*(reinterpret_cast<float2*>(&v[0])), row, col);
+    } else if constexpr (sizeof(T) * N == 16) {
+      store(*(reinterpret_cast<float4*>(&v[0])), row, col);
+    } else {
+      MLX_UNROLL
+      for (int i = 0; i < N; i++) {
+        *ptr(data, row, col + i) = v[i];
+      }
+    }
+  }
+};
+
+/**
+ * Load the tile from global memory by loading 16 bytes at a time and storing
+ * them immediately.
+ */
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void load(Tile& tile, const T* x, int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  x += row * N + col * ELEMENTS_PER_LOAD;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    float4 tmp;
+    tmp = *(reinterpret_cast<const float4*>(&x[i * STEP_ROWS * N]));
+    tile.store(tmp, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+  }
+}
+
+/**
+ * Copy 16 bytes from the globale memory address pointed to by x to the smem
+ * address pointed to by row_address.
+ *
+ * A simple wrapper over the PTX.
+ */
+template <typename T>
+__device__ inline void cp_async_16(uint32_t row_address, const T* x) {
+  asm volatile(
+      "cp.async.ca.shared::cta.global [%0], [%1], 16;\n" ::"r"(row_address),
+      "l"(reinterpret_cast<const int4*>(x)));
+}
+
+/**
+ * Submit all the previous async copies to be executed.
+ */
+__device__ inline void cp_async_commit() {
+  asm volatile("cp.async.commit_group;\n" ::);
+}
+
+/**
+ * Wait for all the async copies to finish.
+ */
+__device__ inline void cp_async_wait_all() {
+  asm volatile("cp.async.wait_all;\n" ::);
+}
+
+/**
+ * The asynchronous equivalent of load.
+ *
+ * Loads the tile from global memory by submitting a bunch of async copy
+ * instructions. The copy won't start until commit is called and we don't have
+ * a guarantee it will finish until wait is called.
+ *
+ * It should be used as follows
+ *
+ *    load(...)
+ *    load(...)
+ *    cp_async_commit()
+ *    do_other_stuff()
+ *    cp_async_wait_all()
+ *    do_stuff_with_shmem()
+ */
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void
+load_async(Tile& tile, uint32_t base_address, const T* x, int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  x += row * N + col * ELEMENTS_PER_LOAD;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    cp_async_16(
+        tile.loc(base_address, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD),
+        x + i * STEP_ROWS * N);
+  }
+}
+
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void load_async_safe(
+    Tile& tile,
+    uint32_t base_address,
+    const T* x,
+    int N,
+    int max_rows) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  x += row * N + col * ELEMENTS_PER_LOAD;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    if (row + i * STEP_ROWS < max_rows) {
+      cp_async_16(
+          tile.loc(base_address, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD),
+          x + i * STEP_ROWS * N);
+    } else {
+      float4 tmp = {0, 0, 0, 0};
+      tile.store(tmp, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+    }
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/primitives.cu

@@ -24,22 +24,21 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   if (out.size() == 0) {
     return;
   }
-  auto& s = stream();
-  auto& encoder = cu::get_command_encoder(s);
+  auto& encoder = cu::get_command_encoder(stream());
   encoder.set_output_array(out);
-  encoder.launch_kernel([&, this](cudaStream_t stream) {
-    MLX_SWITCH_INT_FLOAT_TYPES_CHECKED(out.dtype(), "Arange", CTYPE, {
-      using OutType = cuda_type_t<CTYPE>;
-      CTYPE step =
-          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
-      thrust::transform(
-          cu::thrust_policy(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)});
-    });
+  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)});
   });
 }
 
@@ -71,10 +70,8 @@ bool fast::ScaledDotProductAttention::use_fallback(
     throw std::runtime_error(#func " has no CUDA implementation.");   \
   }
 
-NO_GPU(ArgPartition)
 NO_GPU(BlockMaskedMM)
 NO_GPU(Convolution)
-NO_GPU_MULTI(DivMod)
 NO_GPU(DynamicSlice)
 NO_GPU(DynamicSliceUpdate)
 NO_GPU(FFT)
@@ -83,10 +80,8 @@ NO_GPU(GatherQMM)
 NO_GPU(Hadamard)
 NO_GPU(Load)
 NO_GPU_MULTI(LUF)
-NO_GPU(Partition)
 NO_GPU_MULTI(QRF)
-NO_GPU(QuantizedMatmul)
-NO_GPU(Scan)
+NO_GPU(SegmentedMM)
 NO_GPU_MULTI(SVD)
 NO_GPU(Inverse)
 NO_GPU(Cholesky)
@@ -95,7 +90,6 @@ NO_GPU_MULTI(Eigh)
 
 namespace fast {
 NO_GPU(ScaledDotProductAttention)
-NO_GPU_MULTI(AffineQuantize)
 NO_GPU_MULTI(CustomKernel)
 } // namespace fast
 

mlx/backend/cuda/quantized/affine_quantize.cu

@@ -0,0 +1,331 @@
+// 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(kernel, 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(kernel, 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

mlx/backend/cuda/quantized/qmm.cu

@@ -0,0 +1,228 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/matmul/mma.cuh"
+#include "mlx/backend/cuda/matmul/tiles.cuh"
+#include "mlx/backend/cuda/quantized/quantized_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+namespace mlx::core {
+
+namespace cu {
+
+template <int NUM_WARPS, int group_size, int bits, typename T, typename Tile>
+__device__ inline void load_quantized(
+    Tile& tile,
+    const uint8_t* x,
+    const T* scales,
+    const T* biases,
+    int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(uint32_t) * get_pack_factor<bits>();
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+  constexpr int MASK = (1 << bits) - 1;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  const int Nx = N / get_pack_factor<bits>();
+  const int Ng = N / group_size;
+
+  x += row * Nx + col * (ELEMENTS_PER_LOAD / get_pack_factor<bits>());
+  scales += row * Ng + col * ELEMENTS_PER_LOAD / group_size;
+  biases += row * Ng + col * ELEMENTS_PER_LOAD / group_size;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    T vs[ELEMENTS_PER_LOAD];
+    uint32_t w = *reinterpret_cast<const uint32_t*>(x + i * STEP_ROWS * Nx);
+    T s = scales[i * STEP_ROWS * Ng];
+    T b = biases[i * STEP_ROWS * Ng];
+    MLX_UNROLL
+    for (int j = 0; j < ELEMENTS_PER_LOAD; j++) {
+      vs[j] = static_cast<T>((w >> (j * bits)) & MASK) * s + b;
+    }
+    tile.store(vs, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+  }
+}
+
+template <
+    typename T,
+    int BM,
+    int BN,
+    int BK,
+    int group_size,
+    int bits,
+    bool aligned_M>
+__global__ void qmm_t(
+    const T* x,
+    const uint8_t* w,
+    const T* scales,
+    const T* biases,
+    T* y,
+    int M,
+    int N,
+    int K) {
+  constexpr int WARPS_M = 2;
+  constexpr int WARPS_N = 4;
+  constexpr int NUM_WARPS = WARPS_M * WARPS_N;
+  constexpr int WARP_STEP_M = BM / WARPS_M;
+  constexpr int WARP_STEP_N = BN / WARPS_N;
+
+  const int warpid = threadIdx.x / 32;
+  const int laneid = threadIdx.x % 32;
+  const int wm = warpid / WARPS_N;
+  const int wn = warpid % WARPS_N;
+  const int offset_m = wm * WARP_STEP_M;
+  const int offset_n = wn * WARP_STEP_N;
+
+  extern __shared__ char shmem[];
+  SharedTile<T, BM, BK>(&xs)[1] = *(SharedTile<T, BM, BK>(*)[1])(&shmem[0]);
+  SharedTile<T, BN, BK>(&ws)[1] =
+      *(SharedTile<T, BN, BK>(*)[1])(&shmem[1 * sizeof(T) * BM * BK]);
+
+  RegisterTile<float, BM / WARPS_M, BN / WARPS_N> C;
+  RegisterTile<T, BM / WARPS_M, 16> A;
+  RegisterTile<T, BN / WARPS_N, 16> B;
+
+  const int max_rows = M - blockIdx.y * BM;
+
+  x += blockIdx.y * BM * K;
+  w += blockIdx.x * BN * K / get_pack_factor<bits>();
+  scales += blockIdx.x * BN * K / group_size;
+  biases += blockIdx.x * BN * K / group_size;
+  y += blockIdx.y * BM * N + blockIdx.x * BN;
+
+  C.fill(0);
+
+  int tic = 0;
+  uint32_t base_addr_xs[1], base_addr_ws[1];
+  base_addr_xs[0] = __cvta_generic_to_shared(&xs[0].data[0]);
+  base_addr_ws[0] = __cvta_generic_to_shared(&ws[0].data[0]);
+
+  if (aligned_M || max_rows >= BM) {
+    for (int k_block = 0; k_block < K; k_block += BK) {
+      load_async<NUM_WARPS>(xs[tic], base_addr_xs[tic], x + k_block, K);
+      cp_async_commit();
+      load_quantized<NUM_WARPS, group_size, bits>(
+          ws[tic],
+          w + k_block / get_pack_factor<bits>(),
+          scales + k_block / group_size,
+          biases + k_block / group_size,
+          K);
+      cp_async_wait_all();
+      __syncthreads();
+
+      MLX_UNROLL
+      for (int k = 0; k < BK / 16; k++) {
+        A.load(
+            xs[tic],
+            base_addr_xs[tic],
+            offset_m + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        B.load(
+            ws[tic],
+            base_addr_ws[tic],
+            offset_n + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        mma_t(C, A, B);
+      }
+    }
+    C.store_global(y, N, offset_m, offset_n);
+  } else {
+    for (int k_block = 0; k_block < K; k_block += BK) {
+      load_async_safe<NUM_WARPS>(
+          xs[tic], base_addr_xs[tic], x + k_block, K, max_rows);
+      cp_async_commit();
+      load_quantized<NUM_WARPS, group_size, bits>(
+          ws[tic],
+          w + k_block / get_pack_factor<bits>(),
+          scales + k_block / group_size,
+          biases + k_block / group_size,
+          K);
+      cp_async_wait_all();
+      __syncthreads();
+
+      MLX_UNROLL
+      for (int k = 0; k < BK / 16; k++) {
+        A.load(
+            xs[tic],
+            base_addr_xs[tic],
+            offset_m + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        B.load(
+            ws[tic],
+            base_addr_ws[tic],
+            offset_n + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        mma_t(C, A, B);
+      }
+    }
+    C.store_global_safe(y, N, offset_m, offset_n, max_rows);
+  }
+}
+
+} // namespace cu
+
+void qmm(
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& biases,
+    array& out,
+    bool transpose_,
+    int group_size_,
+    int bits_,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  if (x.dtype() != bfloat16) {
+    throw std::invalid_argument("[qmm] Only bfloat16 is supported for now");
+  }
+  if (!transpose_) {
+    throw std::invalid_argument(
+        "[qmm] Only transposed matmul is supported for now");
+  }
+
+  dispatch_float_types(x.dtype(), "qmm", [&](auto type_tag) {
+    dispatch_groups(group_size_, [&](auto group_size) {
+      dispatch_bits(bits_, [&](auto bits) {
+        using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+
+        constexpr int BM = 128;
+        constexpr int BN = 128;
+        constexpr int BK = 32;
+        auto kernel =
+            cu::qmm_t<DataType, BM, BN, BK, group_size.value, bits.value, true>;
+        if (M % BM != 0) {
+          kernel = cu::
+              qmm_t<DataType, BM, BN, BK, group_size.value, bits.value, false>;
+        }
+
+        dim3 grid((N + BN - 1) / BN, (M + BM - 1) / BM);
+
+        enc.add_kernel_node(
+            kernel,
+            grid,
+            2 * 4 * 32,
+            1 * sizeof(DataType) * (BM * BK + BN * BK),
+            x.data<DataType>(),
+            w.data<uint8_t>(),
+            scales.data<DataType>(),
+            biases.data<DataType>(),
+            out.data<DataType>(),
+            M,
+            N,
+            K);
+      });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/quantized/quantized.cu

@@ -0,0 +1,113 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/quantized/quantized.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/fast_primitives.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace {
+
+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) {
+  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;
+  } else {
+    array x_copy = contiguous_copy_gpu(x, s);
+    enc.add_temporary(x_copy);
+    return x_copy;
+  }
+}
+
+} // namespace
+
+void QuantizedMatmul::eval_gpu(const std::vector<array>& inputs, array& out) {
+  auto& s = stream();
+  auto& d = cu::device(s.device);
+  auto& enc = d.get_command_encoder(s);
+
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  // Make sure the last two dims of x and w, s, b are contiguous. This should
+  // be relaxed for x.
+  array x = ensure_row_contiguous_matrix(inputs[0], enc, s);
+  array w = ensure_row_contiguous_matrix(inputs[1], enc, s);
+  array scales = ensure_row_contiguous_matrix(inputs[2], enc, s);
+  array biases = ensure_row_contiguous_matrix(inputs[3], enc, s);
+
+  // Extract the matmul shapes
+  bool non_batched = w.ndim() == 2 && x.flags().row_contiguous;
+  int K = x.shape(-1);
+  int M = non_batched ? x.size() / K : x.shape(-2);
+  int N = out.shape(-1);
+
+  qmm(x,
+      w,
+      scales,
+      biases,
+      out,
+      transpose_,
+      group_size_,
+      bits_,
+      M,
+      N,
+      K,
+      enc,
+      s);
+}
+
+void fast::AffineQuantize::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  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

mlx/backend/cuda/quantized/quantized.cuh

@@ -0,0 +1,42 @@
+// 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);
+
+void qmm(
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& biases,
+    array& out,
+    bool transpose_,
+    int group_size_,
+    int bits_,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc,
+    const Stream& s);
+
+} // namespace mlx::core

mlx/backend/cuda/quantized/quantized_utils.cuh

@@ -0,0 +1,59 @@
+// 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

mlx/backend/cuda/random.cu

@@ -156,34 +156,41 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(keys);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    dim3 grid_dims{num_keys, half_size + odd};
-    int64_t total = grid_dims.x * grid_dims.y;
-    int32_t threads_y = 1;
-    while ((total / threads_y) >= (1U << 31)) {
-      threads_y *= 2;
-    }
-    int32_t threads_x = cuda::ceil_div(total, threads_y);
-    auto [grid, block] = get_grid_and_block(threads_x, threads_y, 1);
-    if (keys.flags().row_contiguous) {
-      cu::rbitsc<<<grid, block, 0, stream>>>(
-          keys.data<uint32_t>(),
-          out.data<uint8_t>(),
-          grid_dims,
-          odd,
-          bytes_per_key);
-    } else {
-      cu::rbits<<<grid, block, 0, stream>>>(
-          keys.data<uint32_t>(),
-          out.data<uint8_t>(),
-          grid_dims,
-          odd,
-          bytes_per_key,
-          keys.ndim(),
-          const_param(keys.shape()),
-          const_param(keys.strides()));
-    }
-  });
+  dim3 grid_dims{num_keys, half_size + odd};
+  int64_t total = grid_dims.x * grid_dims.y;
+  int32_t threads_y = 1;
+  while ((total / threads_y) >= (1U << 31)) {
+    threads_y *= 2;
+  }
+  int32_t threads_x = cuda::ceil_div(total, threads_y);
+  auto [grid, block] = get_grid_and_block(threads_x, threads_y, 1);
+  auto& stream = encoder.stream();
+  if (keys.flags().row_contiguous) {
+    encoder.add_kernel_node(
+        cu::rbitsc,
+        grid,
+        block,
+        0,
+        keys.data<uint32_t>(),
+        out.data<uint8_t>(),
+        grid_dims,
+        odd,
+        bytes_per_key);
+  } else {
+    encoder.add_kernel_node(
+        cu::rbits,
+        grid,
+        block,
+        0,
+        keys.data<uint32_t>(),
+        out.data<uint8_t>(),
+        grid_dims,
+        odd,
+        bytes_per_key,
+        keys.ndim(),
+        const_param(keys.shape()),
+        const_param(keys.strides()));
+  }
 }
 
 } // namespace mlx::core

mlx/backend/cuda/reduce.cu

@@ -21,28 +21,11 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(!axes_.empty());
   assert(out.size() != in.size());
 
-  out.set_data(allocator::malloc(out.nbytes()));
-
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  encoder.set_input_array(in);
-  encoder.set_output_array(out);
 
-  // Fill out with init value.
   if (in.size() == 0) {
-    encoder.launch_kernel([&](cudaStream_t stream) {
-      MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
-        MLX_SWITCH_REDUCE_OPS(reduce_type_, OP, {
-          using InType = cuda_type_t<CTYPE>;
-          using OutType = cu::ReduceResult<OP, InType>::type;
-          thrust::fill_n(
-              cu::thrust_policy(stream),
-              thrust::device_pointer_cast(out.data<OutType>()),
-              out.data_size(),
-              cu::ReduceInit<OP, InType>::value());
-        });
-      });
-    });
+    init_reduce(encoder, in, out, reduce_type_);
     return;
   }
 
@@ -51,17 +34,27 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   // If it is a general reduce then copy the input to a contiguous array and
   // recompute the plan.
-  if (plan.type == GeneralReduce) {
-    array in_copy(in.shape(), in.dtype(), nullptr, {});
-    copy_gpu(in, in_copy, CopyType::General, s);
+  //
+  // TODO: Instead of copying we can use elem-to-loc to deal with broadcasting
+  //       like we do in Metal. When it comes to broadcasted reduction axes
+  //       some can be ignored eg for min/max.
+  bool broadcasted = false;
+  for (int i = 0, j = 0; i < in.ndim() && !broadcasted; i++) {
+    if (j < axes_.size() && axes_[j] == i) {
+      j++;
+    } else {
+      broadcasted = in.strides(i) == 0;
+    }
+  }
+  if (plan.type == GeneralReduce || broadcasted || !in.flags().contiguous) {
+    array in_copy = contiguous_copy_gpu(in, s);
     encoder.add_temporary(in_copy);
     in = in_copy;
     plan = get_reduction_plan(in, axes_);
   }
 
-  if ((plan.type == ContiguousAllReduce) ||
-      (plan.type == ContiguousReduce && plan.shape.size() == 1)) {
-    segmented_reduce(encoder, in, out, reduce_type_, axes_, plan);
+  if (plan.type == ContiguousAllReduce) {
+    all_reduce(encoder, in, out, reduce_type_);
     return;
   }
 

mlx/backend/cuda/reduce/all_reduce.cu

@@ -0,0 +1,159 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/reduce/reduce.cuh"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <cub/block/block_load.cuh>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T, typename U, typename ReduceOp, int N = 4>
+__global__ void all_reduce(T* in, U* out, size_t block_step, size_t size) {
+  // TODO: Process multiple "rows" in each thread
+  constexpr int M = 1;
+
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  const U init = cu::ReduceInit<ReduceOp, T>::value();
+  ReduceOp op;
+
+  T vals[N];
+  U accs[M];
+  accs[0] = init;
+
+  size_t start = grid.block_rank() * block_step;
+  size_t end = start + block_step;
+  size_t check = min(end, size);
+
+  size_t i = start;
+  for (; i + block.size() * N <= check; i += block.size() * N) {
+    cub::LoadDirectBlockedVectorized<T, N>(block.thread_rank(), in + i, vals);
+    for (int j = 0; j < N; j++) {
+      accs[0] = op(accs[0], cast_to<U>(vals[j]));
+    }
+  }
+
+  if (i < check) {
+    cub::LoadDirectBlocked(
+        block.thread_rank(), in + i, vals, check - i, cast_to<T>(init));
+    for (int i = 0; i < N; i++) {
+      accs[0] = op(accs[0], cast_to<U>(vals[i]));
+    }
+  }
+
+  __shared__ U shared_accumulators[32];
+  block_reduce(block, warp, accs, shared_accumulators, op, init);
+
+  if (block.thread_rank() == 0) {
+    out[grid.block_rank()] = accs[0];
+  }
+}
+
+} // namespace cu
+
+void all_reduce(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type) {
+  constexpr int N_READS = 8;
+
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto get_args = [](size_t size, int N) {
+    int threads = std::min(512UL, (size + N - 1) / N);
+    threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+    int reductions_per_step = threads * N;
+    size_t steps_needed =
+        (size + reductions_per_step - 1) / reductions_per_step;
+
+    int blocks;
+    if (steps_needed < 32) {
+      blocks = 1;
+    } else if (steps_needed < 128) {
+      blocks = 32;
+    } else if (steps_needed < 512) {
+      blocks = 128;
+    } else if (steps_needed < 1024) {
+      blocks = 512;
+    } else {
+      blocks = 1024;
+    }
+
+    size_t steps_per_block = (steps_needed + blocks - 1) / blocks;
+    size_t block_step = steps_per_block * reductions_per_step;
+
+    return std::make_tuple(blocks, threads, block_step);
+  };
+
+  int blocks, threads;
+  size_t block_step;
+  size_t insize = in.size();
+  Dtype dt = in.dtype();
+
+  // Cub doesn't like const pointers for load (sigh).
+  void* indata = const_cast<void*>(in.data<void>());
+
+  // Large array so allocate an intermediate and accumulate there
+  std::tie(blocks, threads, block_step) = get_args(insize, N_READS);
+  encoder.set_input_array(in);
+  if (blocks > 1) {
+    array intermediate({blocks}, out.dtype(), nullptr, {});
+    intermediate.set_data(allocator::malloc(intermediate.nbytes()));
+    encoder.add_temporary(intermediate);
+    encoder.set_output_array(intermediate);
+    dispatch_all_types(dt, [&](auto type_tag) {
+      dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+        using OP = MLX_GET_TYPE(reduce_type_tag);
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        using U = typename cu::ReduceResult<OP, T>::type;
+        auto kernel = cu::all_reduce<T, U, OP, N_READS>;
+        encoder.add_kernel_node(
+            kernel,
+            blocks,
+            threads,
+            0,
+            static_cast<T*>(indata),
+            intermediate.data<U>(),
+            block_step,
+            insize);
+      });
+    });
+
+    // Set the input for the next step and recalculate the blocks
+    indata = intermediate.data<void>();
+    dt = intermediate.dtype();
+    insize = intermediate.size();
+    std::tie(blocks, threads, block_step) = get_args(insize, N_READS);
+    encoder.set_input_array(intermediate);
+  }
+
+  encoder.set_output_array(out);
+  dispatch_all_types(dt, [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      using OP = MLX_GET_TYPE(reduce_type_tag);
+      using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      using U = typename cu::ReduceResult<OP, T>::type;
+      auto kernel = cu::all_reduce<T, U, OP, N_READS>;
+      encoder.add_kernel_node(
+          kernel,
+          blocks,
+          threads,
+          0,
+          static_cast<T*>(indata),
+          out.data<U>(),
+          block_step,
+          insize);
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/reduce/col_reduce.cu

@@ -1,7 +1,8 @@
 // Copyright © 2025 Apple Inc.
 
+#include <numeric>
+
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cast_op.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 
 #include <cooperative_groups.h>
@@ -36,19 +37,36 @@ struct ColReduceArgs {
       const array& in,
       const ReductionPlan& plan,
       const std::vector<int>& axes) {
+    using ShapeVector = decltype(plan.shape);
+    using StridesVector = decltype(plan.strides);
+
+    ShapeVector shape_vec;
+    StridesVector strides_vec;
+
     assert(!plan.shape.empty());
     reduction_size = plan.shape.back();
     reduction_stride = plan.strides.back();
 
     int64_t stride_back = 1;
-    auto [shape_vec, strides_vec] = shapes_without_reduction_axes(in, axes);
+    std::tie(shape_vec, strides_vec) = shapes_without_reduction_axes(in, axes);
     while (!shape_vec.empty() && stride_back < reduction_stride) {
       stride_back *= shape_vec.back();
       shape_vec.pop_back();
       strides_vec.pop_back();
     }
+    std::vector<int> indices(shape_vec.size());
+    std::iota(indices.begin(), indices.end(), 0);
+    std::sort(indices.begin(), indices.end(), [&](int left, int right) {
+      return strides_vec[left] > strides_vec[right];
+    });
+    ShapeVector sorted_shape;
+    StridesVector sorted_strides;
+    for (auto idx : indices) {
+      sorted_shape.push_back(shape_vec[idx]);
+      sorted_strides.push_back(strides_vec[idx]);
+    }
     std::tie(shape_vec, strides_vec) =
-        collapse_contiguous_dims(shape_vec, strides_vec);
+        collapse_contiguous_dims(sorted_shape, sorted_strides);
     shape = const_param(shape_vec);
     strides = const_param(strides_vec);
     ndim = shape_vec.size();
@@ -64,86 +82,6 @@ struct ColReduceArgs {
   }
 };
 
-template <typename T, typename U, typename Op, int NDIM, int N_READS = 4>
-__global__ void col_reduce_small(
-    const T* in,
-    U* out,
-    const __grid_constant__ ColReduceArgs args) {
-  auto grid = cg::this_grid();
-  auto block = cg::this_thread_block();
-
-  int column =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  if (column * N_READS >= args.reduction_stride) {
-    return;
-  }
-
-  int out_idx = grid.block_rank() / grid.dim_blocks().x;
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
-
-  Op op;
-  U totals[N_READS];
-  for (int i = 0; i < N_READS; i++) {
-    totals[i] = ReduceInit<Op, T>::value();
-  }
-
-  // Read input to local.
-  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-  loop.next(
-      block.thread_index().y,
-      args.reduce_shape.data(),
-      args.reduce_strides.data());
-  for (size_t r = block.thread_index().y;
-       r < args.non_col_reductions * args.reduction_size;
-       r += block.dim_threads().y) {
-    U vals[N_READS];
-    cub::LoadDirectBlocked(
-        column,
-        make_cast_iterator<U>(in + loop.location()),
-        vals,
-        args.reduction_stride,
-        ReduceInit<Op, T>::value());
-    for (int i = 0; i < N_READS; i++) {
-      totals[i] = op(vals[i], totals[i]);
-    }
-    loop.next(
-        block.dim_threads().y,
-        args.reduce_shape.data(),
-        args.reduce_strides.data());
-  }
-
-  // Do block reduce when each column has more than 1 element to reduce.
-  if (block.dim_threads().y > 1) {
-    __shared__ U shared_vals[32 * 8 * N_READS];
-    size_t col =
-        block.thread_index().y * block.dim_threads().x + block.thread_index().x;
-    for (int i = 0; i < N_READS; i++) {
-      shared_vals[col * N_READS + i] = totals[i];
-    }
-    block.sync();
-    if (block.thread_index().y == 0) {
-      for (int i = 0; i < N_READS; i++) {
-        totals[i] = shared_vals[block.thread_index().x * N_READS + i];
-      }
-      for (int j = 1; j < block.dim_threads().y; j++) {
-        col = j * block.dim_threads().x + block.thread_index().x;
-        for (int i = 0; i < N_READS; i++) {
-          totals[i] = op(shared_vals[col * N_READS + i], totals[i]);
-        }
-      }
-    }
-  }
-
-  // Write result.
-  if (block.thread_index().y == 0) {
-    cub::StoreDirectBlocked(
-        column,
-        out + out_idx * args.reduction_stride,
-        totals,
-        args.reduction_stride);
-  }
-}
-
 template <
     typename T,
     typename U,
@@ -152,67 +90,94 @@ template <
     int BM,
     int BN,
     int N_READS = 4>
-__global__ void col_reduce_looped(
-    const T* in,
-    U* out,
-    const __grid_constant__ ColReduceArgs args) {
+__global__ void
+col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<WARP_SIZE>(block);
 
-  constexpr int n_warps = BN / N_READS;
+  constexpr int threads_per_row = BN / N_READS;
 
-  int out_idx = grid.block_rank() / grid.dim_blocks().x;
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
+  // Compute the indices for the tile
+  size_t tile_idx = grid.block_rank();
+  size_t tile_x = tile_idx % ((args.reduction_stride + BN - 1) / BN);
+  size_t tile_y = tile_idx / ((args.reduction_stride + BN - 1) / BN);
 
+  // Compute the indices for the thread within the tile
+  short thread_x = block.thread_rank() % threads_per_row;
+  short thread_y = block.thread_rank() / threads_per_row;
+
+  // Move the input pointer
+  in += elem_to_loc(tile_y, args.shape.data(), args.strides.data(), args.ndim) +
+      tile_x * BN;
+
+  // Initialize the running totals
   Op op;
   U totals[N_READS];
   for (int i = 0; i < N_READS; i++) {
     totals[i] = ReduceInit<Op, T>::value();
   }
 
-  // Read input to local.
-  int r = block.thread_rank() / n_warps;
-  int column = block.thread_rank() % n_warps;
-  int in_offset = grid.block_index().x * BN;
   LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-  loop.next(r, args.reduce_shape.data(), args.reduce_strides.data());
-  for (; r < args.non_col_reductions * args.reduction_size; r += BM) {
-    U vals[N_READS];
-    cub::LoadDirectBlocked(
-        column,
-        make_cast_iterator<U>(in + loop.location() + in_offset),
-        vals,
-        args.reduction_stride - in_offset,
-        ReduceInit<Op, T>::value());
-    for (int i = 0; i < N_READS; i++) {
-      totals[i] = op(vals[i], totals[i]);
+  loop.next(thread_y, args.reduce_shape.data(), args.reduce_strides.data());
+  size_t total = args.non_col_reductions * args.reduction_size;
+  if (tile_x * BN + BN <= args.reduction_stride) {
+    if (args.reduction_stride % N_READS == 0) {
+      for (size_t r = thread_y; r < total; r += BM) {
+        T vals[N_READS];
+        cub::LoadDirectBlockedVectorized(thread_x, in + loop.location(), vals);
+        for (int i = 0; i < N_READS; i++) {
+          totals[i] = op(totals[i], cast_to<U>(vals[i]));
+        }
+        loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
+      }
+    } else {
+      for (size_t r = thread_y; r < total; r += BM) {
+        T vals[N_READS];
+        cub::LoadDirectBlocked(thread_x, in + loop.location(), vals);
+        for (int i = 0; i < N_READS; i++) {
+          totals[i] = op(totals[i], cast_to<U>(vals[i]));
+        }
+        loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
+      }
+    }
+  } else {
+    for (size_t r = thread_y; r < total; r += BM) {
+      T vals[N_READS];
+      cub::LoadDirectBlocked(
+          thread_x,
+          in + loop.location(),
+          vals,
+          args.reduction_stride - tile_x * BN,
+          cast_to<T>(ReduceInit<Op, T>::value()));
+      for (int i = 0; i < N_READS; i++) {
+        totals[i] = op(totals[i], cast_to<U>(vals[i]));
+      }
+      loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
     }
-    loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
   }
 
   // Do warp reduce for each output.
-  constexpr int n_outputs = BN / n_warps;
+  constexpr int n_outputs = BN / threads_per_row;
   static_assert(BM == 32 && n_outputs == N_READS);
   __shared__ U shared_vals[BM * BN];
-  size_t col = block.thread_index().y * BN + block.thread_index().x * N_READS;
+  short s_idx = thread_y * BN + thread_x * N_READS;
   for (int i = 0; i < N_READS; i++) {
-    shared_vals[col + i] = totals[i];
+    shared_vals[s_idx + i] = totals[i];
   }
   block.sync();
-  col = warp.thread_rank() * BN + warp.meta_group_rank() * n_outputs;
+  s_idx = warp.thread_rank() * BN + warp.meta_group_rank() * n_outputs;
   for (int i = 0; i < n_outputs; i++) {
-    totals[i] = cg::reduce(warp, shared_vals[col + i], op);
+    totals[i] = cg::reduce(warp, shared_vals[s_idx + i], op);
   }
 
   // Write result.
   if (warp.thread_rank() == 0) {
-    size_t out_offset = grid.block_index().x * BN;
     cub::StoreDirectBlocked(
         warp.meta_group_rank(),
-        out + out_idx * args.reduction_stride + out_offset,
+        out + tile_y * args.reduction_stride + tile_x * BN,
         totals,
-        args.reduction_stride - out_offset);
+        args.reduction_stride - tile_x * BN);
   }
 }
 
@@ -220,14 +185,55 @@ __global__ void col_reduce_looped(
 
 inline auto output_grid_for_col_reduce(
     const array& out,
-    const cu::ColReduceArgs& args) {
-  auto out_shape = out.shape();
-  auto out_strides = out.strides();
-  while (!out_shape.empty() && out_strides.back() < args.reduction_stride) {
-    out_shape.pop_back();
-    out_strides.pop_back();
+    const cu::ColReduceArgs& args,
+    int bn) {
+  int gx, gy = 1;
+  size_t n_inner_blocks = cuda::ceil_div(args.reduction_stride, bn);
+  size_t n_outer_blocks = out.size() / args.reduction_stride;
+  size_t n_blocks = n_outer_blocks * n_inner_blocks;
+  while (n_blocks / gy > INT32_MAX) {
+    gy *= 2;
   }
-  return get_2d_grid_dims(out_shape, out_strides);
+  gx = cuda::ceil_div(n_blocks, gy);
+
+  return dim3(gx, gy, 1);
+}
+
+void col_reduce_looped(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type,
+    const std::vector<int>& axes,
+    const ReductionPlan& plan,
+    cu::ColReduceArgs args) {
+  // Allocate data for the output using in's layout to access them as
+  // contiguously as possible.
+  allocate_same_layout(out, in, axes);
+
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
+        using OP = MLX_GET_TYPE(reduce_type_tag);
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        using U = typename cu::ReduceResult<OP, T>::type;
+        // Cub doesn't like const pointers for vectorized loads. (sigh)
+        T* indata = const_cast<T*>(in.data<T>());
+
+        constexpr int N_READS = 4;
+        constexpr int BM = 32;
+        constexpr int BN = 32;
+        dim3 grid = output_grid_for_col_reduce(out, args, BN);
+        int blocks = BM * BN / N_READS;
+        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);
+      });
+    });
+  });
 }
 
 void col_reduce(
@@ -237,42 +243,23 @@ void col_reduce(
     Reduce::ReduceType reduce_type,
     const std::vector<int>& axes,
     const ReductionPlan& plan) {
+  // Current col reduce options
+  //
+  // - col_reduce_looped
+  //
+  //   It is a general strided reduce. Each threadblock computes the output for
+  //   a subrow of the fast moving axis. For instance 32 elements.
+  //
+  // Notes: As in row reduce we opt to read as much in order as possible and
+  //        leave transpositions as they are (contrary to our Metal backend).
+  //
+  //        Moreover we need different kernels for short rows and tuning
+
+  // Make the args struct to help route to the best kernel
   cu::ColReduceArgs args(in, plan, axes);
 
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
-      using InType = cuda_type_t<CTYPE>;
-      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
-        using OutType = cu::ReduceResult<OP, InType>::type;
-        MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
-          constexpr int N_READS = 4;
-          dim3 block_dims;
-          dim3 num_blocks = output_grid_for_col_reduce(out, args);
-          num_blocks.z = num_blocks.y;
-          num_blocks.y = num_blocks.x;
-          auto kernel =
-              cu::col_reduce_small<InType, OutType, OP, NDIM, N_READS>;
-          size_t total = args.non_col_reductions * args.reduction_size;
-          if (total < 32) {
-            size_t stride_blocks =
-                cuda::ceil_div(args.reduction_stride, N_READS);
-            block_dims.x = std::min(stride_blocks, 32ul);
-            block_dims.y = std::min(total, 8ul);
-            num_blocks.x = cuda::ceil_div(stride_blocks, block_dims.x);
-          } else {
-            constexpr int BM = 32;
-            constexpr int BN = 32;
-            block_dims.x = BM * BN / N_READS;
-            num_blocks.x = cuda::ceil_div(args.reduction_stride, BN);
-            kernel = cu::
-                col_reduce_looped<InType, OutType, OP, NDIM, BM, BN, N_READS>;
-          }
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in.data<InType>(), out.data<OutType>(), args);
-        });
-      });
-    });
-  });
+  // Fallback col reduce
+  col_reduce_looped(encoder, in, out, reduce_type, axes, plan, args);
 }
 
 } // namespace mlx::core

mlx/backend/cuda/reduce/init_reduce.cu

@@ -0,0 +1,50 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/reduce/reduce.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T, typename U, typename Op>
+__global__ void init_reduce(U* out, size_t size) {
+  auto index = cg::this_grid().thread_rank();
+  if (index < size) {
+    out[index] = ReduceInit<Op, T>::value();
+  }
+}
+
+} // namespace cu
+
+void init_reduce(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type) {
+  // Allocate if needed
+  if (out.data_shared_ptr() == nullptr) {
+    out.set_data(allocator::malloc(out.nbytes()));
+  }
+
+  encoder.set_output_array(out);
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      using OP = MLX_GET_TYPE(reduce_type_tag);
+      using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      using U = typename cu::ReduceResult<OP, T>::type;
+      auto kernel = cu::init_reduce<T, U, OP>;
+      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());
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/reduce/reduce.cuh

@@ -1,7 +1,8 @@
 // Copyright © 2025 Apple Inc.
 
+#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"
@@ -9,51 +10,41 @@
 
 namespace mlx::core {
 
-// Dispatch dynamic ndim to constexpr.
-// The behavior follows get_kernel_reduce_ndim in metal/reduce.cpp file.
-#define MLX_SWITCH_REDUCE_NDIM(ndim, NDIM, ...) \
-  if (ndim == 1) {                              \
-    constexpr uint32_t NDIM = 1;                \
-    __VA_ARGS__;                                \
-  } else if (ndim == 2) {                       \
-    constexpr uint32_t NDIM = 2;                \
-    __VA_ARGS__;                                \
-  } else {                                      \
-    constexpr uint32_t NDIM = 5;                \
-    __VA_ARGS__;                                \
+template <typename F>
+void dispatch_reduce_ndim(int ndim, F&& f) {
+  if (ndim == 1) {
+    f(std::integral_constant<int, 1>{});
+  } else if (ndim == 2) {
+    f(std::integral_constant<int, 2>{});
+  } else {
+    f(std::integral_constant<int, 5>{});
   }
+}
 
-// Dispatch reduce ops to constexpr.
-#define MLX_SWITCH_REDUCE_OPS(REDUCE, OP, ...)           \
-  if (REDUCE == Reduce::ReduceType::And) {               \
-    using OP = cu::And;                                  \
-    __VA_ARGS__;                                         \
-  } else if (REDUCE == Reduce::ReduceType::Or) {         \
-    using OP = cu::Or;                                   \
-    __VA_ARGS__;                                         \
-  } else if (REDUCE == Reduce::ReduceType::Sum) {        \
-    using OP = cu::Sum;                                  \
-    __VA_ARGS__;                                         \
-  } else if (REDUCE == Reduce::ReduceType::Prod) {       \
-    using OP = cu::Prod;                                 \
-    __VA_ARGS__;                                         \
-  } else if (REDUCE == Reduce::ReduceType::Max) {        \
-    using OP = cu::Max;                                  \
-    __VA_ARGS__;                                         \
-  } else if (REDUCE == Reduce::ReduceType::Min) {        \
-    using OP = cu::Min;                                  \
-    __VA_ARGS__;                                         \
-  } else {                                               \
-    throw std::invalid_argument("Unknown reduce type."); \
+template <typename F>
+void dispatch_reduce_ops(Reduce::ReduceType reduce_type, F&& f) {
+  if (reduce_type == Reduce::ReduceType::And) {
+    f(type_identity<cu::And>{});
+  } else if (reduce_type == Reduce::ReduceType::Or) {
+    f(type_identity<cu::Or>{});
+  } else if (reduce_type == Reduce::ReduceType::Sum) {
+    f(type_identity<cu::Sum>{});
+  } else if (reduce_type == Reduce::ReduceType::Prod) {
+    f(type_identity<cu::Prod>{});
+  } else if (reduce_type == Reduce::ReduceType::Max) {
+    f(type_identity<cu::Max>{});
+  } else if (reduce_type == Reduce::ReduceType::Min) {
+    f(type_identity<cu::Min>{});
+  } else {
+    throw std::invalid_argument("Unknown reduce type.");
   }
+}
 
-void segmented_reduce(
+void all_reduce(
     cu::CommandEncoder& encoder,
     const array& in,
     array& out,
-    Reduce::ReduceType reduce_type,
-    const std::vector<int>& axes,
-    const ReductionPlan& plan);
+    Reduce::ReduceType reduce_type);
 
 void row_reduce(
     cu::CommandEncoder& encoder,
@@ -71,4 +62,10 @@ void col_reduce(
     const std::vector<int>& axes,
     const ReductionPlan& plan);
 
+void init_reduce(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type);
+
 } // namespace mlx::core

mlx/backend/cuda/reduce/reduce_ops.cuh

@@ -2,49 +2,116 @@
 
 #pragma once
 
+#include "mlx/backend/cuda/device/atomic_ops.cuh"
+#include "mlx/backend/cuda/device/cast_op.cuh"
 #include "mlx/backend/cuda/device/utils.cuh"
+#include "mlx/backend/cuda/reduce/reduce_utils.cuh"
 
 namespace mlx::core::cu {
 
 // Reduce ops.
 struct And {
-  __device__ bool operator()(bool a, bool b) {
+  __device__ __forceinline__ bool operator()(bool a, bool b) {
     return a && b;
   }
+
+  __device__ void atomic_update(bool* x, bool y) {
+    atomic_reduce<bool, And>(x, y);
+  }
 };
 
 struct Or {
-  __device__ bool operator()(bool a, bool b) {
+  __device__ __forceinline__ bool operator()(bool a, bool b) {
     return a || b;
   }
+
+  __device__ void atomic_update(bool* x, bool y) {
+    atomic_reduce<bool, Or>(x, y);
+  }
 };
 
 struct Sum {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __device__ __forceinline__ T operator()(T a, T b) {
     return a + b;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Sum>(x, y);
+  }
+
+  __device__ void atomic_update(__nv_bfloat16* x, __nv_bfloat16 y) {
+    atomic_add(x, y);
+  }
+
+  __device__ void atomic_update(int* x, int y) {
+    atomic_add(x, y);
+  }
+
+  __device__ void atomic_update(float* x, float y) {
+    atomic_add(x, y);
+  }
 };
 
 struct Prod {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __device__ __forceinline__ T operator()(T a, T b) {
     return a * b;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Prod>(x, y);
+  }
 };
 
 struct Min {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __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;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Min>(x, y);
+  }
 };
 
 struct Max {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __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;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Max>(x, y);
+  }
 };
 
 // Traits to get the result type of reduce op.
@@ -108,10 +175,10 @@ struct ReduceInit<Or, T> {
 template <typename T>
 struct ReduceInit<Sum, T> {
   static constexpr __host__ __device__ auto value() {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
+    if constexpr (is_complex_v<T>) {
       return T{0, 0};
     } else {
-      return typename ReduceResult<Sum, T>::type{0};
+      return cast_to<typename ReduceResult<Sum, T>::type>(0);
     }
   }
 };
@@ -119,10 +186,10 @@ struct ReduceInit<Sum, T> {
 template <typename T>
 struct ReduceInit<Prod, T> {
   static constexpr __host__ __device__ auto value() {
-    if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return T{1, 1};
+    if constexpr (is_complex_v<T>) {
+      return T{1, 0};
     } else {
-      return typename ReduceResult<Prod, T>::type{1};
+      return cast_to<typename ReduceResult<Prod, T>::type>(1);
     }
   }
 };

mlx/backend/cuda/reduce/reduce_utils.cuh

@@ -0,0 +1,143 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <numeric>
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <size_t N>
+struct uint_by_size;
+template <>
+struct uint_by_size<2> {
+  using type = uint16_t;
+};
+template <>
+struct uint_by_size<4> {
+  using type = uint32_t;
+};
+template <>
+struct uint_by_size<8> {
+  using type = unsigned long long int;
+};
+
+template <typename T, typename Op>
+__device__ void atomic_reduce(T* x, T y) {
+  if constexpr (sizeof(T) == 1) {
+    using U = uint16_t;
+    U* x_int = (U*)((char*)x - ((size_t)x % 2));
+    int shift = ((char*)x - (char*)x_int) * 8;
+    int mask = 0xff << shift;
+    U old_val, new_val;
+    do {
+      old_val = *x_int;
+      T result = Op{}(static_cast<T>((old_val >> shift) & 0xff), y);
+      new_val = (old_val & ~mask) | (result << shift);
+    } while (atomicCAS(x_int, old_val, new_val) != old_val);
+  } else {
+    using U = typename uint_by_size<sizeof(T)>::type;
+    U* x_int = (U*)(x);
+    U old_val, new_val;
+    do {
+      old_val = *x_int;
+      T result = Op{}(*((T*)&old_val), y);
+      new_val = *((U*)&result);
+    } while (atomicCAS(x_int, old_val, new_val) != old_val);
+  }
+}
+
+template <typename T, int N, typename Block, typename Warp, typename Op>
+inline __device__ void
+block_reduce(Block block, Warp warp, T (&vals)[N], T* smem, Op op, T init) {
+  // First reduce in the current warp
+  for (int i = 0; i < N; i++) {
+    vals[i] = cg::reduce(warp, vals[i], op);
+  }
+
+  // Reduce across warps
+  if (warp.meta_group_size() > 1) {
+    if (warp.thread_rank() == 0) {
+      for (int i = 0; i < N; i++) {
+        smem[warp.meta_group_rank() * N + i] = vals[i];
+      }
+    }
+    block.sync();
+    if (warp.thread_rank() < warp.meta_group_size()) {
+      for (int i = 0; i < N; i++) {
+        vals[i] = smem[warp.thread_rank() * N + i];
+      }
+    } else {
+      for (int i = 0; i < N; i++) {
+        vals[i] = init;
+      }
+    }
+    for (int i = 0; i < N; i++) {
+      vals[i] = cg::reduce(warp, vals[i], op);
+    }
+  }
+}
+
+} // namespace cu
+
+inline void allocate_same_layout(
+    array& out,
+    const array& in,
+    const std::vector<int>& axes) {
+  if (in.flags().row_contiguous) {
+    out.set_data(allocator::malloc(out.nbytes()));
+    return;
+  }
+
+  if (out.ndim() < in.ndim()) {
+    throw std::runtime_error(
+        "Reduction without keepdims only supported for row-contiguous inputs");
+  }
+
+  // Calculate the transpositions applied to in in order to apply them to out.
+  std::vector<int> axis_order(in.ndim());
+  std::iota(axis_order.begin(), axis_order.end(), 0);
+  std::sort(axis_order.begin(), axis_order.end(), [&](int left, int right) {
+    return in.strides(left) > in.strides(right);
+  });
+
+  // Transpose the shape and calculate the strides
+  Shape out_shape(in.ndim());
+  Strides out_strides(in.ndim(), 1);
+  for (int i = 0; i < in.ndim(); i++) {
+    out_shape[i] = out.shape(axis_order[i]);
+  }
+  for (int i = in.ndim() - 2; i >= 0; i--) {
+    out_strides[i] = out_shape[i + 1] * out_strides[i + 1];
+  }
+
+  // Reverse the axis order to get the final strides
+  Strides final_strides(in.ndim());
+  for (int i = 0; i < in.ndim(); i++) {
+    final_strides[axis_order[i]] = out_strides[i];
+  }
+
+  // Calculate the resulting contiguity and do the memory allocation
+  auto [data_size, rc, cc] = check_contiguity(out.shape(), final_strides);
+  auto fl = in.flags();
+  fl.row_contiguous = rc;
+  fl.col_contiguous = cc;
+  fl.contiguous = true;
+  out.set_data(
+      allocator::malloc(out.nbytes()),
+      data_size,
+      final_strides,
+      fl,
+      allocator::free);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/reduce/row_reduce.cu

@@ -1,7 +1,8 @@
 // Copyright © 2025 Apple Inc.
 
+#include <numeric>
+
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cast_op.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 
 #include <cooperative_groups.h>
@@ -55,84 +56,108 @@ struct RowReduceArgs {
       non_row_reductions *= reduce_shape[i];
     }
   }
+
+  // Convert shape and strides as if in was contiguous
+  void sort_access_pattern(const array& in, const std::vector<int>& axes) {
+    auto shape_vec = in.shape();
+    auto strides_vec = in.strides();
+    std::tie(shape_vec, strides_vec) =
+        shapes_without_reduction_axes(shape_vec, strides_vec, axes);
+    std::vector<int> indices(shape_vec.size());
+    std::iota(indices.begin(), indices.end(), 0);
+    std::sort(indices.begin(), indices.end(), [&](int left, int right) {
+      return strides_vec[left] > strides_vec[right];
+    });
+    decltype(shape_vec) sorted_shape;
+    decltype(strides_vec) sorted_strides;
+    for (auto idx : indices) {
+      sorted_shape.push_back(shape_vec[idx]);
+      sorted_strides.push_back(strides_vec[idx]);
+    }
+    std::tie(shape_vec, strides_vec) =
+        collapse_contiguous_dims(sorted_shape, sorted_strides);
+    shape = const_param(shape_vec);
+    strides = const_param(strides_vec);
+    ndim = shape_vec.size();
+  }
 };
 
-template <typename T, typename U, typename Op, int NDIM, int N_READS = 4>
-__global__ void row_reduce_small(
-    const T* in,
-    U* out,
-    size_t out_size,
-    const __grid_constant__ RowReduceArgs args) {
-  size_t out_idx = cg::this_grid().thread_rank();
-  if (out_idx >= out_size) {
-    return;
-  }
-
-  Op op;
-
-  U total_val = ReduceInit<Op, T>::value();
-  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
-
-  for (size_t n = 0; n < args.non_row_reductions; n++) {
-    for (int r = 0; r < cuda::ceil_div(args.row_size, N_READS); r++) {
-      U vals[N_READS];
-      cub::LoadDirectBlocked(
-          r,
-          make_cast_iterator<U>(in + loop.location()),
-          vals,
-          args.row_size,
-          ReduceInit<Op, T>::value());
-      total_val = op(total_val, cub::ThreadReduce(vals, op));
-    }
-    loop.next(args.reduce_shape.data(), args.reduce_strides.data());
-  }
-
-  out[out_idx] = total_val;
-}
-
-template <typename T, typename U, typename Op, int NDIM, int N_READS = 4>
-__global__ void row_reduce_small_warp(
-    const T* in,
-    U* out,
-    size_t out_size,
-    const __grid_constant__ RowReduceArgs args) {
+template <typename T, typename U, typename ReduceOp, int N = 4, int M = 1>
+__global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<WARP_SIZE>(block);
 
-  size_t out_idx = grid.thread_rank() / WARP_SIZE;
-  if (out_idx >= out_size) {
-    return;
+  const U init = cu::ReduceInit<ReduceOp, T>::value();
+  ReduceOp op;
+
+  T vals[M][N];
+  U accs[M];
+  for (int i = 0; i < M; i++) {
+    accs[i] = init;
   }
 
-  Op op;
+  const size_t start_row =
+      min(n_rows - M, static_cast<size_t>(grid.block_rank() * M));
+  const size_t full_blocks = size / (block.size() * N);
+  const size_t final_offset = full_blocks * (block.size() * N);
+  in += start_row * size;
+  out += start_row;
 
-  U total_val = ReduceInit<Op, T>::value();
-  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
-
-  for (size_t n = warp.thread_rank(); n < args.non_row_reductions;
-       n += WARP_SIZE) {
-    for (int r = 0; r < cuda::ceil_div(args.row_size, N_READS); r++) {
-      U vals[N_READS];
-      cub::LoadDirectBlocked(
-          r,
-          make_cast_iterator<U>(in + loop.location()),
-          vals,
-          args.row_size,
-          ReduceInit<Op, T>::value());
-      total_val = op(total_val, cub::ThreadReduce(vals, op));
+  if (size % N == 0) {
+    for (size_t r = 0; r < full_blocks; r++) {
+      for (int k = 0; k < M; k++) {
+        cub::LoadDirectBlockedVectorized<T, N>(
+            block.thread_rank(),
+            in + k * size + r * (block.size() * N),
+            vals[k]);
+        for (int j = 0; j < N; j++) {
+          accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
+        }
+      }
+    }
+  } else {
+    for (size_t r = 0; r < full_blocks; r++) {
+      for (int k = 0; k < M; k++) {
+        cub::LoadDirectBlocked(
+            block.thread_rank(),
+            in + k * size + r * (block.size() * N),
+            vals[k]);
+        for (int j = 0; j < N; j++) {
+          accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
+        }
+      }
     }
-    loop.next(WARP_SIZE, args.reduce_shape.data(), args.reduce_strides.data());
   }
 
-  total_val = cg::reduce(warp, total_val, op);
+  if (final_offset < size) {
+    for (int k = 0; k < M; k++) {
+      cub::LoadDirectBlocked(
+          block.thread_rank(),
+          in + k * size + final_offset,
+          vals[k],
+          size,
+          cast_to<T>(init));
+      for (int j = 0; j < N; j++) {
+        accs[k] = op(accs[k], cast_to<U>(vals[k][j]));
+      }
+    }
+  }
 
-  if (warp.thread_rank() == 0) {
-    out[out_idx] = total_val;
+  __shared__ U shared_accumulators[32 * M];
+  block_reduce(block, warp, accs, shared_accumulators, op, init);
+
+  if (block.thread_rank() == 0) {
+    if (grid.block_rank() * M + M <= n_rows) {
+      for (int i = 0; i < M; i++) {
+        out[i] = accs[i];
+      }
+    } else {
+      short offset = grid.block_rank() * M + M - n_rows;
+      for (int i = offset; i < M; i++) {
+        out[i] = accs[i];
+      }
+    }
   }
 }
 
@@ -141,55 +166,167 @@ template <
     typename U,
     typename Op,
     int NDIM,
-    int BLOCK_DIM_X,
+    int BLOCK_DIM,
     int N_READS = 4>
 __global__ void row_reduce_looped(
-    const T* in,
+    T* in,
     U* out,
     size_t out_size,
     const __grid_constant__ RowReduceArgs args) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
 
-  size_t out_idx = grid.thread_rank() / BLOCK_DIM_X;
-  if (out_idx >= out_size) {
-    return;
-  }
+  size_t out_idx = grid.block_rank();
 
   Op op;
 
-  U total_val = ReduceInit<Op, T>::value();
+  U total[1];
+  U init = ReduceInit<Op, T>::value();
+  total[0] = init;
   LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
+  size_t full_blocks = args.row_size / (BLOCK_DIM * N_READS);
+  size_t final_offset = full_blocks * BLOCK_DIM * N_READS;
 
   in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
 
   for (size_t n = 0; n < args.non_row_reductions; n++) {
-    for (size_t r = 0; r < cuda::ceil_div(args.row_size, BLOCK_DIM_X * N_READS);
-         r++) {
-      U vals[N_READS];
-      cub::LoadDirectBlocked(
-          r * BLOCK_DIM_X + block.thread_index().x,
-          make_cast_iterator<U>(in + loop.location()),
-          vals,
-          args.row_size,
-          ReduceInit<Op, T>::value());
-      total_val = op(total_val, cub::ThreadReduce(vals, op));
+    for (size_t r = 0; r < full_blocks; r++) {
+      T vals[N_READS];
+      cub::LoadDirectBlockedVectorized<T, N_READS>(
+          block.thread_rank(),
+          in + loop.location() + r * BLOCK_DIM * N_READS,
+          vals);
+      for (int i = 0; i < N_READS; i++) {
+        total[0] = op(total[0], cast_to<U>(vals[i]));
+      }
     }
+    if (final_offset < args.row_size) {
+      T vals[N_READS];
+      cub::LoadDirectBlocked(
+          block.thread_rank(),
+          in + loop.location() + final_offset,
+          vals,
+          args.row_size - final_offset,
+          cast_to<T>(init));
+      for (int i = 0; i < N_READS; i++) {
+        total[0] = op(total[0], cast_to<U>(vals[i]));
+      }
+    }
+    // TODO: Maybe block.sync() here?
     loop.next(args.reduce_shape.data(), args.reduce_strides.data());
   }
 
-  typedef cub::BlockReduce<U, BLOCK_DIM_X> BlockReduceT;
-  __shared__ typename BlockReduceT::TempStorage temp;
-
-  total_val = BlockReduceT(temp).Reduce(total_val, op);
+  __shared__ U shared_accumulators[32];
+  block_reduce(block, warp, total, shared_accumulators, op, init);
 
   if (block.thread_rank() == 0) {
-    out[out_idx] = total_val;
+    out[out_idx] = total[0];
   }
 }
 
 } // namespace cu
 
+void row_reduce_simple(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type,
+    const std::vector<int>& axes,
+    const ReductionPlan& plan) {
+  constexpr int N_READS = 8;
+
+  // Allocate data for the output using in's layout to avoid elem_to_loc in the
+  // kernel.
+  allocate_same_layout(out, in, axes);
+
+  // TODO: If out.size() < 1024 which will be a common case then write this in
+  //       2 passes. Something like 32 * out.size() and then do a warp reduce.
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      using OP = MLX_GET_TYPE(reduce_type_tag);
+      using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      using U = typename cu::ReduceResult<OP, T>::type;
+
+      // Cub doesn't like const pointers for vectorized loads. (sigh)
+      T* indata = const_cast<T*>(in.data<T>());
+
+      // Calculate the grid and block dims
+      size_t reductions = (plan.shape.back() + N_READS - 1) / N_READS;
+      dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
+      int threads = std::min(1024UL, reductions);
+      threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+      dim3 block(threads, 1, 1);
+
+      // Pick the kernel
+      auto kernel = cu::row_reduce_simple<T, U, OP, N_READS>;
+      if (grid.x >= 1024) {
+        grid.x = (grid.x + 1) / 2;
+        kernel = cu::row_reduce_simple<T, U, OP, N_READS, 2>;
+      }
+
+      int size = plan.shape.back();
+      encoder.add_kernel_node(
+          kernel, grid, block, 0, indata, out.data<U>(), out.size(), size);
+    });
+  });
+}
+
+void row_reduce_looped(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type,
+    const std::vector<int>& axes,
+    const ReductionPlan& plan,
+    cu::RowReduceArgs args) {
+  constexpr int N_READS = 8;
+
+  // Allocate data for the output using in's layout to access them as
+  // contiguously as possible.
+  allocate_same_layout(out, in, axes);
+
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    dispatch_reduce_ops(reduce_type, [&](auto reduce_type_tag) {
+      using OP = MLX_GET_TYPE(reduce_type_tag);
+      using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      using U = typename cu::ReduceResult<OP, T>::type;
+      // Cub doesn't like const pointers for vectorized loads. (sigh)
+      T* indata = const_cast<T*>(in.data<T>());
+
+      // Calculate the grid and block dims
+      args.sort_access_pattern(in, axes);
+      dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
+      size_t reductions = (args.row_size + N_READS - 1) / N_READS;
+      int threads = std::min(1024UL, reductions);
+      threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+      dim3 block(threads, 1, 1);
+
+      // Pick the kernel
+      auto kernel = cu::row_reduce_looped<T, U, OP, 1, 32, N_READS>;
+      dispatch_reduce_ndim(args.reduce_ndim, [&](auto reduce_ndim) {
+        dispatch_block_dim(threads, [&](auto threads_constant) {
+          kernel = cu::row_reduce_looped<
+              T,
+              U,
+              OP,
+              reduce_ndim.value,
+              threads_constant.value,
+              N_READS>;
+          block.x = threads_constant.value;
+        });
+      });
+
+      encoder.add_kernel_node(
+          kernel, grid, block, 0, indata, out.data<U>(), out.size(), args);
+    });
+  });
+}
+
 void row_reduce(
     cu::CommandEncoder& encoder,
     const array& in,
@@ -197,54 +334,35 @@ void row_reduce(
     Reduce::ReduceType reduce_type,
     const std::vector<int>& axes,
     const ReductionPlan& plan) {
+  // Current row reduction options
+  //
+  // - row_reduce_simple
+  //
+  //   That means that we are simply reducing across the fastest moving axis.
+  //   We are reducing 1 or 2 rows per threadblock depending on the size of
+  //   output.
+  //
+  // - row_reduce_looped
+  //
+  //   It is a general row reduction. We are computing 1 output per
+  //   threadblock. We read the fastest moving axis vectorized and loop over
+  //   the rest of the axes.
+  //
+  // Notes: We opt to read as much in order as possible and leave
+  //        transpositions as they are (contrary to our Metal backend).
+
+  // Simple row reduce means that we have 1 axis that we are reducing over and
+  // it has stride 1.
+  if (plan.shape.size() == 1) {
+    row_reduce_simple(encoder, in, out, reduce_type, axes, plan);
+    return;
+  }
+
+  // Make the args struct to help route to the best kernel
   cu::RowReduceArgs args(in, plan, axes);
 
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
-      using InType = cuda_type_t<CTYPE>;
-      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
-        using OutType = cu::ReduceResult<OP, InType>::type;
-        MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
-          constexpr size_t N_READS = 4;
-          dim3 out_dims = get_2d_grid_dims(out.shape(), out.strides());
-          dim3 block_dims, num_blocks;
-          auto kernel =
-              cu::row_reduce_small<InType, OutType, OP, NDIM, N_READS>;
-          if (args.row_size <= 64) {
-            if ((args.non_row_reductions < 32 && args.row_size <= 8) ||
-                (args.non_row_reductions <= 8)) {
-              block_dims.x = std::min(out_dims.x, 1024u);
-              num_blocks.x = cuda::ceil_div(out_dims.x, block_dims.x);
-              num_blocks.y = out_dims.y;
-            } else {
-              block_dims.x = WARP_SIZE;
-              num_blocks.y = out_dims.x;
-              num_blocks.z = out_dims.y;
-              kernel =
-                  cu::row_reduce_small_warp<InType, OutType, OP, NDIM, N_READS>;
-            }
-          } else {
-            size_t num_threads = cuda::ceil_div(args.row_size, N_READS);
-            num_threads = cuda::ceil_div(num_threads, WARP_SIZE) * WARP_SIZE;
-            MLX_SWITCH_BLOCK_DIM(num_threads, BLOCK_DIM_X, {
-              num_blocks.y = out_dims.x;
-              num_blocks.z = out_dims.y;
-              block_dims.x = BLOCK_DIM_X;
-              kernel = cu::row_reduce_looped<
-                  InType,
-                  OutType,
-                  OP,
-                  NDIM,
-                  BLOCK_DIM_X,
-                  N_READS>;
-            });
-          }
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in.data<InType>(), out.data<OutType>(), out.size(), args);
-        });
-      });
-    });
-  });
+  // Fallback row reduce
+  row_reduce_looped(encoder, in, out, reduce_type, axes, plan, std::move(args));
 }
 
 } // namespace mlx::core

mlx/backend/cuda/reduce/segmented_reduce.cu

@@ -1,84 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cast_op.cuh"
-#include "mlx/backend/cuda/reduce/reduce.cuh"
-
-#include <thrust/device_ptr.h>
-#include <cub/device/device_reduce.cuh>
-#include <cub/device/device_segmented_reduce.cuh>
-
-namespace mlx::core {
-
-template <typename... Args>
-void cub_all_reduce(cu::CommandEncoder& encoder, Args&&... args) {
-  // Allocate temporary storage.
-  size_t size;
-  CHECK_CUDA_ERROR(cub::DeviceReduce::Reduce(nullptr, size, args...));
-  array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
-  encoder.add_temporary(temp);
-  // Run op.
-  CHECK_CUDA_ERROR(cub::DeviceReduce::Reduce(temp.data<void>(), size, args...));
-}
-
-template <typename... Args>
-void cub_segmented_reduce(cu::CommandEncoder& encoder, Args&&... args) {
-  // Allocate temporary storage.
-  size_t size;
-  CHECK_CUDA_ERROR(cub::DeviceSegmentedReduce::Reduce(nullptr, size, args...));
-  array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
-  encoder.add_temporary(temp);
-  // Run op.
-  CHECK_CUDA_ERROR(
-      cub::DeviceSegmentedReduce::Reduce(temp.data<void>(), size, args...));
-}
-
-struct MultiplyOp {
-  int factor;
-  __device__ int operator()(int i) {
-    return i * factor;
-  }
-};
-
-void segmented_reduce(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    array& out,
-    Reduce::ReduceType reduce_type,
-    const std::vector<int>& axes,
-    const ReductionPlan& plan) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
-      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
-        using InType = cuda_type_t<CTYPE>;
-        using OutType = cu::ReduceResult<OP, InType>::type;
-        auto in_iter = cu::make_cast_iterator<OutType>(
-            thrust::device_pointer_cast(in.data<InType>()));
-        auto out_ptr = thrust::device_pointer_cast(out.data<OutType>());
-        auto init = cu::ReduceInit<OP, InType>::value();
-
-        if (plan.type == ContiguousAllReduce) {
-          cub_all_reduce(
-              encoder, in_iter, out_ptr, in.data_size(), OP(), init, stream);
-        } else if (plan.type == ContiguousReduce) {
-          auto offsets = thrust::make_transform_iterator(
-              thrust::make_counting_iterator(0), MultiplyOp{plan.shape.back()});
-          cub_segmented_reduce(
-              encoder,
-              in_iter,
-              out_ptr,
-              out.size(),
-              offsets,
-              offsets + 1,
-              OP(),
-              init,
-              stream);
-        } else {
-          throw std::runtime_error("Unsupported plan in segmented_reduce.");
-        }
-      });
-    });
-  });
-}
-
-} // namespace mlx::core

mlx/backend/cuda/rms_norm.cu

@@ -74,7 +74,7 @@ __global__ void rms_norm(
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
     T xn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size, 0);
+    cub::LoadDirectBlocked(index, x, xn, axis_size, cast_to<T>(0));
     for (int i = 0; i < N_READS; ++i) {
       float t = static_cast<float>(xn[i]);
       normalizer += t * t;
@@ -130,7 +130,7 @@ __global__ void rms_norm_vjp(
     T wn[N_READS] = {};
     T gn[N_READS] = {};
     auto index = r * BLOCK_DIM + block.thread_rank();
-    cub::LoadDirectBlocked(index, x, xn, axis_size, 0);
+    cub::LoadDirectBlocked(index, x, xn, axis_size, cast_to<T>(0));
     cub::LoadDirectBlocked(index, g, gn, axis_size);
     cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
     for (int i = 0; i < N_READS; i++) {
@@ -206,8 +206,7 @@ void RMSNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -224,20 +223,22 @@ void RMSNorm::eval_gpu(
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "rms_norm", CTYPE, {
-      using DataType = cuda_type_t<CTYPE>;
-      constexpr uint32_t N_READS = 4;
-      MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-        auto kernel = cu::rms_norm<DataType, BLOCK_DIM, N_READS>;
-        kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
-            x.data<DataType>(),
-            w.data<DataType>(),
-            out.data<DataType>(),
-            eps_,
-            axis_size,
-            w_stride);
-      });
+  dispatch_float_types(out.dtype(), "rms_norm", [&](auto type_tag) {
+    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>(),
+          eps_,
+          axis_size,
+          w_stride);
     });
   });
 }
@@ -252,20 +253,22 @@ void RMSNormVJP::eval_gpu(
   // Ensure row contiguity. We could relax this step by checking that the array
   // is contiguous (no broadcasts or holes) and that the input strides are the
   // same as the cotangent strides but for now this is simpler.
-  auto check_input = [&s](const array& x) -> std::pair<array, bool> {
+  auto check_input = [&s](const array& x, bool& copied) {
     if (x.flags().row_contiguous) {
-      return {x, false};
+      copied = false;
+      return x;
     }
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
-    copy_gpu(x, x_copy, CopyType::General, s);
-    return {x_copy, true};
+    copied = true;
+    return contiguous_copy_gpu(x, s);
   };
   bool donate_x = inputs[0].is_donatable();
   bool donate_g = inputs[2].is_donatable();
-  auto [x, copied] = check_input(inputs[0]);
+  bool copied;
+  auto x = check_input(inputs[0], copied);
   donate_x |= copied;
   const array& w = inputs[1];
-  auto [g, g_copied] = check_input(inputs[2]);
+  bool g_copied;
+  auto g = check_input(inputs[2], g_copied);
   donate_g |= g_copied;
   array& gx = outputs[0];
   array& gw = outputs[1];
@@ -303,31 +306,38 @@ void RMSNormVJP::eval_gpu(
       encoder.add_temporary(gw_temp);
     }
   }
-  gw.set_data(allocator::malloc(gw.nbytes()));
 
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(g);
   encoder.set_output_array(gx);
   encoder.set_output_array(gw_temp);
-  encoder.launch_kernel([&, x = x, g = g](cudaStream_t stream) {
-    MLX_SWITCH_FLOAT_TYPES_CHECKED(gx.dtype(), "rms_norm_vjp", CTYPE, {
-      using DataType = cuda_type_t<CTYPE>;
+  dispatch_float_types(gx.dtype(), "rms_norm_vjp", [&](auto type_tag) {
+    dispatch_bool(has_w, [&](auto has_w_constant) {
       constexpr int N_READS = 4;
-      MLX_SWITCH_BOOL(has_w, HAS_W, {
-        MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-          auto kernel = cu::rms_norm_vjp<DataType, HAS_W, BLOCK_DIM, N_READS>;
-          kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
-              x.data<DataType>(),
-              w.data<DataType>(),
-              g.data<DataType>(),
-              gx.data<DataType>(),
-              gw_temp.data<DataType>(),
-              eps_,
-              axis_size,
-              w_stride);
-        });
-      });
+      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,
+                block_dim(),
+                N_READS>;
+            encoder.add_kernel_node(
+                kernel,
+                n_rows,
+                block_dim(),
+                0,
+                x.data<DataType>(),
+                w.data<DataType>(),
+                g.data<DataType>(),
+                gx.data<DataType>(),
+                gw_temp.data<DataType>(),
+                eps_,
+                axis_size,
+                w_stride);
+          });
     });
   });
 

mlx/backend/cuda/rope.cu

@@ -308,73 +308,93 @@ void RoPE::eval_gpu(
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(donated ? out : in);
   encoder.set_input_array(offset);
+  if (with_freqs) {
+    encoder.set_input_array(inputs[2]);
+  }
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_FLOAT_TYPES_CHECKED(in.dtype(), "rope", CTYPE, {
-      using DataType = cuda_type_t<CTYPE>;
-      MLX_SWITCH_BOOL(traditional_, TRADITIONAL, {
-        MLX_SWITCH_BOOL(forward_, FORWARD, {
-          if (single && !with_freqs) {
-            auto kernel = cu::rope_single<DataType, TRADITIONAL, FORWARD>;
-            uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
-            auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
-            kernel<<<grid, block, 0, stream>>>(
-                (donated ? out : in).data<DataType>(),
-                out.data<DataType>(),
-                offset.data<int32_t>(),
-                scale_,
-                std::log2(base_),
-                mat_size,
-                dims);
-          } else if (single) {
-            auto kernel = cu::rope_single_freqs<DataType, TRADITIONAL, FORWARD>;
-            uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
-            auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
-            kernel<<<grid, block, 0, stream>>>(
-                (donated ? out : in).data<DataType>(),
-                out.data<DataType>(),
-                offset.data<int32_t>(),
-                inputs[2].data<float>(),
-                scale_,
-                mat_size,
-                dims,
-                inputs[2].strides(0));
-          } else if (with_freqs) {
-            auto kernel = cu::rope_freqs<DataType, TRADITIONAL, FORWARD>;
-            uint3 dims =
-                make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
-            dims.z = (dims.z + 3) / 4;
-            auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
-            kernel<<<grid, block, 0, stream>>>(
-                (donated ? out : in).data<DataType>(),
-                out.data<DataType>(),
-                offset.data<int32_t>(),
-                inputs[2].data<float>(),
-                scale_,
-                std::log2(base_),
-                strides,
-                out_strides,
-                in.size() / mat_size,
-                dims,
-                inputs[2].strides(0));
-          } else {
-            auto kernel = cu::rope<DataType, TRADITIONAL, FORWARD>;
-            uint3 dims =
-                make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
-            dims.z = (dims.z + 3) / 4;
-            auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
-            kernel<<<grid, block, 0, stream>>>(
-                (donated ? out : in).data<DataType>(),
-                out.data<DataType>(),
-                offset.data<int32_t>(),
-                scale_,
-                std::log2(base_),
-                strides,
-                out_strides,
-                in.size() / mat_size,
-                dims);
-          }
-        });
+  dispatch_float_types(out.dtype(), "rope", [&](auto type_tag) {
+    dispatch_bool(traditional_, [&](auto traditional) {
+      dispatch_bool(forward_, [&](auto forward) {
+        using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        if (single && !with_freqs) {
+          auto kernel =
+              cu::rope_single<DataType, traditional.value, forward.value>;
+          uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
+          auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
+          encoder.add_kernel_node(
+              kernel,
+              grid,
+              block,
+              0,
+              (donated ? out : in).data<DataType>(),
+              out.data<DataType>(),
+              offset.data<int32_t>(),
+              scale_,
+              std::log2(base_),
+              mat_size,
+              dims);
+        } else if (single) {
+          auto kernel =
+              cu::rope_single_freqs<DataType, traditional.value, forward.value>;
+          uint2 dims = make_uint2(dims_ / 2, in.size() / mat_size);
+          auto [grid, block] = get_grid_and_block(dims.x, dims.y, 1);
+          encoder.add_kernel_node(
+              kernel,
+              grid,
+              block,
+              0,
+              (donated ? out : in).data<DataType>(),
+              out.data<DataType>(),
+              offset.data<int32_t>(),
+              inputs[2].data<float>(),
+              scale_,
+              mat_size,
+              dims,
+              inputs[2].strides(0));
+        } else if (with_freqs) {
+          auto kernel =
+              cu::rope_freqs<DataType, traditional.value, forward.value>;
+          uint3 dims =
+              make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
+          dims.z = (dims.z + 3) / 4;
+          auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
+          encoder.add_kernel_node(
+              kernel,
+              grid,
+              block,
+              0,
+              (donated ? out : in).data<DataType>(),
+              out.data<DataType>(),
+              offset.data<int32_t>(),
+              inputs[2].data<float>(),
+              scale_,
+              std::log2(base_),
+              strides,
+              out_strides,
+              in.size() / mat_size,
+              dims,
+              inputs[2].strides(0));
+        } else {
+          auto kernel = cu::rope<DataType, traditional.value, forward.value>;
+          uint3 dims =
+              make_uint3(dims_ / 2, in.shape(-2), in.size() / mat_size);
+          dims.z = (dims.z + 3) / 4;
+          auto [grid, block] = get_grid_and_block(dims.x, dims.y, dims.z);
+          encoder.add_kernel_node(
+              kernel,
+              grid,
+              block,
+              0,
+              (donated ? out : in).data<DataType>(),
+              out.data<DataType>(),
+              offset.data<int32_t>(),
+              scale_,
+              std::log2(base_),
+              strides,
+              out_strides,
+              in.size() / mat_size,
+              dims);
+        }
       });
     });
   });

mlx/backend/cuda/scan.cu

@@ -0,0 +1,467 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/reduce/reduce_ops.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/scan.h>
+#include <nvtx3/nvtx3.hpp>
+
+#include <cassert>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename T>
+struct ScanResult {
+  using type = T;
+};
+
+template <>
+struct ScanResult<Sum, bool> {
+  using type = int32_t;
+};
+
+template <typename T>
+struct ReduceInit<LogAddExp, T> {
+  static constexpr __host__ __device__ T value() {
+    return Limits<T>::min();
+  }
+};
+
+template <bool reverse, typename T, typename U, int N_READS>
+inline __device__ void
+load_values(int index, const T* in, U (&values)[N_READS], int size, U init) {
+  int remaining = size - index * N_READS;
+  if constexpr (reverse) {
+    in += remaining - N_READS;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        values[N_READS - i - 1] =
+            (N_READS - i - 1 < remaining) ? cast_to<U>(in[i]) : init;
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        values[N_READS - i - 1] = cast_to<U>(in[i]);
+      }
+    }
+  } else {
+    in += index * N_READS;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        values[i] = (i < remaining) ? cast_to<U>(in[i]) : init;
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        values[i] = cast_to<U>(in[i]);
+      }
+    }
+  }
+}
+
+template <bool reverse, int offset, typename T, int N_READS>
+inline __device__ void
+store_values(int index, T* out, T (&values)[N_READS], int size) {
+  int start = index * N_READS + offset;
+  int remaining = size - start;
+  if constexpr (reverse) {
+    out += remaining - N_READS;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        if (N_READS - i - 1 < remaining) {
+          out[i] = values[N_READS - i - 1];
+        }
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        out[i] = values[N_READS - i - 1];
+      }
+    }
+  } else {
+    out += start;
+    if (remaining < N_READS) {
+      for (int i = 0; i < N_READS; ++i) {
+        if (i < remaining) {
+          out[i] = values[i];
+        }
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        out[i] = values[i];
+      }
+    }
+  }
+}
+
+template <
+    typename T,
+    typename U,
+    typename Op,
+    int N_READS,
+    bool inclusive,
+    bool reverse>
+__global__ void contiguous_scan(const T* in, U* out, int32_t axis_size) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  in += grid.block_rank() * axis_size;
+  out += grid.block_rank() * axis_size;
+
+  __shared__ U warp_sums[WARP_SIZE];
+
+  Op op;
+  U init = ReduceInit<Op, T>::value();
+  U prefix = init;
+
+  // Scan per block.
+  for (int r = 0; r < cuda::ceil_div(axis_size, block.size() * N_READS); ++r) {
+    int32_t index = r * block.size() + block.thread_rank();
+    U values[N_READS];
+    load_values<reverse>(index, in, values, axis_size, init);
+
+    // Compute an inclusive scan per thread.
+    for (int i = 1; i < N_READS; ++i) {
+      values[i] = op(values[i], values[i - 1]);
+    }
+
+    // Compute exclusive scan of thread sums.
+    U prev_thread_sum = cg::exclusive_scan(warp, values[N_READS - 1], op);
+    if (warp.thread_rank() == 0) {
+      prev_thread_sum = init;
+    }
+
+    // Write wrap's sum to shared memory.
+    if (warp.thread_rank() == WARP_SIZE - 1) {
+      warp_sums[warp.meta_group_rank()] =
+          op(prev_thread_sum, values[N_READS - 1]);
+    }
+    block.sync();
+
+    // Compute exclusive scan of warp sums.
+    if (warp.meta_group_rank() == 0) {
+      U prev_warp_sum =
+          cg::exclusive_scan(warp, warp_sums[warp.thread_rank()], op);
+      if (warp.thread_rank() == 0) {
+        prev_warp_sum = init;
+      }
+      warp_sums[warp.thread_rank()] = prev_warp_sum;
+    }
+    block.sync();
+
+    // Compute the output.
+    for (int i = 0; i < N_READS; ++i) {
+      values[i] = op(values[i], prefix);
+      values[i] = op(values[i], warp_sums[warp.meta_group_rank()]);
+      values[i] = op(values[i], prev_thread_sum);
+    }
+
+    // Write the values.
+    if (inclusive) {
+      store_values<reverse, 0>(index, out, values, axis_size);
+    } else {
+      store_values<reverse, 1>(index, out, values, axis_size);
+      if (reverse) {
+        if (block.thread_rank() == 0 && index == 0) {
+          out[axis_size - 1] = init;
+        }
+      } else {
+        if (block.thread_rank() == 0 && index == 0) {
+          out[0] = init;
+        }
+      }
+    }
+    block.sync();
+
+    // Share the prefix.
+    if ((warp.meta_group_rank() == warp.meta_group_size() - 1) &&
+        (warp.thread_rank() == WARP_SIZE - 1)) {
+      warp_sums[0] = values[N_READS - 1];
+    }
+    block.sync();
+    prefix = warp_sums[0];
+  }
+}
+
+template <
+    typename T,
+    typename U,
+    typename Op,
+    int N_READS,
+    int BM,
+    int BN,
+    bool inclusive,
+    bool reverse>
+__global__ void strided_scan(
+    const T* in,
+    U* out,
+    int32_t axis_size,
+    int64_t stride,
+    int64_t stride_blocks) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  constexpr int BN_pad = WARP_SIZE + 16 / sizeof(U);
+  constexpr int n_warps = BN / N_READS;
+  constexpr int n_scans = BN / n_warps;
+
+  __shared__ U read_buffer[BM * BN_pad];
+
+  Op op;
+  U init = ReduceInit<Op, T>::value();
+  U values[n_scans];
+  U prefix[n_scans];
+  for (int i = 0; i < n_scans; ++i) {
+    prefix[i] = init;
+  }
+
+  // Compute offsets.
+  int64_t offset = (grid.block_rank() / stride_blocks) * axis_size * stride;
+  int64_t global_index_x = (grid.block_rank() % stride_blocks) * BN;
+  uint read_offset_y = (block.thread_rank() * N_READS) / BN;
+  uint read_offset_x = (block.thread_rank() * N_READS) % BN;
+  uint scan_offset_y = warp.thread_rank();
+  uint scan_offset_x = warp.meta_group_rank() * n_scans;
+
+  uint stride_limit = stride - global_index_x;
+  in += offset + global_index_x + read_offset_x;
+  out += offset + global_index_x + read_offset_x;
+  U* read_into = read_buffer + read_offset_y * BN_pad + read_offset_x;
+  U* read_from = read_buffer + scan_offset_y * BN_pad + scan_offset_x;
+
+  for (uint j = 0; j < axis_size; j += BM) {
+    // Calculate the indices for the current thread.
+    uint index_y = j + read_offset_y;
+    uint check_index_y = index_y;
+    if (reverse) {
+      index_y = axis_size - 1 - index_y;
+    }
+
+    // Read in SM.
+    if (check_index_y < axis_size && (read_offset_x + N_READS) < stride_limit) {
+      for (int i = 0; i < N_READS; ++i) {
+        read_into[i] = in[index_y * stride + i];
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        if (check_index_y < axis_size && (read_offset_x + i) < stride_limit) {
+          read_into[i] = in[index_y * stride + i];
+        } else {
+          read_into[i] = init;
+        }
+      }
+    }
+    block.sync();
+
+    // Read strided into registers.
+    for (int i = 0; i < n_scans; ++i) {
+      values[i] = read_from[i];
+    }
+
+    // Perform the scan.
+    for (int i = 0; i < n_scans; ++i) {
+      values[i] = cg::inclusive_scan(warp, values[i], op);
+      values[i] = op(values[i], prefix[i]);
+      prefix[i] = warp.shfl(values[i], WARP_SIZE - 1);
+    }
+
+    // Write to SM.
+    for (int i = 0; i < n_scans; ++i) {
+      read_from[i] = values[i];
+    }
+    block.sync();
+
+    // Write to device memory.
+    if (!inclusive) {
+      if (check_index_y == 0) {
+        if ((read_offset_x + N_READS) < stride_limit) {
+          for (int i = 0; i < N_READS; ++i) {
+            out[index_y * stride + i] = init;
+          }
+        } else {
+          for (int i = 0; i < N_READS; ++i) {
+            if ((read_offset_x + i) < stride_limit) {
+              out[index_y * stride + i] = init;
+            }
+          }
+        }
+      }
+      if (reverse) {
+        index_y -= 1;
+        check_index_y += 1;
+      } else {
+        index_y += 1;
+        check_index_y += 1;
+      }
+    }
+    if (check_index_y < axis_size && (read_offset_x + N_READS) < stride_limit) {
+      for (int i = 0; i < N_READS; ++i) {
+        out[index_y * stride + i] = read_into[i];
+      }
+    } else {
+      for (int i = 0; i < N_READS; ++i) {
+        if (check_index_y < axis_size && (read_offset_x + i) < stride_limit) {
+          out[index_y * stride + i] = read_into[i];
+        }
+      }
+    }
+  }
+}
+
+} // namespace cu
+
+template <typename F>
+void dispatch_scan_ops(Scan::ReduceType scan_op, F&& f) {
+  if (scan_op == Scan::ReduceType::Max) {
+    f(type_identity<cu::Max>{});
+  } else if (scan_op == Scan::ReduceType::Min) {
+    f(type_identity<cu::Min>{});
+  } else if (scan_op == Scan::ReduceType::Sum) {
+    f(type_identity<cu::Sum>{});
+  } else if (scan_op == Scan::ReduceType::Prod) {
+    f(type_identity<cu::Prod>{});
+  } else if (scan_op == Scan::ReduceType::LogAddExp) {
+    f(type_identity<cu::LogAddExp>{});
+  } else {
+    throw std::invalid_argument("Unknown reduce type.");
+  }
+}
+
+template <typename Op>
+const char* op_to_string() {
+  if (cuda::std::is_same_v<Op, cu::Max>) {
+    return "Max";
+  } else if (cuda::std::is_same_v<Op, cu::Min>) {
+    return "Min";
+  } else if (cuda::std::is_same_v<Op, cu::Sum>) {
+    return "Sum";
+  } else if (cuda::std::is_same_v<Op, cu::Prod>) {
+    return "Prod";
+  } else if (cuda::std::is_same_v<Op, cu::LogAddExp>) {
+    return "LogAddExp";
+  } else {
+    throw std::invalid_argument("Unknown op.");
+  }
+}
+
+template <typename Op, typename T>
+constexpr bool supports_scan_op() {
+  if constexpr (cuda::std::is_same_v<Op, LogAddExp>) {
+    return is_inexact_v<T>;
+  } else {
+    return true;
+  }
+}
+
+void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Scan::eval_gpu");
+  assert(inputs.size() == 1);
+  auto in = inputs[0];
+  auto& s = stream();
+
+  if (in.flags().contiguous && in.strides()[axis_] != 0) {
+    if (in.is_donatable() && in.itemsize() == out.itemsize()) {
+      out.copy_shared_buffer(in);
+    } else {
+      out.set_data(
+          allocator::malloc(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+    }
+  } else {
+    in = contiguous_copy_gpu(in, s);
+    out.copy_shared_buffer(in);
+  }
+
+  constexpr int N_READS = 4;
+  int32_t axis_size = in.shape(axis_);
+  bool contiguous = in.strides()[axis_] == 1;
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    dispatch_scan_ops(reduce_type_, [&](auto scan_op_tag) {
+      using Op = MLX_GET_TYPE(scan_op_tag);
+      if constexpr (supports_scan_op<Op, T>) {
+        using U = typename cu::ScanResult<Op, T>::type;
+        dispatch_bool(inclusive_, [&](auto inclusive) {
+          dispatch_bool(reverse_, [&](auto reverse) {
+            if (contiguous) {
+              auto kernel = cu::contiguous_scan<
+                  T,
+                  U,
+                  Op,
+                  N_READS,
+                  inclusive.value,
+                  reverse.value>;
+              int block_dim = cuda::ceil_div(axis_size, N_READS);
+              block_dim = cuda::ceil_div(block_dim, WARP_SIZE) * WARP_SIZE;
+              block_dim = std::min(block_dim, WARP_SIZE * WARP_SIZE);
+              encoder.add_kernel_node(
+                  kernel,
+                  in.data_size() / axis_size,
+                  block_dim,
+                  0,
+                  in.data<T>(),
+                  out.data<U>(),
+                  axis_size);
+            } else {
+              constexpr int BM = WARP_SIZE;
+              constexpr int BN = WARP_SIZE;
+              auto kernel = cu::strided_scan<
+                  T,
+                  U,
+                  Op,
+                  N_READS,
+                  BM,
+                  BN,
+                  inclusive.value,
+                  reverse.value>;
+              int64_t stride = in.strides()[axis_];
+              int64_t stride_blocks = cuda::ceil_div(stride, BN);
+              dim3 num_blocks = get_2d_grid_dims(
+                  in.shape(), in.strides(), axis_size * stride);
+              if (num_blocks.x * stride_blocks <= UINT32_MAX) {
+                num_blocks.x *= stride_blocks;
+              } else {
+                num_blocks.y *= stride_blocks;
+              }
+              int block_dim = (BN / N_READS) * WARP_SIZE;
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dim,
+                  0,
+                  in.data<T>(),
+                  out.data<U>(),
+                  axis_size,
+                  stride,
+                  stride_blocks);
+            }
+          });
+        });
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do scan op {} on inputs of {} with result of {}.",
+            op_to_string<Op>(),
+            dtype_to_string(in.dtype()),
+            dtype_to_string(out.dtype())));
+      }
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/softmax.cu

@@ -43,7 +43,7 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
   // Thread reduce.
   AccT prevmax;
   AccT maxval = Limits<AccT>::finite_min();
-  AccT normalizer = 0;
+  AccT normalizer = cast_to<AccT>(0);
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
     AccT vals[N_READS];
     cub::LoadDirectBlocked(
@@ -51,7 +51,7 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
         make_cast_iterator<AccT>(in),
         vals,
         axis_size,
-        Limits<AccT>::finite_min());
+        Limits<AccT>::min());
     prevmax = maxval;
     maxval = max_op(maxval, cub::ThreadReduce(vals, max_op));
     // Online normalizer calculation for softmax:
@@ -79,7 +79,7 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
   block.sync();
   maxval = warp.thread_rank() < warp.meta_group_size()
       ? local_max[warp.thread_rank()]
-      : Limits<AccT>::finite_min();
+      : Limits<AccT>::min();
   maxval = cg::reduce(warp, maxval, max_op);
   normalizer = normalizer * softmax_exp(prevmax - maxval);
   if (warp.thread_rank() == 0) {
@@ -125,8 +125,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy_gpu(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_gpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -141,18 +140,22 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_FLOAT_TYPES_CHECKED(out.dtype(), "softmax", CTYPE, {
-      using DataType = cuda_type_t<CTYPE>;
-      constexpr int N_READS = 4;
-      MLX_SWITCH_BLOCK_DIM(cuda::ceil_div(axis_size, N_READS), BLOCK_DIM, {
-        auto kernel = cu::softmax<DataType, DataType, BLOCK_DIM, N_READS>;
-        if (precise) {
-          kernel = cu::softmax<DataType, float, BLOCK_DIM, N_READS>;
-        }
-        kernel<<<n_rows, BLOCK_DIM, 0, stream>>>(
-            in.data<DataType>(), out.data<DataType>(), axis_size);
-      });
+  dispatch_float_types(out.dtype(), "softmax", [&](auto type_tag) {
+    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>;
+      }
+      encoder.add_kernel_node(
+          kernel,
+          n_rows,
+          block_dim(),
+          0,
+          in.data<DataType>(),
+          out.data<DataType>(),
+          axis_size);
     });
   });
 }

mlx/backend/cuda/sort.cu

@@ -50,43 +50,21 @@ array swapaxes_in_eval(const array& in, int axis1, int axis2) {
   return out;
 }
 
-template <typename... Args>
-void segmented_sort_pairs(cu::CommandEncoder& encoder, Args&&... args) {
-  // Allocate temporary storage.
-  size_t size;
-  CHECK_CUDA_ERROR(
-      cub::DeviceSegmentedSort::StableSortPairs(nullptr, size, args...));
-  array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
-  encoder.add_temporary(temp);
-  // Run op.
-  CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
-      temp.data<void>(), size, args...));
-}
+struct OffsetTransform {
+  int nsort;
 
-template <typename... Args>
-void segmented_sort(cu::CommandEncoder& encoder, Args&&... args) {
-  // Allocate temporary storage.
-  size_t size;
-  CHECK_CUDA_ERROR(
-      cub::DeviceSegmentedSort::StableSortKeys(nullptr, size, args...));
-  array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
-  encoder.add_temporary(temp);
-  // Run op.
-  CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
-      temp.data<void>(), size, args...));
-}
+  int __device__ operator()(int i) {
+    return i * nsort;
+  }
+};
 
 void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
   array out = out_;
   auto& encoder = cu::get_command_encoder(s);
-  encoder.set_input_array(in);
-  encoder.set_output_array(out);
-
   if (axis < 0) {
     axis += in.ndim();
   }
   int nsort = in.shape(axis);
-  int nsegments = in.data_size() / nsort;
   int last_dim = in.ndim() - 1;
 
   // If we are not sorting the innermost dimension of a contiguous array,
@@ -94,66 +72,108 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
   bool is_segmented_sort = in.flags().contiguous && in.strides()[axis] == 1;
   if (!is_segmented_sort) {
     array trans = swapaxes_in_eval(in, axis, last_dim);
-    in = array(trans.shape(), trans.dtype(), nullptr, {});
-    copy_gpu(trans, in, CopyType::General, s);
+    in = contiguous_copy_gpu(trans, s);
     encoder.add_temporary(in);
     out = array(allocator::malloc(out.nbytes()), in.shape(), out.dtype());
     encoder.add_temporary(out);
   } else {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(
+        allocator::malloc(in.data_size() * out.itemsize()),
+        in.data_size(),
+        in.strides(),
+        in.flags());
   }
 
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
-      if constexpr (!std::is_same_v<CTYPE, complex64_t>) {
-        using Type = cuda_type_t<CTYPE>;
-        auto offsets = thrust::make_transform_iterator(
-            thrust::make_counting_iterator(0),
-            [nsort] __device__(int i) { return i * nsort; });
-        if (argsort) {
-          // Indices in the sorted dimension.
-          array indices(
-              allocator::malloc(out.nbytes()), in.shape(), out.dtype());
-          encoder.add_temporary(indices);
-          thrust::transform(
-              cu::thrust_policy(stream),
-              thrust::counting_iterator<uint32_t>(0),
-              thrust::counting_iterator<uint32_t>(indices.data_size()),
-              thrust::device_pointer_cast(indices.data<uint32_t>()),
-              ModOp<uint32_t>{static_cast<uint32_t>(nsort)});
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  dispatch_all_types(in.dtype(), [&](auto type_tag) {
+    using CTYPE = MLX_GET_TYPE(type_tag);
+    auto& stream = encoder.stream();
+    if constexpr (!std::is_same_v<CTYPE, complex64_t>) {
+      using Type = cuda_type_t<CTYPE>;
+      auto offsets = thrust::make_transform_iterator(
+          thrust::make_counting_iterator(0), OffsetTransform{nsort});
+      if (argsort) {
+        // Indices in the sorted dimension.
+        array indices(allocator::malloc(out.nbytes()), in.shape(), out.dtype());
+        encoder.add_temporary(indices);
 
-          // In argsort though we don't need the result of sorted values, the
-          // API requires us to provide an array to store it.
-          array discard(allocator::malloc(in.nbytes()), in.shape(), in.dtype());
-          encoder.add_temporary(discard);
+        // In argsort though we don't need the result of sorted values, the
+        // API requires us to provide an array to store it.
+        array discard(allocator::malloc(in.nbytes()), in.shape(), in.dtype());
+        encoder.add_temporary(discard);
 
-          segmented_sort_pairs(
-              encoder,
-              in.data<Type>(),
-              discard.data<Type>(),
-              indices.data<uint32_t>(),
-              out.data<uint32_t>(),
-              in.data_size(),
-              nsegments,
-              offsets,
-              offsets + 1,
-              stream);
-        } else {
-          segmented_sort(
-              encoder,
-              in.data<Type>(),
-              out.data<Type>(),
-              in.data_size(),
-              nsegments,
-              offsets,
-              offsets + 1,
-              stream);
-        }
+        size_t size;
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
+            nullptr,
+            size,
+            in.data<Type>(),
+            discard.data<Type>(),
+            indices.data<uint32_t>(),
+            out.data<uint32_t>(),
+            in.data_size(),
+            in.data_size() / nsort,
+            offsets,
+            offsets + 1,
+            stream));
+
+        array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
+        encoder.add_temporary(temp);
+
+        // Start capturing after allocations
+        auto capture = encoder.capture_context();
+        thrust::transform(
+            cu::thrust_policy(stream),
+            thrust::counting_iterator<uint32_t>(0),
+            thrust::counting_iterator<uint32_t>(indices.data_size()),
+            thrust::device_pointer_cast(indices.data<uint32_t>()),
+            ModOp<uint32_t>{static_cast<uint32_t>(nsort)});
+
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
+            temp.data<void>(),
+            size,
+            in.data<Type>(),
+            discard.data<Type>(),
+            indices.data<uint32_t>(),
+            out.data<uint32_t>(),
+            in.data_size(),
+            in.data_size() / nsort,
+            offsets,
+            offsets + 1,
+            stream));
       } else {
-        throw std::runtime_error(
-            "CUDA backend does not support sorting complex numbers");
+        size_t size;
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
+            nullptr,
+            size,
+            in.data<Type>(),
+            out.data<Type>(),
+            in.data_size(),
+            in.data_size() / nsort,
+            offsets,
+            offsets + 1,
+            stream));
+
+        array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
+        encoder.add_temporary(temp);
+
+        // Start capturing after allocations
+        auto capture = encoder.capture_context();
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
+            temp.data<void>(),
+            size,
+            in.data<Type>(),
+            out.data<Type>(),
+            in.data_size(),
+            in.data_size() / nsort,
+            offsets,
+            offsets + 1,
+            stream));
       }
-    });
+    } else {
+      throw std::runtime_error(
+          "CUDA backend does not support sorting complex numbers");
+    }
   });
 
   if (!is_segmented_sort) {
@@ -177,4 +197,14 @@ void Sort::eval_gpu(const std::vector<array>& inputs, array& out) {
   gpu_sort(stream(), inputs[0], out, axis_, false);
 }
 
+void ArgPartition::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("ArgPartition::eval_gpu");
+  gpu_sort(stream(), inputs[0], out, axis_, true);
+}
+
+void Partition::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Partition::eval_gpu");
+  gpu_sort(stream(), inputs[0], out, axis_, false);
+}
+
 } // namespace mlx::core

mlx/backend/cuda/ternary.cu

@@ -15,12 +15,27 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename Op, typename T, typename IdxT>
+template <typename Op, typename T, typename IdxT, int N_READS>
 __global__ void
 ternary_v(const bool* a, const T* b, const T* c, T* out, IdxT size) {
   IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    out[index] = Op{}(a[index], b[index], c[index]);
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(a[i], b[i], c[i]);
+    }
+  } else {
+    auto a_vec = load_vector<N_READS>(a, index);
+    auto b_vec = load_vector<N_READS>(b, index);
+    auto c_vec = load_vector<N_READS>(c, index);
+
+    AlignedVector<T, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(a_vec.val[i], b_vec.val[i], c_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
   }
 }
 
@@ -91,68 +106,90 @@ void ternary_op_gpu_inplace(
   encoder.set_input_array(b);
   encoder.set_input_array(c);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE, {
-      using DType = cuda_type_t<CTYPE>;
+  dispatch_all_types(out.dtype(), [&](auto type_tag) {
+    using DType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
 
-      auto topt = get_ternary_op_type(a, b, c);
-      if (topt == TernaryOpType::General) {
-        auto [shape, strides] = collapse_contiguous_dims(a, b, c, out);
-        auto& a_strides = strides[0];
-        auto& b_strides = strides[1];
-        auto& c_strides = strides[2];
-        bool large = a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
-            c.data_size() > INT32_MAX || out.data_size() > INT32_MAX;
-        MLX_SWITCH_BOOL(large, LARGE, {
-          using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
-          int ndim = shape.size();
-          if (ndim <= 3) {
-            MLX_SWITCH_1_2_3(ndim, NDIM, {
-              auto kernel = cu::ternary_g_nd<Op, DType, IdxT, NDIM>;
+    auto topt = get_ternary_op_type(a, b, c);
+    if (topt == TernaryOpType::General) {
+      dispatch_bool(
+          a.data_size() > INT32_MAX || b.data_size() > INT32_MAX ||
+              c.data_size() > INT32_MAX || out.data_size() > INT32_MAX,
+          [&](auto large) {
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+            Shape shape;
+            std::vector<Strides> strides;
+            std::tie(shape, strides) = collapse_contiguous_dims(a, b, c, out);
+            auto& a_strides = strides[0];
+            auto& b_strides = strides[1];
+            auto& c_strides = strides[2];
+            int ndim = shape.size();
+            if (ndim <= 3) {
+              dispatch_1_2_3(ndim, [&](auto dims_constant) {
+                auto kernel =
+                    cu::ternary_g_nd<Op, DType, IdxT, dims_constant()>;
+                auto [num_blocks, block_dims] =
+                    get_launch_args(kernel, out, large());
+                encoder.add_kernel_node(
+                    kernel,
+                    num_blocks,
+                    block_dims,
+                    0,
+                    a.data<bool>(),
+                    b.data<DType>(),
+                    c.data<DType>(),
+                    out.data<DType>(),
+                    out.size(),
+                    const_param<dims_constant()>(shape),
+                    const_param<dims_constant()>(a_strides),
+                    const_param<dims_constant()>(b_strides),
+                    const_param<dims_constant()>(c_strides));
+              });
+            } else {
+              auto kernel = cu::ternary_g<Op, DType, IdxT>;
               auto [num_blocks, block_dims] =
-                  get_launch_args(kernel, out, large);
-              kernel<<<num_blocks, block_dims, 0, stream>>>(
+                  get_launch_args(kernel, out, large());
+              encoder.add_kernel_node(
+                  kernel,
+                  num_blocks,
+                  block_dims,
+                  0,
                   a.data<bool>(),
                   b.data<DType>(),
                   c.data<DType>(),
                   out.data<DType>(),
-                  out.size(),
-                  const_param<NDIM>(shape),
-                  const_param<NDIM>(a_strides),
-                  const_param<NDIM>(b_strides),
-                  const_param<NDIM>(c_strides));
-            });
-          } else {
-            auto kernel = cu::ternary_g<Op, DType, IdxT>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-            kernel<<<num_blocks, block_dims, 0, stream>>>(
-                a.data<bool>(),
-                b.data<DType>(),
-                c.data<DType>(),
-                out.data<DType>(),
-                out.data_size(),
-                const_param(shape),
-                const_param(a_strides),
-                const_param(b_strides),
-                const_param(c_strides),
-                ndim);
-          }
-        });
-      } else {
-        MLX_SWITCH_BOOL(out.data_size() > UINT32_MAX, LARGE, {
-          using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
-          auto kernel = cu::ternary_v<Op, DType, IdxT>;
-          auto [num_blocks, block_dims] = get_launch_args(
-              kernel, out.data_size(), out.shape(), out.strides(), LARGE);
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              a.data<bool>(),
-              b.data<DType>(),
-              c.data<DType>(),
-              out.data<DType>(),
-              out.data_size());
-        });
-      }
-    });
+                  out.data_size(),
+                  const_param(shape),
+                  const_param(a_strides),
+                  const_param(b_strides),
+                  const_param(c_strides),
+                  ndim);
+            }
+          });
+    } else {
+      dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
+        using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+        // TODO: Choose optimized value based on type size.
+        constexpr int N_READS = 4;
+        auto kernel = cu::ternary_v<Op, DType, IdxT, N_READS>;
+        auto [num_blocks, block_dims] = get_launch_args(
+            kernel,
+            out.data_size(),
+            out.shape(),
+            out.strides(),
+            large(),
+            N_READS);
+        encoder.add_kernel_node(
+            kernel,
+            num_blocks,
+            block_dims,
+            0,
+            a.data<bool>(),
+            b.data<DType>(),
+            c.data<DType>(),
+            out.data<DType>(),
+            out.data_size());
+      });
+    }
   });
 }
 

mlx/backend/cuda/unary.cu

@@ -2,59 +2,92 @@
 
 #include "mlx/backend/common/unary.h"
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cucomplex_math.cuh"
 #include "mlx/backend/cuda/device/unary_ops.cuh"
 #include "mlx/backend/cuda/iterators/general_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
+#include <cooperative_groups.h>
 #include <nvtx3/nvtx3.hpp>
-#include <thrust/device_ptr.h>
-#include <thrust/transform.h>
 
 namespace mlx::core {
 
 namespace cu {
 
+namespace cg = cooperative_groups;
+
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+__global__ void unary_v(const In* in, Out* out, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+
+  if ((index + 1) * N_READS > size) {
+    for (IdxT i = index * N_READS; i < size; ++i) {
+      out[i] = Op{}(in[i]);
+    }
+  } else {
+    auto in_vec = load_vector<N_READS>(in, index);
+
+    AlignedVector<Out, N_READS> out_vec;
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      out_vec.val[i] = Op{}(in_vec.val[i]);
+    }
+
+    store_vector<N_READS>(out, index, out_vec);
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void unary_g(
+    const In* in,
+    Out* out,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides strides,
+    int ndim) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto idx = elem_to_loc_4d(index, shape.data(), strides.data(), ndim);
+    out[index] = Op{}(in[idx]);
+  }
+}
+
 template <typename Op, typename In, typename Out>
 constexpr bool supports_unary_op() {
   if (std::is_same_v<Op, Abs> || std::is_same_v<Op, Negative> ||
-      std::is_same_v<Op, Sign>) {
+      std::is_same_v<Op, Sign> || std::is_same_v<Op, Square>) {
     return std::is_same_v<In, Out>;
   }
-  if (std::is_same_v<Op, ArcCos> || std::is_same_v<Op, ArcCosh> ||
-      std::is_same_v<Op, ArcSin> || std::is_same_v<Op, ArcSinh> ||
-      std::is_same_v<Op, ArcTan> || std::is_same_v<Op, ArcTanh> ||
-      std::is_same_v<Op, Erf> || std::is_same_v<Op, ErfInv> ||
-      std::is_same_v<Op, Expm1> || std::is_same_v<Op, Sigmoid> ||
-      std::is_same_v<Op, Sqrt> || std::is_same_v<Op, Rsqrt>) {
+  if (std::is_same_v<Op, ArcCosh> || std::is_same_v<Op, ArcSinh> ||
+      std::is_same_v<Op, ArcTanh> || std::is_same_v<Op, Erf> ||
+      std::is_same_v<Op, ErfInv> || std::is_same_v<Op, Expm1> ||
+      std::is_same_v<Op, Sigmoid>) {
     return std::is_same_v<In, Out> && is_floating_v<In>;
   }
-  if (std::is_same_v<Op, Log> || std::is_same_v<Op, Log2> ||
-      std::is_same_v<Op, Log10> || std::is_same_v<Op, Log1p>) {
-    return std::is_same_v<In, Out> && is_inexact_v<In>;
-  }
   if (std::is_same_v<Op, BitwiseInvert>) {
     return std::is_same_v<In, Out> && std::is_integral_v<In> &&
         !std::is_same_v<In, bool>;
   }
-  if (std::is_same_v<Op, Ceil> || std::is_same_v<Op, Floor> ||
-      std::is_same_v<Op, Square>) {
-    return std::is_same_v<In, Out> && !std::is_same_v<In, complex64_t>;
+  if (std::is_same_v<Op, Ceil> || std::is_same_v<Op, Floor>) {
+    return std::is_same_v<In, Out> && !mlx::core::is_complex_v<In>;
   }
   if (std::is_same_v<Op, Conjugate>) {
-    return std::is_same_v<In, Out> && std::is_same_v<In, complex64_t>;
+    return std::is_same_v<In, Out> && mlx::core::is_complex_v<In>;
   }
-  if (std::is_same_v<Op, Cos> || std::is_same_v<Op, Cosh> ||
-      std::is_same_v<Op, Exp> || std::is_same_v<Op, Round> ||
-      std::is_same_v<Op, Sin> || std::is_same_v<Op, Sinh> ||
-      std::is_same_v<Op, Tan> || std::is_same_v<Op, Tanh>) {
-    return std::is_same_v<In, Out> &&
-        (is_floating_v<In> || std::is_same_v<In, complex64_t>);
+  if (std::is_same_v<Op, ArcCos> || std::is_same_v<Op, ArcSin> ||
+      std::is_same_v<Op, ArcTan> || std::is_same_v<Op, Cos> ||
+      std::is_same_v<Op, Cosh> || std::is_same_v<Op, Exp> ||
+      std::is_same_v<Op, Log> || std::is_same_v<Op, Log2> ||
+      std::is_same_v<Op, Log10> || std::is_same_v<Op, Log1p> ||
+      std::is_same_v<Op, Round> || std::is_same_v<Op, Rsqrt> ||
+      std::is_same_v<Op, Sqrt> || std::is_same_v<Op, Sin> ||
+      std::is_same_v<Op, Sinh> || std::is_same_v<Op, Tan> ||
+      std::is_same_v<Op, Tanh>) {
+    return std::is_same_v<In, Out> && is_inexact_v<In>;
   }
   if (std::is_same_v<Op, Imag> || std::is_same_v<Op, Real>) {
-    return std::is_same_v<In, complex64_t> && std::is_same_v<Out, float>;
+    return mlx::core::is_complex_v<In> && std::is_same_v<Out, float>;
   }
   if (std::is_same_v<Op, LogicalNot>) {
     return std::is_same_v<In, Out> && std::is_same_v<In, bool>;
@@ -68,42 +101,76 @@ template <typename Op>
 void unary_op_gpu_inplace(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   auto& in = inputs[0];
   if (in.size() == 0) {
     return;
   }
+  bool contig = in.flags().contiguous;
+  bool large;
+  if (!contig) {
+    large = in.data_size() > INT32_MAX || out.size() > INT32_MAX;
+  } else {
+    large = in.data_size() > UINT32_MAX;
+  }
 
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE_IN, {
-      MLX_SWITCH_ALL_TYPES(out.dtype(), CTYPE_OUT, {
-        if constexpr (cu::supports_unary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+  dispatch_all_types(in.dtype(), [&](auto in_type_tag) {
+    dispatch_all_types(out.dtype(), [&](auto out_type_tag) {
+      using CTYPE_IN = MLX_GET_TYPE(in_type_tag);
+      using CTYPE_OUT = MLX_GET_TYPE(out_type_tag);
+      if constexpr (cu::supports_unary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+        dispatch_bool(large, [&](auto large) {
           using InType = cuda_type_t<CTYPE_IN>;
           using OutType = cuda_type_t<CTYPE_OUT>;
-          auto policy = cu::thrust_policy(stream);
-          auto in_ptr = thrust::device_pointer_cast(in.data<InType>());
-          auto out_ptr = thrust::device_pointer_cast(out.data<OutType>());
-          if (in.flags().contiguous) {
-            thrust::transform(
-                policy, in_ptr, in_ptr + in.data_size(), out_ptr, Op());
+          if (contig) {
+            using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
+            auto kernel = cu::unary_v<Op, InType, OutType, IdxT, N_READS>;
+            auto [num_blocks, block_dims] = get_launch_args(
+                kernel,
+                out.data_size(),
+                out.shape(),
+                out.strides(),
+                large,
+                N_READS);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                in.data<InType>(),
+                out.data<OutType>(),
+                out.data_size());
           } else {
+            using IdxT = std::conditional_t<large(), int64_t, int32_t>;
             auto [shape, strides] = collapse_contiguous_dims(in);
-            auto [in_begin, in_end] = cu::make_general_iterators<int64_t>(
-                in_ptr, in.size(), shape, strides);
-            thrust::transform(policy, in_begin, in_end, out_ptr, Op());
+            auto kernel = cu::unary_g<Op, InType, OutType, IdxT>;
+            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
+            encoder.add_kernel_node(
+                kernel,
+                num_blocks,
+                block_dims,
+                0,
+                in.data<InType>(),
+                out.data<OutType>(),
+                out.data_size(),
+                const_param(shape),
+                const_param(strides),
+                shape.size());
           }
-        } else {
-          throw std::runtime_error(fmt::format(
-              "Can not do unary op {} on input of {} with output of {}.",
-              op,
-              dtype_to_string(in.dtype()),
-              dtype_to_string(out.dtype())));
-        }
-      });
+        });
+      } else {
+        throw std::runtime_error(fmt::format(
+            "Can not do unary op {} on input of {} with output of {}.",
+            op,
+            dtype_to_string(in.dtype()),
+            dtype_to_string(out.dtype())));
+      }
     });
   });
 }
@@ -112,17 +179,17 @@ template <typename Op>
 void unary_op_gpu(
     const std::vector<array>& inputs,
     array& out,
-    const std::string& op,
+    const char* op,
     const Stream& s) {
   set_unary_output_data(inputs[0], out);
   unary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 
-#define UNARY_GPU(func)                                                 \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) {   \
-    nvtx3::scoped_range r(#func "::eval_gpu");                          \
-    auto& s = out.primitive().stream();                                 \
-    unary_op_gpu<cu::func>(inputs, out, get_primitive_string(this), s); \
+#define UNARY_GPU(func)                                               \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    nvtx3::scoped_range r(#func "::eval_gpu");                        \
+    auto& s = out.primitive().stream();                               \
+    unary_op_gpu<cu::func>(inputs, out, name(), s);                   \
   }
 
 UNARY_GPU(Abs)
@@ -158,16 +225,15 @@ UNARY_GPU(Tanh)
 void Log::eval_gpu(const std::vector<array>& inputs, array& out) {
   nvtx3::scoped_range r("Log::eval_gpu");
   auto& s = out.primitive().stream();
-  auto op = get_primitive_string(this);
   switch (base_) {
     case Base::e:
-      unary_op_gpu<cu::Log>(inputs, out, op, s);
+      unary_op_gpu<cu::Log>(inputs, out, name(), s);
       break;
     case Base::two:
-      unary_op_gpu<cu::Log2>(inputs, out, op, s);
+      unary_op_gpu<cu::Log2>(inputs, out, name(), s);
       break;
     case Base::ten:
-      unary_op_gpu<cu::Log10>(inputs, out, op, s);
+      unary_op_gpu<cu::Log10>(inputs, out, name(), s);
       break;
   }
 }
@@ -178,7 +244,7 @@ void Round::eval_gpu(const std::vector<array>& inputs, array& out) {
   const auto& in = inputs[0];
   auto& s = out.primitive().stream();
   if (issubdtype(in.dtype(), inexact)) {
-    unary_op_gpu<cu::Round>(inputs, out, get_primitive_string(this), s);
+    unary_op_gpu<cu::Round>(inputs, out, name(), s);
   } else {
     // No-op integer types
     out.copy_shared_buffer(in);

mlx/backend/cuda/utils.cpp

@@ -24,23 +24,47 @@ void check_cuda_error(const char* name, cudaError_t err) {
   }
 }
 
+void check_cuda_error(const char* name, CUresult err) {
+  if (err != CUDA_SUCCESS) {
+    const char* err_str = "Unknown error";
+    cuGetErrorString(err, &err_str);
+    throw std::runtime_error(fmt::format("{} failed: {}", name, err_str));
+  }
+}
+
 const char* dtype_to_cuda_type(const Dtype& dtype) {
-  if (dtype == float16) {
-    return "__half";
+  switch (dtype) {
+    case bool_:
+      return "bool";
+    case int8:
+      return "int8_t";
+    case int16:
+      return "int16_t";
+    case int32:
+      return "int32_t";
+    case int64:
+      return "int64_t";
+    case uint8:
+      return "uint8_t";
+    case uint16:
+      return "uint16_t";
+    case uint32:
+      return "uint32_t";
+    case uint64:
+      return "uint64_t";
+    case float16:
+      return "__half";
+    case bfloat16:
+      return "__nv_bfloat16";
+    case float32:
+      return "float";
+    case float64:
+      return "double";
+    case complex64:
+      return "complex64_t";
+    default:
+      return "unknown";
   }
-  if (dtype == bfloat16) {
-    return "__nv_bfloat16";
-  }
-  if (dtype == complex64) {
-    return "cuComplex";
-  }
-#define SPECIALIZE_DtypeToString(CPP_TYPE, DTYPE) \
-  if (dtype == DTYPE) {                           \
-    return #CPP_TYPE;                             \
-  }
-  MLX_FORALL_DTYPES(SPECIALIZE_DtypeToString)
-#undef SPECIALIZE_DtypeToString
-  return nullptr;
 }
 
 } // namespace mlx::core

mlx/backend/cuda/utils.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <cuda.h>
 #include <cuda_runtime.h>
 
 namespace mlx::core {
@@ -33,6 +34,7 @@ class CudaStream {
 
 // Throw exception if the cuda API does not succeed.
 void check_cuda_error(const char* name, cudaError_t err);
+void check_cuda_error(const char* name, CUresult err);
 
 // The macro version that prints the command that failed.
 #define CHECK_CUDA_ERROR(cmd) check_cuda_error(#cmd, (cmd))

mlx/backend/cuda/worker.cpp

@@ -80,7 +80,9 @@ void Worker::thread_fn() {
       }
       worker_tasks_.erase(worker_tasks_.begin(), end);
     }
-    for (auto& task : tasks) {
+    // Make sure tasks are cleared before the next wait
+    for (int i = 0; i < tasks.size(); ++i) {
+      auto task = std::move(tasks[i]);
       task();
     }
     worker_event_.wait(batch + 1);