Experimenting with a gemm based on the cuda steel utils

Co-authored-by: Chen-Chen Yeh <ge96noj@mytum.de>

.circleci/config.yml

@@ -7,6 +7,9 @@ parameters:
   nightly_build:
     type: boolean
     default: false
+  test_release:
+    type: boolean
+    default: false
 
 jobs:
   build_documentation:
@@ -78,23 +81,24 @@ jobs:
             export DEBIAN_FRONTEND=noninteractive
             export NEEDRESTART_MODE=a
             sudo apt-get update
-            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
+            curl -LsSf https://astral.sh/uv/install.sh | sh
       - run:
           name: Install Python package
           command: |
-            pip install -e ".[dev]"
+            uv venv
+            uv pip install cmake
+            uv pip install -e ".[dev]" -v
       - run:
           name: Generate package stubs
           command: |
-            echo "stubs"
-            pip install typing_extensions
-            python setup.py generate_stubs
+            uv pip install typing_extensions
+            uv run --no-project setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
+            source .venv/bin/activate
             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 -v 2> >(tee -a stderr.log >&2)
@@ -102,6 +106,7 @@ jobs:
       - run:
           name: Build CPP only
           command: |
+            source .venv/bin/activate
             mkdir -p build && cd build
             cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
             make -j `nproc`
@@ -127,33 +132,30 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            brew install python@3.9
-            brew install openmpi
-            python3.9 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
-            pip install torch
-            pip install tensorflow
-            pip install unittest-xml-reporting
+            HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
+              brew install openmpi uv
       - run:
           name: Install Python package
           command: |
-            source env/bin/activate
+            uv venv --python 3.9
+            uv pip install \
+              nanobind==2.4.0 \
+              cmake \
+              numpy \
+              torch \
+              tensorflow \
+              unittest-xml-reporting
             DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              pip install -e . -v
+              uv pip install -e . -v
       - run:
           name: Generate package stubs
           command: |
-            source env/bin/activate
-            pip install typing_extensions
-            python setup.py generate_stubs
+            uv pip install typing_extensions
+            uv run --no-project setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
             mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
@@ -162,16 +164,17 @@ jobs:
       - run:
           name: Build example extension
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             cd examples/extensions
-            pip install -r requirements.txt
-            python setup.py build_ext -j8
+            uv pip install -r requirements.txt
+            uv run --no-project setup.py build_ext --inplace
+            uv run --no-project python test.py
       - store_test_results:
           path: test-results
       - run:
           name: Build CPP only
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
       - run:
           name: Run CPP tests
@@ -180,7 +183,7 @@ jobs:
       - run:
           name: Build small binary
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             cd build/
             cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
               -DBUILD_SHARED_LIBS=ON \
@@ -192,34 +195,60 @@ jobs:
       - run:
           name: Run Python tests with JIT
           command: |
-            source env/bin/activate
             CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              pip install -e . -v
+              uv pip install -e .
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
-              python -m xmlrunner discover -v python/tests -o test-results/gpu_jit
+              uv run --no-project python -m xmlrunner discover \
+                -v python/tests \
+                -o test-results/gpu_jit
 
   cuda_build_and_test:
+    parameters:
+      image_date:
+        type: string
+        default: "2023.11.1"
     machine:
-      image: linux-cuda-12:2023.11.1
+      image: "linux-cuda-12:<< parameters.image_date >>"
       resource_class: gpu.nvidia.small.gen2
     steps:
       - checkout
+      - restore_cache:
+          keys:
+            - cuda-<< parameters.image_date >>-{{ arch }}-
+      - run:
+          name: Install dependencies
+          command: |
+            sudo apt-get update
+            sudo apt-get install libcudnn9-dev-cuda-12
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
+            sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
+            rm -rf ccache-4.11.3-linux-x86_64
+            curl -LsSf https://astral.sh/uv/install.sh | sh
       - run:
           name: Install Python package
           command: |
-            sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            python3 -m venv env
-            source env/bin/activate
+            uv venv
             CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              pip install -e ".[dev]"
+              uv pip install -e ".[dev]" -v
       - run:
           name: Run Python tests
           command: |
-            source env/bin/activate
+            source .venv/bin/activate
             LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
             LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
+      - run:
+          name: CCache report
+          command: |
+            ccache --show-stats
+            ccache --zero-stats
+            ccache --max-size 400MB
+            ccache --cleanup
+      - save_cache:
+          key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
+          paths:
+            - /home/circleci/.cache/ccache
 
   build_release:
     parameters:
@@ -315,14 +344,10 @@ jobs:
             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
             $PYTHON -m venv env
             source env/bin/activate
             pip install --upgrade pip
@@ -366,22 +391,27 @@ jobs:
         type: string
         default: ""
     machine:
-      image: linux-cuda-12:default
-      resource_class: gpu.nvidia.small.gen2
+      image: ubuntu-2204:current
+      resource_class: large
     steps:
       - checkout
       - run:
           name: Build wheel
           command: |
+            export DEBIAN_FRONTEND=noninteractive
+            export NEEDRESTART_MODE=a
+            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
+            sudo dpkg -i cuda-keyring_1.1-1_all.deb
             sudo apt-get update
+            sudo apt-get install cuda-toolkit-12-9 libcudnn9-dev-cuda-12
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
             sudo apt-get install zip
-            python -m venv env
-            source env/bin/activate
             pip install auditwheel
             pip install patchelf
             pip install build
             pip install twine
+            export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
+            export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
             << parameters.build_env >> MLX_BUILD_STAGE=2 \
               CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               python -m build -w
@@ -392,7 +422,6 @@ jobs:
             - run:
                 name: Upload package
                 command: |
-                  source env/bin/activate
                   twine upload wheelhouse/*.whl
       - store_artifacts:
           path: wheelhouse/
@@ -405,19 +434,24 @@ workflows:
             pattern: "^(?!pull/)[-\\w]+$"
             value: << pipeline.git.branch >>
         - not: << pipeline.parameters.nightly_build >>
+        - not: << pipeline.parameters.test_release >>
     jobs:
       - mac_build_and_test:
           matrix:
             parameters:
               macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test
-      - cuda_build_and_test 
+      - cuda_build_and_test:
+          matrix:
+            parameters:
+              image_date: ["2023.11.1", "2025.05.1"]
       - build_documentation 
 
   build_pypi_release:
     when:
       and:
         - not: << pipeline.parameters.nightly_build >>
+        - not: << pipeline.parameters.test_release >>
     jobs:
       - build_release:
           filters:
@@ -538,6 +572,9 @@ workflows:
           requires: [ hold ]
       - cuda_build_and_test:
           requires: [ hold ]
+          matrix:
+            parameters:
+              image_date: ["2023.11.1", "2025.05.1"]
   nightly_build:
     when:
       and:
@@ -601,3 +638,87 @@ workflows:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
       - build_cuda_release
+
+  build_dev_release:
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.test_release >>
+    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"
+      - build_linux_release:
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              build_env: ["DEV_RELEASE=1"]
+      - build_cuda_release:
+          matrix:
+            parameters:
+              build_env: ["DEV_RELEASE=1"]

CMakeLists.txt

@@ -41,7 +41,9 @@ option(MLX_BUILD_GGUF "Include support for GGUF format" ON)
 option(MLX_BUILD_SAFETENSORS "Include support for safetensors format" ON)
 option(MLX_BUILD_BLAS_FROM_SOURCE "Build OpenBLAS from source code" OFF)
 option(MLX_METAL_JIT "Use JIT compilation for Metal kernels" OFF)
+option(MLX_USE_CCACHE "Use CCache for compilation cache when available" ON)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
+option(USE_SYSTEM_FMT "Use system's provided fmt library" OFF)
 
 # --------------------- Processor tests -------------------------
 message(
@@ -68,6 +70,15 @@ else()
   set(MLX_BUILD_METAL OFF)
 endif()
 
+if(MLX_USE_CCACHE)
+  find_program(CCACHE_PROGRAM ccache)
+  if(CCACHE_PROGRAM)
+    set(CMAKE_C_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
+    set(CMAKE_CXX_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
+    set(CMAKE_CUDA_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
+  endif()
+endif()
+
 # ----------------------------- Lib -----------------------------
 
 include(FetchContent)
@@ -232,12 +243,16 @@ target_include_directories(
 # Do not add mlx_EXPORTS define for shared library.
 set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")
 
-FetchContent_Declare(
-  fmt
-  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
-  GIT_TAG 10.2.1
-  EXCLUDE_FROM_ALL)
-FetchContent_MakeAvailable(fmt)
+if(USE_SYSTEM_FMT)
+  find_package(fmt REQUIRED)
+else()
+  FetchContent_Declare(
+    fmt
+    GIT_REPOSITORY https://github.com/fmtlib/fmt.git
+    GIT_TAG 10.2.1
+    EXCLUDE_FROM_ALL)
+  FetchContent_MakeAvailable(fmt)
+endif()
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:fmt::fmt-header-only>)
 
 if(MLX_BUILD_PYTHON_BINDINGS)

README.md

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

docs/requirements.txt

@@ -1,4 +1,5 @@
 sphinx
 breathe
 sphinx-book-theme
+sphinx-copybutton
 mlx

docs/src/conf.py

@@ -18,6 +18,7 @@ release = version
 # -- General configuration ---------------------------------------------------
 
 extensions = [
+    "sphinx_copybutton",
     "sphinx.ext.autodoc",
     "sphinx.ext.autosummary",
     "sphinx.ext.intersphinx",

docs/src/dev/extensions.rst

@@ -394,14 +394,14 @@ below.
         out.set_data(allocator::malloc(out.nbytes()));
 
         // Resolve name of kernel
-        std::ostringstream kname;
-        kname << "axpby_" << "general_" << type_to_name(out);
+        std::stream kname;
+        kname = "axpby_general_" + type_to_name(out);
 
         // Load the metal library
-        auto lib = d.get_library("mlx_ext");
+        auto lib = d.get_library("mlx_ext", current_binary_dir());
 
         // Make a kernel from this metal library
-        auto kernel = d.get_kernel(kname.str(), lib);
+        auto kernel = d.get_kernel(kname, lib);
 
         // Prepare to encode kernel
         auto& compute_encoder = d.get_command_encoder(s.index);

docs/src/install.rst

@@ -13,7 +13,7 @@ silicon computer is
 
     pip install mlx
 
-To install from PyPI you must meet the following requirements:
+To install from PyPI your system must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
 - Using a native Python >= 3.9
@@ -26,12 +26,21 @@ To install from PyPI you must meet the following requirements:
 CUDA
 ^^^^
 
-MLX has a CUDA backend which you can use on any Linux platform with CUDA 12
-and SM 7.0 (Volta) and up. To install MLX with CUDA support, run:
+MLX has a CUDA backend which you can install with:
 
 .. code-block:: shell
 
-    pip install "mlx[cuda]"
+    pip install mlx[cuda]
+
+To install the CUDA package from PyPi your system must meet the following
+requirements:
+
+- Nvidia architecture >= SM 7.0 (Volta)
+- Nvidia driver >= 550.54.14
+- CUDA toolkit >= 12.0
+- Linux distribution with glibc >= 2.35
+- Python >= 3.9
+
 
 CPU-only (Linux)
 ^^^^^^^^^^^^^^^^
@@ -40,7 +49,14 @@ For a CPU-only version of MLX that runs on Linux use:
 
 .. code-block:: shell
 
-    pip install "mlx[cpu]"
+    pip install mlx[cpu]
+
+To install the CPU-only package from PyPi your system must meet the following
+requirements:
+
+- Linux distribution with glibc >= 2.35
+- Python >= 3.9
+
 
 Troubleshooting
 ^^^^^^^^^^^^^^^

docs/src/python/optimizers.rst

@@ -51,14 +51,14 @@ the saved state. Here's a simple example:
    optimizer.update(model, grads)
 
    # Save the state
-   state = tree_flatten(optimizer.state)
-   mx.save_safetensors("optimizer.safetensors", dict(state))
+   state = tree_flatten(optimizer.state, destination={})
+   mx.save_safetensors("optimizer.safetensors", state)
 
    # Later on, for example when loading from a checkpoint,
    # recreate the optimizer and load the state
    optimizer = optim.Adam(learning_rate=1e-2)
 
-   state = tree_unflatten(list(mx.load("optimizer.safetensors").items()))
+   state = tree_unflatten(mx.load("optimizer.safetensors"))
    optimizer.state = state
 
 Note, not every optimizer configuation parameter is saved in the state. For

docs/src/usage/export.rst

@@ -7,17 +7,17 @@ Exporting Functions
 
 MLX has an API to export and import functions to and from a file. This lets you
 run computations written in one MLX front-end (e.g. Python) in another MLX
-front-end (e.g. C++). 
+front-end (e.g. C++).
 
 This guide walks through the basics of the MLX export API with some examples.
 To see the full list of functions check-out the :ref:`API documentation
 <export>`.
 
-Basics of Exporting 
+Basics of Exporting
 -------------------
 
 Let's start with a simple example:
- 
+
 .. code-block:: python
 
   def fun(x, y):
@@ -67,7 +67,7 @@ specified as variable positional arguments or as a tuple of arrays:
 
   x = mx.array(1.0)
   y = mx.array(1.0)
-   
+
   # Both arguments to fun are positional
   mx.export_function("add.mlxfn", fun, x, y)
 
@@ -133,7 +133,7 @@ parameters are also saved to the ``model.mlxfn`` file.
    For enclosed arrays inside an exported function, be extra careful to ensure
    they are evaluated. The computation graph that gets exported will include
    the computation that produces enclosed inputs.
-  
+
    If the above example was missing ``mx.eval(model.parameters()``, the
    exported function would include the random initialization of the
    :obj:`mlx.nn.Module` parameters.
@@ -150,8 +150,8 @@ parameters, pass them as inputs to the ``call`` wrapper:
      # Set the model's parameters to the input parameters
      model.update(tree_unflatten(list(params.items())))
      return model(x)
- 
-   params = dict(tree_flatten(model.parameters()))
+
+   params = tree_flatten(model.parameters(), destination={})
    mx.export_function("model.mlxfn", call, (mx.zeros(4),), params)
 
 
@@ -169,8 +169,8 @@ to export a function which can be used for inputs with variable shapes:
 
   # Ok
   out, = imported_abs(mx.array(-1.0))
-  
-  # Also ok 
+
+  # Also ok
   out, = imported_abs(mx.array([-1.0, -2.0]))
 
 With ``shapeless=False`` (which is the default), the second call to
@@ -197,7 +197,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
   def fun(x, y=None):
       constant = mx.array(3.0)
       if y is not None:
-        x += y 
+        x += y
       return x + constant
 
   with mx.exporter("fun.mlxfn", fun) as exporter:
@@ -215,7 +215,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
   print(out)
 
 In the above example the function constant data, (i.e. ``constant``), is only
-saved once. 
+saved once.
 
 Transformations with Imported Functions
 ---------------------------------------
@@ -238,7 +238,7 @@ on imported functions just like regular Python functions:
   # Prints: array(1, dtype=float32)
   print(dfdx(x))
 
-  # Compile the imported function 
+  # Compile the imported function
   mx.compile(imported_fun)
   # Prints: array(0, dtype=float32)
   print(compiled_fun(x)[0])
@@ -275,7 +275,7 @@ Import and run the function in C++ with only a few lines of code:
   // Prints: array(2, dtype=float32)
   std::cout << outputs[0] << std::endl;
 
-Imported functions can be transformed in C++ just like in Python. Use 
+Imported functions can be transformed in C++ just like in Python. Use
 ``std::vector<mx::array>`` for positional arguments and ``std::map<std::string,
 mx::array>`` for keyword arguments when calling imported functions in C++.
 

examples/extensions/axpby/axpby.cpp

@@ -1,5 +1,6 @@
 // Copyright © 2023-2025 Apple Inc.
 
+#include <dlfcn.h>
 #include <iostream>
 #include <sstream>
 
@@ -16,6 +17,19 @@
 
 namespace my_ext {
 
+// A helper function to find the location of the current binary on disk.
+// The Metal library ("mlx_ext.mtllib"), should be in the same directory.
+std::string current_binary_dir() {
+  static std::string binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path().string();
+  }();
+  return binary_dir;
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Operation Implementation
 ///////////////////////////////////////////////////////////////////////////////
@@ -167,16 +181,15 @@ void Axpby::eval_gpu(
   }
 
   // Resolve name of kernel (corresponds to axpby.metal)
-  std::ostringstream kname;
-  kname << "axpby_";
-  kname << (contiguous_kernel ? "contiguous_" : "general_");
-  kname << type_to_name(out);
+  std::string kname = "axpby_";
+  kname += (contiguous_kernel ? "contiguous_" : "general_");
+  kname += type_to_name(out);
 
   // Load the metal library
-  auto lib = d.get_library("mlx_ext");
+  auto lib = d.get_library("mlx_ext", current_binary_dir());
 
   // Make a kernel from this metal library
-  auto kernel = d.get_kernel(kname.str(), lib);
+  auto kernel = d.get_kernel(kname, lib);
 
   // Prepare to encode kernel
   auto& compute_encoder = d.get_command_encoder(s.index);

examples/extensions/requirements.txt

@@ -1,4 +1,4 @@
 setuptools>=42
 cmake>=3.25
 mlx>=0.21.0
-nanobind==2.2.0
+nanobind==2.4.0

examples/extensions/test.py

@@ -3,8 +3,10 @@ from mlx_sample_extensions import axpby
 
 a = mx.ones((3, 4))
 b = mx.ones((3, 4))
-c = axpby(a, b, 4.0, 2.0, stream=mx.cpu)
+c_cpu = axpby(a, b, 4.0, 2.0, stream=mx.cpu)
+c_gpu = axpby(a, b, 4.0, 2.0, stream=mx.gpu)
 
-print(f"c shape: {c.shape}")
-print(f"c dtype: {c.dtype}")
-print(f"c correct: {mx.all(c == 6.0).item()}")
+print(f"c shape: {c_cpu.shape}")
+print(f"c dtype: {c_cpu.dtype}")
+print(f"c_cpu correct: {mx.all(c_cpu == 6.0).item()}")
+print(f"c_gpu correct: {mx.all(c_gpu == 6.0).item()}")

mlx/array.h

@@ -10,6 +10,7 @@
 #include "mlx/allocator.h"
 #include "mlx/dtype.h"
 #include "mlx/event.h"
+#include "mlx/small_vector.h"
 
 namespace mlx::core {
 
@@ -18,8 +19,8 @@ class Primitive;
 
 using Deleter = std::function<void(allocator::Buffer)>;
 using ShapeElem = int32_t;
-using Shape = std::vector<ShapeElem>;
-using Strides = std::vector<int64_t>;
+using Shape = SmallVector<ShapeElem>;
+using Strides = SmallVector<int64_t>;
 
 class array {
   /* An array is really a node in a graph. It contains a shared ArrayDesc

mlx/backend/common/utils.cpp

@@ -228,4 +228,31 @@ std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
       std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
 }
 
+array swapaxes_in_eval(const array& x, int axis1, int axis2) {
+  int ndim = x.ndim();
+  if (axis1 < 0) {
+    axis1 += ndim;
+  }
+  if (axis2 < 0) {
+    axis2 += ndim;
+  }
+
+  auto shape = x.shape();
+  std::swap(shape[axis1], shape[axis2]);
+  auto strides = x.strides();
+  std::swap(strides[axis1], strides[axis2]);
+
+  auto [data_size, row_contiguous, col_contiguous] =
+      check_contiguity(shape, strides);
+  bool contiguous = data_size == x.data_size();
+
+  array out(std::move(shape), x.dtype(), nullptr, {});
+  out.copy_shared_buffer(
+      x,
+      std::move(strides),
+      {contiguous, row_contiguous, col_contiguous},
+      x.data_size());
+  return out;
+}
+
 } // namespace mlx::core

mlx/backend/common/utils.h

@@ -196,8 +196,11 @@ void shared_buffer_reshape(
     const Strides& out_strides,
     array& out);
 
+// Like the swapaxes op but safe to call in eval_gpu.
+array swapaxes_in_eval(const array& x, int axis1, int axis2);
+
 template <typename T>
-inline std::vector<T> remove_index(std::vector<T> vec, size_t index) {
+inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
   vec.erase(std::next(vec.begin(), index));
   return vec;
 }

mlx/backend/cpu/compiled.cpp

@@ -288,6 +288,14 @@ void Compiled::eval_cpu(
   auto [contiguous, shape, strides] =
       compiled_collapse_contiguous_dims(inputs, outputs[0], is_constant_);
 
+  // Force allocating shape/strides on heap so we can take their data() first
+  // and then std::move them.
+  // TODO: Refactor code to avoid heap allocation.
+  shape.grow();
+  for (auto& s : strides) {
+    s.grow();
+  }
+
   // Collect function input arguments.
   std::vector<void*> args;
   int strides_index = 1;

mlx/backend/cpu/copy.cpp

@@ -377,4 +377,10 @@ void copy_cpu_inplace(
       });
 }
 
+array contiguous_copy_cpu(const array& arr, Stream stream) {
+  array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+  copy_cpu(arr, arr_copy, CopyType::General, stream);
+  return arr_copy;
+}
+
 } // namespace mlx::core

mlx/backend/cpu/copy.h

@@ -30,4 +30,7 @@ void copy_cpu_inplace(
     const std::optional<array>& dynamic_i_offset = std::nullopt,
     const std::optional<array>& dynamic_o_offset = std::nullopt);
 
+// Return a contiguous array with same shape that copies the data of |arr|.
+array contiguous_copy_cpu(const array& arr, Stream stream);
+
 } // namespace mlx::core

mlx/backend/cpu/distributed.cpp

@@ -13,9 +13,7 @@ std::pair<array, bool> ensure_row_contiguous(const array& arr, Stream stream) {
   if (arr.flags().row_contiguous) {
     return {arr, false};
   } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy_cpu(arr, arr_copy, CopyType::General, stream);
-    return {arr_copy, true};
+    return {contiguous_copy_cpu(arr, stream), true};
   }
 };
 
@@ -34,8 +32,7 @@ void AllReduce::eval_cpu(
       }
       return in;
     } else {
-      array arr_copy(in.shape(), in.dtype(), nullptr, {});
-      copy_cpu(in, arr_copy, CopyType::General, s);
+      array arr_copy = contiguous_copy_cpu(in, s);
       out.copy_shared_buffer(arr_copy);
       return arr_copy;
     }

mlx/backend/cpu/jit_compiler.cpp

@@ -2,6 +2,7 @@
 
 #include "mlx/backend/cpu/jit_compiler.h"
 
+#include <algorithm>
 #include <sstream>
 #include <vector>
 

mlx/backend/cpu/logsumexp.cpp

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

mlx/backend/cpu/masked_mm.cpp

@@ -136,9 +136,8 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
           }
           return std::make_tuple(true, sty, arr, false);
         } else {
-          array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy_cpu(arr, arr_copy, CopyType::General, s);
           int64_t stx = arr.shape(-1);
+          array arr_copy = contiguous_copy_cpu(arr, s);
           return std::make_tuple(false, stx, arr_copy, true);
         }
       };

mlx/backend/cpu/quantized.cpp

@@ -712,9 +712,7 @@ void fast::AffineQuantize::eval_cpu(
     if (arr.flags().row_contiguous) {
       return std::make_pair(arr, false);
     } else {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, arr_copy, CopyType::General, s);
-      return std::make_pair(arr_copy, true);
+      return std::make_pair(contiguous_copy_cpu(arr, s), true);
     }
   };
 

mlx/backend/cpu/reduce.cpp

@@ -491,19 +491,27 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
         switch (in.dtype()) {
           case bool_:
           case uint8:
+            reduce_dispatch_sum_prod<uint8_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint16:
+            reduce_dispatch_sum_prod<uint16_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint32:
+            reduce_dispatch_sum_prod<uint32_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint64:
+            reduce_dispatch_sum_prod<uint64_t>(in, out, reduce_type_, axes_);
+            break;
           case int8:
             reduce_dispatch_sum_prod<int8_t>(in, out, reduce_type_, axes_);
             break;
           case int16:
-          case uint16:
             reduce_dispatch_sum_prod<int16_t>(in, out, reduce_type_, axes_);
             break;
           case int32:
-          case uint32:
             reduce_dispatch_sum_prod<int32_t>(in, out, reduce_type_, axes_);
             break;
           case int64:
-          case uint64:
             reduce_dispatch_sum_prod<int64_t>(in, out, reduce_type_, axes_);
             break;
           case float16:

mlx/backend/cpu/scan.cpp

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

mlx/backend/cpu/softmax.cpp

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

mlx/backend/cpu/sort.cpp

@@ -8,7 +8,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
-
+#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
@@ -333,47 +333,24 @@ void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
+  int axis = axis_;
+  if (axis < 0) {
+    axis += in.ndim();
+  }
+
   // Copy input to output
-  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
+  CopyType 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);
-  encoder.dispatch(
-      [out = array::unsafe_weak_copy(out), axis_ = axis_]() mutable {
-        switch (out.dtype()) {
-          case bool_:
-            return sort<bool>(out, axis_);
-          case uint8:
-            return sort<uint8_t>(out, axis_);
-          case uint16:
-            return sort<uint16_t>(out, axis_);
-          case uint32:
-            return sort<uint32_t>(out, axis_);
-          case uint64:
-            return sort<uint64_t>(out, axis_);
-          case int8:
-            return sort<int8_t>(out, axis_);
-          case int16:
-            return sort<int16_t>(out, axis_);
-          case int32:
-            return sort<int32_t>(out, axis_);
-          case int64:
-            return sort<int64_t>(out, axis_);
-          case float32:
-            return sort<float>(out, axis_);
-          case float64:
-            return sort<double>(out, axis_);
-          case float16:
-            return sort<float16_t>(out, axis_);
-          case bfloat16:
-            return sort<bfloat16_t>(out, axis_);
-          case complex64:
-            return sort<complex64_t>(out, axis_);
-        }
-      });
+  encoder.dispatch([out = array::unsafe_weak_copy(out), axis]() mutable {
+    dispatch_all_types(out.dtype(), [&](auto type_tag) {
+      sort<MLX_GET_TYPE(type_tag)>(out, axis);
+    });
+  });
 }
 
 void ArgPartition::eval_cpu(const std::vector<array>& inputs, array& out) {

mlx/backend/cuda/CMakeLists.txt

@@ -6,6 +6,7 @@
 target_sources(
   mlx
   PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/arange.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
@@ -15,18 +16,22 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_dynamic.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_general_input.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/gemv.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/steel_gemm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/jit_module.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/kernel_utils.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/layer_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/logsumexp.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/random.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/all_reduce.cu
@@ -35,6 +40,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
@@ -42,9 +48,18 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.9.0)
+  target_sources(
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_9.cu)
+else()
+  target_sources(
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_0.cpp)
+endif()
+
 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
 
 # Embed kernel sources in binary for JIT compilation.
@@ -87,11 +102,18 @@ endif()
 target_compile_options(
   mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--Wno-deprecated-gpu-targets>")
 
-# Compute capability 7 is required for synchronization between CPU/GPU with
-# managed memory. TODO: Add more architectures for potential performance gain.
-set(MLX_CUDA_ARCHITECTURES
-    "70;80"
-    CACHE STRING "CUDA architectures")
+# Use stronger binaries compression. This feature was introduced in CUDA 12.8
+# and requires drivers released after CUDA 12.4.
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
+  target_compile_options(
+    mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--compress-mode=size>")
+endif()
+
+# Compute capability >= 7.0 is required for synchronization between CPU/GPU with
+# managed memory.
+if(NOT DEFINED MLX_CUDA_ARCHITECTURES)
+  set(MLX_CUDA_ARCHITECTURES "native")
+endif()
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
                                      "${MLX_CUDA_ARCHITECTURES}")
@@ -123,6 +145,23 @@ target_link_libraries(mlx PRIVATE CUDA::cublasLt)
 # Use NVRTC and driver APIs.
 target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 
+# Use the frontend APIs of cuDNN.
+FetchContent_Declare(
+  cudnn
+  GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
+  GIT_TAG v1.12.1
+  GIT_SHALLOW TRUE
+  EXCLUDE_FROM_ALL)
+set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)
+set(CUDNN_FRONTEND_BUILD_SAMPLES OFF)
+set(CUDNN_FRONTEND_BUILD_TESTS OFF)
+set(CUDNN_FRONTEND_BUILD_PYTHON_BINDINGS OFF)
+FetchContent_MakeAvailable(cudnn)
+target_link_libraries(mlx PRIVATE cudnn_frontend)
+# Link with the actual cuDNN libraries.
+include(${cudnn_frontend_SOURCE_DIR}/cmake/cuDNN.cmake)
+target_link_libraries(mlx PRIVATE CUDNN::cudnn_all)
+
 # Suppress nvcc warnings on MLX headers.
 target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
                                    --diag_suppress=997>)

mlx/backend/cuda/allocator.cpp

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

mlx/backend/cuda/allocator.h

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

mlx/backend/cuda/arange.cu

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

mlx/backend/cuda/arg_reduce.cu

@@ -1,8 +1,8 @@
 // Copyright © 2025 Apple Inc.
+
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/device/fp16_math.cuh"
-#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
@@ -44,8 +44,11 @@ struct ArgMin {
   }
 
   template <int N>
-  __device__ IndexValPair<T>
-  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
+  __device__ IndexValPair<T> reduce_many(
+      IndexValPair<T> best,
+      const AlignedVector<T, N>& vals,
+      uint32_t offset) {
+#pragma unroll
     for (int i = 0; i < N; i++) {
       if (vals[i] < best.val) {
         best.val = vals[i];
@@ -74,8 +77,11 @@ struct ArgMax {
   }
 
   template <int N>
-  __device__ IndexValPair<T>
-  reduce_many(IndexValPair<T> best, T (&vals)[N], uint32_t offset) {
+  __device__ IndexValPair<T> reduce_many(
+      IndexValPair<T> best,
+      const AlignedVector<T, N>& vals,
+      uint32_t offset) {
+#pragma unroll
     for (int i = 0; i < N; i++) {
       if (vals[i] > best.val) {
         best.val = vals[i];
@@ -106,16 +112,15 @@ __global__ void arg_reduce_general(
 
   int64_t in_idx = elem_to_loc(index, shape.data(), in_strides.data(), ndim);
   int64_t out_idx = elem_to_loc(index, shape.data(), out_strides.data(), ndim);
+  in += in_idx;
 
   Op op;
   T init = op.init();
   IndexValPair<T> best{0, init};
 
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
-    T vals[N_READS];
     auto tid = r * BLOCK_DIM + block.thread_index().x;
-    cub::LoadDirectBlocked(
-        tid, strided_iterator(in + in_idx, axis_stride), vals, axis_size, init);
+    auto vals = load_vector<N_READS>(in, tid, axis_size, axis_stride, init);
     best = op.reduce_many(best, vals, tid * N_READS);
   }
 
@@ -166,6 +171,7 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
           kernel,
           num_blocks,
           block_dim(),
+          0,
           in.data<T>(),
           out.data<uint32_t>(),
           out.size(),

mlx/backend/cuda/binary.cu

@@ -28,7 +28,7 @@ __global__ void binary_ss(const In* a, const In* b, Out* out, IdxT size) {
     AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      out_vec.val[i] = Op{}(a[0], b[0]);
+      out_vec[i] = Op{}(a[0], b[0]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -49,7 +49,7 @@ __global__ void binary_sv(const In* a, const In* b, Out* out, IdxT size) {
     AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      out_vec.val[i] = Op{}(a[0], b_vec.val[i]);
+      out_vec[i] = Op{}(a[0], b_vec[i]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -70,7 +70,7 @@ __global__ void binary_vs(const In* a, const In* b, Out* out, IdxT size) {
     AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      out_vec.val[i] = Op{}(a_vec.val[i], b[0]);
+      out_vec[i] = Op{}(a_vec[i], b[0]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -92,7 +92,7 @@ __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
     AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      out_vec.val[i] = Op{}(a_vec.val[i], b_vec.val[i]);
+      out_vec[i] = Op{}(a_vec[i], b_vec[i]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -128,7 +128,7 @@ __global__ void binary_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_4d(
+    auto [a_idx, b_idx] = elem_to_loc(
         index, shape.data(), a_strides.data(), b_strides.data(), ndim);
     out[index] = Op{}(a[a_idx], b[b_idx]);
   }
@@ -211,14 +211,18 @@ void binary_op_gpu_inplace(
                 int ndim = shape.size();
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::
-                        binary_g_nd<Op, InType, OutType, IdxT, dims_constant()>;
                     auto [num_blocks, block_dims] =
-                        get_launch_args(kernel, out, large());
+                        get_launch_args(out, large());
                     encoder.add_kernel_node(
-                        kernel,
+                        cu::binary_g_nd<
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
                         num_blocks,
                         block_dims,
+                        0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out.data<OutType>(),
@@ -228,13 +232,12 @@ void binary_op_gpu_inplace(
                         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());
+                  auto [num_blocks, block_dims] = get_launch_args(out, large());
                   encoder.add_kernel_node(
-                      kernel,
+                      cu::binary_g<Op, InType, OutType, IdxT>,
                       num_blocks,
                       block_dims,
+                      0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out.data<OutType>(),
@@ -248,8 +251,7 @@ void binary_op_gpu_inplace(
         } else {
           dispatch_bool(out.data_size() > UINT32_MAX, [&](auto large) {
             using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            // TODO: Choose optimized value based on type size.
-            constexpr int N_READS = 4;
+            constexpr int N_READS = 16 / sizeof(InType);
             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>;
@@ -259,16 +261,12 @@ void binary_op_gpu_inplace(
               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);
+                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>(),

mlx/backend/cuda/binary_two.cu

@@ -33,8 +33,8 @@ binary_two_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
       auto out = Op{}(a[0], b[0]);
-      out_a_vec.val[i] = out[0];
-      out_b_vec.val[i] = out[1];
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
     }
 
     store_vector<N_READS>(out_a, index, out_a_vec);
@@ -60,9 +60,9 @@ binary_two_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
     AlignedVector<Out, N_READS> out_b_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      auto out = Op{}(a[0], b_vec.val[i]);
-      out_a_vec.val[i] = out[0];
-      out_b_vec.val[i] = out[1];
+      auto out = Op{}(a[0], b_vec[i]);
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
     }
 
     store_vector<N_READS>(out_a, index, out_a_vec);
@@ -88,9 +88,9 @@ binary_two_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
     AlignedVector<Out, N_READS> out_b_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      auto out = Op{}(a_vec.val[i], b[0]);
-      out_a_vec.val[i] = out[0];
-      out_b_vec.val[i] = out[1];
+      auto out = Op{}(a_vec[i], b[0]);
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
     }
 
     store_vector<N_READS>(out_a, index, out_a_vec);
@@ -117,9 +117,9 @@ binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
     AlignedVector<Out, N_READS> out_b_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      auto out = Op{}(a_vec.val[i], b_vec.val[i]);
-      out_a_vec.val[i] = out[0];
-      out_b_vec.val[i] = out[1];
+      auto out = Op{}(a_vec[i], b_vec[i]);
+      out_a_vec[i] = out[0];
+      out_b_vec[i] = out[1];
     }
 
     store_vector<N_READS>(out_a, index, out_a_vec);
@@ -160,7 +160,7 @@ __global__ void binary_two_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_4d(
+    auto [a_idx, b_idx] = elem_to_loc(
         index, shape.data(), a_strides.data(), b_strides.data(), ndim);
     auto out = Op{}(a[a_idx], b[b_idx]);
     out_a[index] = out[0];
@@ -227,18 +227,18 @@ void binary_two_op_gpu_inplace(
                 int ndim = shape.size();
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::binary_two_g_nd<
-                        Op,
-                        InType,
-                        OutType,
-                        IdxT,
-                        dims_constant()>;
                     auto [num_blocks, block_dims] =
-                        get_launch_args(kernel, out_a, large());
+                        get_launch_args(out_a, large());
                     encoder.add_kernel_node(
-                        kernel,
+                        cu::binary_two_g_nd<
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
                         num_blocks,
                         block_dims,
+                        0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out_a.data<OutType>(),
@@ -249,13 +249,13 @@ void binary_two_op_gpu_inplace(
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT>;
                   auto [num_blocks, block_dims] =
-                      get_launch_args(kernel, out_a, large());
+                      get_launch_args(out_a, large());
                   encoder.add_kernel_node(
-                      kernel,
+                      cu::binary_two_g<Op, InType, OutType, IdxT>,
                       num_blocks,
                       block_dims,
+                      0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out_a.data<OutType>(),
@@ -270,8 +270,7 @@ void binary_two_op_gpu_inplace(
         } else {
           dispatch_bool(out_a.data_size() > UINT32_MAX, [&](auto large) {
             using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            // TODO: Choose optimized value based on type size.
-            constexpr int N_READS = 4;
+            constexpr int N_READS = 16 / sizeof(InType);
             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>;
@@ -281,7 +280,6 @@ void binary_two_op_gpu_inplace(
               kernel = cu::binary_two_vv<Op, InType, OutType, IdxT, N_READS>;
             }
             auto [num_blocks, block_dims] = get_launch_args(
-                kernel,
                 out_a.data_size(),
                 out_a.shape(),
                 out_a.strides(),
@@ -291,6 +289,7 @@ void binary_two_op_gpu_inplace(
                 kernel,
                 num_blocks,
                 block_dims,
+                0,
                 a.data<InType>(),
                 b.data<InType>(),
                 out_a.data<OutType>(),

mlx/backend/cuda/compiled.cpp

@@ -104,10 +104,41 @@ struct FusedKernelBuilder {
           "  }\n";
     }
 
+    // Vectorized read loop
+    if (contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        const std::string& xname = namer.get_name(x);
+        std::string type = dtype_to_cuda_type(x.dtype());
+        os += fmt::format(
+            "  auto vec_{0} = load_vector<work_per_thread, {1}>({0} + index, 0, size - index, 0);\n",
+            xname,
+            type);
+      }
+    }
+
+    // Create some space for the outputs
+    for (const auto& x : outputs) {
+      const std::string& xname = namer.get_name(x);
+      std::string type = dtype_to_cuda_type(x.dtype());
+      os += fmt::format(
+          "  AlignedVector<{}, work_per_thread> vec_{};\n", type, xname);
+    }
+
     // Work loop
-    os +=
-        "\n"
-        "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
+    if (!contiguous) {
+      os +=
+          "\n"
+          "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
+    } else {
+      os +=
+          "\n"
+          "  #pragma unroll\n"
+          "  for (int i = 0; i < work_per_thread; i++) {\n";
+    }
 
     // Read inputs.
     for (size_t i = 0; i < inputs.size(); ++i) {
@@ -122,7 +153,7 @@ struct FusedKernelBuilder {
       } else if (is_scalar(x)) {
         value = fmt::format("{}[0]", xname);
       } else if (contiguous) {
-        value = fmt::format("{}[index]", xname);
+        value = fmt::format("vec_{}[i]", xname);
       } else {
         value = fmt::format("{}[{}_idx]", xname, xname);
       }
@@ -150,25 +181,30 @@ struct FusedKernelBuilder {
 
     // Write output.
     for (const auto& x : outputs) {
-      os += fmt::format("    {0}[index] = tmp_{0};\n", namer.get_name(x));
+      os += fmt::format("    vec_{0}[i] = tmp_{0};\n", namer.get_name(x));
     }
 
     // End of work loop
-    os +=
-        "\n"
-        "    index++;\n";
     if (!contiguous) {
+      os += "\n";
       for (size_t i = 0; i < inputs.size(); ++i) {
         const auto& x = inputs[i];
         const std::string& xname = namer.get_name(x);
         if (is_scalar(x) || is_constant(i)) {
           continue;
         }
-        os += "    " + xname + "_idx += " + xname + "_strides[NDIM - 1];\n";
+        os += fmt::format("    {0}_idx += {0}_strides[NDIM - 1];\n", xname);
       }
     }
     os += "  }\n";
 
+    // Store the output to global memory
+    for (const auto& x : outputs) {
+      os += fmt::format(
+          "  store_vector({0} + index, 0, vec_{0}, size - index);\n",
+          namer.get_name(x));
+    }
+
     os += "}\n";
   }
 };
@@ -192,6 +228,15 @@ void Compiled::eval_gpu(
   nvtx3::scoped_range r("Compiled::eval_gpu");
   auto& s = stream();
 
+  // Determine the work per thread for the vectorized reads/writes. We take it
+  // as 16 over the max itemsize for the outputs. Another heuristic could be
+  // over the max itemsize of all arrays.
+  int max_size = 1;
+  for (const auto& x : outputs) {
+    max_size = (max_size > x.itemsize()) ? max_size : x.itemsize();
+  }
+  int work_per_thread = 16 / max_size;
+
   cu::JitModule& mod = cu::get_jit_module(s.device, lib_name(), [&]() {
     // Build source code.
     cu::FusedKernelBuilder builder{
@@ -205,28 +250,23 @@ void Compiled::eval_gpu(
     builder.os += "\n} // namespace mlx::core::cu\n";
     // Build kernel names.
     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::{}_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));
+    kernel_names.push_back(fmt::format(
+        "mlx::core::cu::{}_contiguous<uint32_t, {}>",
+        lib_name(),
+        work_per_thread));
+    kernel_names.push_back(fmt::format(
+        "mlx::core::cu::{}_contiguous<int64_t, {}>",
+        lib_name(),
+        work_per_thread));
+    for (auto wpt : std::array<int, 2>{1, work_per_thread}) {
       for (int i = 1; i <= MAX_NDIM; ++i) {
         kernel_names.push_back(fmt::format(
-            "mlx::core::cu::{}_strided<{}, uint32_t, {}>",
-            lib_name(),
-            i,
-            work_per_thread));
+            "mlx::core::cu::{}_strided<{}, uint32_t, {}>", lib_name(), i, wpt));
         kernel_names.push_back(fmt::format(
-            "mlx::core::cu::{}_strided<{}, int64_t, {}>",
-            lib_name(),
-            i,
-            work_per_thread));
+            "mlx::core::cu::{}_strided<{}, int64_t, {}>", lib_name(), i, wpt));
       }
     }
+
     return std::make_pair(std::move(builder.os), std::move(kernel_names));
   });
 
@@ -269,7 +309,6 @@ void Compiled::eval_gpu(
   }
 
   // Choose work per thread
-  int work_per_thread = 4;
   if (!contiguous && shape.back() % work_per_thread != 0) {
     work_per_thread = 1;
   }
@@ -294,8 +333,8 @@ void Compiled::eval_gpu(
 
   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, args.args());
+      get_launch_args(outputs[0], large, work_per_thread);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 } // namespace mlx::core

mlx/backend/cuda/conv.cpp

@@ -0,0 +1,546 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/config.h"
+#include "mlx/backend/cuda/lru_cache.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+// cudnn_frontend.h redefines this macro.
+#undef CHECK_CUDA_ERROR
+
+#include <cudnn_frontend.h>
+#include <cudnn_frontend_find_plan.h>
+#include <fmt/format.h>
+#include <nvtx3/nvtx3.hpp>
+
+#include <cassert>
+
+namespace mlx::core {
+
+namespace {
+
+// Not all engines support it so can not use this API now.
+#define MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API 0
+
+// Alias for better readability.
+#define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
+#define CONV_BACKWARD_INPUT \
+  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DATA_DESCRIPTOR
+#define CONV_BACKWARD_WEIGHT \
+  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
+
+struct ConvCacheKey {
+  int device_id;
+  cudnnDataType_t cudnn_dtype;
+  std::array<int, MAX_NDIM> input_shape;
+  std::array<int, MAX_NDIM> weight_shape;
+  std::array<int, MAX_NDIM> stride;
+  std::array<int, MAX_NDIM> padding_lo;
+  std::array<int, MAX_NDIM> padding_hi;
+  std::array<int, MAX_NDIM> dilation;
+  int groups;
+  bool flip;
+  uint8_t input_alignment;
+  uint8_t weight_alignment;
+  uint8_t output_alignment;
+};
+
+auto& conv_cache() {
+  static LRUBytesKeyCache<
+      ConvCacheKey,
+      std::pair<cudnnBackendDescriptorType_t, cudnn_frontend::ExecutionPlan>>
+      cache(/* capacity */ 128);
+  return cache;
+}
+
+template <typename T, typename Vec>
+inline SmallVector<T> convert_vector(const Vec& vec) {
+  return SmallVector<T>(vec.begin(), vec.end());
+}
+
+template <typename T, template <typename U> class Vec>
+inline std::array<T, MAX_NDIM> fixed_vector(const Vec<T>& vec) {
+  if (vec.size() > MAX_NDIM) {
+    throw std::runtime_error(
+        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
+  }
+  std::array<T, MAX_NDIM> result = {};
+  std::copy_n(vec.begin(), vec.size(), result.begin());
+  return result;
+}
+
+auto nhwc_to_nchw(const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  shape.insert(shape.begin() + 1, shape.back());
+  shape.erase(shape.end() - 1);
+  auto strides = convert_vector<int64_t>(x.strides());
+  strides.insert(strides.begin() + 1, strides.back());
+  strides.erase(strides.end() - 1);
+  return std::make_tuple(std::move(shape), std::move(strides));
+}
+
+inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return CUDNN_DATA_INT8;
+    case int32:
+      return CUDNN_DATA_INT32;
+    case uint8:
+      return CUDNN_DATA_UINT8;
+    case float16:
+      return CUDNN_DATA_HALF;
+    case bfloat16:
+      return CUDNN_DATA_BFLOAT16;
+    case float32:
+      return CUDNN_DATA_FLOAT;
+    case float64:
+      return CUDNN_DATA_DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+  }
+}
+
+inline uint8_t get_alignment(const array& x) {
+  uint8_t alignment = 1;
+  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
+  for (; alignment < 32; alignment *= 2) {
+    if (address % (alignment * 2)) {
+      return alignment;
+    }
+  }
+  return alignment;
+}
+
+inline cudnn_frontend::Tensor build_tensor(int64_t id, const array& x) {
+  auto [shape, strides] = nhwc_to_nchw(x);
+  return cudnn_frontend::TensorBuilder()
+      .setDim(shape.size(), shape.data())
+      .setStrides(strides.size(), strides.data())
+      .setId(id)
+      .setAlignment(get_alignment(x))
+      .setDataType(dtype_to_cudnn_type(x.dtype()))
+      .build();
+}
+
+cudnn_frontend::EngineConfigList get_engine_configs(
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph,
+    bool use_fallback = false) {
+  cudnn_frontend::GeneratorSource source;
+  if (use_fallback) {
+    source = [&backend_type](cudnn_frontend::OperationGraph& op_graph) {
+      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
+                          .setOperationGraph(op_graph)
+                          .setOperation(backend_type)
+                          .build();
+      return fallback.getFallbackList();
+    };
+  } else {
+    source = [](cudnn_frontend::OperationGraph& op_graph) {
+      auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
+                            .setOperationGraph(op_graph)
+                            .setHeurMode(CUDNN_HEUR_MODE_A)
+                            .build();
+      return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
+    };
+  }
+
+  cudnn_frontend::EngineConfigGenerator generator(1, &source);
+  auto configs = generator.generate_engine_config(op_graph);
+
+  cudnn_frontend::EngineConfigList filtered_configs;
+  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
+    if (cudnn_frontend::hasNumericalNote<
+            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
+      return true;
+    }
+    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
+        dtype == float32 && !env::enable_tf32()) {
+      return true;
+    }
+    return false;
+  });
+  return filtered_configs;
+}
+
+bool execute_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    array& x,
+    array& w,
+    array& y) {
+  int workspace_size = plan.getWorkspaceSize();
+  array workspace(allocator::malloc(workspace_size), {workspace_size}, uint8);
+
+  int64_t uids[3] = {'x', 'w', 'y'};
+  void* data_ptrs[3] = {
+      x.data<void>(),
+      w.data<void>(),
+      y.data<void>(),
+  };
+
+  auto variantPack = cudnn_frontend::VariantPackBuilder()
+                         .setWorkspacePointer(workspace.data<void>())
+                         .setDataPointers(3, data_ptrs)
+                         .setUids(3, uids)
+                         .build();
+
+  auto handle = encoder.device().cudnn_handle();
+  cudnnSetStream(handle, encoder.stream());
+
+#if CUDNN_VERSION >= 90500 && MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API
+  cudaGraph_t graph;
+  cudaGraphCreate(&graph, 0);
+  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
+      &graph, [](cudaGraph_t* p) { cudaGraphDestroy(*p); });
+  if (cudnnBackendPopulateCudaGraph(
+          handle, plan.get_raw_desc(), variantPack.get_raw_desc(), graph) !=
+      CUDNN_STATUS_SUCCESS) {
+    return false;
+  }
+  encoder.add_graph_node(graph);
+#else
+  auto capture = encoder.capture_context();
+  if (cudnnBackendExecute(
+          handle, plan.get_raw_desc(), variantPack.get_raw_desc()) !=
+      CUDNN_STATUS_SUCCESS) {
+    // Discard the captured graph when failed.
+    capture.discard = true;
+    return false;
+  }
+#endif
+
+  encoder.add_temporary(workspace);
+  return true;
+}
+
+bool try_engines(
+    cu::CommandEncoder& encoder,
+    const ConvCacheKey& cache_key,
+    cudnnBackendDescriptorType_t backend_type,
+    cudnn_frontend::EngineConfigList& configs,
+    const std::string& op_graph_tag,
+    array& x,
+    array& w,
+    array& y) {
+  for (auto& config : configs) {
+    try {
+      auto plan = cudnn_frontend::ExecutionPlanBuilder()
+                      .setHandle(encoder.device().cudnn_handle())
+                      .setEngineConfig(config, op_graph_tag)
+                      .build();
+      if (execute_plan(encoder, plan, x, w, y)) {
+        conv_cache().emplace(
+            cache_key, std::make_pair(backend_type, std::move(plan)));
+        return true;
+      }
+    } catch (cudnn_frontend::cudnnException& error) {
+      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
+        throw;
+      }
+    }
+  }
+  return false;
+}
+
+auto get_conv_op_settings(
+    cudnnBackendDescriptorType_t backend_type,
+    array& x,
+    array& w,
+    array& y,
+    const std::vector<int>& kernel_strides,
+    const std::vector<int>& padding_lo_,
+    const std::vector<int>& padding_hi_,
+    const std::vector<int>& kernel_dilation,
+    const std::vector<int>& input_dilation) {
+  auto padding_lo = convert_vector<int64_t>(padding_lo_);
+  auto padding_hi = convert_vector<int64_t>(padding_hi_);
+
+  if (backend_type == CONV_BACKWARD_INPUT) {
+    for (int i = 0; i < padding_lo.size(); ++i) {
+      int wt_size = 1 + kernel_dilation[i] * (w.shape(1 + i) - 1);
+      padding_lo[i] = wt_size - padding_lo[i] - 1;
+      int in_size = 1 + kernel_strides[i] * (x.shape(1 + i) - 1);
+      int out_size = 1 + input_dilation[i] * (y.shape(1 + i) - 1);
+      padding_hi[i] = out_size - in_size + padding_hi[i];
+    }
+    return std::make_tuple(
+        convert_vector<int64_t>(input_dilation),
+        std::move(padding_lo),
+        std::move(padding_hi),
+        convert_vector<int64_t>(kernel_dilation));
+
+  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
+    padding_hi = padding_lo;
+    return std::make_tuple(
+        convert_vector<int64_t>(kernel_dilation),
+        std::move(padding_lo),
+        std::move(padding_hi),
+        convert_vector<int64_t>(kernel_strides));
+
+  } else {
+    return std::make_tuple(
+        convert_vector<int64_t>(kernel_strides),
+        std::move(padding_lo),
+        std::move(padding_hi),
+        convert_vector<int64_t>(kernel_dilation));
+  }
+}
+
+std::optional<cudnn_frontend::OperationGraph> build_op_graph(
+    cu::CommandEncoder& encoder,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    array& x,
+    array& w,
+    array& y,
+    const SmallVector<int64_t>& stride,
+    const SmallVector<int64_t>& padding_lo,
+    const SmallVector<int64_t>& padding_hi,
+    const SmallVector<int64_t>& dilation) {
+  try {
+    auto compute_dtype = (dtype == float16 || dtype == bfloat16)
+        ? CUDNN_DATA_FLOAT
+        : dtype_to_cudnn_type(dtype);
+    auto conv_desc = cudnn_frontend::ConvDescBuilder()
+                         .setDataType(compute_dtype)
+                         .setMathMode(CUDNN_CROSS_CORRELATION)
+                         .setNDims(stride.size())
+                         .setStrides(stride.size(), stride.data())
+                         .setPrePadding(padding_lo.size(), padding_lo.data())
+                         .setPostPadding(padding_hi.size(), padding_hi.data())
+                         .setDilation(dilation.size(), dilation.data())
+                         .build();
+
+    auto op = cudnn_frontend::OperationBuilder(backend_type)
+                  .setxDesc(build_tensor('x', x))
+                  .setwDesc(build_tensor('w', w))
+                  .setyDesc(build_tensor('y', y))
+                  .setcDesc(conv_desc)
+                  .build();
+
+    std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
+    return cudnn_frontend::OperationGraphBuilder()
+        .setHandle(encoder.device().cudnn_handle())
+        .setOperationGraph(ops.size(), ops.data())
+        .build();
+  } catch (cudnn_frontend::cudnnException& error) {
+    if (error.getCudnnStatus() != CUDNN_STATUS_BAD_PARAM) {
+      throw;
+    }
+    return std::nullopt;
+  }
+}
+
+// Do necessary transposes and copies to prepare the inputs and outputs for
+// building the cuDNN conv op. It is safe to be called multiple times in one
+// eval_gpu, with cost of possible redundant copies.
+std::tuple<array, array, array> prepare_args(
+    cu::CommandEncoder& encoder,
+    cudnnBackendDescriptorType_t backend_type,
+    array in,
+    array wt,
+    array out,
+    Stream s) {
+  // Transpose the args depending on the backend type.
+  // TODO: Handle groups.
+  if (backend_type == CONV_BACKWARD_INPUT) {
+    wt = swapaxes_in_eval(wt, 0, -1);
+  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
+    in = swapaxes_in_eval(in, 0, -1);
+    wt = swapaxes_in_eval(wt, 0, -1);
+    // Create a contiguous array that shares the data with |out|, but with dim
+    // C_in and C_out swapped.
+    Shape shape(out.shape());
+    std::swap(shape.front(), shape.back());
+    Strides strides(shape.size(), 1);
+    for (int i = shape.size() - 2; i >= 0; --i) {
+      strides[i] = shape[i + 1] * strides[i + 1];
+    }
+    array intermediate(std::move(shape), out.dtype(), nullptr, {});
+    intermediate.copy_shared_buffer(
+        out, std::move(strides), {true, true, false}, out.data_size());
+    out = intermediate;
+  }
+
+  // cuDNN requires contiguous input.
+  if (!in.flags().row_contiguous) {
+    in = contiguous_copy_gpu(in, s);
+    encoder.add_temporary(in);
+  }
+  if (!wt.flags().row_contiguous) {
+    wt = contiguous_copy_gpu(wt, s);
+    encoder.add_temporary(wt);
+  }
+
+  return {std::move(in), std::move(wt), std::move(out)};
+}
+
+// Get the x/w/y args from the in/wt/out args depending on backend type.
+inline std::tuple<array&, array&, array&> dispatch_args(
+    cudnnBackendDescriptorType_t backend_type,
+    array& in,
+    array& wt,
+    array& out) {
+  switch (backend_type) {
+    case CONV_BACKWARD_INPUT:
+      return {out, wt, in};
+    case CONV_BACKWARD_WEIGHT:
+      return {in, out, wt};
+    default:
+      return {in, wt, out};
+  }
+}
+
+// Register inputs and outputs before actually running conv op. Can only be
+// called once per eval_gpu.
+void register_args(
+    cu::CommandEncoder& encoder,
+    cudnnBackendDescriptorType_t backend_type,
+    array& in,
+    array& wt,
+    array& intermediate_out,
+    array& final_out) {
+  encoder.set_input_array(in);
+  encoder.set_input_array(wt);
+  encoder.set_output_array(final_out);
+
+  if (backend_type == CONV_BACKWARD_WEIGHT) {
+    // Turn |out| into a strided array, which will have C_in and C_out swapped
+    // in vjp and the final |grad_weight| will then be contiguous.
+    Strides strides = intermediate_out.strides();
+    std::swap(strides.front(), strides.back());
+    final_out.copy_shared_buffer(
+        intermediate_out,
+        std::move(strides),
+        {false, false, false},
+        intermediate_out.data_size());
+  }
+}
+
+} // namespace
+
+void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
+  nvtx3::scoped_range r("Convolution::eval_gpu");
+  if (out_.size() == 0) {
+    return;
+  }
+
+  assert(inputs.size() == 2);
+  array in = inputs[0];
+  array wt = inputs[1];
+  array out = out_;
+  out.set_data(allocator::malloc(out.nbytes()));
+  Dtype dtype = out.dtype();
+
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  // Search cache.
+  ConvCacheKey cache_key{
+      encoder.device().cuda_device(),
+      dtype_to_cudnn_type(dtype),
+      fixed_vector(in.shape()),
+      fixed_vector(wt.shape()),
+      fixed_vector(kernel_strides_),
+      fixed_vector(padding_lo_),
+      fixed_vector(padding_hi_),
+      fixed_vector(kernel_dilation_),
+      groups_,
+      flip_,
+      get_alignment(in),
+      get_alignment(wt),
+      get_alignment(out)};
+  if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
+    auto& [backend_type, plan] = it->second;
+    std::tie(in, wt, out) = prepare_args(encoder, backend_type, in, wt, out, s);
+    register_args(encoder, backend_type, in, wt, out, out_);
+    auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
+    if (!execute_plan(encoder, plan, x, w, y)) {
+      throw std::runtime_error("[conv] Cached plan failed to execute.");
+    }
+    return;
+  }
+
+  // There is no reliable way to deduce the proper cuDNN backend for the
+  // convolution, so we make a best guess and then try.
+  SmallVector<cudnnBackendDescriptorType_t, 2> try_backends;
+  if (flip_) {
+    // When weight is flipped, we assume it is backward input convolution.
+    try_backends.push_back(CONV_BACKWARD_INPUT);
+  } else {
+    // Otherwise it could be backward weight convolution or forward convolution,
+    // mathematically there is no difference so we have to use heuristics.
+    // Empirically backward convolutions have large kernel dimensions, and
+    // usually have |in| and |wt| transposed.
+    if (!in.flags().row_contiguous && !wt.flags().row_contiguous &&
+        wt.shape(2) > out.shape(2)) {
+      try_backends = {CONV_BACKWARD_WEIGHT, CONV_FORWARD};
+    } else {
+      try_backends = {CONV_FORWARD, CONV_BACKWARD_WEIGHT};
+    }
+  }
+
+  // Try to build op graph.
+  cudnnBackendDescriptorType_t backend_type;
+  std::optional<cudnn_frontend::OperationGraph> op_graph;
+  for (auto try_backend : try_backends) {
+    auto [in_copy, wt_copy, out_copy] =
+        prepare_args(encoder, try_backend, in, wt, out, s);
+    auto [x, w, y] = dispatch_args(try_backend, in_copy, wt_copy, out_copy);
+    auto [stride, padding_lo, padding_hi, dilation] = get_conv_op_settings(
+        try_backend,
+        x,
+        w,
+        y,
+        kernel_strides_,
+        padding_lo_,
+        padding_hi_,
+        kernel_dilation_,
+        input_dilation_);
+    op_graph = build_op_graph(
+        encoder,
+        try_backend,
+        dtype,
+        x,
+        w,
+        y,
+        stride,
+        padding_lo,
+        padding_hi,
+        dilation);
+    if (op_graph) {
+      backend_type = try_backend;
+      in = std::move(in_copy);
+      wt = std::move(wt_copy);
+      out = std::move(out_copy);
+      break;
+    }
+  }
+  if (!op_graph) {
+    throw std::runtime_error("[conv] Can not build op graph.");
+  }
+
+  // Get ready to execute the graph.
+  register_args(encoder, backend_type, in, wt, out, out_);
+
+  // Try to run plans based on heuristics.
+  auto configs = get_engine_configs(backend_type, dtype, *op_graph);
+  auto tag = op_graph->getTag();
+  auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
+  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
+    return;
+  }
+  // Then try fallback plans.
+  configs = get_engine_configs(backend_type, dtype, *op_graph);
+  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
+    return;
+  }
+  throw std::runtime_error("[conv] Unable to find a working engine.");
+}
+
+} // namespace mlx::core

mlx/backend/cuda/copy/copy_contiguous.cu

@@ -22,7 +22,7 @@ __global__ void copy_s(const In* in, Out* out, IdxT size) {
     AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      out_vec.val[i] = cast_to<Out>(in[0]);
+      out_vec[i] = cast_to<Out>(in[0]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -43,7 +43,7 @@ __global__ void copy_v(const In* in, Out* out, IdxT size) {
     AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      out_vec.val[i] = cast_to<Out>(in_vec.val[i]);
+      out_vec[i] = cast_to<Out>(in_vec[i]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -65,23 +65,18 @@ void copy_contiguous(
         using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
         using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
         using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-        // TODO: Choose optimized value based on type size.
-        constexpr int N_READS = 4;
+        constexpr int N_READS = 16 / sizeof(InType);
         auto kernel = cu::copy_s<InType, OutType, IdxT, N_READS>;
         if (ctype == CopyType::Vector) {
           kernel = cu::copy_v<InType, OutType, IdxT, N_READS>;
         }
         auto [num_blocks, block_dims] = get_launch_args(
-            kernel,
-            out.data_size(),
-            out.shape(),
-            out.strides(),
-            large(),
-            N_READS);
+            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

@@ -37,7 +37,7 @@ __global__ void copy_gg(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_4d(
+    auto [idx_in, idx_out] = elem_to_loc(
         index, shape.data(), strides_in.data(), strides_out.data(), ndim);
     out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
   }
@@ -71,14 +71,13 @@ void copy_general(
               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());
+                auto [num_blocks, block_dims] =
+                    get_launch_args(data_size, shape, out.strides(), large());
                 encoder.add_kernel_node(
-                    kernel,
+                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>,
                     num_blocks,
                     block_dims,
+                    0,
                     in_ptr,
                     out_ptr,
                     data_size,
@@ -87,13 +86,13 @@ void copy_general(
                     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());
+              auto [num_blocks, block_dims] =
+                  get_launch_args(data_size, shape, out.strides(), large());
               encoder.add_kernel_node(
-                  kernel,
+                  cu::copy_gg<InType, OutType, IdxT>,
                   num_blocks,
                   block_dims,
+                  0,
                   in_ptr,
                   out_ptr,
                   data_size,

mlx/backend/cuda/copy/copy_general_dynamic.cu

@@ -41,7 +41,7 @@ __global__ void copy_gg_dynamic(
     const int64_t* offset_out) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_4d(
+    auto [idx_in, idx_out] = elem_to_loc(
         index, shape.data(), strides_in.data(), strides_out.data(), ndim);
     out[idx_out + *offset_out] = CastOp<In, Out>{}(in[idx_in + *offset_in]);
   }
@@ -74,14 +74,16 @@ void copy_general_dynamic(
             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());
+                auto [num_blocks, block_dims] = get_launch_args(out, large());
                 encoder.add_kernel_node(
-                    kernel,
+                    cu::copy_gg_dynamic_nd<
+                        InType,
+                        OutType,
+                        IdxT,
+                        dims_constant()>,
                     num_blocks,
                     block_dims,
+                    0,
                     in_ptr,
                     out_ptr,
                     out.size(),
@@ -92,13 +94,12 @@ void copy_general_dynamic(
                     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());
+              auto [num_blocks, block_dims] = get_launch_args(out, large());
               encoder.add_kernel_node(
-                  kernel,
+                  cu::copy_gg_dynamic<InType, OutType, IdxT>,
                   num_blocks,
                   block_dims,
+                  0,
                   in_ptr,
                   out_ptr,
                   out.size(),

mlx/backend/cuda/copy/copy_general_input.cu

@@ -34,7 +34,7 @@ __global__ void copy_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    IdxT idx_in = elem_to_loc_4d(index, shape.data(), strides_in.data(), ndim);
+    IdxT idx_in = elem_to_loc(index, shape.data(), strides_in.data(), ndim);
     out[index] = CastOp<In, Out>{}(in[idx_in]);
   }
 }
@@ -63,14 +63,12 @@ void copy_general_input(
             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());
+                auto [num_blocks, block_dims] = get_launch_args(out, large());
                 encoder.add_kernel_node(
-                    kernel,
+                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>,
                     num_blocks,
                     block_dims,
+                    0,
                     in_ptr,
                     out_ptr,
                     out.size(),
@@ -78,13 +76,12 @@ void copy_general_input(
                     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());
+              auto [num_blocks, block_dims] = get_launch_args(out, large());
               encoder.add_kernel_node(
-                  kernel,
+                  cu::copy_g<InType, OutType, IdxT>,
                   num_blocks,
                   block_dims,
+                  0,
                   in_ptr,
                   out_ptr,
                   out.size(),

mlx/backend/cuda/device.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/jit_module.h"
 #include "mlx/backend/cuda/worker.h"
 #include "mlx/utils.h"
 
@@ -9,12 +10,23 @@
 #include <future>
 #include <unordered_set>
 
-namespace mlx::core {
+namespace mlx::core::cu {
+
+namespace {
 
 // Can be tuned with MLX_MAX_OPS_PER_BUFFER
 // This should be less than 255
 constexpr int default_max_nodes_per_graph = 20;
 
+#define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
+
+void check_cudnn_error(const char* name, cudnnStatus_t err) {
+  if (err != CUDNN_STATUS_SUCCESS) {
+    throw std::runtime_error(
+        fmt::format("{} failed: {}.", name, cudnnGetErrorString(err)));
+  }
+}
+
 int cuda_graph_cache_size() {
   static int cache_size = []() {
     return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 100);
@@ -22,7 +34,7 @@ int cuda_graph_cache_size() {
   return cache_size;
 }
 
-namespace cu {
+} // namespace
 
 Device::Device(int device) : device_(device) {
   CHECK_CUDA_ERROR(cudaDeviceGetAttribute(
@@ -40,11 +52,18 @@ Device::Device(int device) : device_(device) {
   }
   // The cublasLt handle is used by matmul.
   make_current();
-  cublasLtCreate(&lt_);
+  CHECK_CUBLAS_ERROR(cublasLtCreate(&lt_));
+  // The cudnn handle is used by Convolution.
+  CHECK_CUDNN_ERROR(cudnnCreate(&cudnn_));
+
+  // Initialize the jit module cache here ensures it is not
+  // unloaded before any evaluation is done
+  get_jit_module_cache();
 }
 
 Device::~Device() {
-  cublasLtDestroy(lt_);
+  CHECK_CUDNN_ERROR(cudnnDestroy(cudnn_));
+  CHECK_CUBLAS_ERROR(cublasLtDestroy(lt_));
 }
 
 void Device::make_current() {
@@ -66,29 +85,19 @@ CommandEncoder& Device::get_command_encoder(Stream s) {
 }
 
 CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
+  enc.device().make_current();
   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'});
+  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
+      &graph, [](cudaGraph_t* p) { CHECK_CUDA_ERROR(cudaGraphDestroy(*p)); });
+  if (discard) {
+    return;
   }
-  CHECK_CUDA_ERROR(cudaGraphDestroy(graph));
+  enc.add_graph_node(graph);
 }
 
 CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
@@ -175,21 +184,11 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
   }
 }
 
-CommandEncoder::CommandEncoder(Device& d) : device_(d), stream_(d) {
+CommandEncoder::CommandEncoder(Device& d)
+    : device_(d), stream_(d), graph_cache_(cuda_graph_cache_size()) {
   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));
 }
@@ -215,22 +214,22 @@ void CommandEncoder::add_kernel_node(
     void* func,
     dim3 grid_dim,
     dim3 block_dim,
+    uint32_t smem_bytes,
     void** params) {
   cudaKernelNodeParams kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDim = grid_dim;
   kernel_params.blockDim = block_dim;
   kernel_params.kernelParams = params;
-  cudaGraphNode_t node;
-  CHECK_CUDA_ERROR(
-      cudaGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
-  insert_graph_dependencies(GraphNode{node, 'K'});
+  kernel_params.sharedMemBytes = smem_bytes;
+  add_kernel_node(kernel_params);
 }
 
 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;
@@ -241,13 +240,30 @@ void CommandEncoder::add_kernel_node(
   kernel_params.blockDimY = block_dim.y;
   kernel_params.blockDimZ = block_dim.z;
   kernel_params.kernelParams = params;
-  CUgraphNode node;
-  CHECK_CUDA_ERROR(
-      cuGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
+  kernel_params.sharedMemBytes = smem_bytes;
+  add_kernel_node(kernel_params);
+}
+
+void CommandEncoder::add_kernel_node(const cudaKernelNodeParams& params) {
+  cudaGraphNode_t node;
+  CHECK_CUDA_ERROR(cudaGraphAddKernelNode(&node, graph_, NULL, 0, &params));
   insert_graph_dependencies(GraphNode{node, 'K'});
 }
 
+void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
+  CUgraphNode node;
+  CHECK_CUDA_ERROR(cuGraphAddKernelNode(&node, graph_, NULL, 0, &params));
+  insert_graph_dependencies(GraphNode{node, 'K'});
+}
+
+void CommandEncoder::add_graph_node(cudaGraph_t child) {
+  cudaGraphNode_t node;
+  CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
+  insert_graph_dependencies(GraphNode{node, 'G'});
+}
+
 void CommandEncoder::commit() {
+  nvtx3::scoped_range r("CommandEncoder::commit");
   if (!temporaries_.empty()) {
     add_completed_handler([temporaries = std::move(temporaries_)]() {});
   }
@@ -264,7 +280,7 @@ void CommandEncoder::commit() {
     graph_key_ += ".";
     graph_key_ += std::to_string(empty_node_count_);
 
-    cudaGraphExec_t& graph_exec = graph_cache_[graph_key_];
+    CudaGraphExec& graph_exec = graph_cache_[graph_key_];
 
     if (graph_exec != nullptr) {
       cudaGraphExecUpdateResult update_result;
@@ -278,25 +294,19 @@ void CommandEncoder::commit() {
 #endif // CUDART_VERSION >= 12000
       if (update_result != cudaGraphExecUpdateSuccess) {
         cudaGetLastError(); // reset error
-        CHECK_CUDA_ERROR(cudaGraphExecDestroy(graph_exec));
-        graph_exec = nullptr;
+        graph_exec.reset();
       }
     }
     if (graph_exec == nullptr) {
-      CHECK_CUDA_ERROR(
-          cudaGraphInstantiate(&graph_exec, graph_, NULL, NULL, 0));
+      graph_exec.instantiate(graph_);
     }
     device_.make_current();
     CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
 
-    // TODO smarter cache policy
-    if (graph_cache_.size() > cuda_graph_cache_size()) {
-      clear_graphs(graph_cache_);
-    }
-
     // Reset state
     node_count_ = 0;
     graph_node_count_ = 0;
+    empty_node_count_ = 0;
     from_nodes_.clear();
     to_nodes_.clear();
     graph_key_.clear();
@@ -306,7 +316,6 @@ void CommandEncoder::commit() {
   }
 
   // Put completion handlers in a batch.
-  worker_.end_batch();
   worker_.commit(stream_);
 }
 
@@ -315,7 +324,6 @@ void CommandEncoder::synchronize() {
   auto p = std::make_shared<std::promise<void>>();
   std::future<void> f = p->get_future();
   add_completed_handler([p = std::move(p)]() { p->set_value(); });
-  worker_.end_batch();
   commit();
   f.wait();
 }
@@ -333,6 +341,4 @@ CommandEncoder& get_command_encoder(Stream s) {
   return device(s.device).get_command_encoder(s);
 }
 
-} // namespace cu
-
-} // namespace mlx::core
+} // namespace mlx::core::cu

mlx/backend/cuda/device.h

@@ -3,11 +3,13 @@
 #pragma once
 
 #include "mlx/array.h"
+#include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/cuda/worker.h"
 #include "mlx/stream.h"
 
 #include <cublasLt.h>
 #include <cuda.h>
+#include <cudnn.h>
 #include <thrust/execution_policy.h>
 
 #include <unordered_map>
@@ -21,6 +23,7 @@ class CommandEncoder {
     ~CaptureContext();
     cudaGraph_t graph;
     CommandEncoder& enc;
+    bool discard{false};
   };
   struct ConcurrentContext {
     ConcurrentContext(CommandEncoder& enc);
@@ -29,7 +32,6 @@ class CommandEncoder {
   };
 
   explicit CommandEncoder(Device& d);
-  ~CommandEncoder();
 
   CommandEncoder(const CommandEncoder&) = delete;
   CommandEncoder& operator=(const CommandEncoder&) = delete;
@@ -45,25 +47,39 @@ class CommandEncoder {
   void set_output_array(const array& arr);
 
   template <typename F, typename... Params>
-  void
-  add_kernel_node(F* func, dim3 grid_dim, dim3 block_dim, Params&&... 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, ptrs);
+    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, void** params);
+  void add_kernel_node(
+      void* func,
+      dim3 grid_dim,
+      dim3 block_dim,
+      uint32_t smem_bytes,
+      void** params);
+
+  // Low-level graph helpers.
+  void add_kernel_node(const cudaKernelNodeParams& params);
+  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
+  void add_graph_node(cudaGraph_t child);
 
   void add_temporary(const array& arr) {
     temporaries_.push_back(arr.data_shared_ptr());
@@ -73,6 +89,10 @@ class CommandEncoder {
   void maybe_commit();
   void commit();
 
+  Device& device() {
+    return device_;
+  }
+
   CudaStream& stream() {
     return stream_;
   }
@@ -106,7 +126,7 @@ class CommandEncoder {
   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_;
+  LRUCache<std::string, CudaGraphExec> 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_;
@@ -137,12 +157,16 @@ class Device {
   cublasLtHandle_t lt_handle() const {
     return lt_;
   }
+  cudnnHandle_t cudnn_handle() const {
+    return cudnn_;
+  }
 
  private:
   int device_;
   int compute_capability_major_;
   int compute_capability_minor_;
   cublasLtHandle_t lt_;
+  cudnnHandle_t cudnn_;
   std::unordered_map<int, CommandEncoder> encoders_;
 };
 

mlx/backend/cuda/device/arange.cuh

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

mlx/backend/cuda/device/atomic_ops.cuh

@@ -49,11 +49,7 @@ inline __device__ void atomic_add(__half* out, __half val) {
 }
 
 inline __device__ void atomic_add(complex64_t* out, complex64_t val) {
-#if __CUDA_ARCH__ < 900
   atomic_add_general(out, val);
-#else
-  atomicAdd(out, val);
-#endif
 }
 
 inline __device__ void atomic_add(__nv_bfloat16* out, __nv_bfloat16 val) {

mlx/backend/cuda/device/utils.cuh

@@ -32,23 +32,120 @@ using Strides = cuda::std::array<int64_t, MAX_NDIM>;
 template <typename T, int N>
 struct alignas(sizeof(T) * N) AlignedVector {
   T val[N];
+
+  __device__ T& operator[](int i) {
+    return val[i];
+  }
+
+  __device__ T operator[](int i) const {
+    return val[i];
+  }
 };
 
 template <int N, typename T>
-inline __device__ AlignedVector<T, N> load_vector(
+inline __host__ __device__ bool is_aligned(T* x) {
+  return (reinterpret_cast<uintptr_t>(x) % (N * sizeof(T))) == 0;
+}
+
+template <int N, typename T>
+inline __device__ AlignedVector<T, N> unsafe_load_vector(
     const T* ptr,
     uint32_t offset) {
   auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
   return from[offset];
 }
 
+template <int N, typename T>
+inline __device__ AlignedVector<T, N> load_vector(
+    const T* ptr,
+    uint32_t offset) {
+  if (is_aligned<N>(ptr)) {
+    auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
+    return from[offset];
+  } else {
+    AlignedVector<T, N> v;
+#pragma unroll
+    for (int i = 0; i < N; ++i) {
+      v[i] = ptr[offset * N + i];
+    }
+    return v;
+  }
+}
+
+template <int N, typename T, typename SizeT>
+inline __device__ AlignedVector<T, N>
+load_vector(const T* ptr, uint32_t offset, SizeT size, T fallback) {
+  if (is_aligned<N>(ptr) && (offset + 1) * N <= size) {
+    auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
+    return from[offset];
+  } else {
+    AlignedVector<T, N> v;
+#pragma unroll
+    for (int i = 0; i < N; ++i) {
+      v[i] = (N * offset + i) < size ? ptr[offset * N + i] : fallback;
+    }
+    return v;
+  }
+}
+
+template <int N, typename T, typename SizeT>
+inline __device__ AlignedVector<T, N> load_vector(
+    const T* ptr,
+    uint32_t offset,
+    SizeT size,
+    int64_t stride,
+    T fallback) {
+  if (is_aligned<N>(ptr) && stride == 1 && (offset + 1) * N <= size) {
+    auto* from = reinterpret_cast<const AlignedVector<T, N>*>(ptr);
+    return from[offset];
+  } else {
+    AlignedVector<T, N> v;
+#pragma unroll
+    for (int i = 0; i < N; ++i) {
+      v[i] =
+          (N * offset + i) < size ? ptr[stride * (offset * N + i)] : fallback;
+    }
+    return v;
+  }
+}
+
 template <int N, typename T>
 inline __device__ void
-store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
+unsafe_store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
   auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
   to[offset] = vec;
 }
 
+template <int N, typename T>
+inline __device__ void
+store_vector(T* ptr, uint32_t offset, const AlignedVector<T, N>& vec) {
+  if (is_aligned<N>(ptr)) {
+    auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
+    to[offset] = vec;
+  } else {
+#pragma unroll
+    for (int i = 0; i < N; ++i) {
+      ptr[offset * N + i] = vec[i];
+    }
+  }
+}
+
+template <int N, typename T, typename SizeT>
+inline __device__ void store_vector(
+    T* ptr,
+    uint32_t offset,
+    const AlignedVector<T, N>& vec,
+    SizeT size) {
+  if (is_aligned<N>(ptr) && (offset + 1) * N <= size) {
+    auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
+    to[offset] = vec;
+  } else {
+    for (int i = 0; (offset * N + i) < size && i < N; ++i) {
+      ptr[offset * N + i] = vec[i];
+    }
+  }
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Type limits utils
 ///////////////////////////////////////////////////////////////////////////////
@@ -204,20 +301,8 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_nd(
   return cuda::std::make_tuple(a_loc, b_loc, c_loc);
 }
 
-// Optimized version when ndim is larger than 4.
 template <typename IdxT = int64_t>
-inline __host__ __device__ IdxT
-elem_to_loc_4d(IdxT elem, const int* shape, const int64_t* strides, int ndim) {
-  IdxT loc = 0;
-  for (int i = ndim - 1; i >= 0; --i) {
-    loc += (elem % shape[i]) * IdxT(strides[i]);
-    elem /= shape[i];
-  }
-  return loc;
-}
-
-template <typename IdxT = int64_t>
-inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_4d(
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc(
     IdxT elem,
     const int* shape,
     const int64_t* a_strides,
@@ -235,7 +320,7 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_4d(
 }
 
 template <typename IdxT = int64_t>
-inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_4d(
+inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc(
     IdxT elem,
     const int* shape,
     const int64_t* a_strides,

mlx/backend/cuda/eval.cpp

@@ -19,8 +19,6 @@ void new_stream(Stream s) {
   cudaFree(nullptr);
   // Ensure the static stream objects get created.
   cu::get_command_encoder(s);
-  // The main thread is safe to free buffers.
-  cu::allocator().register_this_thread();
 }
 
 void eval(array& arr) {
@@ -38,18 +36,15 @@ void eval(array& arr) {
 
   auto& encoder = cu::get_command_encoder(arr.primitive().stream());
   // Keep used buffers alive until kernel finishes running.
-  std::unordered_set<std::shared_ptr<array::Data>> buffers;
   for (auto& in : arr.inputs()) {
-    buffers.insert(in.data_shared_ptr());
+    // Except for the donated one.
+    if (in.data_shared_ptr() != arr.data_shared_ptr()) {
+      encoder.add_temporary(in);
+    }
   }
   for (auto& s : arr.siblings()) {
-    buffers.insert(s.data_shared_ptr());
+    encoder.add_temporary(s);
   }
-  // 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();
 }
 

mlx/backend/cuda/event.cu

@@ -110,24 +110,26 @@ __global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
   event_signal(ac, value);
 }
 
+SharedEvent::Atomic* to_atomic(std::shared_ptr<Buffer> buf) {
+  return static_cast<SharedEvent::Atomic*>(buf->raw_ptr());
+}
+
 SharedEvent::SharedEvent() {
-  // Allocate cuda::atomic on managed memory.
-  Atomic* ac;
-  CHECK_CUDA_ERROR(cudaMallocManaged(&ac, sizeof(Atomic)));
-  new (ac) Atomic(0);
-  ac_ = std::shared_ptr<Atomic>(ac, [](Atomic* ptr) {
-    ptr->~Atomic();
-    allocator().cuda_free(ptr);
-  });
+  buf_ = std::shared_ptr<Buffer>(
+      new Buffer{allocator().malloc(sizeof(Atomic))}, [](Buffer* ptr) {
+        allocator().free(*ptr);
+        delete ptr;
+      });
+  *static_cast<uint64_t*>(buf_->raw_ptr()) = 0;
 }
 
 void SharedEvent::wait(uint64_t value) {
   nvtx3::scoped_range r("cu::SharedEvent::wait");
-  event_wait(ac_.get(), value);
+  event_wait(to_atomic(buf_), value);
 }
 
 void SharedEvent::wait(cudaStream_t stream, uint64_t value) {
-  event_wait_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
+  event_wait_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
 }
 
 void SharedEvent::wait(Stream s, uint64_t value) {
@@ -138,17 +140,17 @@ void SharedEvent::wait(Stream s, uint64_t value) {
     auto& encoder = get_command_encoder(s);
     encoder.commit();
     wait(encoder.stream(), value);
-    encoder.add_completed_handler([ac = ac_]() {});
+    encoder.add_completed_handler([buf = buf_]() {});
   }
 }
 
 void SharedEvent::signal(uint64_t value) {
   nvtx3::scoped_range r("cu::SharedEvent::signal");
-  event_signal(ac_.get(), value);
+  event_signal(to_atomic(buf_), value);
 }
 
 void SharedEvent::signal(cudaStream_t stream, uint64_t value) {
-  event_signal_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
+  event_signal_kernel<<<1, 1, 0, stream>>>(to_atomic(buf_), value);
 }
 
 void SharedEvent::signal(Stream s, uint64_t value) {
@@ -162,18 +164,18 @@ void SharedEvent::signal(Stream s, uint64_t value) {
     auto& encoder = get_command_encoder(s);
     encoder.commit();
     signal(encoder.stream(), value);
-    encoder.add_completed_handler([ac = ac_]() {});
+    encoder.add_completed_handler([buf = buf_]() {});
   }
 }
 
 bool SharedEvent::is_signaled(uint64_t value) const {
   nvtx3::scoped_range r("cu::SharedEvent::is_signaled");
-  return ac_->load() >= value;
+  return to_atomic(buf_)->load() >= value;
 }
 
 uint64_t SharedEvent::value() const {
   nvtx3::scoped_range r("cu::SharedEvent::value");
-  return ac_->load();
+  return to_atomic(buf_)->load();
 }
 
 } // namespace cu

mlx/backend/cuda/event.h

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

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -0,0 +1,329 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/utils.h"
+
+#include <fmt/format.h>
+
+namespace mlx::core {
+
+namespace {
+
+struct CublasPreference {
+  CublasPreference(cu::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(cu::Device& device) {
+  static CublasPreference pref(device);
+  return pref.pref_;
+}
+
+cublasComputeType_t dtype_to_compute_type(Dtype dtype) {
+  switch (dtype) {
+    case float16:
+      return CUBLAS_COMPUTE_32F;
+    case bfloat16:
+      return CUBLAS_COMPUTE_32F;
+    case float32:
+      return mlx::core::env::enable_tf32() ? CUBLAS_COMPUTE_32F_FAST_TF32
+                                           : CUBLAS_COMPUTE_32F;
+    case float64:
+    case complex64:
+      return CUBLAS_COMPUTE_64F;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in CublasGemm: {}.", dtype_to_string(dtype)));
+  }
+}
+
+cudaDataType_t dtype_to_cublas_type(Dtype dtype) {
+  switch (dtype) {
+    case float16:
+      return CUDA_R_16F;
+    case bfloat16:
+      return CUDA_R_16BF;
+    case float32:
+      return CUDA_R_32F;
+    case float64:
+      return CUDA_R_64F;
+    case complex64:
+      return CUDA_C_32F;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in CublasGemm: {}.", dtype_to_string(dtype)));
+  }
+}
+
+cublasLtMatrixLayout_t create_matrix_layout(
+    cudaDataType_t type,
+    uint64_t rows,
+    uint64_t cols,
+    bool transposed,
+    int64_t ld,
+    int32_t batch_count,
+    int64_t batch_stride) {
+  cublasLtMatrixLayout_t desc;
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutCreate(&desc, type, rows, cols, ld));
+  cublasLtOrder_t order = transposed ? CUBLASLT_ORDER_COL : CUBLASLT_ORDER_ROW;
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc, CUBLASLT_MATRIX_LAYOUT_ORDER, &order, sizeof(cublasLtOrder_t)));
+  if (batch_count > 1) {
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+        desc,
+        CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT,
+        &batch_count,
+        sizeof(int32_t)));
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+        desc,
+        CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET,
+        &batch_stride,
+        sizeof(int64_t)));
+  }
+  return desc;
+}
+
+} // namespace
+
+CublasGemm::CublasGemm(
+    cu::Device& device,
+    Dtype dtype,
+    bool a_transposed,
+    uint64_t a_rows,
+    uint64_t a_cols,
+    int64_t lda,
+    bool b_transposed,
+    uint64_t b_rows,
+    uint64_t b_cols,
+    int64_t ldb,
+    int32_t batch_count,
+    int64_t a_batch_stride,
+    int64_t b_batch_stride)
+    : handle_(device.lt_handle()),
+      pref_(cublas_preference(device)),
+      M_(a_rows),
+      N_(b_cols) {
+  heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
+
+  auto scale_type = dtype_to_cublas_type(dtype);
+  if (dtype == bfloat16 || dtype == float16) {
+    scale_type = CUDA_R_32F;
+  }
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescCreate(
+      &matmul_desc_, dtype_to_compute_type(dtype), scale_type));
+  int32_t pointer_mode = CUBLASLT_POINTER_MODE_HOST;
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_POINTER_MODE,
+      &pointer_mode,
+      sizeof(int32_t)));
+  cublasOperation_t op = CUBLAS_OP_N;
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_TRANSA,
+      &op,
+      sizeof(cublasOperation_t)));
+  CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
+      matmul_desc_,
+      CUBLASLT_MATMUL_DESC_TRANSB,
+      &op,
+      sizeof(cublasOperation_t)));
+
+  auto type = dtype_to_cublas_type(dtype);
+  a_desc_ = create_matrix_layout(
+      type, a_rows, a_cols, a_transposed, lda, batch_count, a_batch_stride);
+  b_desc_ = create_matrix_layout(
+      type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
+  out_desc_ = create_matrix_layout(
+      type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
+}
+
+CublasGemm::CublasGemm(
+    cu::Device& device,
+    Dtype dtype,
+    bool a_transposed,
+    uint64_t a_rows,
+    uint64_t a_cols,
+    int64_t lda,
+    bool b_transposed,
+    uint64_t b_rows,
+    uint64_t b_cols,
+    int64_t ldb,
+    int64_t ldc,
+    int32_t batch_count,
+    int64_t a_batch_stride,
+    int64_t b_batch_stride,
+    int64_t c_batch_stride)
+    : CublasGemm(
+          device,
+          dtype,
+          a_transposed,
+          a_rows,
+          a_cols,
+          lda,
+          b_transposed,
+          b_rows,
+          b_cols,
+          ldb,
+          batch_count,
+          a_batch_stride,
+          b_batch_stride) {
+  auto type = dtype_to_cublas_type(dtype);
+  c_desc_ = create_matrix_layout(
+      type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
+}
+
+CublasGemm::~CublasGemm() {
+  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 CublasGemm::run(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides) {
+  int batch_count = out.size() / (M_ * N_);
+  if (batch_count / batch_shape.back() > 1) {
+    run_batched(
+        encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides);
+    return;
+  }
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+
+  execute(encoder, out.data<void>(), a.data<void>(), b.data<void>(), nullptr);
+}
+
+void CublasGemm::run(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const array& c,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides,
+    const Strides& c_batch_strides,
+    float alpha,
+    float beta) {
+  int batch_count = out.size() / (M_ * N_);
+  if (batch_count / batch_shape.back() > 1) {
+    run_batched(
+        encoder,
+        out,
+        a,
+        b,
+        c,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        c_batch_strides,
+        alpha,
+        beta);
+    return;
+  }
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  execute(
+      encoder,
+      out.data<void>(),
+      a.data<void>(),
+      b.data<void>(),
+      c.data<void>(),
+      alpha,
+      beta);
+}
+
+void CublasGemm::execute(
+    cu::CommandEncoder& encoder,
+    void* out,
+    const void* a,
+    const void* b,
+    const void* c,
+    float alpha /* = 1 */,
+    float beta /* = 0 */) {
+  if (heuristic_.state != CUBLAS_STATUS_SUCCESS) {
+    int ret = 0;
+    CHECK_CUBLAS_ERROR(cublasLtMatmulAlgoGetHeuristic(
+        handle_,
+        matmul_desc_,
+        a_desc_,
+        b_desc_,
+        c ? c_desc_ : out_desc_,
+        out_desc_,
+        pref_,
+        1,
+        &heuristic_,
+        &ret));
+    if (ret == 0) {
+      throw std::runtime_error("Can not find algorithm for matmul.");
+    }
+  }
+
+  void* workspace_ptr = nullptr;
+  if (heuristic_.workspaceSize > 0) {
+    // Ensure workspace is 256-byte aligned
+    int nbytes = cuda::ceil_div(heuristic_.workspaceSize, 256) * 256;
+    array workspace(
+        allocator::malloc(nbytes),
+        {static_cast<int>(heuristic_.workspaceSize)},
+        int8);
+    encoder.add_temporary(workspace);
+    workspace_ptr = workspace.data<void>();
+  }
+
+  auto capture = encoder.capture_context();
+  CHECK_CUBLAS_ERROR(cublasLtMatmul(
+      handle_,
+      matmul_desc_,
+      &alpha,
+      a,
+      a_desc_,
+      b,
+      b_desc_,
+      &beta,
+      c ? c : out,
+      c ? c_desc_ : out_desc_,
+      out,
+      out_desc_,
+      &heuristic_.algo,
+      workspace_ptr,
+      heuristic_.workspaceSize,
+      encoder.stream()));
+}
+
+} // namespace mlx::core

mlx/backend/cuda/gemms/cublas_gemm.h

@@ -0,0 +1,113 @@
+// Copyright © 2025 Apple Inc.
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/cuda/device.h"
+
+#include <cublasLt.h>
+
+namespace mlx::core {
+
+class CublasGemm {
+ public:
+  CublasGemm(
+      cu::Device& device,
+      Dtype dtype,
+      bool a_transposed,
+      uint64_t a_rows,
+      uint64_t a_cols,
+      int64_t lda,
+      bool b_transposed,
+      uint64_t b_rows,
+      uint64_t b_cols,
+      int64_t ldb,
+      int32_t batch_count,
+      int64_t a_batch_stride,
+      int64_t b_batch_stride);
+
+  CublasGemm(
+      cu::Device& device,
+      Dtype dtype,
+      bool a_transposed,
+      uint64_t a_rows,
+      uint64_t a_cols,
+      int64_t lda,
+      bool b_transposed,
+      uint64_t b_rows,
+      uint64_t b_cols,
+      int64_t ldb,
+      int64_t ldc,
+      int32_t batch_count,
+      int64_t a_batch_stride,
+      int64_t b_batch_stride,
+      int64_t c_batch_stride);
+
+  ~CublasGemm();
+
+  void run(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const Shape& batch_shape,
+      const Strides& a_batch_strides,
+      const Strides& b_batch_strides);
+
+  void run(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const array& c,
+      const Shape& batch_shape,
+      const Strides& a_batch_strides,
+      const Strides& b_batch_strides,
+      const Strides& c_batch_strides,
+      float alpha,
+      float beta);
+
+ private:
+  void run_batched(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const Shape& batch_shape,
+      const Strides& a_batch_strides,
+      const Strides& b_batch_strides);
+
+  void run_batched(
+      cu::CommandEncoder& encoder,
+      array& out,
+      const array& a,
+      const array& b,
+      const array& c,
+      const Shape& batch_shape,
+      const Strides& a_batch_strides,
+      const Strides& b_batch_strides,
+      const Strides& c_batch_strides,
+      float alpha,
+      float beta);
+
+  void execute(
+      cu::CommandEncoder& encoder,
+      void* out,
+      const void* a,
+      const void* b,
+      const void* c,
+      float alpha = 1,
+      float beta = 0);
+
+  uint64_t M_;
+  uint64_t N_;
+  cublasLtMatmulPreference_t pref_{nullptr};
+  cublasLtHandle_t handle_{nullptr};
+  cublasLtMatmulDesc_t matmul_desc_{nullptr};
+  cublasLtMatrixLayout_t a_desc_{nullptr};
+  cublasLtMatrixLayout_t b_desc_{nullptr};
+  cublasLtMatrixLayout_t c_desc_{nullptr};
+  cublasLtMatrixLayout_t out_desc_{nullptr};
+  cublasLtMatmulHeuristicResult_t heuristic_;
+};
+
+} // namespace mlx::core

mlx/backend/cuda/gemms/cublas_gemm_batched_12_0.cpp

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

mlx/backend/cuda/gemms/cublas_gemm_batched_12_9.cu

@@ -0,0 +1,327 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <int NDIM>
+__global__ void set_mm_device_pointers_nd(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> batch_shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_batch_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_batch_strides,
+    int64_t batch_stride,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset] = elem_to_loc_nd<NDIM>(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data());
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+__global__ void set_mm_device_pointers_g(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides a_batch_strides,
+    const __grid_constant__ Strides b_batch_strides,
+    int64_t batch_stride,
+    int batch_ndim,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset] = elem_to_loc(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      batch_ndim);
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+template <int NDIM>
+__global__ void set_addmm_device_pointers_nd(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* c_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> batch_shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_batch_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_batch_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> c_batch_strides,
+    int64_t batch_stride,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset, c_offset] = elem_to_loc_nd<NDIM>(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      c_batch_strides.data());
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] = c_start + item_size * c_offset;
+  pointers[index + 3 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+__global__ void set_addmm_device_pointers_g(
+    int8_t** pointers,
+    int8_t* a_start,
+    int8_t* b_start,
+    int8_t* c_start,
+    int8_t* out_start,
+    int item_size,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides a_batch_strides,
+    const __grid_constant__ Strides b_batch_strides,
+    const __grid_constant__ Strides c_batch_strides,
+    int64_t batch_stride,
+    int batch_ndim,
+    int batch_count) {
+  auto index = cg::this_grid().thread_rank();
+  if (index >= batch_count) {
+    return;
+  }
+  auto [a_offset, b_offset, c_offset] = elem_to_loc(
+      index,
+      batch_shape.data(),
+      a_batch_strides.data(),
+      b_batch_strides.data(),
+      c_batch_strides.data(),
+      batch_ndim);
+  pointers[index] = a_start + item_size * a_offset;
+  pointers[index + batch_count] = b_start + item_size * b_offset;
+  pointers[index + 2 * batch_count] = c_start + item_size * c_offset;
+  pointers[index + 3 * batch_count] =
+      out_start + item_size * index * batch_stride;
+}
+
+} // namespace cu
+
+namespace {
+
+void set_pointer_mode(cublasLtMatrixLayout_t desc, int batch_count) {
+  auto batch_mode = CUBLASLT_BATCH_MODE_POINTER_ARRAY;
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc,
+      CUBLASLT_MATRIX_LAYOUT_BATCH_MODE,
+      &batch_mode,
+      sizeof(batch_mode)));
+  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
+      desc, CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT, &batch_count, sizeof(int32_t)));
+}
+
+} // namespace
+
+void CublasGemm::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides) {
+  int batch_count = out.size() / (M_ * N_);
+  set_pointer_mode(a_desc_, batch_count);
+  set_pointer_mode(b_desc_, batch_count);
+  set_pointer_mode(out_desc_, batch_count);
+
+  // Launch kernel to set device offsets
+  auto pointers = array(
+      allocator::malloc(batch_count * sizeof(void*) * 3),
+      {batch_count * 3},
+      uint64);
+
+  encoder.add_temporary(pointers);
+  encoder.set_output_array(pointers);
+
+  int block_dims = std::min(batch_count, 256);
+  int num_blocks = cuda::ceil_div(batch_count, block_dims);
+  int64_t batch_stride = M_ * N_;
+  int item_size = out.itemsize();
+
+  int ndim = batch_shape.size();
+  if (ndim <= 3) {
+    dispatch_1_2_3(ndim, [&](auto ndim_constant) {
+      encoder.add_kernel_node(
+          cu::set_mm_device_pointers_nd<ndim_constant()>,
+          num_blocks,
+          block_dims,
+          0,
+          pointers.data<int8_t*>(),
+          a.data<int8_t>(),
+          b.data<int8_t>(),
+          out.data<int8_t>(),
+          item_size,
+          const_param<ndim_constant()>(batch_shape),
+          const_param<ndim_constant()>(a_batch_strides),
+          const_param<ndim_constant()>(b_batch_strides),
+          batch_stride,
+          batch_count);
+    });
+  } else {
+    encoder.add_kernel_node(
+        cu::set_mm_device_pointers_g,
+        num_blocks,
+        block_dims,
+        0,
+        pointers.data<int8_t*>(),
+        a.data<int8_t>(),
+        b.data<int8_t>(),
+        out.data<int8_t>(),
+        item_size,
+        const_param(batch_shape),
+        const_param(a_batch_strides),
+        const_param(b_batch_strides),
+        batch_stride,
+        ndim,
+        batch_count);
+  }
+
+  // Run matmul
+  encoder.set_input_array(pointers);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+
+  auto a_pointers = pointers.data<int8_t*>();
+  auto b_pointers = a_pointers + batch_count;
+  auto out_pointers = b_pointers + batch_count;
+  execute(
+      encoder,
+      reinterpret_cast<void*>(out_pointers),
+      reinterpret_cast<void*>(a_pointers),
+      reinterpret_cast<void*>(b_pointers),
+      nullptr);
+}
+
+void CublasGemm::run_batched(
+    cu::CommandEncoder& encoder,
+    array& out,
+    const array& a,
+    const array& b,
+    const array& c,
+    const Shape& batch_shape,
+    const Strides& a_batch_strides,
+    const Strides& b_batch_strides,
+    const Strides& c_batch_strides,
+    float alpha,
+    float beta) {
+  int batch_count = out.size() / (M_ * N_);
+  set_pointer_mode(a_desc_, batch_count);
+  set_pointer_mode(b_desc_, batch_count);
+  set_pointer_mode(c_desc_, batch_count);
+  set_pointer_mode(out_desc_, batch_count);
+
+  // Launch kernel to set device offsets
+  auto pointers = array(
+      allocator::malloc(batch_count * sizeof(uint64_t) * 4),
+      {batch_count * 4},
+      uint64);
+
+  encoder.add_temporary(pointers);
+  encoder.set_output_array(pointers);
+
+  int block_dims = std::min(batch_count, 256);
+  int num_blocks = cuda::ceil_div(batch_count, block_dims);
+  int64_t batch_stride = M_ * N_;
+  int item_size = out.itemsize();
+
+  int ndim = batch_shape.size();
+  if (ndim <= 3) {
+    dispatch_1_2_3(ndim, [&](auto ndim_constant) {
+      encoder.add_kernel_node(
+          cu::set_addmm_device_pointers_nd<ndim_constant()>,
+          num_blocks,
+          block_dims,
+          0,
+          pointers.data<int8_t*>(),
+          a.data<int8_t>(),
+          b.data<int8_t>(),
+          c.data<int8_t>(),
+          out.data<int8_t>(),
+          item_size,
+          const_param<ndim_constant()>(batch_shape),
+          const_param<ndim_constant()>(a_batch_strides),
+          const_param<ndim_constant()>(b_batch_strides),
+          const_param<ndim_constant()>(c_batch_strides),
+          batch_stride,
+          batch_count);
+    });
+  } else {
+    encoder.add_kernel_node(
+        cu::set_addmm_device_pointers_g,
+        num_blocks,
+        block_dims,
+        0,
+        pointers.data<int8_t*>(),
+        a.data<int8_t>(),
+        b.data<int8_t>(),
+        c.data<int8_t>(),
+        out.data<int8_t>(),
+        item_size,
+        const_param(batch_shape),
+        const_param(a_batch_strides),
+        const_param(b_batch_strides),
+        const_param(c_batch_strides),
+        batch_stride,
+        ndim,
+        batch_count);
+  }
+
+  // Run matmul
+  encoder.set_input_array(pointers);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  auto a_pointers = pointers.data<int8_t*>();
+  auto b_pointers = a_pointers + batch_count;
+  auto c_pointers = b_pointers + batch_count;
+  auto out_pointers = c_pointers + batch_count;
+  execute(
+      encoder,
+      reinterpret_cast<void*>(out_pointers),
+      reinterpret_cast<void*>(a_pointers),
+      reinterpret_cast<void*>(b_pointers),
+      reinterpret_cast<void*>(c_pointers),
+      alpha,
+      beta);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/gemms/gemv.cu

@@ -0,0 +1,173 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/gemms/gemv.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+
+namespace mlx::core::cu {
+
+namespace cg = cooperative_groups;
+
+static constexpr int rows_per_block = 8;
+
+template <typename T, int rows_per_block, int n_per_thread>
+__device__ void
+gemv_impl(const T* mat, const T* vec, T* out, int rows, int cols) {
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  auto g_idx = block.group_index();
+  auto t_idx = block.thread_index();
+  int row = g_idx.x * rows_per_block + t_idx.y;
+
+  if (row < rows) {
+    float sum = 0.0f;
+    for (int col = n_per_thread * warp.thread_rank(); col < cols;
+         col += (WARP_SIZE * n_per_thread)) {
+      auto local_mat =
+          unsafe_load_vector<n_per_thread>(mat + row * cols + col, 0);
+      auto local_vec = unsafe_load_vector<n_per_thread>(vec + col, 0);
+#pragma unroll
+      for (int j = 0; j < n_per_thread; ++j) {
+        sum +=
+            static_cast<float>(local_mat[j]) * static_cast<float>(local_vec[j]);
+      }
+    }
+
+    sum = cg::reduce(warp, sum, cg::plus<float>{});
+    if (warp.thread_rank() == 0) {
+      out[row] = static_cast<T>(sum);
+    }
+  }
+}
+
+template <typename T, int rows_per_block, int n_per_thread>
+__global__ void
+gemv_single(const T* mat, const T* vec, T* out, int rows, int cols) {
+  gemv_impl<T, rows_per_block, n_per_thread>(mat, vec, out, rows, cols);
+}
+
+template <typename T, int rows_per_block, int n_per_thread>
+__global__ void gemv_batched(
+    const T* mat,
+    const T* vec,
+    T* out,
+    int rows,
+    int cols,
+    const __grid_constant__ Shape batch_shape,
+    const __grid_constant__ Strides mat_batch_strides,
+    const __grid_constant__ Strides vec_batch_strides,
+    int batch_ndim) {
+  auto block = cg::this_thread_block();
+  auto batch_idx = block.group_index().y;
+  auto [vec_offset, mat_offset] = elem_to_loc(
+      batch_idx,
+      batch_shape.data(),
+      vec_batch_strides.data(),
+      mat_batch_strides.data(),
+      batch_ndim);
+  gemv_impl<T, rows_per_block, n_per_thread>(
+      mat + mat_offset, vec + vec_offset, out + batch_idx * rows, rows, cols);
+}
+
+bool can_use_gemv(int M, int N, int K, bool a_transposed, bool b_transposed) {
+  return K % 32 == 0 && ((M == 1 && b_transposed) || (N == 1 && !a_transposed));
+}
+
+template <typename F>
+void dispatch_n_per_thread(int n_per_thread, F&& f) {
+  switch (n_per_thread) {
+    case 1:
+      f(std::integral_constant<int, 1>{});
+      break;
+    case 2:
+      f(std::integral_constant<int, 2>{});
+      break;
+    case 4:
+      f(std::integral_constant<int, 4>{});
+      break;
+  }
+}
+
+void gemv(
+    const array& a,
+    const array& b,
+    array& out,
+    int M,
+    int N,
+    int K,
+    uint32_t batch_count,
+    const mlx::core::Shape& batch_shape,
+    const mlx::core::Strides& a_batch_strides,
+    const mlx::core::Strides& b_batch_strides,
+    CommandEncoder& encoder) {
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  dispatch_float_types(out.dtype(), "gemv", [&](auto type_tag) {
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    dim3 block_dims{WARP_SIZE, rows_per_block};
+    const DataType* mat;
+    const DataType* vec;
+    int rows;
+    int cols = K;
+    auto mat_strides = const_param(a_batch_strides);
+    auto vec_strides = const_param(b_batch_strides);
+
+    if (M == 1) {
+      mat = b.data<DataType>();
+      vec = a.data<DataType>();
+      rows = N;
+      std::swap(mat_strides, vec_strides);
+    } else {
+      mat = a.data<DataType>();
+      vec = b.data<DataType>();
+      rows = M;
+    }
+    uint32_t num_blocks_x = (rows + rows_per_block - 1) / rows_per_block;
+    int n_per_t;
+    if (K % 128 == 0 && is_aligned<4>(mat) && is_aligned<4>(vec)) {
+      n_per_t = 4;
+    } else if (K % 64 == 0 && is_aligned<2>(mat) && is_aligned<2>(vec)) {
+      n_per_t = 2;
+    } else {
+      n_per_t = 1;
+    }
+    dispatch_n_per_thread(n_per_t, [&](auto n_per_thread) {
+      if (batch_count == 1) {
+        auto kernel = gemv_single<DataType, rows_per_block, n_per_thread()>;
+        encoder.add_kernel_node(
+            kernel,
+            num_blocks_x,
+            block_dims,
+            0,
+            mat,
+            vec,
+            out.data<DataType>(),
+            rows,
+            cols);
+      } else {
+        auto kernel = gemv_batched<DataType, rows_per_block, n_per_thread()>;
+        encoder.add_kernel_node(
+            kernel,
+            dim3{num_blocks_x, batch_count},
+            block_dims,
+            0,
+            mat,
+            vec,
+            out.data<DataType>(),
+            rows,
+            cols,
+            const_param(batch_shape),
+            mat_strides,
+            vec_strides,
+            batch_shape.size());
+      }
+    });
+  });
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/gemv.h

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

mlx/backend/cuda/gemms/steel_gemm.cu

@@ -0,0 +1,301 @@
+#include "mlx/backend/common/matmul.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/utils.cuh"
+#include "mlx/backend/cuda/gemms/steel_gemm.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+#include <numeric>
+
+#include <cooperative_groups.h>
+
+#include "mlx/backend/cuda/steel/gemm.cuh"
+#include "mlx/backend/cuda/steel/mma.cuh"
+#include "mlx/backend/cuda/steel/tiles.cuh"
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+struct GemmParams {
+  int M;
+  int N;
+  int K;
+  int lda;
+  int ldb;
+  int ldd;
+
+  int NblockM;
+  int NblockN;
+  int NblockK;
+};
+
+template <
+    typename T,
+    int BM,
+    int BN,
+    int BK,
+    int WM,
+    int WN,
+    bool transpose_a,
+    bool transpose_b,
+    int SL,
+    int Nstages>
+__global__ void kernel_steel_gemm(
+    const T* a,
+    const T* b,
+    T* d,
+    __grid_constant__ const GemmParams params) {
+  const int bM_idx = (blockIdx.y << SL) + (blockIdx.x & ((1 << SL) - 1));
+  const int bN_idx = blockIdx.x >> SL;
+
+  if (params.NblockN <= bN_idx || params.NblockM <= bM_idx) {
+    return;
+  }
+
+  const int d_row = bM_idx * BM;
+  const int d_col = bN_idx * BN;
+  const size_t d_row_long = size_t(d_row);
+  const size_t d_col_long = size_t(d_col);
+
+  a += transpose_a ? d_row_long : d_row_long * params.K;
+  b += transpose_b ? d_col_long * params.K : d_col_long;
+  d += d_row_long * params.ldd + d_col_long;
+
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<32>(block);
+
+  const int lane_idx = warp.thread_rank();
+  const int warp_idx = warp.meta_group_rank();
+
+  const int wm = warp_idx / WN;
+  const int wn = warp_idx % WN;
+
+  constexpr int SM = BM / WM;
+  constexpr int SN = BN / WN;
+  constexpr int SK = BK;
+  constexpr int TK = SK / 16;
+
+  constexpr int NUM_WARPS = WM * WN;
+
+  // Allocate shared memory
+  extern __shared__ char shmem[];
+  SharedTile<T, BM, BK>(&as)[Nstages] =
+      *(SharedTile<T, BM, BK>(*)[Nstages])(&shmem[0]);
+  SharedTile<T, BN, BK>(&bs)[Nstages] = *(SharedTile<T, BN, BK>(*)[Nstages])(
+      &shmem[sizeof(T) * Nstages * BM * BK]);
+
+  // Allocate registers for the MMA
+  RegisterTile<float, SM, SN> C;
+  RegisterTile<T, SM, 16> A[TK];
+  RegisterTile<T, SN, 16> B[TK];
+
+  // Zero the accumulators
+  C.fill(0);
+
+  // Start gmem -> smem copies
+  int k_block_read = 0;
+
+  MLX_UNROLL
+  for (int bk = 0; bk < (Nstages - 1); bk++) {
+    load_async<NUM_WARPS>(
+        as[bk], as[bk].base_addr(), a + k_block_read, params.K);
+    load_async<NUM_WARPS>(
+        bs[bk], bs[bk].base_addr(), b + k_block_read, params.K);
+    k_block_read += BK;
+    cp_async_commit();
+  }
+
+  int smem_pipe_read = 0;
+  int smem_pipe_write = Nstages - 1;
+
+  // Wait till only 1 remains laoding
+  cp_async_wait<1>();
+  block.sync();
+
+  const int offset_m = wm * SM;
+  const int offset_n = wn * SN;
+
+  // Start smem -> register copy
+  A[0].load(
+      as[smem_pipe_read],
+      as[smem_pipe_read].base_addr(),
+      offset_m + lane_idx % 16,
+      lane_idx / 16 * 8);
+  B[0].load(
+      bs[smem_pipe_read],
+      bs[smem_pipe_read].base_addr(),
+      offset_n + lane_idx % 16,
+      lane_idx / 16 * 8);
+
+  // Main loop
+  for (int kb = 0; kb < params.NblockK; kb++) {
+    // Prepare next registers
+    {
+      A[1].load(
+          as[smem_pipe_read],
+          as[smem_pipe_read].base_addr(),
+          offset_m + lane_idx % 16,
+          16 + lane_idx / 16 * 8);
+      B[1].load(
+          bs[smem_pipe_read],
+          bs[smem_pipe_read].base_addr(),
+          offset_n + lane_idx % 16,
+          16 + lane_idx / 16 * 8);
+    }
+
+    // Prepare next smem
+    if ((kb + Nstages - 1) < params.NblockK) {
+      load_async<NUM_WARPS>(
+          as[smem_pipe_write],
+          as[smem_pipe_write].base_addr(),
+          a + k_block_read,
+          params.K);
+      load_async<NUM_WARPS>(
+          bs[smem_pipe_write],
+          bs[smem_pipe_write].base_addr(),
+          b + k_block_read,
+          params.K);
+    }
+    k_block_read += BK;
+
+    cp_async_commit();
+
+    smem_pipe_write = smem_pipe_read;
+    smem_pipe_read = smem_pipe_read + 1;
+    smem_pipe_read = (smem_pipe_read == Nstages) ? 0 : smem_pipe_read;
+
+    // Do current gemm
+    mma_t(C, A[0], B[0]);
+
+    // Do wait for next register
+    cp_async_wait<1>();
+    block.sync();
+
+    // Prepare next register (smem_pipe_read has moved to the next)
+    {
+      A[0].load(
+          as[smem_pipe_read],
+          as[smem_pipe_read].base_addr(),
+          offset_m + lane_idx % 16,
+          lane_idx / 16 * 8);
+      B[0].load(
+          bs[smem_pipe_read],
+          bs[smem_pipe_read].base_addr(),
+          offset_n + lane_idx % 16,
+          lane_idx / 16 * 8);
+    }
+
+    // Do current gemm
+    mma_t(C, A[1], B[1]);
+  }
+
+  // Wait and clear
+  cp_async_wait_all();
+  block.sync();
+
+  C.store_global(d, params.ldd, offset_m, offset_n);
+}
+
+} // namespace cu
+
+void dispatch_steel_gemm(
+    const Stream& s,
+    cu::CommandEncoder& encoder,
+    const array& a,
+    const array& b,
+    array& d,
+    int M,
+    int N,
+    int K,
+    int lda,
+    int ldb,
+    int ldd,
+    bool a_transposed,
+    bool b_transposed) {
+  using DataType = cuda_type_t<float16_t>;
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(d);
+
+  constexpr int BM = 128;
+  constexpr int BN = 128;
+  constexpr int BK = 32;
+
+  constexpr int WM = 2;
+  constexpr int WN = 2;
+
+  constexpr int SL = 0;
+  constexpr int Nstages = 3;
+
+  constexpr uint32_t smem_bytes = BK * (BM + BN) * Nstages * sizeof(DataType);
+
+  const int NblockM = (M + BM - 1) / BM;
+  const int NblockN = (N + BN - 1) / BN;
+  const int NblockK = (K + BK - 1) / BK;
+
+  cu::GemmParams params{
+      /* int M = */ M,
+      /* int N = */ N,
+      /* int K = */ K,
+      /* int lda = */ lda,
+      /* int ldb = */ ldb,
+      /* int ldd = */ ldd,
+
+      /* int NblockM = */ NblockM,
+      /* int NblockN = */ NblockN,
+      /* int NblockK = */ NblockK,
+  };
+
+  // Prepare launch grid params
+  int tile = 1 << SL;
+  int tm = (NblockM + tile - 1) / tile;
+  int tn = NblockN * tile;
+
+  dim3 grid_dim(tn, tm, 1);
+  dim3 block_dim(32 * WM * WN, 1, 1);
+
+  dispatch_bool(a_transposed, [&](auto ta_) {
+    dispatch_bool(b_transposed, [&](auto tb_) {
+      constexpr bool ta = ta_.value;
+      constexpr bool tb = tb_.value;
+
+      auto kernel = cu::ab_t_aligned<DataType, BM, BN, BK>;
+      cudaFuncSetAttribute(
+          kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_bytes);
+
+      encoder.add_kernel_node(
+          kernel,
+          grid_dim,
+          block_dim,
+          smem_bytes,
+          a.data<DataType>(),
+          b.data<DataType>(),
+          d.data<DataType>(),
+          N,
+          K);
+
+      //   auto kernel = cu::kernel_steel_gemm<DataType, BM, BN, BK, WM, WN, ta,
+      //   tb, SL, Nstages>;
+
+      //   cudaFuncSetAttribute(kernel,
+      //   cudaFuncAttributeMaxDynamicSharedMemorySize, smem_bytes);
+
+      //   encoder.add_kernel_node(
+      //       kernel,
+      //       grid_dim,
+      //       block_dim,
+      //       smem_bytes,
+      //       a.data<DataType>(),
+      //       b.data<DataType>(),
+      //       d.data<DataType>(),
+      //       params);
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/gemms/steel_gemm.h

@@ -0,0 +1,27 @@
+#pragma once
+
+#include "mlx/backend/common/matmul.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+#include <numeric>
+
+namespace mlx::core {
+
+void dispatch_steel_gemm(
+    const Stream& s,
+    cu::CommandEncoder& encoder,
+    const array& a,
+    const array& b,
+    array& d,
+    int M,
+    int N,
+    int K,
+    int lda,
+    int ldb,
+    int ldd,
+    bool a_transposed,
+    bool b_transposed);
+
+} // namespace mlx::core

mlx/backend/cuda/indexing.cpp

@@ -29,12 +29,12 @@ void append_indices_arg(
     const std::vector<array>& inputs,
     int nidx,
     int idx_ndim) {
-  std::vector<const void*> indices(nidx);
+  SmallVector<const void*> indices(nidx);
   for (int i = 0; i < nidx; ++i) {
     indices[i] = inputs[i + 1].data<void>();
   }
   args.append(std::move(indices));
-  std::vector<int32_t> indices_shape(nidx * idx_ndim);
+  SmallVector<int32_t> indices_shape(nidx * idx_ndim);
   for (int i = 0; i < nidx; ++i) {
     std::copy_n(
         inputs[i + 1].shape().begin(),
@@ -42,7 +42,7 @@ void append_indices_arg(
         indices_shape.data() + i * idx_ndim);
   }
   args.append(std::move(indices_shape));
-  std::vector<int64_t> indices_strides(nidx * idx_ndim);
+  SmallVector<int64_t> indices_strides(nidx * idx_ndim);
   for (int i = 0; i < nidx; ++i) {
     std::copy_n(
         inputs[i + 1].strides().begin(),
@@ -110,7 +110,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append<int32_t>(src.ndim());
   args.append_ndim(slice_sizes_);
   args.append(slice_size);
-  args.append(axes_);
+  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
@@ -128,8 +128,8 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   encoder.set_output_array(out);
 
   auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(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) {
@@ -211,7 +211,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append_ndim(out.shape());
   args.append_ndim(out.strides());
   args.append<int32_t>(out.ndim());
-  args.append(axes_);
+  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
@@ -229,8 +229,8 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
   encoder.set_output_array(out);
   auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(kernel, upd, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(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) {
@@ -317,8 +317,8 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
   encoder.set_output_array(out);
   auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(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) {
@@ -421,8 +421,8 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
   encoder.set_output_array(out);
   auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  auto [num_blocks, block_dims] = get_launch_args(idx, large);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 } // namespace mlx::core

mlx/backend/cuda/iterators/general_iterator.cuh

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

mlx/backend/cuda/iterators/strided_iterator.cuh

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

mlx/backend/cuda/jit_module.cpp

@@ -9,7 +9,6 @@
 #include <cstdlib>
 #include <filesystem>
 #include <fstream>
-#include <unordered_map>
 
 #include <fmt/format.h>
 #include <nvrtc.h>
@@ -330,11 +329,16 @@ CUfunction JitModule::get_kernel(const std::string& kernel_name) {
   return it->second;
 }
 
+std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
+  static std::unordered_map<std::string, JitModule> map;
+  return map;
+}
+
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,
     const KernelBuilder& builder) {
-  static std::unordered_map<std::string, JitModule> map;
+  auto& map = get_jit_module_cache();
   auto it = map.find(name);
   if (it == map.end()) {
     it = map.try_emplace(name, cu::device(device), name, builder).first;

mlx/backend/cuda/jit_module.h

@@ -40,19 +40,14 @@ struct KernelArgs {
   }
 
   template <typename T>
-  void append(std::vector<T> vec) {
-    if (vec.empty()) {
-      // The nullptr can not be used as arg, pass something not null.
-      append(std::monostate{});
-    } else {
-      append_ptr(vec.data());
-      storage_.emplace_back(std::move(vec));
-    }
+  void append(SmallVector<T> vec) {
+    storage_.emplace_back(std::move(vec));
+    append_ptr(std::get<SmallVector<T>>(storage_.back()).data());
   }
 
   // Make sure the arg is copied to an array with size of NDIM.
   template <size_t NDIM = MAX_NDIM, typename T>
-  void append_ndim(std::vector<T> vec) {
+  void append_ndim(SmallVector<T> vec) {
     if (vec.size() > NDIM) {
       throw std::runtime_error(
           fmt::format("ndim can not be larger than {}.", NDIM));
@@ -76,9 +71,9 @@ struct KernelArgs {
       int32_t,
       uint32_t,
       int64_t,
-      std::vector<const void*>,
-      std::vector<int32_t>,
-      std::vector<int64_t>>;
+      SmallVector<const void*>,
+      SmallVector<int32_t>,
+      SmallVector<int64_t>>;
   std::deque<Arg> storage_;
 };
 
@@ -99,6 +94,8 @@ class JitModule {
   std::unordered_map<std::string, CUfunction> kernels_;
 };
 
+std::unordered_map<std::string, JitModule>& get_jit_module_cache();
+
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,

mlx/backend/cuda/kernel_utils.cu

@@ -30,4 +30,25 @@ std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2) {
   return std::make_pair(dim3(gx, gy, gz), dim3(bx, by, bz));
 }
 
+std::tuple<dim3, uint> get_launch_args(
+    size_t size,
+    const Shape& shape,
+    const Strides& strides,
+    bool large,
+    int work_per_thread) {
+  size_t nthreads = cuda::ceil_div(size, work_per_thread);
+  uint block_dim = 1024;
+  if (block_dim > nthreads) {
+    block_dim = nthreads;
+  }
+  dim3 num_blocks;
+  if (large) {
+    num_blocks = get_2d_grid_dims(shape, strides, work_per_thread);
+    num_blocks.x = cuda::ceil_div(num_blocks.x, block_dim);
+  } else {
+    num_blocks.x = cuda::ceil_div(nthreads, block_dim);
+  }
+  return std::make_tuple(num_blocks, block_dim);
+}
+
 } // namespace mlx::core

mlx/backend/cuda/kernel_utils.cuh

@@ -101,7 +101,7 @@ 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>
-inline cuda::std::array<T, NDIM> const_param(const std::vector<T>& vec) {
+inline cuda::std::array<T, NDIM> const_param(const SmallVector<T>& vec) {
   if (vec.size() > NDIM) {
     throw std::runtime_error(
         fmt::format("ndim can not be larger than {}.", NDIM));
@@ -120,53 +120,19 @@ dim3 get_2d_grid_dims(
     size_t divisor);
 std::pair<dim3, dim3> get_grid_and_block(int dim0, int dim1, int dim2);
 
-// Return a block size that achieves maximum potential occupancy for kernel.
-template <typename T>
-inline uint max_occupancy_block_dim(T kernel) {
-  int _, block_dim;
-  if constexpr (std::is_same_v<T, CUfunction>) {
-    CHECK_CUDA_ERROR(
-        cuOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel, 0, 0, 0));
-  } else {
-    CHECK_CUDA_ERROR(
-        cudaOccupancyMaxPotentialBlockSize(&_, &block_dim, kernel));
-  }
-  return block_dim;
-}
-
 // Get the num_blocks and block_dims that maximize occupancy for |kernel|,
 // assuming each thread handles |work_per_thread| elements of |arr|.
-template <typename T>
-inline std::tuple<dim3, uint> get_launch_args(
-    T kernel,
+std::tuple<dim3, uint> get_launch_args(
     size_t size,
     const Shape& shape,
     const Strides& strides,
     bool large,
-    int work_per_thread = 1) {
-  size_t nthreads = cuda::ceil_div(size, work_per_thread);
-  uint block_dim = max_occupancy_block_dim(kernel);
-  if (block_dim > nthreads) {
-    block_dim = nthreads;
-  }
-  dim3 num_blocks;
-  if (large) {
-    num_blocks = get_2d_grid_dims(shape, strides, work_per_thread);
-    num_blocks.x = cuda::ceil_div(num_blocks.x, block_dim);
-  } else {
-    num_blocks.x = cuda::ceil_div(nthreads, block_dim);
-  }
-  return std::make_tuple(num_blocks, block_dim);
-}
+    int work_per_thread = 1);
 
-template <typename T>
-inline std::tuple<dim3, uint> get_launch_args(
-    T kernel,
-    const array& arr,
-    bool large,
-    int work_per_thread = 1) {
+inline std::tuple<dim3, uint>
+get_launch_args(const array& arr, bool large, int work_per_thread = 1) {
   return get_launch_args(
-      kernel, arr.size(), arr.shape(), arr.strides(), large, work_per_thread);
+      arr.size(), arr.shape(), arr.strides(), large, work_per_thread);
 }
 
 } // namespace mlx::core

mlx/backend/cuda/layer_norm.cu

@@ -1,7 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 #include "mlx/backend/gpu/copy.h"
@@ -11,8 +10,6 @@
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 #include <nvtx3/nvtx3.hpp>
-#include <cub/block/block_load.cuh>
-#include <cub/block/block_reduce.cuh>
 
 namespace mlx::core {
 
@@ -75,9 +72,11 @@ __global__ void layer_norm(
   float sum = 0;
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS] = {};
-    cub::LoadDirectBlocked(index, x, xn, axis_size);
-    sum += static_cast<float>(cub::ThreadReduce(xn, cuda::std::plus<>{}));
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      sum += static_cast<float>(xn[i]);
+    }
   }
   sum = BlockReduceT{block, temp}.Sum(sum);
 
@@ -88,11 +87,18 @@ __global__ void layer_norm(
   float normalizer = 0;
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size, mean);
-    for (int i = 0; i < N_READS; ++i) {
-      float t = static_cast<float>(xn[i]) - mean;
-      normalizer += t * t;
+    if ((index + 1) * N_READS <= axis_size) {
+      auto xn = load_vector<N_READS>(x, index);
+#pragma unroll
+      for (int i = 0; i < N_READS; ++i) {
+        float t = static_cast<float>(xn[i]) - mean;
+        normalizer += t * t;
+      }
+    } else {
+      for (int i = index * N_READS; i < axis_size; ++i) {
+        float t = static_cast<float>(x[i]) - mean;
+        normalizer += t * t;
+      }
     }
   }
   normalizer = BlockReduceT{block, temp}.Sum(normalizer);
@@ -101,17 +107,15 @@ __global__ void layer_norm(
   // Outputs.
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS];
-    T wn[N_READS];
-    T bn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(b, b_stride), bn, axis_size);
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+    auto bn = load_vector<N_READS>(b, index, axis_size, b_stride, T(0));
+#pragma unroll
     for (int i = 0; i < N_READS; ++i) {
       float norm = (static_cast<float>(xn[i]) - mean) * normalizer;
       xn[i] = wn[i] * static_cast<T>(norm) + bn[i];
     }
-    cub::StoreDirectBlocked(index, out, xn, axis_size);
+    store_vector<N_READS>(out, index, xn, axis_size);
   }
 }
 
@@ -144,9 +148,11 @@ __global__ void layer_norm_vjp(
   float sum = 0;
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS] = {};
-    cub::LoadDirectBlocked(index, x, xn, axis_size);
-    sum += static_cast<float>(cub::ThreadReduce(xn, cuda::std::plus<>{}));
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      sum += static_cast<float>(xn[i]);
+    }
   }
   sum = BlockReduceF{block, temp.f}.Sum(sum);
 
@@ -156,19 +162,28 @@ __global__ void layer_norm_vjp(
   // Normalizer.
   float3 factors = {};
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
-    T xn[N_READS];
-    T wn[N_READS] = {};
-    T gn[N_READS] = {};
     auto index = r * BLOCK_DIM + block.thread_rank();
-    cub::LoadDirectBlocked(index, x, xn, axis_size, mean);
-    cub::LoadDirectBlocked(index, g, gn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
-    for (int i = 0; i < N_READS; i++) {
-      float t = static_cast<float>(xn[i]) - mean;
-      float wi = wn[i];
-      float gi = gn[i];
-      float wg = wi * gi;
-      factors = plus_f3(factors, {wg, wg * t, t * t});
+    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
+    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+
+    if ((index + 1) * N_READS <= axis_size) {
+      auto xn = load_vector<N_READS>(x, index);
+#pragma unroll
+      for (int i = 0; i < N_READS; ++i) {
+        float t = static_cast<float>(xn[i]) - mean;
+        float wi = wn[i];
+        float gi = gn[i];
+        float wg = wi * gi;
+        factors = plus_f3(factors, {wg, wg * t, t * t});
+      }
+    } else {
+      for (int i = index * N_READS; i < axis_size; ++i) {
+        float t = static_cast<float>(x[i]) - mean;
+        float wi = wn[i];
+        float gi = gn[i];
+        float wg = wi * gi;
+        factors = plus_f3(factors, {wg, wg * t, t * t});
+      }
     }
   }
   factors = BlockReduceF3{block, temp.f3}.Reduce(factors, plus_f3, {});
@@ -180,12 +195,10 @@ __global__ void layer_norm_vjp(
   // Outputs.
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS];
-    T wn[N_READS];
-    T gn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size);
-    cub::LoadDirectBlocked(index, g, gn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
+    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+
     for (int i = 0; i < N_READS; i++) {
       float xi = (static_cast<float>(xn[i]) - mean) * normalizer;
       float wi = wn[i];
@@ -195,9 +208,9 @@ __global__ void layer_norm_vjp(
         wn[i] = gi * xi;
       }
     }
-    cub::StoreDirectBlocked(index, gx, xn, axis_size);
+    store_vector<N_READS>(gx, index, xn, axis_size);
     if constexpr (HAS_W) {
-      cub::StoreDirectBlocked(index, gw, wn, axis_size);
+      store_vector<N_READS>(gw, index, wn, axis_size);
     }
   }
 }
@@ -258,14 +271,15 @@ void LayerNorm::eval_gpu(
   encoder.set_input_array(b);
   encoder.set_output_array(out);
   dispatch_float_types(out.dtype(), "layernorm", [&](auto type_tag) {
-    constexpr uint32_t N_READS = 4;
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr int N_READS = 16 / sizeof(DataType);
     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>(),
@@ -365,10 +379,10 @@ void LayerNormVJP::eval_gpu(
   encoder.set_output_array(gw_temp);
   dispatch_float_types(gx.dtype(), "layernorm_vjp", [&](auto type_tag) {
     dispatch_bool(has_w, [&](auto has_w_constant) {
-      constexpr int N_READS = 4;
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      constexpr int N_READS = 16 / sizeof(DataType);
       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,
@@ -378,6 +392,7 @@ void LayerNormVJP::eval_gpu(
                 kernel,
                 n_rows,
                 block_dim(),
+                0,
                 x.data<DataType>(),
                 w.data<DataType>(),
                 g.data<DataType>(),

mlx/backend/cuda/logsumexp.cu

@@ -43,20 +43,19 @@ __global__ void logsumexp(const T* in, T* out, int axis_size) {
   AccT maxval = Limits<AccT>::finite_min();
   AccT normalizer = 0;
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
-    AccT vals[N_READS];
-    cub::LoadDirectBlocked(
-        r * BLOCK_DIM + block.thread_rank(),
-        make_cast_iterator<AccT>(in),
-        vals,
-        axis_size,
-        Limits<AccT>::min());
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    auto vals = load_vector<N_READS>(in, index, axis_size, Limits<T>::min());
     prevmax = maxval;
-    maxval = max_op(maxval, cub::ThreadReduce(vals, max_op));
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      maxval = max_op(maxval, static_cast<AccT>(vals[i]));
+    }
     // Online normalizer calculation for softmax:
     // https://github.com/NVIDIA/online-softmax
     normalizer = normalizer * softmax_exp(prevmax - maxval);
     for (int i = 0; i < N_READS; i++) {
-      normalizer = normalizer + softmax_exp(vals[i] - maxval);
+      normalizer =
+          normalizer + softmax_exp(static_cast<AccT>(vals[i]) - maxval);
     }
   }
 
@@ -143,14 +142,15 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
   encoder.set_input_array(in);
   encoder.set_output_array(out);
   dispatch_float_types(out.dtype(), "logsumexp", [&](auto type_tag) {
-    constexpr int N_READS = 4;
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr int N_READS = 16 / sizeof(DataType);
     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/lru_cache.h

@@ -0,0 +1,159 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <cstring>
+#include <list>
+#include <unordered_map>
+#include <utility>
+
+namespace mlx::core {
+
+template <
+    typename K,
+    typename V,
+    template <typename...> typename M = std::unordered_map>
+class LRUCache {
+ public:
+  using value_type = std::pair<K, V>;
+  using list_type = std::list<value_type>;
+  using iterator = typename list_type::iterator;
+  using const_iterator = typename list_type::const_iterator;
+  using map_type = M<K, iterator>;
+
+  explicit LRUCache(size_t capacity) : capacity_(capacity) {
+    if (capacity == 0) {
+      throw std::runtime_error("LRUCache requires capacity > 0.");
+    }
+  }
+
+  size_t size() const {
+    return map_.size();
+  }
+  size_t capacity() const {
+    return capacity_;
+  }
+  bool empty() const {
+    return vlist_.empty();
+  }
+
+  void resize(size_t new_capacity) {
+    capacity_ = new_capacity;
+    trim();
+  }
+
+  iterator begin() {
+    return vlist_.begin();
+  }
+  const_iterator begin() const {
+    return vlist_.begin();
+  }
+  iterator end() {
+    return vlist_.end();
+  }
+  const_iterator end() const {
+    return vlist_.end();
+  }
+
+  void clear() {
+    map_.clear();
+    vlist_.clear();
+  }
+
+  iterator find(const K& key) {
+    auto it = map_.find(key);
+    if (it == map_.end())
+      return end();
+    vlist_.splice(vlist_.begin(), vlist_, it->second);
+    return it->second;
+  }
+
+  template <typename U>
+  std::pair<iterator, bool> emplace(const K& key, U&& value) {
+    auto it = map_.find(key);
+    if (it != map_.end()) {
+      vlist_.splice(vlist_.begin(), vlist_, it->second);
+      return {it->second, false};
+    }
+
+    vlist_.emplace_front(key, std::forward<U>(value));
+    map_[key] = vlist_.begin();
+
+    trim();
+
+    return {vlist_.begin(), true};
+  }
+
+  iterator erase(iterator pos) {
+    map_.erase(pos->first);
+    return vlist_.erase(pos);
+  }
+
+  V& operator[](const K& key) {
+    auto it = find(key);
+    if (it == end()) {
+      it = emplace(key, V{}).first;
+    }
+    return it->second;
+  }
+
+ private:
+  void trim() {
+    while (map_.size() > capacity_) {
+      auto last = std::prev(vlist_.end());
+      map_.erase(last->first);
+      vlist_.pop_back();
+    }
+  }
+
+  list_type vlist_;
+  map_type map_;
+  size_t capacity_;
+};
+
+// Turn a POD struct into a container key by doing bytes compare.
+template <typename T>
+struct BytesKey {
+  T pod;
+  static_assert(std::is_standard_layout_v<T>, "T is not POD");
+
+  BytesKey(T pod) : pod(std::move(pod)) {}
+
+  BytesKey(const BytesKey& other) {
+    memcpy(&pod, &other.pod, sizeof(T));
+  }
+
+  BytesKey(BytesKey&& other) {
+    memcpy(&pod, &other.pod, sizeof(T));
+  }
+
+  bool operator==(const BytesKey& other) const {
+    auto* ptr1 = reinterpret_cast<const uint8_t*>(&pod);
+    auto* ptr2 = reinterpret_cast<const uint8_t*>(&other.pod);
+    return memcmp(ptr1, ptr2, sizeof(T)) == 0;
+  }
+};
+
+// Compute hash according to the bytes value of T.
+template <typename T>
+struct BytesHash {
+  static_assert(std::is_standard_layout_v<T>, "T is not POD");
+
+  size_t operator()(const T& pod) const {
+    auto* ptr = reinterpret_cast<const uint8_t*>(&pod);
+    uint32_t value = 0x811C9DC5;
+    for (int i = 0; i < sizeof(T); ++i) {
+      value ^= ptr[i];
+      value *= 0x01000193;
+    }
+    return value;
+  }
+};
+
+template <typename K, typename V>
+using BytesKeyHashMap = std::unordered_map<K, V, BytesHash<K>>;
+
+template <typename K, typename V>
+using LRUBytesKeyCache = LRUCache<BytesKey<K>, V, BytesKeyHashMap>;
+
+} // namespace mlx::core

mlx/backend/cuda/matmul.cpp

@@ -2,289 +2,17 @@
 
 #include "mlx/backend/common/matmul.h"
 #include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/gemms/gemv.h"
 #include "mlx/backend/gpu/copy.h"
-#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
-#include "mlx/utils.h"
 
-#include <cublasLt.h>
-#include <fmt/format.h>
+#include "mlx/backend/cuda/gemms/steel_gemm.h"
+
 #include <nvtx3/nvtx3.hpp>
-
 #include <numeric>
 
 namespace mlx::core {
-
-namespace cu {
-
-#define CHECK_CUBLAS_ERROR(cmd) check_cublas_error(#cmd, (cmd))
-
-void check_cublas_error(const char* name, cublasStatus_t err) {
-  if (err != CUBLAS_STATUS_SUCCESS) {
-    // TODO: Use cublasGetStatusString when it is widely available.
-    throw std::runtime_error(
-        fmt::format("{} failed with code: {}.", name, static_cast<int>(err)));
-  }
-}
-
-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(
-      Device& device,
-      Dtype dtype,
-      bool a_transposed,
-      uint64_t a_rows,
-      uint64_t a_cols,
-      int64_t lda,
-      bool b_transposed,
-      uint64_t b_rows,
-      uint64_t b_cols,
-      int64_t ldb,
-      int32_t batch_count,
-      int64_t a_batch_stride,
-      int64_t b_batch_stride)
-      : handle_(device.lt_handle()), pref_(cublas_preference(device)) {
-    heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
-
-    auto scale_type = dtype_to_cuda_type(dtype);
-    if (dtype == bfloat16 || dtype == float16) {
-      scale_type = CUDA_R_32F;
-    }
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescCreate(
-        &matmul_desc_, dtype_to_compute_type(dtype), scale_type));
-    int32_t pointer_mode = CUBLASLT_POINTER_MODE_HOST;
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
-        matmul_desc_,
-        CUBLASLT_MATMUL_DESC_POINTER_MODE,
-        &pointer_mode,
-        sizeof(int32_t)));
-    cublasOperation_t op = CUBLAS_OP_N;
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
-        matmul_desc_,
-        CUBLASLT_MATMUL_DESC_TRANSA,
-        &op,
-        sizeof(cublasOperation_t)));
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
-        matmul_desc_,
-        CUBLASLT_MATMUL_DESC_TRANSB,
-        &op,
-        sizeof(cublasOperation_t)));
-
-    auto type = dtype_to_cuda_type(dtype);
-    a_desc_ = create_matrix_layout(
-        type, a_rows, a_cols, a_transposed, lda, batch_count, a_batch_stride);
-    b_desc_ = create_matrix_layout(
-        type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
-    out_desc_ = create_matrix_layout(
-        type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
-  }
-
-  MatMul(
-      Device& device,
-      Dtype dtype,
-      bool a_transposed,
-      uint64_t a_rows,
-      uint64_t a_cols,
-      int64_t lda,
-      bool b_transposed,
-      uint64_t b_rows,
-      uint64_t b_cols,
-      int64_t ldb,
-      int64_t ldc,
-      int32_t batch_count,
-      int64_t a_batch_stride,
-      int64_t b_batch_stride,
-      int64_t c_batch_stride)
-      : MatMul(
-            device,
-            dtype,
-            a_transposed,
-            a_rows,
-            a_cols,
-            lda,
-            b_transposed,
-            b_rows,
-            b_cols,
-            ldb,
-            batch_count,
-            a_batch_stride,
-            b_batch_stride) {
-    auto type = dtype_to_cuda_type(dtype);
-    c_desc_ = create_matrix_layout(
-        type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
-  }
-
-  ~MatMul() {
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(a_desc_));
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(b_desc_));
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(c_desc_));
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(out_desc_));
-    CHECK_CUBLAS_ERROR(cublasLtMatmulDescDestroy(matmul_desc_));
-  }
-
-  void run(
-      cu::CommandEncoder& encoder,
-      void* out,
-      void* a,
-      void* b,
-      void* c = nullptr,
-      float alpha = 1,
-      float beta = 0) {
-    if (heuristic_.state != CUBLAS_STATUS_SUCCESS) {
-      int ret = 0;
-      CHECK_CUBLAS_ERROR(cublasLtMatmulAlgoGetHeuristic(
-          handle_,
-          matmul_desc_,
-          a_desc_,
-          b_desc_,
-          out_desc_,
-          out_desc_,
-          pref_,
-          1,
-          &heuristic_,
-          &ret));
-      if (ret == 0) {
-        throw std::runtime_error("Can not find algorithm for matmul.");
-      }
-    }
-
-    void* workspace_ptr = nullptr;
-    if (heuristic_.workspaceSize > 0) {
-      array workspace(
-          allocator::malloc(heuristic_.workspaceSize),
-          {static_cast<int>(heuristic_.workspaceSize)},
-          int8);
-      encoder.add_temporary(workspace);
-      workspace_ptr = workspace.data<void>();
-    }
-
-    auto capture = encoder.capture_context();
-    CHECK_CUBLAS_ERROR(cublasLtMatmul(
-        handle_,
-        matmul_desc_,
-        &alpha,
-        a,
-        a_desc_,
-        b,
-        b_desc_,
-        &beta,
-        c ? c : out,
-        c ? c_desc_ : out_desc_,
-        out,
-        out_desc_,
-        &heuristic_.algo,
-        workspace_ptr,
-        heuristic_.workspaceSize,
-        encoder.stream()));
-  }
-
- private:
-  cublasComputeType_t dtype_to_compute_type(Dtype dtype) {
-    switch (dtype) {
-      case float16:
-        return CUBLAS_COMPUTE_32F;
-      case bfloat16:
-        return CUBLAS_COMPUTE_32F;
-      case float32:
-        return mlx::core::env::enable_tf32() ? CUBLAS_COMPUTE_32F_FAST_TF32
-                                             : CUBLAS_COMPUTE_32F;
-      case float64:
-      case complex64:
-        return CUBLAS_COMPUTE_64F;
-      default:
-        throw std::runtime_error(fmt::format(
-            "Unsupported dtype in MatMul: {}.", dtype_to_string(dtype)));
-    }
-  }
-
-  cudaDataType_t dtype_to_cuda_type(Dtype dtype) {
-    switch (dtype) {
-      case float16:
-        return CUDA_R_16F;
-      case bfloat16:
-        return CUDA_R_16BF;
-      case float32:
-        return CUDA_R_32F;
-      case float64:
-        return CUDA_R_64F;
-      case complex64:
-        return CUDA_C_32F;
-      default:
-        throw std::runtime_error(fmt::format(
-            "Unsupported dtype in MatMul: {}.", dtype_to_string(dtype)));
-    }
-  }
-
-  cublasLtMatrixLayout_t create_matrix_layout(
-      cudaDataType_t type,
-      uint64_t rows,
-      uint64_t cols,
-      bool transposed,
-      int64_t ld,
-      int32_t batch_count,
-      int64_t batch_stride) {
-    cublasLtMatrixLayout_t desc;
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutCreate(&desc, type, rows, cols, ld));
-    cublasLtOrder_t order =
-        transposed ? CUBLASLT_ORDER_COL : CUBLASLT_ORDER_ROW;
-    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-        desc, CUBLASLT_MATRIX_LAYOUT_ORDER, &order, sizeof(cublasLtOrder_t)));
-    if (batch_count > 1) {
-      CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-          desc,
-          CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT,
-          &batch_count,
-          sizeof(int32_t)));
-      CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutSetAttribute(
-          desc,
-          CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET,
-          &batch_stride,
-          sizeof(int64_t)));
-    }
-    return desc;
-  }
-
-  cublasLtMatmulPreference_t pref_{nullptr};
-  cublasLtHandle_t handle_{nullptr};
-  cublasLtMatmulDesc_t matmul_desc_{nullptr};
-  cublasLtMatrixLayout_t a_desc_{nullptr};
-  cublasLtMatrixLayout_t b_desc_{nullptr};
-  cublasLtMatrixLayout_t c_desc_{nullptr};
-  cublasLtMatrixLayout_t out_desc_{nullptr};
-  cublasLtMatmulHeuristicResult_t heuristic_;
-};
-
-} // namespace cu
-
 namespace {
 
 std::tuple<bool, int64_t, array>
@@ -353,10 +81,43 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     batch_shape = {1};
   }
 
+  if (cu::can_use_gemv(M, N, K, a_transposed, b_transposed)) {
+    cu::gemv(
+        a,
+        b,
+        out,
+        M,
+        N,
+        K,
+        batch_count,
+        batch_shape,
+        a_batch_strides,
+        b_batch_strides,
+        encoder);
+    return;
+  }
+
+  if (out.dtype() == float16 && batch_count == 1 && !a_transposed &&
+      b_transposed) {
+    return dispatch_steel_gemm(
+        /* const Stream& s = */ s,
+        /* cu::CommandEncoder& encoder = */ encoder,
+        /* const array& a = */ a,
+        /* const array& b = */ b,
+        /* array& d = */ out,
+        /* int M = */ M,
+        /* int N = */ N,
+        /* int K = */ K,
+        /* int lda = */ lda,
+        /* int ldb = */ ldb,
+        /* int ldd = */ N,
+        /* bool a_transposed = */ a_transposed,
+        /* bool b_transposed = */ b_transposed);
+  }
+
   /////////////////////////////////////////////////////////////////////////////
   // Invoke cublasLt
-
-  cu::MatMul matmul(
+  CublasGemm gemm(
       cu::device(s.device),
       a.dtype(),
       a_transposed,
@@ -370,28 +131,7 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
       batch_shape.back(),
       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);
-  auto concurrent = encoder.concurrent_context();
-  for (size_t i = 0; i < nbatch; ++i) {
-    matmul.run(
-        encoder,
-        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,
-        a.data<int8_t>() + a.itemsize() * a_it.loc,
-        b.data<int8_t>() + b.itemsize() * b_it.loc);
-    a_it.step();
-    b_it.step();
-  }
+  gemm.run(encoder, out, a, b, batch_shape, a_batch_strides, b_batch_strides);
 }
 
 void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -459,7 +199,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   /////////////////////////////////////////////////////////////////////////////
   // Invoke cublasLt
 
-  cu::MatMul matmul(
+  CublasGemm gemm(
       cu::device(s.device),
       a.dtype(),
       a_transposed,
@@ -475,42 +215,18 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       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);
-  auto concurrent = encoder.concurrent_context();
-  for (size_t i = 0; i < nbatch; ++i) {
-    matmul.run(
-        encoder,
-        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,
-        a.data<int8_t>() + a.itemsize() * a_it.loc,
-        b.data<int8_t>() + b.itemsize() * b_it.loc,
-        c.data<int8_t>() + c.itemsize() * c_it.loc,
-        alpha_,
-        beta_);
-    a_it.step();
-    b_it.step();
-    c_it.step();
-  }
+  gemm.run(
+      encoder,
+      out,
+      a,
+      b,
+      c,
+      batch_shape,
+      a_batch_strides,
+      b_batch_strides,
+      c_batch_strides,
+      alpha_,
+      beta_);
 }
 
 } // namespace mlx::core

mlx/backend/cuda/primitives.cpp

@@ -0,0 +1,55 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/distributed/primitives.h"
+#include "mlx/fast_primitives.h"
+#include "mlx/primitives.h"
+
+namespace mlx::core {
+
+#define NO_GPU_MULTI(func)                                             \
+  void func::eval_gpu(                                                 \
+      const std::vector<array>& inputs, std::vector<array>& outputs) { \
+    throw std::runtime_error(#func " has no CUDA implementation.");    \
+  }
+
+#define NO_GPU_USE_FALLBACK(func)     \
+  bool func::use_fallback(Stream s) { \
+    return true;                      \
+  }                                   \
+  NO_GPU_MULTI(func)
+
+#define NO_GPU(func)                                                  \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    throw std::runtime_error(#func " has no CUDA implementation.");   \
+  }
+
+NO_GPU(BlockMaskedMM)
+NO_GPU(DynamicSlice)
+NO_GPU(DynamicSliceUpdate)
+NO_GPU(FFT)
+NO_GPU(GatherMM)
+NO_GPU(GatherQMM)
+NO_GPU(Hadamard)
+NO_GPU(Load)
+NO_GPU_MULTI(LUF)
+NO_GPU_MULTI(QRF)
+NO_GPU(QuantizedMatmul)
+NO_GPU(SegmentedMM)
+NO_GPU_MULTI(SVD)
+NO_GPU(Inverse)
+NO_GPU(Cholesky)
+NO_GPU_MULTI(Eig)
+NO_GPU_MULTI(Eigh)
+
+namespace fast {
+NO_GPU_MULTI(CustomKernel)
+} // namespace fast
+
+namespace distributed {
+NO_GPU_MULTI(AllReduce)
+NO_GPU_MULTI(AllGather)
+NO_GPU_MULTI(Send)
+NO_GPU_MULTI(Recv)
+} // namespace distributed
+
+} // namespace mlx::core

mlx/backend/cuda/primitives.cu

@@ -1,104 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/arange.cuh"
-#include "mlx/backend/cuda/device/fp16_math.cuh"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/distributed/primitives.h"
-#include "mlx/dtype_utils.h"
-#include "mlx/fast_primitives.h"
-#include "mlx/primitives.h"
-
-#include <nvtx3/nvtx3.hpp>
-#include <thrust/device_ptr.h>
-#include <thrust/transform.h>
-
-#include <cassert>
-
-namespace mlx::core {
-
-void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Arange::eval_gpu");
-  assert(inputs.size() == 0);
-  out.set_data(allocator::malloc(out.nbytes()));
-  if (out.size() == 0) {
-    return;
-  }
-  auto& encoder = cu::get_command_encoder(stream());
-  encoder.set_output_array(out);
-  auto capture = encoder.capture_context();
-  dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
-    using CTYPE = MLX_GET_TYPE(type_tag);
-    using OutType = cuda_type_t<CTYPE>;
-    CTYPE step =
-        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
-    thrust::transform(
-        cu::thrust_policy(encoder.stream()),
-        thrust::counting_iterator<uint32_t>(0),
-        thrust::counting_iterator<uint32_t>(out.data_size()),
-        thrust::device_pointer_cast(out.data<OutType>()),
-        cu::Arange<OutType>{
-            static_cast<OutType>(start_), static_cast<OutType>(step)});
-  });
-}
-
-bool fast::ScaledDotProductAttention::use_fallback(
-    const array& q,
-    const array& k,
-    const array& v,
-    bool has_mask,
-    bool has_arr_mask,
-    bool do_causal,
-    Stream s) {
-  return true;
-}
-
-#define NO_GPU_MULTI(func)                                             \
-  void func::eval_gpu(                                                 \
-      const std::vector<array>& inputs, std::vector<array>& outputs) { \
-    throw std::runtime_error(#func " has no CUDA implementation.");    \
-  }
-
-#define NO_GPU_USE_FALLBACK(func)     \
-  bool func::use_fallback(Stream s) { \
-    return true;                      \
-  }                                   \
-  NO_GPU_MULTI(func)
-
-#define NO_GPU(func)                                                  \
-  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
-    throw std::runtime_error(#func " has no CUDA implementation.");   \
-  }
-
-NO_GPU(BlockMaskedMM)
-NO_GPU(Convolution)
-NO_GPU(DynamicSlice)
-NO_GPU(DynamicSliceUpdate)
-NO_GPU(FFT)
-NO_GPU(GatherMM)
-NO_GPU(GatherQMM)
-NO_GPU(Hadamard)
-NO_GPU(Load)
-NO_GPU_MULTI(LUF)
-NO_GPU_MULTI(QRF)
-NO_GPU(QuantizedMatmul)
-NO_GPU(SegmentedMM)
-NO_GPU_MULTI(SVD)
-NO_GPU(Inverse)
-NO_GPU(Cholesky)
-NO_GPU_MULTI(Eig)
-NO_GPU_MULTI(Eigh)
-
-namespace fast {
-NO_GPU(ScaledDotProductAttention)
-NO_GPU_MULTI(CustomKernel)
-} // namespace fast
-
-namespace distributed {
-NO_GPU_MULTI(AllReduce)
-NO_GPU_MULTI(AllGather)
-NO_GPU_MULTI(Send)
-NO_GPU_MULTI(Recv)
-} // namespace distributed
-
-} // namespace mlx::core

mlx/backend/cuda/quantized/affine_quantize.cu

@@ -2,30 +2,17 @@
 
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/backend/gpu/copy.h"
+#include "mlx/backend/cuda/quantized/quantized_utils.cuh"
 #include "mlx/dtype_utils.h"
-#include "mlx/fast_primitives.h"
 
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
-#include <nvtx3/nvtx3.hpp>
 
 namespace mlx::core {
 namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <int bits, int wsize = 8>
-inline constexpr __device__ short get_pack_factor() {
-  return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
-}
-
-template <int bits, int wsize = 8>
-inline constexpr __device__ short get_bytes_per_pack() {
-  constexpr int power_of_2_bits = (bits & (bits - 1)) == 0;
-  return power_of_2_bits ? (wsize / 8) : (bits == 5 ? 5 : 3);
-}
-
 template <typename T, int group_size, int bits>
 __global__ void
 affine_quantize(const T* w, uint8_t* out, T* scales, T* biases, size_t size) {
@@ -240,144 +227,102 @@ __global__ void affine_dequantize(
 }
 
 } // namespace cu
-namespace {
 
-inline array ensure_row_contiguous(
-    const array& x,
+void affine_quantize(
+    const array& w,
+    array& wq,
+    array& scales,
+    array& biases,
+    int group_size_,
+    int bits_,
     cu::CommandEncoder& enc,
     const Stream& s) {
-  if (!x.flags().row_contiguous) {
-    array x_copy = contiguous_copy_gpu(x, s);
-    enc.add_temporary(x_copy);
-    return x_copy;
-  } else {
-    return x;
-  }
-}
-
-} // namespace
-
-template <typename F>
-void dispatch_groups(int group_size, F&& f) {
-  switch (group_size) {
-    case 32:
-      f(std::integral_constant<int, 32>{});
-      break;
-    case 64:
-      f(std::integral_constant<int, 64>{});
-      break;
-    case 128:
-      f(std::integral_constant<int, 128>{});
-      break;
-  }
-}
-
-template <typename F>
-void dispatch_bits(int bits, F&& f) {
-  switch (bits) {
-    case 2:
-      f(std::integral_constant<int, 2>{});
-      break;
-    case 3:
-      f(std::integral_constant<int, 3>{});
-      break;
-    case 4:
-      f(std::integral_constant<int, 4>{});
-      break;
-    case 5:
-      f(std::integral_constant<int, 5>{});
-      break;
-    case 6:
-      f(std::integral_constant<int, 6>{});
-      break;
-    case 8:
-      f(std::integral_constant<int, 8>{});
-      break;
-  }
-}
-
-void fast::AffineQuantize::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  auto& w_pre = inputs[0];
-  auto& out = outputs[0];
-  out.set_data(allocator::malloc(out.nbytes()));
-
-  auto& s = stream();
-  auto& d = cu::device(s.device);
-  auto& enc = d.get_command_encoder(s);
-
-  auto w = ensure_row_contiguous(w_pre, enc, s);
-  enc.set_input_array(w);
-  if (dequantize_) {
-    auto scales = ensure_row_contiguous(inputs[1], enc, s);
-    auto biases = ensure_row_contiguous(inputs[2], enc, s);
-    enc.set_input_array(scales);
-    enc.set_input_array(biases);
-    enc.set_output_array(out);
-  } else {
-    auto& scales = outputs[1];
-    auto& biases = outputs[2];
-    scales.set_data(allocator::malloc(scales.nbytes()));
-    biases.set_data(allocator::malloc(biases.nbytes()));
-    enc.set_output_array(out);
-    enc.set_output_array(scales);
-    enc.set_output_array(biases);
-  }
-
-  auto dtype = dequantize_ ? outputs[0].dtype() : inputs[0].dtype();
-
-  // Treat uint32 as uint8 in kernel
-  int uint8_per_uint32 = 4;
-  int packs_per_int = (bits_ == 3 || bits_ == 5) ? 8
-      : bits_ == 6                               ? 4
-                                                 : 8 / bits_;
-  int per_thread = dequantize_ ? packs_per_int : group_size_ / WARP_SIZE;
-  size_t size =
-      dequantize_ ? out.size() / packs_per_int : w.size() / per_thread;
+  // 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;
 
-  if (dequantize_) {
-    grid_shape.back() *= uint8_per_uint32;
-  } else {
-    grid_shape.back() /= per_thread;
-  }
-
-  dispatch_float_types(dtype, "affine_quantize", [&](auto type_tag) {
+  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 DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-        if (dequantize_) {
-          auto kernel =
-              cu::affine_dequantize<DataType, group_size.value, bits.value>;
-          auto [num_blocks, block_dims] =
-              get_launch_args(kernel, size, grid_shape, w.strides(), large);
-          enc.add_kernel_node(
-              kernel,
-              num_blocks,
-              block_dims,
-              w.data<uint8_t>(),
-              inputs[1].data<DataType>(),
-              inputs[2].data<DataType>(),
-              out.data<DataType>(),
-              out.size());
-        } else {
-          auto kernel =
-              cu::affine_quantize<DataType, group_size.value, bits.value>;
-          auto [num_blocks, block_dims] =
-              get_launch_args(kernel, size, grid_shape, w.strides(), large);
-          enc.add_kernel_node(
-              kernel,
-              num_blocks,
-              block_dims,
-              w.data<DataType>(),
-              out.data<uint8_t>(),
-              outputs[1].data<DataType>(),
-              outputs[2].data<DataType>(),
-              w.size());
-        }
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        auto kernel = cu::affine_quantize<T, group_size.value, bits.value>;
+        auto [num_blocks, block_dims] =
+            get_launch_args(size, grid_shape, w.strides(), large);
+        enc.add_kernel_node(
+            kernel,
+            num_blocks,
+            block_dims,
+            0,
+            w.data<T>(),
+            wq.data<uint8_t>(),
+            scales.data<T>(),
+            biases.data<T>(),
+            w.size());
+      });
+    });
+  });
+}
+
+void affine_dequantize(
+    const array& wq,
+    const array& scales,
+    const array& biases,
+    array& w,
+    int group_size_,
+    int bits_,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  // Calculate how many numbers we pack together. For 2, 4, 8 bits we pack in
+  // one uint8, for 3, 6 in 3 uint8 and for 5 in 5 uint8.
+  constexpr int uint8_per_uint32 = 4;
+  int packs_per_int;
+  switch (bits_) {
+    case 3:
+    case 5:
+      packs_per_int = 8;
+      break;
+    case 6:
+      packs_per_int = 4;
+      break;
+    default:
+      packs_per_int = 8 / bits_;
+  }
+
+  size_t size = w.size() / packs_per_int;
+  bool large = size > UINT_MAX;
+  auto grid_shape = w.shape();
+  grid_shape.back() *= uint8_per_uint32;
+
+  enc.set_input_array(wq);
+  enc.set_input_array(scales);
+  enc.set_input_array(biases);
+  enc.set_output_array(w);
+  dispatch_float_types(w.dtype(), "affine_quantize", [&](auto type_tag) {
+    dispatch_groups(group_size_, [&](auto group_size) {
+      dispatch_bits(bits_, [&](auto bits) {
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        auto kernel = cu::affine_dequantize<T, group_size.value, bits.value>;
+        auto [num_blocks, block_dims] =
+            get_launch_args(size, grid_shape, w.strides(), large);
+        enc.add_kernel_node(
+            kernel,
+            num_blocks,
+            block_dims,
+            0,
+            wq.data<uint8_t>(),
+            scales.data<T>(),
+            biases.data<T>(),
+            w.data<T>(),
+            w.size());
       });
     });
   });

mlx/backend/cuda/quantized/quantized.cpp

@@ -0,0 +1,80 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/quantized/quantized.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/fast_primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace {
+
+inline array ensure_row_contiguous(
+    const array& x,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  if (!x.flags().row_contiguous) {
+    array x_copy = contiguous_copy_gpu(x, s);
+    enc.add_temporary(x_copy);
+    return x_copy;
+  } else {
+    return x;
+  }
+}
+
+inline array ensure_row_contiguous_matrix(
+    const array& x,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  if (x.ndim() < 2) {
+    if (x.strides()[0] == 1) {
+      return x;
+    }
+  } else {
+    auto stride_0 = x.strides()[x.ndim() - 2];
+    auto stride_1 = x.strides()[x.ndim() - 1];
+    if (stride_0 == x.shape(-1) && stride_1 == 1) {
+      return x;
+    }
+  }
+  array x_copy = contiguous_copy_gpu(x, s);
+  enc.add_temporary(x_copy);
+  return x_copy;
+}
+
+} // namespace
+
+void fast::AffineQuantize::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("AffineQuantize::eval_gpu");
+  auto& s = stream();
+  auto& d = cu::device(s.device);
+  auto& enc = d.get_command_encoder(s);
+
+  if (dequantize_) {
+    auto wq = ensure_row_contiguous(inputs[0], enc, s);
+    auto scales = ensure_row_contiguous(inputs[1], enc, s);
+    auto biases = ensure_row_contiguous(inputs[2], enc, s);
+    auto& w = outputs[0];
+
+    w.set_data(allocator::malloc(w.nbytes()));
+
+    affine_dequantize(wq, scales, biases, w, group_size_, bits_, enc, s);
+  } else {
+    auto w = ensure_row_contiguous(inputs[0], enc, s);
+    auto& wq = outputs[0];
+    auto& scales = outputs[1];
+    auto& biases = outputs[2];
+
+    wq.set_data(allocator::malloc(wq.nbytes()));
+    scales.set_data(allocator::malloc(scales.nbytes()));
+    biases.set_data(allocator::malloc(biases.nbytes()));
+
+    affine_quantize(w, wq, scales, biases, group_size_, bits_, enc, s);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/quantized/quantized.h

@@ -0,0 +1,27 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core {
+
+void affine_quantize(
+    const array& w,
+    array& wq,
+    array& scales,
+    array& biases,
+    int group_size_,
+    int bits_,
+    cu::CommandEncoder& enc,
+    const Stream& s);
+
+void affine_dequantize(
+    const array& wq,
+    const array& scales,
+    const array& biases,
+    array& w,
+    int group_size_,
+    int bits_,
+    cu::CommandEncoder& enc,
+    const Stream& s);
+
+} // namespace mlx::core

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

@@ -170,6 +170,7 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
         cu::rbitsc,
         grid,
         block,
+        0,
         keys.data<uint32_t>(),
         out.data<uint8_t>(),
         grid_dims,
@@ -180,6 +181,7 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
         cu::rbits,
         grid,
         block,
+        0,
         keys.data<uint32_t>(),
         out.data<uint8_t>(),
         grid_dims,

mlx/backend/cuda/reduce.cu

@@ -5,8 +5,6 @@
 #include "mlx/backend/gpu/copy.h"
 
 #include <nvtx3/nvtx3.hpp>
-#include <thrust/device_ptr.h>
-#include <thrust/fill.h>
 
 #include <cassert>
 

mlx/backend/cuda/reduce/all_reduce.cu

@@ -120,6 +120,7 @@ void all_reduce(
             kernel,
             blocks,
             threads,
+            0,
             static_cast<T*>(indata),
             intermediate.data<U>(),
             block_step,
@@ -146,6 +147,7 @@ void all_reduce(
           kernel,
           blocks,
           threads,
+          0,
           static_cast<T*>(indata),
           out.data<U>(),
           block_step,

mlx/backend/cuda/reduce/col_reduce.cu

@@ -230,7 +230,7 @@ void col_reduce_looped(
         auto kernel =
             cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
         encoder.add_kernel_node(
-            kernel, grid, blocks, indata, out.data<U>(), args);
+            kernel, grid, blocks, 0, indata, out.data<U>(), args);
       });
     });
   });

mlx/backend/cuda/reduce/init_reduce.cu

@@ -41,7 +41,8 @@ void init_reduce(
       dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
       dim3 block(grid.x < 1024 ? grid.x : 1024, 1, 1);
       grid.x = (grid.x + 1023) / 1024;
-      encoder.add_kernel_node(kernel, grid, block, out.data<U>(), out.size());
+      encoder.add_kernel_node(
+          kernel, grid, block, 0, out.data<U>(), out.size());
     });
   });
 }

mlx/backend/cuda/reduce/row_reduce.cu

@@ -269,7 +269,7 @@ void row_reduce_simple(
 
       int size = plan.shape.back();
       encoder.add_kernel_node(
-          kernel, grid, block, indata, out.data<U>(), out.size(), size);
+          kernel, grid, block, 0, indata, out.data<U>(), out.size(), size);
     });
   });
 }
@@ -322,7 +322,7 @@ void row_reduce_looped(
       });
 
       encoder.add_kernel_node(
-          kernel, grid, block, indata, out.data<U>(), out.size(), args);
+          kernel, grid, block, 0, indata, out.data<U>(), out.size(), args);
     });
   });
 }

mlx/backend/cuda/rms_norm.cu

@@ -1,7 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/iterators/strided_iterator.cuh"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/cuda/reduce/reduce.cuh"
 #include "mlx/backend/gpu/copy.h"
@@ -11,8 +10,6 @@
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 #include <nvtx3/nvtx3.hpp>
-#include <cub/block/block_load.cuh>
-#include <cub/block/block_reduce.cuh>
 
 namespace mlx::core {
 
@@ -58,7 +55,7 @@ __global__ void rms_norm(
     const T* w,
     T* out,
     float eps,
-    int32_t axis_size,
+    uint32_t axis_size,
     int64_t w_stride) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
@@ -73,8 +70,8 @@ __global__ void rms_norm(
   float normalizer = 0;
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size, cast_to<T>(0));
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+#pragma unroll
     for (int i = 0; i < N_READS; ++i) {
       float t = static_cast<float>(xn[i]);
       normalizer += t * t;
@@ -86,15 +83,14 @@ __global__ void rms_norm(
   // Outputs.
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS];
-    T wn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+#pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      float norm = static_cast<float>(xn[i]) * normalizer;
-      xn[i] = wn[i] * static_cast<T>(norm);
+      float y = static_cast<float>(xn[i]) * normalizer;
+      xn[i] = wn[i] * static_cast<T>(y);
     }
-    cub::StoreDirectBlocked(index, out, xn, axis_size);
+    store_vector<N_READS>(out, index, xn, axis_size);
   }
 }
 
@@ -126,13 +122,10 @@ __global__ void rms_norm_vjp(
   // Normalizer.
   float2 factors = {};
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
-    T xn[N_READS];
-    T wn[N_READS] = {};
-    T gn[N_READS] = {};
     auto index = r * BLOCK_DIM + block.thread_rank();
-    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);
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
+    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
     for (int i = 0; i < N_READS; i++) {
       float t = static_cast<float>(xn[i]);
       float wi = wn[i];
@@ -149,12 +142,9 @@ __global__ void rms_norm_vjp(
   // Outputs.
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); ++r) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T xn[N_READS];
-    T wn[N_READS];
-    T gn[N_READS];
-    cub::LoadDirectBlocked(index, x, xn, axis_size);
-    cub::LoadDirectBlocked(index, g, gn, axis_size);
-    cub::LoadDirectBlocked(index, strided_iterator(w, w_stride), wn, axis_size);
+    auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+    auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
+    auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
     for (int i = 0; i < N_READS; i++) {
       float xi = xn[i];
       float wi = wn[i];
@@ -164,9 +154,9 @@ __global__ void rms_norm_vjp(
         wn[i] = static_cast<T>(gi * xi * normalizer);
       }
     }
-    cub::StoreDirectBlocked(index, gx, xn, axis_size);
+    store_vector<N_READS>(gx, index, xn, axis_size);
     if constexpr (HAS_W) {
-      cub::StoreDirectBlocked(index, gw, wn, axis_size);
+      store_vector<N_READS>(gw, index, wn, axis_size);
     }
   }
 }
@@ -224,14 +214,15 @@ void RMSNorm::eval_gpu(
   encoder.set_input_array(w);
   encoder.set_output_array(out);
   dispatch_float_types(out.dtype(), "rms_norm", [&](auto type_tag) {
-    constexpr uint32_t N_READS = 4;
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr int N_READS = 16 / sizeof(DataType);
     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>(),
@@ -313,11 +304,10 @@ void RMSNormVJP::eval_gpu(
   encoder.set_output_array(gw_temp);
   dispatch_float_types(gx.dtype(), "rms_norm_vjp", [&](auto type_tag) {
     dispatch_bool(has_w, [&](auto has_w_constant) {
-      constexpr int N_READS = 4;
+      using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+      constexpr int N_READS = 16 / sizeof(DataType);
       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,
@@ -327,6 +317,7 @@ void RMSNormVJP::eval_gpu(
                 kernel,
                 n_rows,
                 block_dim(),
+                0,
                 x.data<DataType>(),
                 w.data<DataType>(),
                 g.data<DataType>(),

mlx/backend/cuda/rope.cu

@@ -325,6 +325,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),
@@ -341,6 +342,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),
@@ -360,6 +362,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),
@@ -381,6 +384,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),

mlx/backend/cuda/scaled_dot_product_attention.cu

@@ -0,0 +1,781 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/config.h"
+#include "mlx/backend/cuda/device/utils.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/lru_cache.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/fast_primitives.h"
+#include "mlx/transforms_impl.h"
+
+#include <nvtx3/nvtx3.hpp>
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+#define PRAGMA_LOOP_UNROLL #pragma unroll
+
+struct AttnParams {
+  int B;
+  int H;
+  int D;
+
+  int qL;
+  int kL;
+
+  int gqa_factor;
+  float scale;
+
+  int64_t Q_strides[3];
+  int64_t K_strides[3];
+  int64_t V_strides[3];
+  int64_t O_strides[3];
+};
+
+template <typename T, bool do_causal, int D>
+__global__ void kernel_sdpav_1pass(
+    const T* Q,
+    const T* K,
+    const T* V,
+    T* O,
+    __grid_constant__ const AttnParams params) {
+  constexpr int BN = 32;
+  constexpr int BD = 32;
+
+  constexpr int v_per_thread = D / BD;
+
+  const int inner_k_stride = BN * int(params.K_strides[2]);
+  const int inner_v_stride = BN * int(params.V_strides[2]);
+
+  typedef float U;
+
+  U q[v_per_thread];
+  U k[v_per_thread];
+  U o[v_per_thread];
+
+  __shared__ U outputs[BN][BD + 1];
+  __shared__ U max_scores[BN];
+  __shared__ U sum_exp_scores[BN];
+
+  const U scale_log2 = params.scale * 1.44269504089f;
+
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<32>(block);
+
+  const int lane_idx = warp.thread_rank();
+  const int warp_idx = warp.meta_group_rank();
+
+  // Adjust to thread block and thread
+  const int batch_idx = blockIdx.z;
+  const int head_idx = blockIdx.x;
+  const int kv_head_idx = head_idx / params.gqa_factor;
+
+  const int q_seq_idx = blockIdx.y;
+  const int kv_seq_idx = warp_idx;
+
+  Q += batch_idx * params.Q_strides[0] + // Batch
+      head_idx * params.Q_strides[1] + // Head
+      q_seq_idx * params.Q_strides[2]; // Sequence
+
+  K += batch_idx * params.K_strides[0] + // Batch
+      kv_head_idx * params.K_strides[1] + // Head
+      kv_seq_idx * params.K_strides[2]; // Sequence
+
+  V += batch_idx * params.V_strides[0] + // Batch
+      kv_head_idx * params.V_strides[1] + // Head
+      kv_seq_idx * params.V_strides[2]; // Sequence
+
+  O += batch_idx * params.O_strides[0] + // Batch
+      head_idx * params.O_strides[1] + // Head
+      q_seq_idx * params.O_strides[2]; // Sequence
+
+  // Read the query and 0 the output accumulator
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    q[i] = scale_log2 * static_cast<U>(Q[v_per_thread * lane_idx + i]);
+  }
+
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    o[i] = 0.f;
+  }
+
+  U max_score = -INFINITY;
+  U sum_exp_score = 0.f;
+
+  // For each key
+  for (int i = kv_seq_idx; i < params.kL; i += BN) {
+    bool use_key = true;
+    if constexpr (do_causal) {
+      use_key = i <= (params.kL - params.qL + q_seq_idx);
+    }
+
+    if (use_key) {
+      // Read the key
+      PRAGMA_LOOP_UNROLL
+      for (int j = 0; j < v_per_thread; j++) {
+        k[j] = K[v_per_thread * lane_idx + j];
+      }
+
+      // Compute the i-th score
+      U score = 0.f;
+      PRAGMA_LOOP_UNROLL
+      for (int j = 0; j < v_per_thread; j++) {
+        score += q[j] * k[j];
+      }
+
+      // Warp sum
+      score = cg::reduce(warp, score, cg::plus<U>());
+
+      // Update the accumulators
+      U new_max = max(max_score, score);
+      U factor = exp2f(max_score - new_max);
+      U exp_score = exp2f(score - new_max);
+
+      max_score = new_max;
+      sum_exp_score = sum_exp_score * factor + exp_score;
+
+      // Update the output accumulator
+      PRAGMA_LOOP_UNROLL
+      for (int j = 0; j < v_per_thread; j++) {
+        o[j] = o[j] * factor +
+            exp_score * static_cast<U>(V[v_per_thread * lane_idx + j]);
+      }
+    }
+
+    // Move the pointers to the next kv
+    K += inner_k_stride;
+    V += inner_v_stride;
+  }
+
+  if (lane_idx == 0) {
+    max_scores[warp_idx] = max_score;
+    sum_exp_scores[warp_idx] = sum_exp_score;
+  }
+  block.sync();
+
+  max_score = max_scores[lane_idx];
+  U new_max = cg::reduce(warp, max_score, cg::greater<U>());
+  U factor = exp2f(max_score - new_max);
+  sum_exp_score =
+      cg::reduce(warp, sum_exp_scores[lane_idx] * factor, cg::plus<U>());
+  sum_exp_score = __frcp_rn(sum_exp_score);
+
+  // Now we need to aggregate all the outputs
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    outputs[lane_idx][warp_idx] = o[i];
+    block.sync();
+    U ot = outputs[warp_idx][lane_idx] * factor;
+    o[i] = cg::reduce(warp, ot, cg::plus<U>()) * sum_exp_score;
+    block.sync();
+  }
+
+  // And write the output
+  if (lane_idx == 0) {
+    PRAGMA_LOOP_UNROLL
+    for (int i = 0; i < v_per_thread; i++) {
+      O[v_per_thread * warp_idx + i] = static_cast<T>(o[i]);
+    }
+  }
+}
+
+template <typename T, bool do_causal, int D>
+__global__ void kernel_sdpav_2pass_1(
+    const T* Q,
+    const T* K,
+    const T* V,
+    float* partials,
+    float* sums,
+    float* maxs,
+    __grid_constant__ const AttnParams params) {
+  constexpr int BN = 8;
+  constexpr int BD = 32;
+  constexpr int blocks = 32;
+
+  constexpr int v_per_thread = D / BD;
+
+  const int inner_k_stride = blocks * BN * int(params.K_strides[2]);
+  const int inner_v_stride = blocks * BN * int(params.V_strides[2]);
+
+  typedef float U;
+
+  U q[v_per_thread];
+  U k[v_per_thread];
+  U o[v_per_thread];
+
+  __shared__ U outputs[BN][BD + 1];
+  __shared__ U max_scores[BN];
+  __shared__ U sum_exp_scores[BN];
+
+  const U scale_log2 = params.scale * 1.44269504089f;
+
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<32>(block);
+
+  const int lane_idx = warp.thread_rank();
+  const int warp_idx = warp.meta_group_rank();
+
+  // Adjust to thread block and thread
+  const int batch_idx = blockIdx.z / blocks;
+  const int block_idx = blockIdx.z % blocks;
+  const int head_idx = blockIdx.x;
+  const int kv_head_idx = head_idx / params.gqa_factor;
+
+  const int q_seq_idx = blockIdx.y;
+  const int kv_seq_idx = block_idx * BN + warp_idx;
+
+  Q += batch_idx * params.Q_strides[0] + // Batch
+      head_idx * params.Q_strides[1] + // Head
+      q_seq_idx * params.Q_strides[2]; // Sequence
+
+  K += batch_idx * params.K_strides[0] + // Batch
+      kv_head_idx * params.K_strides[1] + // Head
+      kv_seq_idx * params.K_strides[2]; // Sequence
+
+  V += batch_idx * params.V_strides[0] + // Batch
+      kv_head_idx * params.V_strides[1] + // Head
+      kv_seq_idx * params.V_strides[2]; // Sequence
+
+  const int p_stride_s = blocks;
+  const int p_stride_h = params.qL * p_stride_s;
+  const int p_stride_b = params.H * p_stride_h;
+  const int p_offset = batch_idx * p_stride_b + // Batch
+      head_idx * p_stride_h + // Head
+      q_seq_idx * p_stride_s + // Sequence
+      block_idx; // Block
+
+  partials += p_offset * D;
+  sums += p_offset;
+  maxs += p_offset;
+
+  // Read the query and 0 the output accumulator
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    q[i] = scale_log2 * static_cast<U>(Q[v_per_thread * lane_idx + i]);
+  }
+
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    o[i] = 0.f;
+  }
+
+  U max_score = -1e9;
+  U sum_exp_score = 0.f;
+
+  // For each key
+  for (int i = kv_seq_idx; i < params.kL; i += blocks * BN) {
+    bool use_key = true;
+    if constexpr (do_causal) {
+      use_key = i <= (params.kL - params.qL + q_seq_idx);
+    }
+
+    if (use_key) {
+      // Read the key
+      PRAGMA_LOOP_UNROLL
+      for (int j = 0; j < v_per_thread; j++) {
+        k[j] = K[v_per_thread * lane_idx + j];
+      }
+
+      // Compute the i-th score
+      U score = 0.f;
+      PRAGMA_LOOP_UNROLL
+      for (int j = 0; j < v_per_thread; j++) {
+        score += q[j] * k[j];
+      }
+
+      // Warp sum
+      score = cg::reduce(warp, score, cg::plus<U>());
+
+      // Update the accumulators
+      U new_max = max(max_score, score);
+      U factor = exp2f(max_score - new_max);
+      U exp_score = exp2f(score - new_max);
+
+      max_score = new_max;
+      sum_exp_score = sum_exp_score * factor + exp_score;
+
+      // Update the output accumulator
+      PRAGMA_LOOP_UNROLL
+      for (int j = 0; j < v_per_thread; j++) {
+        o[j] = o[j] * factor +
+            exp_score * static_cast<U>(V[v_per_thread * lane_idx + j]);
+      }
+    }
+
+    // Move the pointers to the next kv
+    K += inner_k_stride;
+    V += inner_v_stride;
+  }
+
+  if (lane_idx == 0) {
+    max_scores[warp_idx] = max_score;
+    sum_exp_scores[warp_idx] = sum_exp_score;
+  }
+
+  block.sync();
+
+  max_score = (lane_idx < BN) ? max_scores[lane_idx] : -1e9;
+  U new_max = cg::reduce(warp, max_score, cg::greater<U>());
+  U factor = exp2f(max_score - new_max);
+  sum_exp_score = (lane_idx < BN) ? sum_exp_scores[lane_idx] : 0.f;
+  sum_exp_score = cg::reduce(warp, sum_exp_score * factor, cg::plus<U>());
+
+  // Write the sum and new max
+  if (warp_idx == 0) {
+    sums[0] = sum_exp_score;
+    maxs[0] = new_max;
+  }
+
+  // Now we need to aggregate all the outputs
+  auto ff = exp2f(max_scores[warp_idx] - new_max);
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    outputs[warp_idx][lane_idx] = o[i] * ff;
+    block.sync();
+
+    if (warp_idx == 0) {
+      U ot = outputs[0][lane_idx];
+      PRAGMA_LOOP_UNROLL
+      for (int j = 1; j < BN; j++) {
+        ot += outputs[j][lane_idx];
+        warp.sync();
+      }
+      o[i] = ot;
+    }
+    block.sync();
+  }
+
+  if (warp_idx == 0) {
+    PRAGMA_LOOP_UNROLL
+    for (int i = 0; i < v_per_thread; i++) {
+      partials[v_per_thread * lane_idx + i] = o[i];
+    }
+  }
+}
+
+template <typename T, bool do_causal, int D>
+__global__ void kernel_sdpav_2pass_2(
+    const float* partials,
+    const float* sums,
+    const float* maxs,
+    T* O,
+    __grid_constant__ const AttnParams params) {
+  constexpr int BN = 32;
+  constexpr int BD = 32;
+  constexpr int blocks = 32;
+
+  constexpr int v_per_thread = D / BD;
+
+  typedef float U;
+
+  U o[v_per_thread];
+  __shared__ U outputs[BN][BD + 1];
+
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<32>(block);
+
+  const int lane_idx = warp.thread_rank();
+  const int warp_idx = warp.meta_group_rank();
+
+  // Adjust to thread block and thread
+  const int batch_idx = blockIdx.z;
+  const int head_idx = blockIdx.x;
+  const int q_seq_idx = blockIdx.y;
+
+  const int p_stride_s = blocks;
+  const int p_stride_h = params.qL * p_stride_s;
+  const int p_stride_b = params.H * p_stride_h;
+  const int p_offset = batch_idx * p_stride_b + // Batch
+      head_idx * p_stride_h + // Head
+      q_seq_idx * p_stride_s; // Sequence
+
+  partials += p_offset * D + warp_idx * D;
+  sums += p_offset;
+  maxs += p_offset;
+
+  O += batch_idx * params.O_strides[0] + // Batch
+      head_idx * params.O_strides[1] + // Head
+      q_seq_idx * params.O_strides[2]; // Sequence
+
+  U max_score = maxs[lane_idx];
+  U new_max = cg::reduce(warp, max_score, cg::greater<U>());
+  U factor = exp2f(max_score - new_max);
+  U sum_exp_score = cg::reduce(warp, sums[lane_idx] * factor, cg::plus<U>());
+  sum_exp_score = __frcp_rn(sum_exp_score);
+
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    o[i] = partials[v_per_thread * lane_idx + i];
+  }
+
+  // Now we need to aggregate all the outputs
+  PRAGMA_LOOP_UNROLL
+  for (int i = 0; i < v_per_thread; i++) {
+    outputs[lane_idx][warp_idx] = o[i];
+    block.sync();
+    U ot = outputs[warp_idx][lane_idx] * factor;
+    o[i] = cg::reduce(warp, ot, cg::plus<U>()) * sum_exp_score;
+    block.sync();
+  }
+
+  // And write the output
+  if (lane_idx == 0) {
+    PRAGMA_LOOP_UNROLL
+    for (int i = 0; i < v_per_thread; i++) {
+      O[v_per_thread * warp_idx + i] = static_cast<T>(o[i]);
+    }
+  }
+}
+
+} // namespace cu
+
+namespace {
+
+template <typename F>
+void dispatch_headdim(int n, F&& f) {
+  switch (n) {
+    case 64:
+      f(std::integral_constant<int, 64>{});
+      break;
+    case 96:
+      f(std::integral_constant<int, 96>{});
+      break;
+    case 128:
+      f(std::integral_constant<int, 128>{});
+      break;
+  }
+}
+
+void sdpa_vector_1pass_fallback(
+    const Stream& s,
+    cu::CommandEncoder& encoder,
+    const array& q,
+    const array& k,
+    const array& v,
+    const float scale,
+    array& o,
+    bool do_causal_ = false) {
+  encoder.set_input_array(q);
+  encoder.set_input_array(k);
+  encoder.set_input_array(v);
+  encoder.set_output_array(o);
+
+  cu::AttnParams params{
+      /* int B = */ q.shape(0),
+      /* int H = */ q.shape(1),
+      /* int D = */ q.shape(3),
+
+      /* int qL = */ q.shape(2),
+      /* int kL = */ k.shape(2),
+
+      /* int gqa_factor = */ q.shape(1) / k.shape(1),
+      /* float scale = */ scale,
+
+      /* int64_t Q_strides[3] = */ {q.strides(0), q.strides(1), q.strides(2)},
+      /* int64_t K_strides[3] = */ {k.strides(0), k.strides(1), k.strides(2)},
+      /* int64_t V_strides[3] = */ {v.strides(0), v.strides(1), v.strides(2)},
+      /* int64_t O_strides[3] = */ {o.strides(0), o.strides(1), o.strides(2)}};
+
+  dim3 grid_dim(params.H, params.qL, params.B);
+  dim3 block_dim(1024, 1, 1);
+
+  dispatch_float_types(o.dtype(), "kernel_sdpav_1pass", [&](auto type_tag) {
+    dispatch_bool(do_causal_, [&](auto do_causal) {
+      dispatch_headdim(params.D, [&](auto headdim) {
+        using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+
+        auto kernel =
+            cu::kernel_sdpav_1pass<DataType, do_causal.value, headdim.value>;
+        encoder.add_kernel_node(
+            kernel,
+            grid_dim,
+            block_dim,
+            0,
+            q.data<DataType>(),
+            k.data<DataType>(),
+            v.data<DataType>(),
+            o.data<DataType>(),
+            params);
+      });
+    });
+  });
+}
+
+void sdpa_vector_2pass_fallback(
+    const Stream& s,
+    cu::CommandEncoder& encoder,
+    const array& q,
+    const array& k,
+    const array& v,
+    const float scale,
+    array& o,
+    bool do_causal_ = false) {
+  cu::AttnParams params{
+      /* int B = */ q.shape(0),
+      /* int H = */ q.shape(1),
+      /* int D = */ q.shape(3),
+
+      /* int qL = */ q.shape(2),
+      /* int kL = */ k.shape(2),
+
+      /* int gqa_factor = */ q.shape(1) / k.shape(1),
+      /* float scale = */ scale,
+
+      /* int64_t Q_strides[3] = */ {q.strides(0), q.strides(1), q.strides(2)},
+      /* int64_t K_strides[3] = */ {k.strides(0), k.strides(1), k.strides(2)},
+      /* int64_t V_strides[3] = */ {v.strides(0), v.strides(1), v.strides(2)},
+      /* int64_t O_strides[3] = */ {o.strides(0), o.strides(1), o.strides(2)}};
+
+  // Allocate the intermediates
+  int blocks = 32;
+
+  Shape intermediate_shape;
+  intermediate_shape.reserve(o.ndim() + 1);
+  intermediate_shape.insert(
+      intermediate_shape.end(), o.shape().begin(), o.shape().end() - 1);
+  intermediate_shape.push_back(blocks);
+  intermediate_shape.push_back(o.shape().back());
+
+  array intermediate(intermediate_shape, float32, nullptr, {});
+  intermediate_shape.pop_back();
+  array sums(intermediate_shape, float32, nullptr, {});
+  array maxs(std::move(intermediate_shape), float32, nullptr, {});
+
+  intermediate.set_data(allocator::malloc(intermediate.nbytes()));
+  sums.set_data(allocator::malloc(sums.nbytes()));
+  maxs.set_data(allocator::malloc(maxs.nbytes()));
+
+  encoder.add_temporary(intermediate);
+  encoder.add_temporary(sums);
+  encoder.add_temporary(maxs);
+
+  dispatch_float_types(o.dtype(), "kernel_sdpav_2pass", [&](auto type_tag) {
+    dispatch_bool(do_causal_, [&](auto do_causal) {
+      dispatch_headdim(params.D, [&](auto headdim) {
+        using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+
+        {
+          auto kernel = cu::
+              kernel_sdpav_2pass_1<DataType, do_causal.value, headdim.value>;
+
+          encoder.set_input_array(q);
+          encoder.set_input_array(k);
+          encoder.set_input_array(v);
+          encoder.set_output_array(intermediate);
+          encoder.set_output_array(sums);
+          encoder.set_output_array(maxs);
+
+          dim3 grid_dim(params.H, params.qL, params.B * 32);
+          dim3 block_dim(8 * 32, 1, 1);
+
+          encoder.add_kernel_node(
+              kernel,
+              grid_dim,
+              block_dim,
+              0,
+              q.data<DataType>(),
+              k.data<DataType>(),
+              v.data<DataType>(),
+              intermediate.data<float>(),
+              sums.data<float>(),
+              maxs.data<float>(),
+              params);
+        }
+
+        {
+          auto kernel = cu::
+              kernel_sdpav_2pass_2<DataType, do_causal.value, headdim.value>;
+
+          encoder.set_input_array(intermediate);
+          encoder.set_input_array(sums);
+          encoder.set_input_array(maxs);
+          encoder.set_output_array(o);
+
+          dim3 grid_dim(params.H, params.qL, params.B);
+          dim3 block_dim(1024, 1, 1);
+
+          encoder.add_kernel_node(
+              kernel,
+              grid_dim,
+              block_dim,
+              0,
+              intermediate.data<float>(),
+              sums.data<float>(),
+              maxs.data<float>(),
+              o.data<DataType>(),
+              params);
+        }
+      });
+    });
+  });
+}
+
+void sdpa_vector_fallback(
+    const Stream& s,
+    cu::CommandEncoder& encoder,
+    const array& q,
+    const array& k,
+    const array& v,
+    const float scale,
+    array& o,
+    bool do_causal_ = false) {
+  int kL = k.shape(2);
+
+  if (kL > 1024) {
+    return sdpa_vector_2pass_fallback(
+        s, encoder, q, k, v, scale, o, do_causal_);
+  } else {
+    return sdpa_vector_1pass_fallback(
+        s, encoder, q, k, v, scale, o, do_causal_);
+  }
+}
+
+} // namespace
+
+namespace fast {
+
+bool ScaledDotProductAttention::use_fallback(
+    const array& q,
+    const array& k,
+    const array& v,
+    bool has_mask,
+    bool has_arr_mask,
+    bool do_causal,
+    Stream s) {
+  if (detail::in_grad_tracing()) {
+    return true;
+  }
+  if (s.device == Device::cpu) {
+    return true;
+  }
+
+  const int value_head_dim = v.shape(-1);
+  const int query_head_dim = q.shape(-1);
+  const int query_sequence_length = q.shape(2);
+  const int key_sequence_length = k.shape(2);
+
+  const bool sdpa_supported_head_dim = query_head_dim == value_head_dim &&
+      (query_head_dim == 64 || query_head_dim == 96 || query_head_dim == 128);
+
+  const bool supported_vector_config =
+      sdpa_supported_head_dim && query_sequence_length < 4;
+
+  const bool supported_config = supported_vector_config;
+
+  return has_arr_mask || !supported_config;
+}
+
+void ScaledDotProductAttention::eval_gpu(
+    const std::vector<array>& inputs,
+    array& out) {
+  nvtx3::scoped_range r("ScaledDotProductAttention::eval_gpu");
+
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  auto& q_pre = inputs[0];
+  auto& k_pre = inputs[1];
+  auto& v_pre = inputs[2];
+  auto& o = out;
+
+  std::vector<array> copies;
+
+  // Define some copy functions to ensure the layout of the inputs is as
+  // expected.
+  copies.reserve(3);
+  auto copy_unless = [&copies, &s](
+                         auto predicate, const array& arr) -> const array& {
+    if (!predicate(arr)) {
+      array arr_copy = contiguous_copy_gpu(arr, s);
+      copies.push_back(std::move(arr_copy));
+      return copies.back();
+    } else {
+      return arr;
+    }
+  };
+
+  // We are in vector mode ie single query
+  if (q_pre.shape(2) < 4) {
+    auto q_copy_unless = [](const array& arr) {
+      if (arr.flags().row_contiguous) {
+        return true;
+      }
+      auto& strides = arr.strides();
+      auto& shape = arr.shape();
+      if (shape[0] == 1 || shape[1] == 1) {
+        // If either the batch or head dimension is a singleton, the other can
+        // be transposed with the sequence dimension
+        auto bidx = shape[0] == 1 ? 1 : 0;
+        return (strides[3] == 1) && (strides[2] == shape[3] * shape[bidx]) &&
+            (strides[bidx] == shape[3]);
+      }
+      return false;
+    };
+
+    auto kv_copy_unless = [](const array& arr) {
+      // keys and values should be copied if:
+      // - the last dimension is not contiguous
+      // - the batch and head dim are not contiguous
+      auto& strides = arr.strides();
+      auto& shape = arr.shape();
+      if (strides.back() != 1) {
+        return false;
+      }
+      if (shape[0] == 1 || shape[1] == 1) {
+        return true;
+      }
+      return (strides[0] == strides[1] * shape[1]);
+    };
+
+    const auto& q = copy_unless(q_copy_unless, q_pre);
+    const auto& k = copy_unless(kv_copy_unless, k_pre);
+    const auto& v = copy_unless(kv_copy_unless, v_pre);
+
+    for (const auto& cp : copies) {
+      encoder.add_temporary(cp);
+    }
+
+    // Donate the query if possible
+    if (q.is_donatable() && q.flags().row_contiguous && q.size() == o.size()) {
+      o.copy_shared_buffer(q);
+    } else {
+      int64_t str_oD = 1;
+      int64_t str_oH = o.shape(3);
+      int64_t str_oL = o.shape(1) * str_oH;
+      int64_t str_oB = o.shape(2) * str_oL;
+      size_t data_size = o.shape(0) * str_oB;
+
+      array::Flags flags{
+          /* bool contiguous = */ 1,
+          /* bool row_contiguous = */ o.shape(2) == 1,
+          /* bool col_contiguous = */ 0,
+      };
+
+      o.set_data(
+          allocator::malloc(o.nbytes()),
+          data_size,
+          {str_oB, str_oH, str_oL, str_oD},
+          flags);
+    }
+
+    return sdpa_vector_fallback(s, encoder, q, k, v, scale_, o, do_causal_);
+  }
+
+  // Full attention mode should never reach here
+  else {
+    throw std::runtime_error("Doesn't support matrix yet.");
+  }
+}
+
+} // namespace fast
+
+} // namespace mlx::core

mlx/backend/cuda/scan.cu

@@ -414,6 +414,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
                   kernel,
                   in.data_size() / axis_size,
                   block_dim,
+                  0,
                   in.data<T>(),
                   out.data<U>(),
                   axis_size);
@@ -443,6 +444,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
                   kernel,
                   num_blocks,
                   block_dim,
+                  0,
                   in.data<T>(),
                   out.data<U>(),
                   axis_size,

mlx/backend/cuda/softmax.cu

@@ -11,7 +11,6 @@
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 #include <nvtx3/nvtx3.hpp>
-#include <cub/block/block_load.cuh>
 
 #include <cassert>
 
@@ -45,20 +44,21 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
   AccT maxval = Limits<AccT>::finite_min();
   AccT normalizer = cast_to<AccT>(0);
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
-    AccT vals[N_READS];
-    cub::LoadDirectBlocked(
-        r * BLOCK_DIM + block.thread_rank(),
-        make_cast_iterator<AccT>(in),
-        vals,
-        axis_size,
-        Limits<AccT>::min());
+    auto index = r * BLOCK_DIM + block.thread_rank();
+    auto vals = load_vector<N_READS>(in, index, axis_size, Limits<T>::min());
     prevmax = maxval;
-    maxval = max_op(maxval, cub::ThreadReduce(vals, max_op));
+#pragma unroll
+    for (int i = 0; i < N_READS; ++i) {
+      maxval = max_op(maxval, static_cast<AccT>(vals[i]));
+    }
+
     // Online normalizer calculation for softmax:
     // https://github.com/NVIDIA/online-softmax
     normalizer = normalizer * softmax_exp(prevmax - maxval);
+#pragma unroll
     for (int i = 0; i < N_READS; i++) {
-      normalizer = normalizer + softmax_exp(vals[i] - maxval);
+      normalizer =
+          normalizer + softmax_exp(static_cast<AccT>(vals[i]) - maxval);
     }
   }
 
@@ -95,12 +95,11 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
   // Write output.
   for (int r = 0; r < cuda::ceil_div(axis_size, BLOCK_DIM * N_READS); r++) {
     auto index = r * BLOCK_DIM + block.thread_rank();
-    T vals[N_READS];
-    cub::LoadDirectBlocked(index, in, vals, axis_size);
+    auto vals = load_vector<N_READS>(in, index, axis_size, T(0));
     for (int i = 0; i < N_READS; i++) {
       vals[i] = softmax_exp(static_cast<AccT>(vals[i]) - maxval) * normalizer;
     }
-    cub::StoreDirectBlocked(index, out, vals, axis_size);
+    store_vector<N_READS>(out, index, vals, axis_size);
   }
 }
 
@@ -141,9 +140,9 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
   encoder.set_input_array(in);
   encoder.set_output_array(out);
   dispatch_float_types(out.dtype(), "softmax", [&](auto type_tag) {
-    constexpr int N_READS = 4;
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr int N_READS = 16 / sizeof(DataType);
     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>;
@@ -152,6 +151,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
           kernel,
           n_rows,
           block_dim(),
+          0,
           in.data<DataType>(),
           out.data<DataType>(),
           axis_size);

mlx/backend/cuda/sort.cu

@@ -13,7 +13,6 @@
 #include <cub/device/device_segmented_sort.cuh>
 
 #include <cassert>
-#include <numeric>
 
 namespace mlx::core {
 
@@ -27,29 +26,6 @@ struct ModOp {
   }
 };
 
-// We can not use any op in eval, make an utility.
-array swapaxes_in_eval(const array& in, int axis1, int axis2) {
-  std::vector<int> axes(in.ndim());
-  std::iota(axes.begin(), axes.end(), 0);
-  std::swap(axes[axis1], axes[axis2]);
-  // TODO: Share the code with Transpose::eval.
-  Shape shape(axes.size());
-  Strides strides(in.ndim());
-  for (size_t ax = 0; ax < axes.size(); ++ax) {
-    shape[ax] = in.shape()[axes[ax]];
-    strides[ax] = in.strides()[axes[ax]];
-  }
-  auto flags = in.flags();
-  if (flags.contiguous) {
-    auto [_, row_contiguous, col_contiguous] = check_contiguity(shape, strides);
-    flags.row_contiguous = row_contiguous;
-    flags.col_contiguous = col_contiguous;
-  }
-  array out(shape, in.dtype(), nullptr, {});
-  out.copy_shared_buffer(in, strides, flags, in.data_size());
-  return out;
-}
-
 struct OffsetTransform {
   int nsort;
 

mlx/backend/cuda/steel/defines.cuh

@@ -0,0 +1,9 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#define MLX_UNROLL _Pragma("unroll")
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+#define MLX_CUDA_SM_80_ENABLED
+#endif

mlx/backend/cuda/steel/gemm.cuh

@@ -0,0 +1,101 @@
+
+#include "mlx/backend/cuda/steel/mma.cuh"
+#include "mlx/backend/cuda/steel/tiles.cuh"
+
+namespace mlx::core::cu {
+
+/**
+ * An example gemm written with the utils.
+ *
+ * Computes A @ B.T when A and B are all aligned with the block sizes.
+ */
+template <typename T, int BM, int BN, int BK>
+__global__ void ab_t_aligned(const T* a, const T* b, T* y, int N, int K) {
+  constexpr int WARPS_M = 2;
+  constexpr int WARPS_N = 2;
+  constexpr int NUM_WARPS = WARPS_M * WARPS_N;
+  constexpr int WARP_STEP_M = BM / WARPS_M;
+  constexpr int WARP_STEP_N = BN / WARPS_N;
+
+  // Precompute some offsets for each thread
+  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;
+
+  // Allocate shared memory
+  extern __shared__ char shmem[];
+  SharedTile<T, BM, BK>(&as)[2] = *(SharedTile<T, BM, BK>(*)[2])(&shmem[0]);
+  SharedTile<T, BN, BK>(&bs)[2] =
+      *(SharedTile<T, BN, BK>(*)[2])(&shmem[sizeof(T) * 2 * BM * BK]);
+
+  // Allocate registers for the MMA
+  RegisterTile<float, BM / WARPS_M, BN / WARPS_N> C;
+  RegisterTile<T, BM / WARPS_M, 16> A;
+  RegisterTile<T, BN / WARPS_N, 16> B;
+
+  // Move the global pointers to the tile
+  a += blockIdx.y * BM * K;
+  b += blockIdx.x * BN * K;
+  y += blockIdx.y * BM * N + blockIdx.x * BN;
+
+  // Zero the accumulators
+  C.fill(0);
+
+  // Start the SM pipeline
+  load_async<NUM_WARPS>(as[0], as[0].base_addr(), a, K);
+  load_async<NUM_WARPS>(bs[0], bs[0].base_addr(), b, K);
+  cp_async_commit();
+
+  int tic = 0;
+  for (int k_block = BK; k_block < K; k_block += BK) {
+    load_async<NUM_WARPS>(as[tic ^ 1], as[tic ^ 1].base_addr(), a + k_block, K);
+    load_async<NUM_WARPS>(bs[tic ^ 1], bs[tic ^ 1].base_addr(), b + k_block, K);
+    cp_async_commit();
+    cp_async_wait<1>();
+    __syncthreads();
+
+    MLX_UNROLL
+    for (int k = 0; k < BK / 16; k++) {
+      A.load(
+          as[tic],
+          as[tic].base_addr(),
+          offset_m + laneid % 16,
+          k * 16 + laneid / 16 * 8);
+      B.load(
+          bs[tic],
+          bs[tic].base_addr(),
+          offset_n + laneid % 16,
+          k * 16 + laneid / 16 * 8);
+
+      mma_t(C, A, B);
+    }
+
+    tic ^= 1;
+  }
+
+  // Empty the pipeline
+  cp_async_wait_all();
+  __syncthreads();
+  MLX_UNROLL
+  for (int k = 0; k < BK / 16; k++) {
+    A.load(
+        as[tic],
+        as[tic].base_addr(),
+        offset_m + laneid % 16,
+        k * 16 + laneid / 16 * 8);
+    B.load(
+        bs[tic],
+        bs[tic].base_addr(),
+        offset_n + laneid % 16,
+        k * 16 + laneid / 16 * 8);
+
+    mma_t(C, A, B);
+  }
+
+  C.store_global(y, N, offset_m, offset_n);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/steel/mma.cuh

@@ -0,0 +1,117 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/steel/defines.cuh"
+#include "mlx/backend/cuda/steel/tiles.cuh"
+
+namespace mlx::core::cu {
+
+/**
+ * Fallback mma.
+ *
+ * We should probably a) implement a fallback or complain about it to the
+ * compiler.
+ */
+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__ __forceinline__ void mma_t(
+    Tile16x16<float>& C,
+    Tile16x16<__nv_bfloat16>& A,
+    Tile16x16<__nv_bfloat16>& B) {
+#if defined(MLX_CUDA_SM_80_ENABLED)
+  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));
+#endif
+}
+
+/**
+ * Multiply larger register tiles by delegating to mma_t.
+ */
+template <typename U, typename T, int M, int N, int K>
+__device__ __forceinline__ 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/steel/tiles.cuh

@@ -0,0 +1,473 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/steel/utils.cuh"
+
+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__ __forceinline__ 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__ __forceinline__ 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 Tile, typename F>
+//   __device__ __forceinline__ void
+//   load(Tile& tile, F f, 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++) {
+//         f(data[i * TILES_X + j],
+//           tile,
+//           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);
+//       }
+//     }
+//   }
+// };
+
+/**
+ * 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 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. Should be changed when we switch to
+  // cute Layouts.
+  //
+  // 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];
+
+  __device__ inline uint32_t base_addr() const {
+    return __cvta_generic_to_shared(&data[0]);
+  }
+
+  // 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.
+ *
+ * Can also be used as a fallback for architectures before sm_80.
+ */
+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);
+  }
+}
+
+/**
+ * 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);
+  }
+}
+
+/**
+ * Same as load_async but checks if we can load the row.
+ *
+ * NOTE: It should be changed to use a predicated cp async instead.
+ */
+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/steel/utils.cuh

@@ -0,0 +1,89 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device/utils.cuh"
+#include "mlx/backend/cuda/steel/defines.cuh"
+
+namespace mlx::core::cu {
+
+/**
+ * Copy bytes from the global memory address pointed to by x to the smem
+ * address pointed to by row_address.
+ *
+ * A simple wrapper over the PTX.
+ */
+template <int N, typename T>
+__device__ inline void cp_async(uint32_t row_address, const T* x) {
+  static_assert(
+      N == 16 || N == 8 || N == 4,
+      "cp.async is only supported for N in {4, 8, 16}.");
+#if defined(MLX_CUDA_SM_80_ENABLED)
+  if constexpr (N == 16) {
+    asm volatile(
+        "cp.async.ca.shared::cta.global [%0], [%1], 16;\n" ::"r"(row_address),
+        "l"(reinterpret_cast<const int4*>(x)));
+  } else if constexpr (N == 8) {
+    asm volatile(
+        "cp.async.ca.shared::cta.global [%0], [%1], 8;\n" ::"r"(row_address),
+        "l"(reinterpret_cast<const int2*>(x)));
+  } else if constexpr (N == 4) {
+    asm volatile(
+        "cp.async.ca.shared::cta.global [%0], [%1], 4;\n" ::"r"(row_address),
+        "l"(reinterpret_cast<const int*>(x)));
+  }
+#endif
+}
+
+/**
+ * Submit all the previous async copies to be executed.
+ */
+__device__ inline void cp_async_commit() {
+#if defined(MLX_CUDA_SM_80_ENABLED)
+  asm volatile("cp.async.commit_group;\n" ::);
+#endif
+}
+
+/**
+ * Wait for all but N of the async copies to finish.
+ */
+template <int N>
+__device__ inline void cp_async_wait() {
+#if defined(MLX_CUDA_SM_80_ENABLED)
+  if constexpr (N == 0) {
+    asm volatile("cp.async.wait_all;\n" ::);
+  } else {
+    asm volatile("cp.async.wait_group %0;\n" ::"n"(N));
+  }
+#endif
+}
+
+/**
+ * Wait for all the async copies to finish.
+ */
+__device__ inline void cp_async_wait_all() {
+  cp_async_wait<0>();
+}
+
+/**
+ * Extract ``bits`` bits from the 32 bit value.
+ *
+ * Single instruction shift and mask.
+ */
+template <int bits>
+__device__ inline uint32_t extract_bits(uint32_t value, int start_bit) {
+  static_assert(
+      bits == 2 || bits == 4 || bits == 8,
+      "extract_bits only supports 2, 4, 8 for now.");
+  uint32_t result;
+  if constexpr (bits == 2) {
+    asm("bfe.u32 %0, %1, %2, 2;" : "=r"(result) : "r"(value), "r"(start_bit));
+  } else if constexpr (bits == 4) {
+    asm("bfe.u32 %0, %1, %2, 4;" : "=r"(result) : "r"(value), "r"(start_bit));
+  } else if constexpr (bits == 8) {
+    asm("bfe.u32 %0, %1, %2, 8;" : "=r"(result) : "r"(value), "r"(start_bit));
+  }
+  return result;
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/ternary.cu

@@ -32,7 +32,7 @@ ternary_v(const bool* a, const T* b, const T* c, T* out, IdxT size) {
     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]);
+      out_vec[i] = Op{}(a_vec[i], b_vec[i], c_vec[i]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -76,7 +76,7 @@ __global__ void ternary_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto [a_idx, b_idx, c_idx] = elem_to_loc_4d(
+    auto [a_idx, b_idx, c_idx] = elem_to_loc(
         index,
         shape.data(),
         a_strides.data(),
@@ -125,14 +125,12 @@ void ternary_op_gpu_inplace(
             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());
+                auto [num_blocks, block_dims] = get_launch_args(out, large());
                 encoder.add_kernel_node(
-                    kernel,
+                    cu::ternary_g_nd<Op, DType, IdxT, dims_constant()>,
                     num_blocks,
                     block_dims,
+                    0,
                     a.data<bool>(),
                     b.data<DType>(),
                     c.data<DType>(),
@@ -144,13 +142,12 @@ void ternary_op_gpu_inplace(
                     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());
+              auto [num_blocks, block_dims] = get_launch_args(out, large());
               encoder.add_kernel_node(
-                  kernel,
+                  cu::ternary_g<Op, DType, IdxT>,
                   num_blocks,
                   block_dims,
+                  0,
                   a.data<bool>(),
                   b.data<DType>(),
                   c.data<DType>(),
@@ -166,20 +163,14 @@ void ternary_op_gpu_inplace(
     } 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>;
+        constexpr int N_READS = 16 / sizeof(DType);
         auto [num_blocks, block_dims] = get_launch_args(
-            kernel,
-            out.data_size(),
-            out.shape(),
-            out.strides(),
-            large(),
-            N_READS);
+            out.data_size(), out.shape(), out.strides(), large(), N_READS);
         encoder.add_kernel_node(
-            kernel,
+            cu::ternary_v<Op, DType, IdxT, N_READS>,
             num_blocks,
             block_dims,
+            0,
             a.data<bool>(),
             b.data<DType>(),
             c.data<DType>(),
@@ -204,7 +195,7 @@ void ternary_op_gpu(
 }
 
 void Select::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("select::eval_gpu");
+  nvtx3::scoped_range r("Select::eval_gpu");
   auto& s = out.primitive().stream();
   ternary_op_gpu<cu::Select>(inputs, out, s);
 }

mlx/backend/cuda/unary.cu

@@ -3,7 +3,6 @@
 #include "mlx/backend/common/unary.h"
 #include "mlx/backend/cuda/device.h"
 #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"
@@ -31,7 +30,7 @@ __global__ void unary_v(const In* in, Out* out, IdxT size) {
     AlignedVector<Out, N_READS> out_vec;
 #pragma unroll
     for (int i = 0; i < N_READS; ++i) {
-      out_vec.val[i] = Op{}(in_vec.val[i]);
+      out_vec[i] = Op{}(in_vec[i]);
     }
 
     store_vector<N_READS>(out, index, out_vec);
@@ -48,7 +47,7 @@ __global__ void unary_g(
     int ndim) {
   IdxT index = cg::this_grid().thread_rank();
   if (index < size) {
-    auto idx = elem_to_loc_4d(index, shape.data(), strides.data(), ndim);
+    auto idx = elem_to_loc(index, shape.data(), strides.data(), ndim);
     out[index] = Op{}(in[idx]);
   }
 }
@@ -130,30 +129,25 @@ void unary_op_gpu_inplace(
             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);
+                out.data_size(), out.shape(), out.strides(), large, N_READS);
             encoder.add_kernel_node(
-                kernel,
+                cu::unary_v<Op, InType, OutType, IdxT, N_READS>,
                 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 kernel = cu::unary_g<Op, InType, OutType, IdxT>;
-            auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
+            auto [num_blocks, block_dims] = get_launch_args(out, large);
             encoder.add_kernel_node(
-                kernel,
+                cu::unary_g<Op, InType, OutType, IdxT>,
                 num_blocks,
                 block_dims,
+                0,
                 in.data<InType>(),
                 out.data<OutType>(),
                 out.data_size(),

mlx/backend/cuda/utils.cpp

@@ -17,6 +17,35 @@ CudaStream::~CudaStream() {
   CHECK_CUDA_ERROR(cudaStreamDestroy(stream_));
 }
 
+CudaGraphExec::CudaGraphExec(cudaGraphExec_t handle) : handle_(handle) {}
+
+CudaGraphExec::CudaGraphExec(CudaGraphExec&& other) : handle_(other.handle_) {
+  other.handle_ = nullptr;
+};
+
+CudaGraphExec::~CudaGraphExec() {
+  reset();
+}
+
+void CudaGraphExec::instantiate(cudaGraph_t graph) {
+  CHECK_CUDA_ERROR(cudaGraphInstantiate(&handle_, graph, nullptr, nullptr, 0));
+}
+
+void CudaGraphExec::reset() {
+  if (handle_ != nullptr) {
+    CHECK_CUDA_ERROR(cudaGraphExecDestroy(handle_));
+    handle_ = nullptr;
+  }
+}
+
+void check_cublas_error(const char* name, cublasStatus_t err) {
+  if (err != CUBLAS_STATUS_SUCCESS) {
+    // TODO: Use cublasGetStatusString when it is widely available.
+    throw std::runtime_error(
+        fmt::format("{} failed with code: {}.", name, static_cast<int>(err)));
+  }
+}
+
 void check_cuda_error(const char* name, cudaError_t err) {
   if (err != cudaSuccess) {
     throw std::runtime_error(

mlx/backend/cuda/utils.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <cublasLt.h>
 #include <cuda.h>
 #include <cuda_runtime.h>
 
@@ -32,11 +33,34 @@ class CudaStream {
   cudaStream_t stream_;
 };
 
+// Move-able RAII handle of cudaGraphExec_t.
+class CudaGraphExec {
+ public:
+  CudaGraphExec(cudaGraphExec_t handle = nullptr);
+  CudaGraphExec(CudaGraphExec&& other);
+  ~CudaGraphExec();
+
+  CudaGraphExec(const CudaGraphExec&) = delete;
+  CudaGraphExec& operator=(const CudaGraphExec&) = delete;
+
+  void instantiate(cudaGraph_t graph);
+  void reset();
+
+  operator cudaGraphExec_t() const {
+    return handle_;
+  }
+
+ private:
+  cudaGraphExec_t handle_;
+};
+
 // Throw exception if the cuda API does not succeed.
+void check_cublas_error(const char* name, cublasStatus_t err);
 void check_cuda_error(const char* name, cudaError_t err);
 void check_cuda_error(const char* name, CUresult err);
 
 // The macro version that prints the command that failed.
+#define CHECK_CUBLAS_ERROR(cmd) check_cublas_error(#cmd, (cmd))
 #define CHECK_CUDA_ERROR(cmd) check_cuda_error(#cmd, (cmd))
 
 // Convert Dtype to CUDA C++ types.

mlx/backend/cuda/worker.cpp

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

mlx/backend/cuda/worker.h

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

mlx/backend/gpu/primitives.cpp

@@ -133,6 +133,7 @@ void NumberOfElements::eval_gpu(const std::vector<array>& inputs, array& out) {
 }
 
 void Pad::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Pad::eval_gpu");
   // Inputs must be base input array and scalar val array
   assert(inputs.size() == 2);
   auto& in = inputs[0];

mlx/backend/metal/allocator.cpp

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

mlx/backend/metal/device.h

@@ -60,6 +60,16 @@ struct CommandEncoder {
     enc_->updateFence(fence);
   }
 
+  template <typename T>
+  void set_vector_bytes(const SmallVector<T>& vec, size_t nelems, int idx) {
+    enc_->setBytes(vec.data(), nelems * sizeof(T), idx);
+  }
+  template <typename T>
+  void set_vector_bytes(const SmallVector<T>& vec, int idx) {
+    return set_vector_bytes(vec, vec.size(), idx);
+  }
+
+  // TODO: Code is duplicated but they should be deleted soon.
   template <typename T>
   void set_vector_bytes(const std::vector<T>& vec, size_t nelems, int idx) {
     enc_->setBytes(vec.data(), nelems * sizeof(T), idx);
@@ -94,7 +104,7 @@ struct CommandEncoder {
   };
 
   // Outputs of all kernels in the encoder including temporaries
-  std::unordered_set<const void*> outputs() {
+  std::unordered_set<const void*>& outputs() {
     return all_outputs_;
   };
 

mlx/backend/metal/event.cpp

@@ -14,6 +14,10 @@ Event::Event(Stream stream) : stream_(stream) {
   auto p = metal::new_scoped_memory_pool();
   event_ = std::shared_ptr<void>(
       metal::device(Device::gpu).mtl_device()->newSharedEvent(), dtor);
+  if (event_ == nullptr) {
+    throw std::runtime_error(
+        "[Event::Event] Failed to create Metal shared event.");
+  }
 }
 
 void Event::wait() {