Experimenting with a gemm based on the cuda steel utils

.circleci/config.yml

@@ -18,14 +18,13 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "26.0.0"
+      xcode: "16.2.0"
-    resource_class: m4pro.medium
+    resource_class: m2pro.medium
     steps:
       - checkout
       - run:
           name: Install
           command: |
-            xcodebuild -downloadComponent MetalToolchain
             brew install python@3.9
             brew install doxygen
             python3.9 -m venv env
@@ -90,7 +89,6 @@ jobs:
           command: |
             uv venv
             uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
             uv pip install -e ".[dev]" -v
       - run:
           name: Generate package stubs
@@ -120,7 +118,7 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "26.0.0"
+        default: "16.2.0"
       macosx_deployment_target:
         type: string
         default: ""
@@ -128,13 +126,12 @@ jobs:
       xcode: << parameters.xcode_version >>
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m4pro.medium
+    resource_class: m2pro.medium
     steps:
       - checkout
       - run:
           name: Install dependencies
           command: |
-            xcodebuild -downloadComponent MetalToolchain
             HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
               brew install openmpi uv
       - run:
@@ -199,7 +196,7 @@ jobs:
           name: Run Python tests with JIT
           command: |
             CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e . -v
+              uv pip install -e .
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               uv run --no-project python -m xmlrunner discover \
@@ -225,7 +222,6 @@ jobs:
             sudo apt-get update
             sudo apt-get install libcudnn9-dev-cuda-12
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install libnccl2 libnccl-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
@@ -234,8 +230,7 @@ jobs:
           name: Install Python package
           command: |
             uv venv
-            uv pip install cmake
+            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-            DEBUG=1 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
               uv pip install -e ".[dev]" -v
       - run:
           name: Run Python tests
@@ -243,18 +238,6 @@ jobs:
             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: Build CPP only
-          command: |
-            source .venv/bin/activate
-            cmake . -B build \
-              -DMLX_BUILD_CUDA=ON \
-              -DCMAKE_CUDA_COMPILER=`which nvcc` \
-              -DCMAKE_BUILD_TYPE=DEBUG
-            cmake --build build -j `nproc`
-      - run:
-          name: Run CPP tests
-          command: ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
       - run:
           name: CCache report
           command: |
@@ -274,7 +257,7 @@ jobs:
         default: "3.9"
       xcode_version:
         type: string
-        default: "26.0.0"
+        default: "16.2.0"
       build_env:
         type: string
         default: ""
@@ -283,7 +266,7 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: m4pro.medium
+    resource_class: m2pro.medium
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
     steps:
@@ -291,15 +274,11 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            xcodebuild -downloadComponent MetalToolchain
+            brew install python@<< parameters.python_version >>
-            mkdir -p ~/miniconda3
+            brew install openmpi
-            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
+            python<< parameters.python_version >> -m venv env
-            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+            source env/bin/activate
-            rm ~/miniconda3/miniconda.sh
+            pip install --upgrade pip
-            source ~/miniconda3/bin/activate
-            conda init --all
-            conda create -n env python=<< parameters.python_version >> -y
-            conda activate env
             pip install --upgrade cmake
             pip install nanobind==2.4.0
             pip install --upgrade setuptools
@@ -309,19 +288,19 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            conda activate env
+            source env/bin/activate
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
-            conda activate env
+            source env/bin/activate
             pip install typing_extensions
             python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
-            conda activate env
+            source env/bin/activate
             python setup.py clean --all
             << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
       - when:
@@ -331,7 +310,7 @@ jobs:
             - run:
                 name: Build common package
                 command: |
-                  conda activate env
+                  source env/bin/activate
                   python setup.py clean --all
                   << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
@@ -340,7 +319,7 @@ jobs:
             - run:
                 name: Upload package
                 command: |
-                  conda activate env
+                  source env/bin/activate
                   twine upload dist/*
       - store_artifacts:
           path: dist/
@@ -413,7 +392,7 @@ jobs:
         default: ""
     machine:
       image: ubuntu-2204:current
-      resource_class: xlarge
+      resource_class: large
     steps:
       - checkout
       - run:
@@ -460,7 +439,7 @@ workflows:
       - mac_build_and_test:
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test
       - cuda_build_and_test:
           matrix:
@@ -485,7 +464,68 @@ workflows:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["26.0.0"]
+              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: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
       - build_documentation:
           filters:
             tags:
@@ -527,7 +567,7 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
       - linux_build_and_test:
           requires: [ hold ]
       - cuda_build_and_test:
@@ -546,7 +586,53 @@ workflows:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["26.0.0"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
       - build_linux_release:
           matrix:
             parameters:
@@ -565,7 +651,68 @@ workflows:
               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: ["26.0.0"]
+              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:

ACKNOWLEDGMENTS.md

@@ -25,11 +25,6 @@ MLX was developed with contributions from the following individuals:
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>
 
-# Organizations
-
-MLX has received contributions from the following companies:
-- NVIDIA Corporation & Affiliates
-
 # Third-Party Software
 
 MLX leverages several third-party software, listed here together with

CMakeLists.txt

@@ -140,12 +140,6 @@ elseif(MLX_BUILD_METAL)
   target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
 endif()
 
-if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
-  # With newer clang/gcc versions following libs are implicitly linked, but when
-  # building on old distributions they need to be explicitly listed.
-  target_link_libraries(mlx PRIVATE dl pthread)
-endif()
-
 if(WIN32)
   if(MSVC)
     # GGUF does not build with MSVC.

cmake/FindNCCL.cmake

@@ -1,54 +0,0 @@
-# FindNCCL.cmake This module finds the NVIDIA NCCL library and its include
-# directories.
-
-set(NCCL_ROOT_DIR
-    $ENV{NCCL_ROOT_DIR}
-    CACHE PATH "Folder contains NVIDIA NCCL")
-
-find_path(
-  NCCL_INCLUDE_DIRS
-  NAMES nccl.h
-  HINTS ${NCCL_INCLUDE_DIR} ${NCCL_ROOT_DIR} ${NCCL_ROOT_DIR}/include
-        ${CUDA_TOOLKIT_ROOT_DIR}/include)
-
-if($ENV{USE_STATIC_NCCL})
-  message(
-    STATUS "USE_STATIC_NCCL detected. Linking against static NCCL library")
-  set(NCCL_LIBNAME "libnccl_static.a")
-else()
-  set(NCCL_LIBNAME "nccl")
-endif()
-
-find_library(
-  NCCL_LIBRARIES
-  NAMES ${NCCL_LIBNAME}
-  HINTS ${NCCL_LIB_DIR}
-        ${NCCL_ROOT_DIR}
-        ${NCCL_ROOT_DIR}/lib
-        ${NCCL_ROOT_DIR}/lib/x86_64-linux-gnu
-        ${NCCL_ROOT_DIR}/lib64
-        ${CUDA_TOOLKIT_ROOT_DIR}/lib
-        ${CUDA_TOOLKIT_ROOT_DIR}/lib64)
-
-include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS
-                                  NCCL_LIBRARIES)
-
-if(NCCL_FOUND)
-  set(NCCL_HEADER_FILE "${NCCL_INCLUDE_DIRS}/nccl.h")
-  message(
-    STATUS "Determining NCCL version from the header file: ${NCCL_HEADER_FILE}")
-  file(
-    STRINGS ${NCCL_HEADER_FILE} NCCL_MAJOR_VERSION_DEFINED
-    REGEX "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+[0-9]+.*$"
-    LIMIT_COUNT 1)
-  if(NCCL_MAJOR_VERSION_DEFINED)
-    string(REGEX REPLACE "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+" ""
-                         NCCL_MAJOR_VERSION ${NCCL_MAJOR_VERSION_DEFINED})
-    message(STATUS "NCCL_MAJOR_VERSION: ${NCCL_MAJOR_VERSION}")
-  endif()
-  message(
-    STATUS
-      "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
-  mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
-endif()

docs/src/dev/custom_metal_kernels.rst

@@ -127,8 +127,7 @@ relying on a copy from ``ensure_row_contiguous``:
       name="myexp_strided",
       input_names=["inp"],
       output_names=["out"],
-      source=source,
+      source=source
-      ensure_row_contiguous=False,
   )
 
   def exp_elementwise(a: mx.array):
@@ -139,6 +138,7 @@ relying on a copy from ``ensure_row_contiguous``:
           threadgroup=(256, 1, 1),
           output_shapes=[a.shape],
           output_dtypes=[a.dtype],
+          ensure_row_contiguous=False,
       )
       return outputs[0]
 

docs/src/index.rst

@@ -70,7 +70,6 @@ are the CPU and GPU.
    python/fft
    python/linalg
    python/metal
-   python/cuda
    python/memory_management
    python/nn
    python/optimizers

docs/src/install.rst

@@ -271,7 +271,7 @@ and the CUDA toolkit. For example on Ubuntu, run the following:
    dpkg -i cuda-keyring_1.1-1_all.deb
    apt-get update -y
    apt-get -y install cuda-toolkit-12-9
-   apt-get install libblas-dev liblapack-dev liblapacke-dev libcudnn9-dev-cuda-12 -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
 
 
 When building either the Python or C++ APIs make sure to pass the cmake flag

docs/src/python/cuda.rst

@@ -1,9 +0,0 @@
-CUDA
-=====
-
-.. currentmodule:: mlx.core.cuda
-
-.. autosummary::
-  :toctree: _autosummary
-
-  is_available

docs/src/python/fast.rst

@@ -13,4 +13,3 @@ Fast
   rope
   scaled_dot_product_attention
   metal_kernel
-  cuda_kernel

docs/src/usage/compile.rst

@@ -225,7 +225,7 @@ In some cases returning updated state can be pretty inconvenient. Hence,
   def fun(x, y):
       z = x + y
       state.append(z)
-      return mx.exp(z)
+      return mx.exp(z), state
 
   fun(mx.array(1.0), mx.array(2.0))
   # Prints [array(3, dtype=float32)]

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

@@ -196,6 +196,9 @@ void shared_buffer_reshape(
     const Strides& out_strides,
     array& out);
 
+// Like the swapaxes op but safe to call in eval_gpu.
+array swapaxes_in_eval(const array& x, int axis1, int axis2);
+
 template <typename T>
 inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
   vec.erase(std::next(vec.begin(), index));

mlx/backend/cpu/compiled.cpp

@@ -15,7 +15,6 @@
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
-#include "mlx/version.h"
 
 namespace mlx::core {
 
@@ -95,11 +94,7 @@ void* compile(
     kernel_file_name = kernel_name;
   }
 
-  auto output_dir =
+  auto output_dir = std::filesystem::temp_directory_path();
-      std::filesystem::temp_directory_path() / "mlx" / version() / "cpu";
-  if (!std::filesystem::exists(output_dir)) {
-    std::filesystem::create_directories(output_dir);
-  }
 
   std::string shared_lib_name = "lib" + kernel_file_name + ".so";
   auto shared_lib_path = (output_dir / shared_lib_name).string();
@@ -162,12 +157,10 @@ inline void build_kernel(
 #endif
 
   // Start the kernel
-  os << "void " << kernel_name
+  os << "void " << kernel_name << "(void** args) {" << std::endl;
-     << "(int* shape, int64_t** strides, void** args) {" << std::endl;
 
   // Add the input arguments
   int cnt = 0;
-  int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
     // Skip constants from the input list
     if (is_constant(i)) {
@@ -182,8 +175,8 @@ inline void build_kernel(
        << "];" << std::endl;
     // Scalars and contiguous need no strides
     if (!is_scalar(x) && !contiguous) {
-      os << "  const int64_t* " << xname << "_strides = strides["
+      os << "  const size_t* " << xname << "_strides = (size_t*)args[" << cnt++
-         << strides_index++ << "];" << std::endl;
+         << "];" << std::endl;
     }
   }
 
@@ -193,8 +186,10 @@ inline void build_kernel(
     os << "  " << tstr << "* " << namer.get_name(x) << " = (" << tstr
        << "*)args[" << cnt++ << "];" << std::endl;
   }
-  // Add output size
+  // Add output strides and shape to extract the indices.
-  if (contiguous) {
+  if (!contiguous) {
+    os << "  const int* shape = (int*)args[" << cnt++ << "];" << std::endl;
+  } else {
     os << "  const size_t size = (size_t)args[" << cnt++ << "];" << std::endl;
   }
 
@@ -293,8 +288,17 @@ 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;
   for (size_t i = 0; i < inputs.size(); ++i) {
     if (is_constant_(i)) {
       continue;
@@ -302,6 +306,9 @@ void Compiled::eval_cpu(
     const auto& x = inputs[i];
     encoder.set_input_array(x);
     args.push_back((void*)x.data<void>());
+    if (!contiguous && !is_scalar(x)) {
+      args.push_back(strides[strides_index++].data());
+    }
   }
 
   // Get the kernel name from the lib
@@ -336,20 +343,16 @@ void Compiled::eval_cpu(
     args.push_back(x.data<void>());
     encoder.set_output_array(x);
   }
-  if (contiguous) {
+  if (!contiguous) {
+    args.push_back((void*)shape.data());
+  } else {
     args.push_back((void*)outputs[0].data_size());
   }
-  auto fun = reinterpret_cast<void (*)(int*, int64_t**, void**)>(fn_ptr);
+  auto fun = (void (*)(void**))fn_ptr;
   encoder.dispatch([fun,
                     args = std::move(args),
                     strides = std::move(strides),
-                    shape = std::move(shape)]() mutable {
+                    shape = std::move(shape)]() mutable { fun(args.data()); });
-    SmallVector<int64_t*> strides_ptrs;
-    for (auto& s : strides) {
-      strides_ptrs.push_back(s.data());
-    }
-    fun(shape.data(), strides_ptrs.data(), args.data());
-  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/lapack.h

@@ -47,7 +47,7 @@ INSTANTIATE_LAPACK_REAL(orgqr)
 INSTANTIATE_LAPACK_REAL(syevd)
 INSTANTIATE_LAPACK_REAL(geev)
 INSTANTIATE_LAPACK_REAL(potrf)
-INSTANTIATE_LAPACK_REAL(gesdd)
+INSTANTIATE_LAPACK_REAL(gesvdx)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)

mlx/backend/cpu/quantized.cpp

@@ -1,5 +1,7 @@
 // Copyright © 2023 Apple Inc.
 
+#include <cassert>
+
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
@@ -11,35 +13,6 @@ namespace mlx::core {
 
 namespace {
 
-const static float MXFP4_LUT[16] = {
-    +0.0f,
-    +0.5f,
-    +1.0f,
-    +1.5f,
-    +2.0f,
-    +3.0f,
-    +4.0f,
-    +6.0f,
-    -0.0f,
-    -0.5f,
-    -1.0f,
-    -1.5f,
-    -2.0f,
-    -3.0f,
-    -4.0f,
-    -6.0f};
-
-template <typename T>
-static inline T dequantize_scale(uint8_t s) {
-  using FOrI = union {
-    bfloat16_t f;
-    uint16_t i;
-  };
-  FOrI out;
-  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
-  return static_cast<T>(out.f);
-}
-
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
   return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -434,231 +407,6 @@ void _qmm_dispatch(
   }
 }
 
-template <typename T>
-void mxfp4_qmm(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = get_pack_factor(4, 8);
-  constexpr int bytes_per_pack = get_bytes_per_pack(4);
-  constexpr int packs_in_group = group_size / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint8_t* w_local = (const uint8_t*)w;
-    const uint8_t* scales_local = scales;
-
-    std::fill(result, result + N, 0);
-
-    for (int k = 0; k < K; k++) {
-      T* result_local = result;
-      T xi = *x++;
-
-      for (int n = 0; n < N; n += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
-        for (int ng = 0; ng < packs_in_group; ng++) {
-          uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-          for (int p = 0; p < pack_factor; p++) {
-            (*result_local++) +=
-                xi * scale * static_cast<T>(MXFP4_LUT[wi & 0xf]);
-            wi >>= 4;
-          }
-        }
-      }
-    }
-
-    result += N;
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_t(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = get_pack_factor(4, 8);
-  constexpr int bytes_per_pack = get_bytes_per_pack(4);
-  constexpr int packs_in_group = group_size / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint8_t* w_local = (const uint8_t*)w;
-    const uint8_t* scales_local = scales;
-
-    for (int n = 0; n < N; n++) {
-      const T* x_local = x;
-      T sum = 0;
-      for (int k = 0; k < K; k += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
-
-        T gsum = 0;
-        for (int kw = 0; kw < packs_in_group; kw++) {
-          uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-          for (int p = 0; p < pack_factor; p++) {
-            gsum += (*x_local++) * static_cast<T>(MXFP4_LUT[wi & 0xf]);
-            wi >>= 4;
-          }
-        }
-        sum += scale * gsum;
-      }
-      *result = sum;
-      result++;
-    }
-
-    x += K;
-  }
-}
-
-template <int S>
-simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
-  if constexpr (S == 8) {
-    constexpr std::array<uint32_t, 8> shifts_ = {{0, 4, 8, 12, 16, 20, 24, 28}};
-    auto shifts(*(simd::Simd<uint32_t, S>*)&shifts_);
-    auto wi = simd::Simd<uint32_t, S>(*w);
-    wi = wi >> shifts;
-    wi = wi & 0xf;
-    simd::Simd<float, S> w_out;
-    for (int i = 0; i < S; ++i) {
-      w_out[i] = MXFP4_LUT[wi[i]];
-    }
-    return w_out;
-  } else {
-    // Appease compiler.. but should never get here
-    throw std::runtime_error("Unsupported combination for simd qmm.");
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_t_simd(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = 32 / 4;
-  constexpr int packs_in_group = group_size / pack_factor;
-  constexpr int S = simd::max_size<T>;
-  static_assert(
-      S % pack_factor == 0, "SIMD size must be divisible by pack factor");
-  constexpr int packs_per_simd = S / pack_factor;
-
-  for (int m = 0; m < M; m++) {
-    const uint32_t* w_local = w;
-    const uint8_t* scales_local = scales;
-
-    for (int n = 0; n < N; n++) {
-      simd::Simd<float, S> acc(0);
-      auto x_local = x;
-      for (int k = 0; k < K; k += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
-
-        simd::Simd<float, S> g_acc(0);
-        for (int kw = 0; kw < packs_in_group; kw += packs_per_simd) {
-          // Extract bits
-          auto wf = mxfp4_extract_bits_simd<S>(w_local);
-          w_local += packs_per_simd;
-          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
-          g_acc = g_acc + x_simd * wf;
-          x_local += S;
-        }
-        acc = acc + scale * g_acc;
-      }
-
-      *result = T(simd::sum(acc));
-      result++;
-    }
-    x += K;
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_dispatch_transpose(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K,
-    bool transposed_w) {
-  if (transposed_w) {
-    // the simd size must be a multiple of the number of elements per word
-    if constexpr (simd::max_size<T> % 8 == 0) {
-      mxfp4_qmm_t_simd<T>(result, x, w, scales, M, N, K);
-    } else {
-      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
-    }
-  } else {
-    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
-  }
-}
-
-template <typename T>
-void mxfp4_qmm_dispatch_typed(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    bool transposed_w) {
-  int K = x.shape(-1);
-  int M = x.ndim() > 1 ? x.shape(-2) : 1;
-  int N = out.shape(-1);
-  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
-  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
-  int batch_size = x.size() / (K * M);
-
-  auto out_ptr = out.data<T>();
-  auto x_ptr = x.data<T>();
-  auto w_ptr = w.data<uint32_t>();
-  auto scales_ptr = scales.data<uint8_t>();
-  for (int i = 0; i < batch_size; i++) {
-    mxfp4_qmm_dispatch_transpose<T>(
-        out_ptr + i * M * N,
-        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
-        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
-        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
-        M,
-        N,
-        K,
-        transposed_w);
-  }
-}
-
-void mxfp4_qmm_dispatch(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    bool transposed_w) {
-  switch (x.dtype()) {
-    case bfloat16:
-      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
-      break;
-    case float16:
-      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
-      break;
-    case float32:
-      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
-      break;
-    default:
-      throw std::invalid_argument(
-          "[quantized_matmul] only floating types are supported");
-  }
-}
-
 template <typename T>
 void _bs_qmm_dispatch_typed(
     array& out,
@@ -765,106 +513,41 @@ void _bs_qmm_dispatch(
   }
 }
 
-template <typename T>
-void mxfp4_bs_qmm_dispatch_typed(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& lhs_indices,
-    const array& rhs_indices,
-    bool transposed_w) {
-  int K = x.shape(-1);
-  int M = x.shape(-2);
-  int N = out.shape(-1);
-
-  int w_els = w.shape(-1) * w.shape(-2);
-  int g_els = scales.shape(-1) * scales.shape(-2);
-
-  auto out_ptr = out.data<T>();
-  auto x_ptr = x.data<T>();
-  auto w_ptr = w.data<uint32_t>();
-  auto scales_ptr = scales.data<uint8_t>();
-  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
-  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
-
-  for (int i = 0; i < lhs_indices.size(); i++) {
-    int x_idx = lhs_indices_ptr[elem_to_loc(
-        i, lhs_indices.shape(), lhs_indices.strides())];
-    int w_idx = rhs_indices_ptr[elem_to_loc(
-        i, rhs_indices.shape(), rhs_indices.strides())];
-    mxfp4_qmm_dispatch_transpose<T>(
-        out_ptr + i * M * N,
-        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
-        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
-        scales_ptr +
-            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
-        M,
-        N,
-        K,
-        transposed_w);
-  }
-}
-
-void mxfp4_bs_qmm_dispatch(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& lhs_indices,
-    const array& rhs_indices,
-    bool transposed_w) {
-  switch (x.dtype()) {
-    case float32:
-      mxfp4_bs_qmm_dispatch_typed<float>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    case float16:
-      mxfp4_bs_qmm_dispatch_typed<float16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    case bfloat16:
-      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    default:
-      throw std::invalid_argument(
-          "[quantized_matmul] only floating types are supported");
-  }
-}
-
 } // namespace
 
 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 4);
+
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
+  auto& biases_pre = inputs[3];
 
-  auto& encoder = cpu::get_command_encoder(stream());
+  std::vector<array> temps;
-  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
+  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
-      encoder.add_temporary(arr_cpy);
+      return temps.back();
-      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous(x_pre);
   auto w = ensure_row_contiguous(w_pre);
   auto scales = ensure_row_contiguous(scales_pre);
+  auto biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_output_array(out);
-  if (mode_ == QuantizationMode::Affine) {
-    auto biases = ensure_row_contiguous(inputs[3]);
   encoder.set_input_array(biases);
+  encoder.set_output_array(out);
   encoder.dispatch([out = array::unsafe_weak_copy(out),
                     x = array::unsafe_weak_copy(x),
                     w = array::unsafe_weak_copy(w),
@@ -875,54 +558,48 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
                     transpose_ = transpose_]() mutable {
     _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
   });
-  } else {
-    encoder.dispatch([out = array::unsafe_weak_copy(out),
-                      x = array::unsafe_weak_copy(x),
-                      w = array::unsafe_weak_copy(w),
-                      scales = array::unsafe_weak_copy(scales),
-                      transpose_ = transpose_]() mutable {
-      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
-    });
-  }
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 6);
+
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& lhs_indices = inputs[inputs.size() - 2];
+  auto& biases_pre = inputs[3];
-  auto& rhs_indices = inputs[inputs.size() - 1];
+  auto& lhs_indices = inputs[4];
+  auto& rhs_indices = inputs[5];
 
-  auto& encoder = cpu::get_command_encoder(stream());
+  std::vector<array> temps;
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &encoder](const array& arr) {
+                                          &temps](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
-      encoder.add_temporary(arr_cpy);
+      return temps.back();
-      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous_last_dims(x_pre);
   auto w = ensure_row_contiguous_last_dims(w_pre);
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
+  auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
+  encoder.set_input_array(biases);
   encoder.set_input_array(lhs_indices);
   encoder.set_input_array(rhs_indices);
   encoder.set_output_array(out);
-  if (mode_ == QuantizationMode::Affine) {
-    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
-    encoder.set_input_array(biases);
   encoder.dispatch([out = array::unsafe_weak_copy(out),
                     x = array::unsafe_weak_copy(x),
                     w = array::unsafe_weak_copy(w),
@@ -945,18 +622,6 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
         bits_,
         transpose_);
   });
-  } else {
-    encoder.dispatch([out = array::unsafe_weak_copy(out),
-                      x = array::unsafe_weak_copy(x),
-                      w = array::unsafe_weak_copy(w),
-                      scales = array::unsafe_weak_copy(scales),
-                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
-                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
-                      transpose_ = transpose_]() mutable {
-      mxfp4_bs_qmm_dispatch(
-          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
-    });
-  }
 }
 
 template <typename T, typename U>
@@ -1040,7 +705,7 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::Quantize::eval_cpu(
+void fast::AffineQuantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto ensure_row_contiguous = [s = stream()](const array& arr) {
@@ -1099,7 +764,7 @@ void fast::Quantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
+          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
     }
   });
 }

mlx/backend/cpu/simd/accelerate_simd.h

@@ -234,7 +234,6 @@ Simd<T, N> remainder(Simd<T, N> a, Simd<T, N> b) {
 
 template <typename MaskT, typename T1, typename T2, int N>
 Simd<T1, N> select(Simd<MaskT, N> mask, Simd<T1, N> x, Simd<T2, N> y) {
-  static_assert(std::is_same_v<MaskT, bool>);
   if constexpr (sizeof(T1) == 1) {
     return asd::bitselect(y.value, x.value, asd::convert<char>(mask.value));
   } else if constexpr (sizeof(T1) == 2) {
@@ -252,13 +251,9 @@ Simd<T, N> pow(Simd<T, N> base, Simd<T, N> exp) {
     return asd::pow(base.value, exp.value);
   } else {
     Simd<T, N> res = 1;
-    // Raising an integer to a negative power is undefined
+    while (any(exp)) {
-    if (any(exp < 0)) {
+      res = select(exp & 1, res * base, res);
-      return 0;
+      base = select(exp, base * base, base);
-    }
-    while (any(exp > 0)) {
-      res = select((exp & 1) != 0, res * base, res);
-      base = select(exp > 0, base * base, base);
       exp = exp >> 1;
     }
     return res;

mlx/backend/cpu/svd.cpp

@@ -81,7 +81,9 @@ void svd_impl(
     // Vᵀ of shape N x N. (M x M in lapack).
     const int ldvt = M;
 
-    auto jobz = (u_ptr) ? "A" : "N";
+    auto job_u = (u_ptr) ? "V" : "N";
+    auto job_vt = (u_ptr) ? "V" : "N";
+    static constexpr auto range = "A";
 
     // Will contain the number of singular values after the call has returned.
     int ns = 0;
@@ -89,20 +91,30 @@ void svd_impl(
 
     // Will contain the indices of eigenvectors that failed to converge (not
     // used here but required by lapack).
-    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
+    auto iwork = array::Data{allocator::malloc(sizeof(int) * 12 * K)};
 
     static const int lwork_query = -1;
 
+    static const int ignored_int = 0;
+    static const T ignored_float = 0;
+
     int info;
 
     // Compute workspace size.
-    gesdd<T>(
+    gesvdx<T>(
-        /* jobz = */ jobz,
+        /* jobu = */ job_u,
+        /* jobvt = */ job_vt,
+        /* range = */ range,
         // M and N are swapped since lapack expects column-major.
         /* m = */ &N,
         /* n = */ &M,
         /* a = */ nullptr,
         /* lda = */ &lda,
+        /* vl = */ &ignored_float,
+        /* vu = */ &ignored_float,
+        /* il = */ &ignored_int,
+        /* iu = */ &ignored_int,
+        /* ns = */ &ns,
         /* s = */ nullptr,
         /* u = */ nullptr,
         /* ldu = */ &ldu,
@@ -124,13 +136,20 @@ void svd_impl(
 
     // Loop over matrices.
     for (int i = 0; i < num_matrices; i++) {
-      gesdd<T>(
+      gesvdx<T>(
-          /* jobz = */ jobz,
+          /* jobu = */ job_u,
+          /* jobvt = */ job_vt,
+          /* range = */ range,
           // M and N are swapped since lapack expects column-major.
           /* m = */ &N,
           /* n = */ &M,
           /* a = */ in_ptr + M * N * i,
           /* lda = */ &lda,
+          /* vl = */ &ignored_float,
+          /* vu = */ &ignored_float,
+          /* il = */ &ignored_int,
+          /* iu = */ &ignored_int,
+          /* ns = */ &ns,
           /* s = */ s_ptr + K * i,
           // According to the identity above, lapack will write Vᵀᵀ as U.
           /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
@@ -148,6 +167,13 @@ void svd_impl(
         ss << "svd_impl: sgesvdx_ failed with code " << info;
         throw std::runtime_error(ss.str());
       }
+
+      if (ns != K) {
+        std::stringstream ss;
+        ss << "svd_impl: expected " << K << " singular values, but " << ns
+           << " were computed.";
+        throw std::runtime_error(ss.str());
+      }
     }
   });
   encoder.add_temporary(in);

mlx/backend/cuda/CMakeLists.txt

@@ -8,6 +8,7 @@ target_sources(
   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
           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cu
@@ -16,18 +17,14 @@ target_sources(
           ${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}/conv/gemm_conv.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/conv/gemm_grouped_conv.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/cudnn_utils.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/custom_kernel.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cu
           ${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
@@ -49,14 +46,12 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/sort.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/binary)
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/unary)
-
 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)
@@ -154,7 +149,7 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 FetchContent_Declare(
   cudnn
   GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
-  GIT_TAG v1.14.0
+  GIT_TAG v1.12.1
   GIT_SHALLOW TRUE
   EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)

mlx/backend/cuda/allocator.cpp

@@ -30,15 +30,8 @@ SmallSizePool::SmallSizePool() {
   next_free_ = buffer_;
 
   CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
-#if CUDART_VERSION >= 13000
-  cudaMemLocation loc;
-  loc.type = cudaMemLocationTypeDevice;
-  loc.id = 0;
-#else
-  int loc = 0;
-#endif // CUDART_VERSION >= 13000
   CHECK_CUDA_ERROR(
-      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, loc));
+      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, 0));
 
   auto curr = next_free_;
   for (size_t i = 1; i < num_blocks; ++i) {

mlx/backend/cuda/arange.cu

@@ -6,33 +6,23 @@
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
-#include <cooperative_groups.h>
 #include <nvtx3/nvtx3.hpp>
+#include <thrust/device_ptr.h>
+#include <thrust/transform.h>
 
 namespace mlx::core {
 
 namespace cu {
 
-namespace cg = cooperative_groups;
+template <typename T>
+struct Arange {
+  const T start;
+  const T step;
 
-template <typename T, typename IdxT, int N_WRITES>
+  __device__ T operator()(uint32_t i) const {
-__global__ void arange(T* out, IdxT size, T start, T step) {
+    return start + i * step;
-  IdxT index = cg::this_grid().thread_rank();
-
-  if ((index + 1) * N_WRITES > size) {
-    for (IdxT i = index * N_WRITES; i < size; ++i) {
-      out[i] = start + i * step;
-    }
-  } else {
-    AlignedVector<T, N_WRITES> out_vec;
-#pragma unroll
-    for (int i = 0; i < N_WRITES; ++i) {
-      out_vec[i] = start + (index * N_WRITES + i) * step;
-    }
-
-    store_vector<N_WRITES>(out, index, out_vec);
-  }
   }
+};
 
 } // namespace cu
 
@@ -46,23 +36,19 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   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>;
-    constexpr int N_WRITES = 16 / sizeof(OutType);
+    CTYPE step =
-    dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
+        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
-      using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+    thrust::transform(
-      auto [num_blocks, block_dims] = get_launch_args(out, large(), N_WRITES);
+        cu::thrust_policy(encoder.stream()),
-      encoder.add_kernel_node(
+        thrust::counting_iterator<uint32_t>(0),
-          cu::arange<OutType, IdxT, N_WRITES>,
+        thrust::counting_iterator<uint32_t>(out.data_size()),
-          num_blocks,
+        thrust::device_pointer_cast(out.data<OutType>()),
-          block_dims,
+        cu::Arange<OutType>{
-          0,
+            static_cast<OutType>(start_), static_cast<OutType>(step)});
-          out.data<OutType>(),
-          out.data_size(),
-          static_cast<CTYPE>(start_),
-          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));
-    });
   });
 }
 

mlx/backend/cuda/binary.cu

@@ -99,89 +99,39 @@ __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
   }
 }
 
-template <
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
-    typename Op,
-    typename In,
-    typename Out,
-    typename IdxT,
-    int NDIM,
-    int N_READS>
 __global__ void binary_g_nd(
     const In* a,
     const In* b,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[NDIM - 1];
-  auto a_stride_x = a_strides[NDIM - 1];
-  auto b_stride_x = b_strides[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
-      index_rest * shape_x, shape.data(), a_strides.data(), b_strides.data());
+        index, shape.data(), a_strides.data(), b_strides.data());
-  auto a_vec =
+    out[index] = Op{}(a[a_idx], b[b_idx]);
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(a_vec[i], b_vec[i]);
   }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT>
 __global__ void binary_g(
     const In* a,
     const In* b,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto a_stride_x = a_strides[ndim - 1];
-  auto b_stride_x = b_strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [a_idx, b_idx] = elem_to_loc(
-      index_rest * shape_x,
+        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
-      shape.data(),
+    out[index] = Op{}(a[a_idx], b[b_idx]);
-      a_strides.data(),
-      b_strides.data(),
-      ndim);
-  auto a_vec =
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(a_vec[i], b_vec[i]);
   }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -259,61 +209,39 @@ void binary_op_gpu_inplace(
                 auto& a_strides = strides[0];
                 auto& b_strides = strides[1];
                 int ndim = shape.size();
-                int work_per_thread = 1;
-                auto dim0 = ndim > 0 ? shape.back() : 1;
-                auto rest = out.size() / dim0;
-                if (dim0 >= 4) {
-                  work_per_thread = 4;
-                }
-                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-                auto block_dims = get_block_dims(dim0, rest, 1);
-                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::binary_g_nd<
+                    auto [num_blocks, block_dims] =
-                        Op,
+                        get_launch_args(out, large());
-                        InType,
-                        OutType,
-                        IdxT,
-                        dims_constant(),
-                        1>;
-                    if (work_per_thread == 4) {
-                      kernel = cu::binary_g_nd<
-                          Op,
-                          InType,
-                          OutType,
-                          IdxT,
-                          dims_constant(),
-                          4>;
-                    }
                     encoder.add_kernel_node(
-                        kernel,
+                        cu::binary_g_nd<
-                        {num_blocks_x, num_blocks_y},
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
+                        num_blocks,
                         block_dims,
                         0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out.data<OutType>(),
-                        rest,
+                        out.size(),
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT, 1>;
+                  auto [num_blocks, block_dims] = get_launch_args(out, large());
-                  if (work_per_thread == 4) {
-                    kernel = cu::binary_g<Op, InType, OutType, IdxT, 4>;
-                  }
                   encoder.add_kernel_node(
-                      kernel,
+                      cu::binary_g<Op, InType, OutType, IdxT>,
-                      {num_blocks_x, num_blocks_y},
+                      num_blocks,
                       block_dims,
                       0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out.data<OutType>(),
-                      rest,
+                      out.size(),
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),
@@ -376,4 +304,54 @@ void binary_op_gpu(
     binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
   }
 
+BINARY_GPU(Add)
+BINARY_GPU(ArcTan2)
+BINARY_GPU(Divide)
+BINARY_GPU(Remainder)
+BINARY_GPU(Greater)
+BINARY_GPU(GreaterEqual)
+BINARY_GPU(Less)
+BINARY_GPU(LessEqual)
+BINARY_GPU(LogicalAnd)
+BINARY_GPU(LogicalOr)
+BINARY_GPU(LogAddExp)
+BINARY_GPU(Maximum)
+BINARY_GPU(Minimum)
+BINARY_GPU(Multiply)
+BINARY_GPU(NotEqual)
+BINARY_GPU(Power)
+BINARY_GPU(Subtract)
+
+void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Equal::eval_gpu");
+  auto& s = out.primitive().stream();
+  if (equal_nan_) {
+    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
+  } else {
+    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
+  }
+}
+
+void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
+  auto& s = out.primitive().stream();
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Or:
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Xor:
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::RightShift:
+      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
+      break;
+  }
+}
+
 } // namespace mlx::core

mlx/backend/cuda/binary/CMakeLists.txt

@@ -1,21 +0,0 @@
-target_sources(
-  mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/add.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctan2.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/bitwise_binary.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/divide.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_and.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_or.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log_add_exp.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/minimum.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/maximum.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/multiply.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/power.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/remainder.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/not_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/subtract.cu)

mlx/backend/cuda/binary/add.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Add)
-} // namespace mlx::core

mlx/backend/cuda/binary/arctan2.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(ArcTan2)
-} // namespace mlx::core

mlx/backend/cuda/binary/bitwise_binary.cu

@@ -1,27 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
-  auto& s = out.primitive().stream();
-  switch (op_) {
-    case BitwiseBinary::And:
-      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
-      break;
-  }
-}
-} // namespace mlx::core

mlx/backend/cuda/binary/divide.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Divide)
-} // namespace mlx::core

mlx/backend/cuda/binary/equal.cu

@@ -1,15 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Equal::eval_gpu");
-  auto& s = out.primitive().stream();
-  if (equal_nan_) {
-    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
-  } else {
-    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
-  }
-}
-} // namespace mlx::core

mlx/backend/cuda/binary/greater.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Greater)
-} // namespace mlx::core

mlx/backend/cuda/binary/greater_equal.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(GreaterEqual)
-} // namespace mlx::core

mlx/backend/cuda/binary/less.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Less)
-} // namespace mlx::core

mlx/backend/cuda/binary/less_equal.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LessEqual)
-} // namespace mlx::core

mlx/backend/cuda/binary/log_add_exp.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogAddExp)
-} // namespace mlx::core

mlx/backend/cuda/binary/logical_and.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogicalAnd)
-} // namespace mlx::core

mlx/backend/cuda/binary/logical_or.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogicalOr)
-} // namespace mlx::core

mlx/backend/cuda/binary/maximum.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Maximum)
-} // namespace mlx::core

mlx/backend/cuda/binary/minimum.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Minimum)
-} // namespace mlx::core

mlx/backend/cuda/binary/multiply.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Multiply)
-} // namespace mlx::core

mlx/backend/cuda/binary/not_equal.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(NotEqual)
-} // namespace mlx::core

mlx/backend/cuda/binary/power.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Power)
-} // namespace mlx::core

mlx/backend/cuda/binary/remainder.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Remainder)
-} // namespace mlx::core

mlx/backend/cuda/binary/subtract.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Subtract)
-} // namespace mlx::core

mlx/backend/cuda/binary_two.cu

@@ -127,99 +127,45 @@ binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   }
 }
 
-template <
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
-    typename Op,
-    typename In,
-    typename Out,
-    typename IdxT,
-    int NDIM,
-    int N_READS>
 __global__ void binary_two_g_nd(
     const In* a,
     const In* b,
     Out* out_a,
     Out* out_b,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[NDIM - 1];
-  auto a_stride_x = a_strides[NDIM - 1];
-  auto b_stride_x = b_strides[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
-      index_rest * shape_x, shape.data(), a_strides.data(), b_strides.data());
+        index, shape.data(), a_strides.data(), b_strides.data());
-  auto a_vec =
+    auto out = Op{}(a[a_idx], b[b_idx]);
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
+    out_a[index] = out[0];
-  auto b_vec =
+    out_b[index] = out[1];
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec_a;
-  AlignedVector<Out, N_READS> out_vec_b;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    auto out = Op{}(a_vec[i], b_vec[i]);
-    out_vec_a[i] = out[0];
-    out_vec_b[i] = out[1];
   }
-  store_vector(out_a + shape_x * index_rest, index_x, out_vec_a, shape_x);
-  store_vector(out_b + shape_x * index_rest, index_x, out_vec_b, shape_x);
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT>
 __global__ void binary_two_g(
     const In* a,
     const In* b,
     Out* out_a,
     Out* out_b,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto a_stride_x = a_strides[ndim - 1];
-  auto b_stride_x = b_strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [a_idx, b_idx] = elem_to_loc(
-      index_rest * shape_x,
+        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
-      shape.data(),
+    auto out = Op{}(a[a_idx], b[b_idx]);
-      a_strides.data(),
+    out_a[index] = out[0];
-      b_strides.data(),
+    out_b[index] = out[1];
-      ndim);
-  auto a_vec =
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, In(0));
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, In(0));
-
-  AlignedVector<Out, N_READS> out_vec_a;
-  AlignedVector<Out, N_READS> out_vec_b;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    auto out = Op{}(a_vec[i], b_vec[i]);
-    out_vec_a[i] = out[0];
-    out_vec_b[i] = out[1];
   }
-  store_vector(out_a + shape_x * index_rest, index_x, out_vec_a, shape_x);
-  store_vector(out_b + shape_x * index_rest, index_x, out_vec_b, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -279,64 +225,42 @@ void binary_two_op_gpu_inplace(
                 auto& a_strides = strides[0];
                 auto& b_strides = strides[1];
                 int ndim = shape.size();
-                int work_per_thread = 1;
-                auto dim0 = ndim > 0 ? shape.back() : 1;
-                auto rest = out_a.size() / dim0;
-                if (dim0 >= 4) {
-                  work_per_thread = 4;
-                }
-                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-                auto block_dims = get_block_dims(dim0, rest, 1);
-                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
                 if (ndim <= 3) {
                   dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::binary_two_g_nd<
+                    auto [num_blocks, block_dims] =
-                        Op,
+                        get_launch_args(out_a, large());
-                        InType,
-                        OutType,
-                        IdxT,
-                        dims_constant(),
-                        1>;
-                    if (work_per_thread == 4) {
-                      kernel = cu::binary_two_g_nd<
-                          Op,
-                          InType,
-                          OutType,
-                          IdxT,
-                          dims_constant(),
-                          4>;
-                    }
                     encoder.add_kernel_node(
-                        kernel,
+                        cu::binary_two_g_nd<
-                        {num_blocks_x, num_blocks_y},
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
+                        num_blocks,
                         block_dims,
                         0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out_a.data<OutType>(),
                         out_b.data<OutType>(),
-                        rest,
+                        out_a.size(),
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 1>;
+                  auto [num_blocks, block_dims] =
-                  if (work_per_thread == 4) {
+                      get_launch_args(out_a, large());
-                    kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 4>;
-                  }
                   encoder.add_kernel_node(
-                      kernel,
+                      cu::binary_two_g<Op, InType, OutType, IdxT>,
-                      {num_blocks_x, num_blocks_y},
+                      num_blocks,
                       block_dims,
                       0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out_a.data<OutType>(),
                       out_b.data<OutType>(),
-                      rest,
+                      out_a.size(),
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),

mlx/backend/cuda/compiled.cpp

@@ -267,8 +267,7 @@ void Compiled::eval_gpu(
       }
     }
 
-    return std::make_tuple(
+    return std::make_pair(std::move(builder.os), std::move(kernel_names));
-        false, std::move(builder.os), std::move(kernel_names));
   });
 
   // Collapse contiguous dims to route to a faster kernel if possible. Also

mlx/backend/cuda/conv.cpp

@@ -1,12 +1,18 @@
 // Copyright © 2025 Apple Inc.
 
-#include "mlx/backend/cuda/conv/conv.h"
-#include "mlx/backend/cuda/cudnn_utils.h"
 #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>
@@ -15,6 +21,9 @@ 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 \
@@ -22,9 +31,6 @@ namespace {
 #define CONV_BACKWARD_WEIGHT \
   CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
 
-// Custom placeholder representing fallback kernel.
-#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
-
 struct ConvCacheKey {
   int device_id;
   cudnnDataType_t cudnn_dtype;
@@ -44,13 +50,203 @@ struct ConvCacheKey {
 auto& conv_cache() {
   static LRUBytesKeyCache<
       ConvCacheKey,
-      std::pair<
+      std::pair<cudnnBackendDescriptorType_t, cudnn_frontend::ExecutionPlan>>
-          cudnnBackendDescriptorType_t,
-          std::optional<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,
@@ -95,7 +291,7 @@ auto get_conv_op_settings(
   }
 }
 
-std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
+std::optional<cudnn_frontend::OperationGraph> build_op_graph(
     cu::CommandEncoder& encoder,
     cudnnBackendDescriptorType_t backend_type,
     Dtype dtype,
@@ -121,9 +317,9 @@ std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
                          .build();
 
     auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .setxDesc(build_cudnn_tensor_nchw('x', x))
+                  .setxDesc(build_tensor('x', x))
-                  .setwDesc(build_cudnn_tensor_nchw('w', w))
+                  .setwDesc(build_tensor('w', w))
-                  .setyDesc(build_cudnn_tensor_nchw('y', y))
+                  .setyDesc(build_tensor('y', y))
                   .setcDesc(conv_desc)
                   .build();
 
@@ -140,42 +336,6 @@ std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
   }
 }
 
-// Transpose from (C_out, H, W, C_in / groups) to (C_in, H, W, C_out / groups).
-array group_transpose(
-    const array& x,
-    int groups,
-    int group_dim,
-    int axis1,
-    int axis2,
-    Stream s) {
-  if (groups == 1) {
-    return swapaxes_in_eval(x, axis1, axis2);
-  }
-  int ndim = x.ndim();
-  if (group_dim < 0) {
-    group_dim += ndim;
-  }
-  if (axis1 < 0) {
-    axis1 += ndim;
-  }
-  if (axis2 < 0) {
-    axis2 += ndim;
-  }
-  if (group_dim <= axis1) {
-    axis1 += 1;
-  }
-  if (group_dim <= axis2) {
-    axis2 += 1;
-  }
-  auto shape = x.shape();
-  shape.insert(shape.begin() + group_dim, groups);
-  shape[group_dim + 1] = shape[group_dim + 1] / groups;
-  array x_trans = reshape_in_eval(x, std::move(shape), s);
-  x_trans = swapaxes_in_eval(x_trans, axis1, axis2);
-  x_trans = flatten_in_eval(x_trans, group_dim, group_dim + 1, s);
-  return x_trans;
-}
-
 // 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.
@@ -185,14 +345,13 @@ std::tuple<array, array, array> prepare_args(
     array in,
     array wt,
     array out,
-    int groups,
     Stream s) {
   // Transpose the args depending on the backend type.
   // TODO: Handle groups.
   if (backend_type == CONV_BACKWARD_INPUT) {
-    wt = group_transpose(wt, groups, 0, 0, -1, s);
+    wt = swapaxes_in_eval(wt, 0, -1);
   } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    in = group_transpose(in, groups, -1, 0, -1, s);
+    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.
@@ -285,12 +444,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   ConvCacheKey cache_key{
       encoder.device().cuda_device(),
       dtype_to_cudnn_type(dtype),
-      vector_key(in.shape()),
+      fixed_vector(in.shape()),
-      vector_key(wt.shape()),
+      fixed_vector(wt.shape()),
-      vector_key(kernel_strides_),
+      fixed_vector(kernel_strides_),
-      vector_key(padding_lo_),
+      fixed_vector(padding_lo_),
-      vector_key(padding_hi_),
+      fixed_vector(padding_hi_),
-      vector_key(kernel_dilation_),
+      fixed_vector(kernel_dilation_),
       groups_,
       flip_,
       get_alignment(in),
@@ -298,30 +457,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       get_alignment(out)};
   if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
     auto& [backend_type, plan] = it->second;
-    if (plan) {
+    std::tie(in, wt, out) = prepare_args(encoder, backend_type, in, wt, out, s);
-      // Run cached plan.
-      std::tie(in, wt, out) =
-          prepare_args(encoder, backend_type, in, wt, out, groups_, s);
     register_args(encoder, backend_type, in, wt, out, out_);
     auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-      if (!encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+    if (!execute_plan(encoder, plan, x, w, y)) {
       throw std::runtime_error("[conv] Cached plan failed to execute.");
     }
-    } else {
-      // Run fallback kernel.
-      gemm_conv(
-          encoder,
-          in,
-          wt,
-          out,
-          kernel_strides_,
-          padding_lo_,
-          kernel_dilation_,
-          input_dilation_,
-          groups_,
-          flip_,
-          s);
-    }
     return;
   }
 
@@ -349,7 +490,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   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, groups_, s);
+        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,
@@ -361,7 +502,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_hi_,
         kernel_dilation_,
         input_dilation_);
-    op_graph = build_conv_op_graph(
+    op_graph = build_op_graph(
         encoder,
         try_backend,
         dtype,
@@ -380,39 +521,26 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       break;
     }
   }
+  if (!op_graph) {
+    throw std::runtime_error("[conv] Can not build op graph.");
+  }
 
-  if (op_graph) {
+  // Get ready to execute the graph.
-    // Setup inputs and outputs.
   register_args(encoder, backend_type, in, wt, out, out_);
 
-    // Find a plan for the graph and execute it.
+  // Try to run plans based on heuristics.
-    auto plan = find_cudnn_plan_from_op_graph(
+  auto configs = get_engine_configs(backend_type, dtype, *op_graph);
-        encoder.device().cudnn_handle(), backend_type, dtype, *op_graph);
+  auto tag = op_graph->getTag();
-    if (!plan) {
-      throw std::runtime_error("[conv] Unable to find an execution plan.");
-    }
   auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-    if (encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-      conv_cache().emplace(
-          cache_key, std::make_pair(backend_type, std::move(*plan)));
     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.");
-  // Use fallback kernel for settings not supported by cuDNN.
-  gemm_conv(
-      encoder,
-      in,
-      wt,
-      out,
-      kernel_strides_,
-      padding_lo_,
-      kernel_dilation_,
-      input_dilation_,
-      groups_,
-      flip_,
-      s);
-  conv_cache().emplace(cache_key, std::make_pair(CONV_FALLBACK, std::nullopt));
 }
 
 } // namespace mlx::core

mlx/backend/cuda/conv/conv.h

@@ -1,126 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/gpu/copy.h"
-
-namespace mlx::core {
-
-template <int NDIM>
-struct ConvParams {
-  int N; // Batch size
-  int C; // In channels
-  int O; // Out channels
-  int strides[NDIM];
-  int padding[NDIM];
-  int kernel_dilation[NDIM];
-  int input_dilation[NDIM];
-  int groups;
-  bool flip;
-  int in_spatial_dims[NDIM];
-  int wt_spatial_dims[NDIM];
-  int out_spatial_dims[NDIM];
-  int64_t in_strides[NDIM + 2];
-
-  ConvParams(
-      const array& in,
-      const array& wt,
-      const array& out,
-      const std::vector<int>& strides,
-      const std::vector<int>& padding,
-      const std::vector<int>& kernel_dilation,
-      const std::vector<int>& input_dilation,
-      int groups,
-      bool flip)
-      : N(in.shape(0)),
-        C(in.shape(-1)),
-        O(wt.shape(0)),
-        groups(groups),
-        flip(flip) {
-    std::copy_n(strides.begin(), NDIM, this->strides);
-    std::copy_n(padding.begin(), NDIM, this->padding);
-    std::copy_n(kernel_dilation.begin(), NDIM, this->kernel_dilation);
-    std::copy_n(input_dilation.begin(), NDIM, this->input_dilation);
-    std::copy_n(in.shape().begin() + 1, NDIM, this->in_spatial_dims);
-    std::copy_n(wt.shape().begin() + 1, NDIM, this->wt_spatial_dims);
-    std::copy_n(out.shape().begin() + 1, NDIM, this->out_spatial_dims);
-    std::copy_n(in.strides().begin(), NDIM + 2, this->in_strides);
-  }
-};
-
-void gemm_grouped_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s);
-
-void gemm_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    bool flip,
-    Stream s);
-
-inline void gemm_conv(
-    cu::CommandEncoder& encoder,
-    array in,
-    array wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s) {
-  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);
-  }
-
-  if (groups == 1) {
-    gemm_conv(
-        encoder,
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        flip,
-        s);
-  } else {
-    gemm_grouped_conv(
-        encoder,
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        groups,
-        flip,
-        s);
-  }
-}
-
-} // namespace mlx::core

mlx/backend/cuda/conv/gemm_conv.cu

@@ -1,217 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/conv/conv.h"
-#include "mlx/backend/cuda/gemms/cublas_gemm.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-
-#include <cooperative_groups.h>
-
-namespace mlx::core {
-
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-template <typename T, int NDIM>
-__global__ void naive_unfold_nd(
-    const T* in,
-    T* out,
-    int filter_size,
-    int out_pixels,
-    const __grid_constant__ ConvParams<NDIM> params) {
-  auto block = cg::this_thread_block();
-  auto tid = block.group_index();
-  auto lid = block.thread_index();
-
-  int index_batch = tid.z / out_pixels; // [0, N)
-  int index_out_spatial = tid.z % out_pixels; // [0, H_out * W_out)
-  int index_wt_spatial =
-      tid.x * block.dim_threads().x + lid.x; // [0, H_wt * W_wt)
-
-  if (index_wt_spatial >= filter_size / params.C) {
-    return;
-  }
-
-  in += tid.y; // [0, C)
-  out += tid.z * filter_size + index_wt_spatial * params.C + tid.y;
-
-  bool valid = index_batch < params.N;
-
-  // Get the coordinates in input.
-  int index_in[NDIM] = {};
-#pragma unroll
-  for (int i = NDIM - 1; i >= 0; --i) {
-    int index_out = index_out_spatial % params.out_spatial_dims[i];
-    int index_wt = index_wt_spatial % params.wt_spatial_dims[i];
-
-    if (params.flip) {
-      index_wt = params.wt_spatial_dims[i] - index_wt - 1;
-    }
-
-    int index = index_out * params.strides[i] - params.padding[i] +
-        index_wt * params.kernel_dilation[i];
-    int index_max =
-        1 + params.input_dilation[i] * (params.in_spatial_dims[i] - 1);
-
-    valid &= (index >= 0) && (index < index_max) &&
-        (index % params.input_dilation[i] == 0);
-
-    index_in[i] = index / params.input_dilation[i];
-
-    index_out_spatial /= params.out_spatial_dims[i];
-    index_wt_spatial /= params.wt_spatial_dims[i];
-  }
-
-  if (valid) {
-    int in_offset = index_batch * params.in_strides[0];
-#pragma unroll
-    for (int i = 0; i < NDIM; ++i) {
-      in_offset += index_in[i] * params.in_strides[i + 1];
-    }
-    *out = in[in_offset];
-  } else {
-    *out = T{0};
-  }
-}
-
-} // namespace cu
-
-template <int NDIM>
-array unfold_inputs_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    int mat_M,
-    int mat_K,
-    int mat_N,
-    ConvParams<NDIM>& params) {
-  array unfolded({mat_M, mat_K}, in.dtype(), nullptr, {});
-  unfolded.set_data(allocator::malloc(unfolded.nbytes()));
-  encoder.add_temporary(unfolded);
-
-  int filter_size = params.C;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    filter_size *= params.wt_spatial_dims[i];
-  }
-
-  int out_pixels = 1;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    out_pixels *= params.out_spatial_dims[i];
-  }
-
-  int wt_spatial_size = mat_K / params.C;
-  dim3 block_dims;
-  block_dims.x = std::min(std::max(wt_spatial_size, 32), 1024);
-  dim3 num_blocks;
-  num_blocks.x = cuda::ceil_div(wt_spatial_size, block_dims.x);
-  num_blocks.y = params.C;
-  num_blocks.z = mat_M;
-
-  encoder.set_input_array(in);
-  encoder.set_output_array(unfolded);
-  dispatch_float_types(in.dtype(), "unfold", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    encoder.add_kernel_node(
-        cu::naive_unfold_nd<DataType, NDIM>,
-        num_blocks,
-        block_dims,
-        0,
-        in.data<DataType>(),
-        unfolded.data<DataType>(),
-        filter_size,
-        out_pixels,
-        params);
-  });
-
-  return unfolded;
-}
-
-template <int NDIM>
-void gemm_conv_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    ConvParams<NDIM>& params,
-    Stream s) {
-  // Get gemm shapes.
-  int mat_M = out.size() / params.O; // N * H_out * W_out
-  int mat_K = wt.size() / params.O; // C * H_wt * W_wt
-  int mat_N = params.O; // O
-
-  // Unfold input to (N * H_out * W_out, C * H_wt * W_wt) for gemm.
-  array in_unfolded =
-      unfold_inputs_nd<NDIM>(encoder, in, mat_M, mat_K, mat_N, params);
-
-  // Reshape weight to (C * H_wt * W_wt, O) for gemm.
-  array wt_reshaped({mat_K, mat_N}, wt.dtype(), nullptr, {});
-  wt_reshaped.copy_shared_buffer(
-      wt,
-      {1, mat_K},
-      {false, false, /* col_contiguous */ true},
-      wt.data_size());
-
-  // Single batch.
-  Shape batch_shape{1};
-  Strides a_batch_strides{0};
-  Strides b_batch_strides{0};
-
-  // Run matmul.
-  CublasGemm gemm(
-      encoder.device(),
-      in.dtype(),
-      false, // a_transposed
-      mat_M, // a_rows
-      mat_K, // a_cols
-      mat_K, // lda
-      true, // b_transposed
-      mat_K, // b_rows
-      mat_N, // b_cols
-      mat_K, // ldb
-      batch_shape.back(),
-      a_batch_strides.back(),
-      b_batch_strides.back());
-  gemm.run(
-      encoder,
-      out,
-      in_unfolded,
-      wt_reshaped,
-      batch_shape,
-      a_batch_strides,
-      b_batch_strides);
-}
-
-void gemm_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    bool flip,
-    Stream s) {
-  int conv_ndim = in.ndim() - 2;
-  if (conv_ndim < 1 || conv_ndim > 3) {
-    throw std::runtime_error(
-        fmt::format("[conv] Unsupported gemm_conv for {}D conv.", conv_ndim));
-  }
-  dispatch_1_2_3(conv_ndim, [&](auto ndim_constant) {
-    ConvParams<ndim_constant()> params(
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        1, // groups
-        flip);
-    gemm_conv_nd<ndim_constant()>(encoder, in, wt, out, params, s);
-  });
-}
-
-} // namespace mlx::core

mlx/backend/cuda/conv/gemm_grouped_conv.cu

@@ -1,231 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/conv/conv.h"
-#include "mlx/backend/cuda/gemms/cublas_gemm.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-
-#include <cooperative_groups.h>
-
-namespace mlx::core {
-
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-template <typename T, int NDIM>
-__global__ void naive_grouped_unfold_transpose_nd(
-    const T* in,
-    T* out,
-    int filter_size,
-    int out_pixels,
-    const __grid_constant__ ConvParams<NDIM> params) {
-  auto block = cg::this_thread_block();
-  auto tid = block.group_index();
-  auto lid = block.thread_index();
-
-  int index_batch = tid.z / out_pixels; // [0, N)
-  int index_out_spatial = tid.z % out_pixels; // [0, H_out * W_out)
-  int index_wt_spatial =
-      tid.x * block.dim_threads().x + lid.x; // [0, H_wt * W_wt)
-
-  if (index_wt_spatial >= filter_size / params.C) {
-    return;
-  }
-
-  in += tid.y; // [0, C)
-  out += tid.z * filter_size + tid.y * (filter_size / params.C);
-
-  bool valid = index_batch < params.N;
-
-  // Get the coordinates in input.
-  int index_in[NDIM] = {};
-  int wt_stride = 1;
-#pragma unroll
-  for (int i = NDIM - 1; i >= 0; --i) {
-    int index_out = index_out_spatial % params.out_spatial_dims[i];
-    int index_wt = index_wt_spatial % params.wt_spatial_dims[i];
-    out += index_wt * wt_stride;
-
-    if (params.flip) {
-      index_wt = params.wt_spatial_dims[i] - index_wt - 1;
-    }
-
-    int index = index_out * params.strides[i] - params.padding[i] +
-        index_wt * params.kernel_dilation[i];
-    int index_max =
-        1 + params.input_dilation[i] * (params.in_spatial_dims[i] - 1);
-
-    valid &= (index >= 0) && (index < index_max) &&
-        (index % params.input_dilation[i] == 0);
-
-    index_in[i] = index / params.input_dilation[i];
-
-    index_out_spatial /= params.out_spatial_dims[i];
-    index_wt_spatial /= params.wt_spatial_dims[i];
-    wt_stride *= params.wt_spatial_dims[i];
-  }
-
-  if (valid) {
-    int in_offset = index_batch * params.in_strides[0];
-#pragma unroll
-    for (int i = 0; i < NDIM; ++i) {
-      in_offset += index_in[i] * params.in_strides[i + 1];
-    }
-    *out = in[in_offset];
-  } else {
-    *out = T{0};
-  }
-}
-
-} // namespace cu
-
-template <int NDIM>
-array grouped_unfold_transpose_inputs_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    int mat_M,
-    int mat_K,
-    int mat_N,
-    ConvParams<NDIM>& params) {
-  array unfolded({mat_M, mat_K * params.groups}, in.dtype(), nullptr, {});
-  unfolded.set_data(allocator::malloc(unfolded.nbytes()));
-  encoder.add_temporary(unfolded);
-
-  int filter_size = params.C;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    filter_size *= params.wt_spatial_dims[i];
-  }
-
-  int out_pixels = 1;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    out_pixels *= params.out_spatial_dims[i];
-  }
-
-  int wt_spatial_size = (mat_K * params.groups) / params.C;
-  dim3 block_dims;
-  block_dims.x = std::min(std::max(wt_spatial_size, 32), 1024);
-  dim3 num_blocks;
-  num_blocks.x = cuda::ceil_div(wt_spatial_size, block_dims.x);
-  num_blocks.y = params.C;
-  num_blocks.z = mat_M;
-
-  encoder.set_input_array(in);
-  encoder.set_output_array(unfolded);
-  dispatch_float_types(in.dtype(), "unfold", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    encoder.add_kernel_node(
-        cu::naive_grouped_unfold_transpose_nd<DataType, NDIM>,
-        num_blocks,
-        block_dims,
-        0,
-        in.data<DataType>(),
-        unfolded.data<DataType>(),
-        filter_size,
-        out_pixels,
-        params);
-  });
-
-  return unfolded;
-}
-
-template <int NDIM>
-void gemm_grouped_conv_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    ConvParams<NDIM>& params,
-    Stream s) {
-  // Get gemm shapes.
-  int C_per_group = params.C / params.groups;
-  int O_per_group = params.O / params.groups;
-  int mat_M = out.size() / params.O; // N * H_out * W_out
-  int mat_K = wt.size() / params.O; // C_per_group * H_wt * W_wt
-  int mat_N = O_per_group; // O_per_group
-
-  // Unfold input to (N * H_out * W_out, C * H_wt * W_wt) for gemm.
-  array in_unfolded = grouped_unfold_transpose_inputs_nd<NDIM>(
-      encoder, in, mat_M, mat_K, mat_N, params);
-
-  // Reshape weight to (O, C_per_group, H_wt * W_wt) for gemm.
-  int wt_spatial_size = (wt.size() / wt.shape(0)) / wt.shape(-1);
-  array wt_view(
-      {params.O, C_per_group, wt_spatial_size}, wt.dtype(), nullptr, {});
-  wt_view.copy_shared_buffer(
-      wt, {wt.strides(0), 1, C_per_group}, wt.flags(), wt.size());
-  array wt_reshaped = contiguous_copy_gpu(wt_view, s);
-
-  // Batch with size of groups.
-  Shape batch_shape{params.groups};
-  Strides a_batch_strides{mat_K};
-  Strides b_batch_strides{mat_N * mat_K};
-
-  // Run matmul.
-  CublasGemm gemm(
-      encoder.device(),
-      in.dtype(),
-      false, // a_transposed
-      mat_M, // a_rows
-      mat_K, // a_cols
-      mat_K * params.groups, // lda
-      true, // b_transposed
-      mat_K, // b_rows
-      mat_N, // b_cols
-      mat_K, // ldb
-      batch_shape.back(),
-      a_batch_strides.back(),
-      b_batch_strides.back());
-  gemm.set_out(
-      out.dtype(),
-      false, // out_transposed
-      mat_M, // out_rows
-      mat_N, // out_cols
-      mat_N * params.groups, // out_ld
-      params.groups, // batch_count
-      mat_N); // batch_stride
-  gemm.run(
-      encoder,
-      out,
-      in_unfolded,
-      wt_reshaped,
-      batch_shape,
-      a_batch_strides,
-      b_batch_strides);
-}
-
-void gemm_grouped_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s) {
-  int conv_ndim = in.ndim() - 2;
-  if (conv_ndim < 1 || conv_ndim > 3) {
-    throw std::runtime_error(
-        fmt::format("[conv] Unsupported gemm_conv for {}D conv.", conv_ndim));
-  }
-  dispatch_1_2_3(conv_ndim, [&](auto ndim_constant) {
-    ConvParams<ndim_constant()> params(
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        groups,
-        flip);
-    gemm_grouped_conv_nd<ndim_constant()>(encoder, in, wt, out, params, s);
-  });
-}
-
-} // namespace mlx::core

mlx/backend/cuda/copy.cu

@@ -15,8 +15,8 @@ void copy_gpu_inplace(
     int64_t offset_out,
     CopyType ctype,
     const Stream& s,
-    std::optional<array> dynamic_offset_in,
+    const std::optional<array>& dynamic_offset_in,
-    std::optional<array> dynamic_offset_out) {
+    const std::optional<array>& dynamic_offset_out) {
   if (out.size() == 0) {
     return;
   }
@@ -44,16 +44,6 @@ void copy_gpu_inplace(
           strides_vec[0]);
     } else {
       if (dynamic_offset_in || dynamic_offset_out) {
-        if (!dynamic_offset_in) {
-          dynamic_offset_in = array(0, int64);
-          encoder.add_temporary(*dynamic_offset_in);
-        }
-        if (!dynamic_offset_out) {
-          dynamic_offset_out = array(0, int64);
-          encoder.add_temporary(*dynamic_offset_out);
-        }
-        encoder.set_input_array(*dynamic_offset_in);
-        encoder.set_input_array(*dynamic_offset_out);
         copy_general_dynamic(
             encoder,
             ctype,
@@ -64,8 +54,8 @@ void copy_gpu_inplace(
             shape_collapsed,
             strides_vec[0],
             strides_vec[1],
-            *dynamic_offset_in,
+            dynamic_offset_in ? *dynamic_offset_in : array(0, int64),
-            *dynamic_offset_out);
+            dynamic_offset_out ? *dynamic_offset_out : array(0, int64));
       } else {
         copy_general(
             encoder,

mlx/backend/cuda/copy/copy_general.cu

@@ -10,80 +10,37 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
+template <typename In, typename Out, typename IdxT, int NDIM>
 __global__ void copy_gg_nd(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_out) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[NDIM - 1];
-  auto in_stride_x = strides_in[NDIM - 1];
-  auto out_stride_x = strides_out[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
-      index_rest * shape_x,
+        index, shape.data(), strides_in.data(), strides_out.data());
-      shape.data(),
+    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
-      strides_in.data(),
-      strides_out.data());
-
-  auto in_vec =
-      load_vector<N_READS>(in + idx_in, index_x, shape_x, in_stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
   }
-  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
-template <typename In, typename Out, typename IdxT, int N_READS>
+template <typename In, typename Out, typename IdxT>
 __global__ void copy_gg(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides strides_in,
     const __grid_constant__ Strides strides_out,
     int ndim) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto in_stride_x = strides_in[ndim - 1];
-  auto out_stride_x = strides_out[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [idx_in, idx_out] = elem_to_loc(
-      index_rest * shape_x,
+        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
-      shape.data(),
+    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
-      strides_in.data(),
-      strides_out.data(),
-      ndim);
-
-  auto in_vec =
-      load_vector<N_READS>(in + idx_in, index_x, shape_x, in_stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
   }
-  store_vector(out + idx_out, index_x, out_vec, shape_x, out_stride_x);
 }
 
 } // namespace cu
@@ -112,52 +69,33 @@ void copy_general(
             size_t data_size = 1;
             for (auto& s : shape)
               data_size *= s;
-
-            int work_per_thread = 1;
-            auto dim0 = ndim > 0 ? shape.back() : 1;
-            auto rest = data_size / dim0;
-            if (dim0 >= 4) {
-              work_per_thread = 4;
-            }
-
-            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-            auto block_dims = get_block_dims(dim0, rest, 1);
-            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto ndim_constant) {
-                auto kernel =
+                auto [num_blocks, block_dims] =
-                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 1>;
+                    get_launch_args(data_size, shape, out.strides(), large());
-                if (work_per_thread == 4) {
-                  kernel =
-                      cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 4>;
-                }
                 encoder.add_kernel_node(
-                    kernel,
+                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>,
-                    {num_blocks_x, num_blocks_y},
+                    num_blocks,
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    rest,
+                    data_size,
                     const_param<ndim_constant()>(shape),
                     const_param<ndim_constant()>(strides_in),
                     const_param<ndim_constant()>(strides_out));
               });
             } else { // ndim >= 4
-              auto kernel = cu::copy_gg<InType, OutType, IdxT, 1>;
+              auto [num_blocks, block_dims] =
-              if (work_per_thread == 4) {
+                  get_launch_args(data_size, shape, out.strides(), large());
-                kernel = cu::copy_gg<InType, OutType, IdxT, 4>;
-              }
               encoder.add_kernel_node(
-                  kernel,
+                  cu::copy_gg<InType, OutType, IdxT>,
-                  {num_blocks_x, num_blocks_y},
+                  num_blocks,
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  rest,
+                  data_size,
                   const_param(shape),
                   const_param(strides_in),
                   const_param(strides_out),

mlx/backend/cuda/copy/copy_general_input.cu

@@ -10,67 +10,33 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
+template <typename In, typename Out, typename IdxT, int NDIM>
 __global__ void copy_g_nd(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
-    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides) {
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
+    IdxT idx_in = elem_to_loc_nd<NDIM>(index, shape.data(), strides_in.data());
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+    out[index] = CastOp<In, Out>{}(in[idx_in]);
-  if (index_rest >= size_rest) {
+  }
-    return;
 }
 
-  auto shape_x = shape[NDIM - 1];
+template <typename In, typename Out, typename IdxT>
-  auto stride_x = strides[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  auto idx =
-      elem_to_loc_nd<NDIM>(index_rest * shape_x, shape.data(), strides.data());
-  auto in_vec =
-      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
-  }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
-}
-
-template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_g(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
-    const __grid_constant__ Strides strides,
+    const __grid_constant__ Strides strides_in,
     int ndim) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
+    IdxT idx_in = elem_to_loc(index, shape.data(), strides_in.data(), ndim);
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+    out[index] = CastOp<In, Out>{}(in[idx_in]);
-  if (index_rest >= size_rest) {
-    return;
   }
-
-  auto shape_x = shape[ndim - 1];
-  auto stride_x = strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  auto idx =
-      elem_to_loc(index_rest * shape_x, shape.data(), strides.data(), ndim);
-  auto in_vec =
-      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = CastOp<In, Out>{}(in_vec[i]);
-  }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 } // namespace cu
@@ -95,49 +61,30 @@ void copy_general_input(
             const InType* in_ptr = in.data<InType>() + offset_in;
             OutType* out_ptr = out.data<OutType>() + offset_out;
             int ndim = shape.size();
-            int work_per_thread = 1;
-            auto dim0 = ndim > 0 ? shape.back() : 1;
-            auto rest = out.size() / dim0;
-            if (dim0 >= 4) {
-              work_per_thread = 4;
-            }
-            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-            auto block_dims = get_block_dims(dim0, rest, 1);
-            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                auto kernel =
+                auto [num_blocks, block_dims] = get_launch_args(out, large());
-                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 1>;
-                if (work_per_thread == 4) {
-                  kernel =
-                      cu::copy_g_nd<InType, OutType, IdxT, dims_constant(), 4>;
-                }
                 encoder.add_kernel_node(
-                    kernel,
+                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>,
-                    {num_blocks_x, num_blocks_y},
+                    num_blocks,
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    rest,
+                    out.size(),
                     const_param<dims_constant()>(shape),
                     const_param<dims_constant()>(strides_in));
               });
             } else { // ndim >= 4
-              auto kernel = cu::copy_g<InType, OutType, IdxT, 1>;
+              auto [num_blocks, block_dims] = get_launch_args(out, large());
-              if (work_per_thread == 4) {
-                kernel = cu::copy_g<InType, OutType, IdxT, 4>;
-              }
               encoder.add_kernel_node(
-                  kernel,
+                  cu::copy_g<InType, OutType, IdxT>,
-                  {num_blocks_x, num_blocks_y},
+                  num_blocks,
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  rest,
+                  out.size(),
                   const_param(shape),
                   const_param(strides_in),
                   ndim);

mlx/backend/cuda/cudnn_utils.cpp

@@ -1,272 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/cudnn_utils.h"
-#include "mlx/backend/cuda/device.h"
-
-namespace mlx::core {
-
-namespace {
-
-// Create a cudnn tensor descriptor.
-template <typename Vec>
-inline cudnn_frontend::Tensor build_cudnn_tensor(
-    int64_t id,
-    const array& x,
-    const Vec& shape,
-    const Vec& strides) {
-  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();
-}
-
-// In MLX a singleton dim (shape[dim] == 1) can have any stride, but in cuDNN
-// whether a tensor is contiguous is determined with:
-// shape[dim] == shape[dim + 1] * strides[dim + 1]
-// So a contiguous array with singleton dims in MLX may be mistakenly treated
-// as strided in cuDNN, and we work around it by normalizing the strides.
-Strides normalized_strides(const array& x) {
-  if (!x.flags().row_contiguous || x.ndim() < 2) {
-    return x.strides();
-  }
-  Strides strides = x.strides();
-  for (int i = x.ndim() - 2; i >= 0; --i) {
-    if (x.shape(i) == 1) {
-      strides[i] = x.shape(i + 1) * strides[i + 1];
-    }
-  }
-  return strides;
-}
-
-// Return the shape and strides after transposing from NHWC to NCHW.
-auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
-  assert(shape.size() >= 3);
-  shape.insert(shape.begin() + 1, shape.back());
-  shape.erase(shape.end() - 1);
-  strides.insert(strides.begin() + 1, strides.back());
-  strides.erase(strides.end() - 1);
-  return std::make_tuple(std::move(shape), std::move(strides));
-}
-
-inline auto nhwc_to_nchw(const array& x) {
-  return nhwc_to_nchw(
-      convert_vector<int64_t>(x.shape()), normalized_strides(x));
-}
-
-// Return available engines for a |op_graph|.
-cudnn_frontend::EngineConfigList get_cudnn_engine_configs(
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph,
-    bool use_fallback = true) {
-  SmallVector<cudnn_frontend::GeneratorSource, 2> sources;
-  sources.push_back([](auto& op_graph) {
-    auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
-                          .setOperationGraph(op_graph)
-                          .setHeurMode(CUDNN_HEUR_MODE_A)
-                          .build();
-    return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
-  });
-  if (use_fallback) {
-    sources.push_back([&backend_type](auto& op_graph) {
-      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
-                          .setOperationGraph(op_graph)
-                          .setOperation(backend_type)
-                          .build();
-      return fallback.getFallbackList();
-    });
-  }
-
-  auto configs =
-      cudnn_frontend::EngineConfigGenerator(sources.size(), sources.data())
-          .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;
-}
-
-// Take |engine_configs| and |op_graph| and find a working execution plans
-// from them.
-std::optional<cudnn_frontend::ExecutionPlan>
-find_cudnn_plan_from_engine_configs(
-    cudnnHandle_t handle,
-    const cudnn_frontend::EngineConfigList& engine_configs,
-    const cudnn_frontend::OperationGraph& op_graph) {
-  auto op_graph_tag = op_graph.getTag();
-  for (const auto& config : engine_configs) {
-    try {
-      return cudnn_frontend::ExecutionPlanBuilder()
-          .setHandle(handle)
-          .setEngineConfig(config, op_graph_tag)
-          .build();
-    } catch (cudnn_frontend::cudnnException& error) {
-      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
-        throw;
-      }
-    }
-  }
-  return std::nullopt;
-}
-
-// Prepare workspace and args to execute plan.
-template <typename F>
-bool prepare_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs,
-    F&& execute) {
-  int workspace_size = plan.getWorkspaceSize();
-  array workspace(
-      workspace_size > 0 ? allocator::malloc(workspace_size)
-                         : allocator::Buffer(nullptr),
-      {workspace_size},
-      uint8);
-
-  auto args = cudnn_frontend::VariantPackBuilder()
-                  .setWorkspacePointer(workspace.data<void>())
-                  .setDataPointers(num_args, data_ptrs)
-                  .setUids(num_args, uids)
-                  .build();
-
-  auto handle = encoder.device().cudnn_handle();
-  cudnnSetStream(handle, encoder.stream());
-
-  if (!execute(handle, plan.get_raw_desc(), args.get_raw_desc())) {
-    return false;
-  }
-
-  encoder.add_temporary(workspace);
-  return true;
-}
-
-} // namespace
-
-cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x) {
-  auto shape = convert_vector<int64_t>(x.shape());
-  return build_cudnn_tensor(id, x, shape, normalized_strides(x));
-}
-
-cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x) {
-  auto [shape, strides] = nhwc_to_nchw(x);
-  return build_cudnn_tensor(id, x, shape, strides);
-}
-
-cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x) {
-  if (x.ndim() == 0) {
-    SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
-    return build_cudnn_tensor(id, x, scalar_dims, scalar_dims);
-  }
-  if (x.ndim() == 1) {
-    int64_t s = x.shape(0);
-    SmallVector<int64_t, 4> shape = {1, x.shape(0), 1, 1};
-    SmallVector<int64_t, 4> strides = {s, 1, s, s};
-    return build_cudnn_tensor(id, x, shape, strides);
-  }
-  if (x.ndim() == 2) {
-    int64_t s =
-        x.flags().row_contiguous ? x.shape(1) * x.strides(1) : x.strides(0);
-    SmallVector<int64_t, 4> shape = {x.shape(0), x.shape(1), 1, 1};
-    SmallVector<int64_t, 4> strides = {s, x.strides(1), s, s};
-    return build_cudnn_tensor(id, x, shape, strides);
-  }
-  if (x.ndim() == 3 || x.ndim() == 4) {
-    return build_cudnn_tensor_nchw(id, x);
-  }
-  throw std::runtime_error(
-      fmt::format("Unsupported array with {} dims.", x.ndim()));
-}
-
-cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype) {
-  SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
-  return cudnn_frontend::TensorBuilder()
-      .setDim(scalar_dims.size(), scalar_dims.data())
-      .setStrides(scalar_dims.size(), scalar_dims.data())
-      .setId(id)
-      .setAlignment(16)
-      .setDataType(dtype_to_cudnn_type(dtype))
-      .setByValue(true)
-      .build();
-}
-
-std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
-    cudnnHandle_t handle,
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph) {
-  auto engine_configs = get_cudnn_engine_configs(backend_type, dtype, op_graph);
-  return find_cudnn_plan_from_engine_configs(handle, engine_configs, op_graph);
-}
-
-bool encode_cudnn_plan_with_capturing(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs) {
-  return prepare_cudnn_plan(
-      encoder,
-      plan,
-      num_args,
-      uids,
-      data_ptrs,
-      [&](auto handle, auto plan, auto args) {
-        auto capture = encoder.capture_context();
-        if (cudnnBackendExecute(handle, plan, args) != CUDNN_STATUS_SUCCESS) {
-          // Discard the captured graph when failed.
-          capture.discard = true;
-          return false;
-        }
-        return true;
-      });
-}
-
-#if CUDNN_VERSION >= 90500
-bool encode_cudnn_plan_with_graph_api(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs) {
-  return prepare_cudnn_plan(
-      encoder,
-      plan,
-      num_args,
-      uids,
-      data_ptrs,
-      [&](auto handle, auto plan, auto args) {
-        if (!graph) {
-          graph = CudaGraph(encoder.device());
-          if (cudnnBackendPopulateCudaGraph(handle, plan, args, graph) !=
-              CUDNN_STATUS_SUCCESS) {
-            return false;
-          }
-        } else {
-          if (cudnnBackendUpdateCudaGraph(handle, plan, args, graph) !=
-              CUDNN_STATUS_SUCCESS) {
-            return false;
-          }
-        }
-        encoder.add_graph_node(graph);
-        return true;
-      });
-}
-#endif
-
-} // namespace mlx::core

mlx/backend/cuda/cudnn_utils.h

@@ -1,164 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include "mlx/array.h"
-#include "mlx/backend/cuda/device/config.h"
-#include "mlx/backend/cuda/utils.h"
-#include "mlx/dtype_utils.h"
-
-#include <cudnn_frontend.h>
-#include <cudnn_frontend_find_plan.h>
-#include <fmt/format.h>
-
-#include <algorithm>
-#include <array>
-
-namespace mlx::core {
-
-namespace cu {
-class CommandEncoder;
-}
-
-// Return pointer alignment of |x|'s data.
-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;
-}
-
-// Convert the type of elements in |vec| to |T|.
-template <typename T, typename Vec>
-inline SmallVector<T> convert_vector(const Vec& vec) {
-  return SmallVector<T>(vec.begin(), vec.end());
-}
-
-// Return an array that can be used as map key for |vec| with size <= MAX_NDIM.
-//
-// There are 2 differences from the const_param util from kernel_utils.cuh:
-// 1. The rest of array is filled with 0.
-// 2. This util can be used in .cpp files.
-template <typename T, template <typename U> class Vec>
-inline std::array<T, MAX_NDIM> vector_key(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;
-}
-
-// Helpers used by get_data_ptrs to get pointers.
-inline void* get_data_ptr(const array& arr) {
-  return const_cast<void*>(arr.data<void>());
-}
-
-template <typename T, typename = std::enable_if_t<std::is_scalar_v<T>>>
-inline void* get_data_ptr(T& scalar) {
-  return &scalar;
-}
-
-// Return an array filled with data pointers of args.
-template <typename... Args>
-inline std::array<void*, sizeof...(Args)> get_data_ptrs(Args&... args) {
-  return {get_data_ptr(args)...};
-}
-
-// Map dtype to cudnn data type.
-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)));
-  }
-}
-
-// Create a tensor descriptor from |x|.
-cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x);
-
-// Create a tensor descriptor from |x|, and transpose from NHWC to NCHW.
-cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x);
-
-// Create a tensor descriptor from |x|, make sure it is 4D, and transpose it
-// from NHWC to NCHW.
-cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x);
-
-// Create a 4D scalar tensor descriptor, which is passed by value.
-cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype);
-
-// Find a working plan for |op_graph|.
-std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
-    cudnnHandle_t handle,
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph);
-
-// Encode the plan to command buffer by capturing.
-bool encode_cudnn_plan_with_capturing(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs);
-
-#if CUDNN_VERSION >= 90500
-// Encode the plan to command buffer by using native graph api of cudnn. If the
-// |graph| is empty it will be populated, otherwise it will be updated.
-bool encode_cudnn_plan_with_graph_api(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs);
-#endif
-
-// Helpers to make calls like encode_cudnn_plan(..., {'x', 'y', 'z'}, x, y, z).
-template <typename... Args>
-bool encode_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    std::initializer_list<int64_t> uids,
-    Args&... args) {
-  assert(uids.size() == sizeof...(args));
-  auto data_ptrs = get_data_ptrs(args...);
-  return encode_cudnn_plan_with_capturing(
-      encoder, plan, uids.size(), uids.begin(), data_ptrs.data());
-}
-
-#if CUDNN_VERSION >= 90500
-template <typename... Args>
-bool encode_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    std::initializer_list<int64_t> uids,
-    Args&... args) {
-  assert(uids.size() == sizeof...(args));
-  auto data_ptrs = get_data_ptrs(args...);
-  return encode_cudnn_plan_with_graph_api(
-      encoder, plan, graph, uids.size(), uids.begin(), data_ptrs.data());
-}
-#endif
-
-} // namespace mlx::core

mlx/backend/cuda/custom_kernel.cpp

@@ -1,379 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include <iostream>
-
-#include "mlx/backend/common/compiled.h"
-#include "mlx/backend/cuda/jit_module.h"
-#include "mlx/backend/cuda/utils.h"
-#include "mlx/backend/gpu/copy.h"
-#include "mlx/fast.h"
-#include "mlx/fast_primitives.h"
-
-#include <fmt/format.h>
-#include <nvtx3/nvtx3.hpp>
-
-namespace mlx::core::fast {
-
-namespace {
-
-constexpr const char* default_header = R"(
-#include "mlx/backend/cuda/device/utils.cuh"
-
-#include <cooperative_groups.h>
-
-#define inf cuda::std::numeric_limits<float>::infinity()
-
-)";
-
-std::string template_arguments_hash(
-    const std::vector<std::pair<std::string, TemplateArg>>& template_args) {
-  if (template_args.empty()) {
-    return "";
-  }
-
-  std::string hash;
-  hash.reserve(512);
-
-  for (const auto& [name, arg] : template_args) {
-    if (std::holds_alternative<int>(arg)) {
-      hash += fmt::format("_{}", std::get<int>(arg));
-    } else if (std::holds_alternative<bool>(arg)) {
-      hash += (std::get<bool>(arg)) ? "_t" : "_f";
-    } else if (std::holds_alternative<Dtype>(arg)) {
-      hash += "_";
-      hash += get_type_string(std::get<Dtype>(arg));
-    }
-  }
-
-  return hash;
-}
-
-std::string build_kernel(
-    const std::string& func_name,
-    const std::string& header,
-    const std::string& source,
-    const std::vector<std::string>& input_names,
-    const std::vector<array>& inputs,
-    const std::vector<std::string>& output_names,
-    const std::vector<Dtype>& output_dtypes,
-    const std::vector<std::pair<std::string, TemplateArg>>& template_args,
-    const std::vector<CustomKernelShapeInfo>& shape_infos) {
-  std::string kernel_source;
-  kernel_source.reserve(header.size() + source.size() + 8192);
-  kernel_source += default_header;
-  kernel_source += header;
-  kernel_source +=
-      "namespace mlx::core::cu {\n\n"
-      "namespace cg = cooperative_groups;\n\n";
-
-  kernel_source += "__global__ void ";
-  kernel_source += func_name;
-  kernel_source += "(\n";
-
-  // Add inputs
-  for (int i = 0; i < inputs.size(); ++i) {
-    const auto& name = input_names[i];
-    const auto& arr = inputs[i];
-    kernel_source += "    const ";
-    kernel_source += dtype_to_cuda_type(arr.dtype());
-    kernel_source += "* ";
-    kernel_source += name;
-    kernel_source += ",\n";
-    // Add input shape, strides and ndim if present in the source
-    if (arr.ndim() > 0) {
-      if (shape_infos[i].shape) {
-        kernel_source += "    const __grid_constant__ Shape ";
-        kernel_source += name;
-        kernel_source += "_shape,\n";
-      }
-      if (shape_infos[i].strides) {
-        kernel_source += "    const __grid_constant__ Strides ";
-        kernel_source += name;
-        kernel_source += "_strides,\n";
-      }
-      if (shape_infos[i].ndim) {
-        kernel_source += "    const __grid_constant__ int ";
-        kernel_source += name;
-        kernel_source += "_ndim,\n";
-      }
-    }
-  }
-
-  // Add outputs
-  for (int i = 0; i < output_names.size(); ++i) {
-    const auto& name = output_names[i];
-    const auto& dtype = output_dtypes[i];
-    kernel_source += "    ";
-    kernel_source += dtype_to_cuda_type(dtype);
-    kernel_source += "* ";
-    kernel_source += name;
-    if (i < output_names.size() - 1) {
-      kernel_source += ",\n";
-    } else {
-      kernel_source += ") {\n";
-    }
-  }
-
-  // Set compile time constants
-  if (!template_args.empty()) {
-    for (const auto& [name, arg] : template_args) {
-      if (std::holds_alternative<int>(arg)) {
-        kernel_source +=
-            fmt::format("  constexpr int {} = {};\n", name, std::get<int>(arg));
-      } else if (std::holds_alternative<bool>(arg)) {
-        kernel_source += fmt::format(
-            "  constexpr bool {} = {};\n", name, std::get<bool>(arg));
-      } else {
-        kernel_source += fmt::format(
-            "  using {} = {};\n",
-            name,
-            dtype_to_cuda_type(std::get<Dtype>(arg)));
-      }
-    }
-    kernel_source += "\n";
-  }
-
-  kernel_source += source;
-  kernel_source += "\n}\n\n} // namespace mlx::core::cu\n";
-
-  return kernel_source;
-}
-
-} // namespace
-
-CustomKernelFunction cuda_kernel(
-    const std::string& name,
-    const std::vector<std::string>& input_names,
-    const std::vector<std::string>& output_names,
-    const std::string& source,
-    const std::string& header,
-    bool ensure_row_contiguous,
-    int shared_memory) {
-  if (output_names.empty()) {
-    throw std::invalid_argument(
-        "[custom_kernel] Must specify at least one output.");
-  }
-
-  std::vector<CustomKernelShapeInfo> shape_infos;
-  for (auto& n : input_names) {
-    CustomKernelShapeInfo shape_info;
-    shape_info.shape = source.find(n + "_shape") != std::string::npos;
-    shape_info.strides = source.find(n + "_strides") != std::string::npos;
-    shape_info.ndim = source.find(n + "_ndim") != std::string::npos;
-    shape_infos.push_back(shape_info);
-  }
-
-  return [=, shape_infos = std::move(shape_infos)](
-             const std::vector<array>& inputs,
-             const std::vector<Shape>& output_shapes,
-             const std::vector<Dtype>& output_dtypes,
-             std::tuple<int, int, int> grid,
-             std::tuple<int, int, int> threadgroup,
-             const std::vector<std::pair<std::string, TemplateArg>>&
-                 template_args = {},
-             std::optional<float> init_value = std::nullopt,
-             bool verbose = false,
-             StreamOrDevice s_ = {}) {
-    if (inputs.size() != input_names.size()) {
-      std::ostringstream msg;
-      msg << "[custom_kernel] Expected `inputs` to have size "
-          << input_names.size() << " but got size " << inputs.size() << "."
-          << std::endl;
-      throw std::invalid_argument(msg.str());
-    }
-    if (output_shapes.size() != output_names.size()) {
-      std::ostringstream msg;
-      msg << "[custom_kernel] Expected `output_shapes` to have size "
-          << output_names.size() << " but got size " << output_shapes.size()
-          << "." << std::endl;
-      throw std::invalid_argument(msg.str());
-    }
-    if (output_dtypes.size() != output_names.size()) {
-      std::ostringstream msg;
-      msg << "[custom_kernel] Expected `output_dtypes` to have size "
-          << output_names.size() << " but got size " << output_dtypes.size()
-          << "." << std::endl;
-      throw std::invalid_argument(msg.str());
-    }
-
-    auto s = to_stream(s_);
-    if (s.device != Device::gpu) {
-      throw std::invalid_argument("[custom_kernel] Only supports the GPU.");
-    }
-
-    std::string kernel_name =
-        "custom_kernel_" + name + template_arguments_hash(template_args);
-    std::string kernel_source = build_kernel(
-        kernel_name,
-        header,
-        source,
-        input_names,
-        inputs,
-        output_names,
-        output_dtypes,
-        template_args,
-        shape_infos);
-
-    if (verbose) {
-      std::cout << "Generated source code for `" << kernel_name
-                << "`:" << std::endl
-                << "```" << std::endl
-                << kernel_source << std::endl
-                << "```" << std::endl;
-    }
-
-    return array::make_arrays(
-        std::move(output_shapes),
-        std::move(output_dtypes),
-        std::make_shared<CustomKernel>(
-            s,
-            std::move(kernel_name),
-            std::move(kernel_source),
-            grid,
-            threadgroup,
-            shape_infos,
-            ensure_row_contiguous,
-            init_value,
-            std::vector<ScalarArg>{},
-            false,
-            shared_memory),
-        std::move(inputs));
-  };
-}
-
-std::vector<array> precompiled_cuda_kernel(
-    const std::string& name,
-    const std::string& compiled_source,
-    const std::vector<array>& inputs,
-    const std::vector<Shape>& output_shapes,
-    const std::vector<Dtype>& output_dtypes,
-    const std::vector<ScalarArg>& scalars,
-    std::tuple<int, int, int> grid,
-    std::tuple<int, int, int> threadgroup,
-    int shared_memory,
-    std::optional<float> init_value,
-    bool ensure_row_contiguous,
-    StreamOrDevice s) {
-  std::vector<CustomKernelShapeInfo> shape_infos(
-      inputs.size(), CustomKernelShapeInfo{false, false, false});
-  return array::make_arrays(
-      output_shapes,
-      output_dtypes,
-      std::make_shared<CustomKernel>(
-          to_stream(s),
-          name,
-          compiled_source,
-          grid,
-          threadgroup,
-          shape_infos,
-          ensure_row_contiguous,
-          init_value,
-          scalars,
-          true,
-          shared_memory),
-      inputs);
-}
-
-void CustomKernel::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  nvtx3::scoped_range r("CustomKernel::eval_gpu");
-  auto& s = stream();
-
-  std::vector<array> copies;
-
-  // Allocate and initialize the output arrays
-  for (auto& out : outputs) {
-    if (init_value_) {
-      copies.emplace_back(init_value_.value(), out.dtype());
-      fill_gpu(copies.back(), out, s);
-    } else {
-      out.set_data(allocator::malloc(out.nbytes()));
-    }
-  }
-
-  // Create the input arrays and copy if needed
-  auto check_input = [&copies, &s, this](const array& x) -> const array {
-    bool no_copy = x.flags().row_contiguous;
-    if (!ensure_row_contiguous_ || no_copy) {
-      return x;
-    } else {
-      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
-      copy_gpu(x, copies.back(), CopyType::General, s);
-      return copies.back();
-    }
-  };
-  std::vector<array> checked_inputs;
-  for (const array& in : inputs) {
-    checked_inputs.push_back(check_input(in));
-  }
-
-  // Compile the custom kernel
-  std::string kernel_name =
-      (is_precompiled_) ? name_ : "mlx::core::cu::" + name_;
-  cu::JitModule& mod = cu::get_jit_module(
-      s.device,
-      name_,
-      [&]() {
-        return std::make_tuple(
-            is_precompiled_, source_, std::vector{kernel_name});
-      },
-      false);
-
-  // Make the arguments
-  cu::KernelArgs args;
-  for (int i = 0; i < checked_inputs.size(); i++) {
-    const array& in = checked_inputs[i];
-    auto& shape_info = shape_infos_[i];
-    args.append(in);
-    if (shape_info.shape) {
-      args.append_ndim(in.shape());
-    }
-    if (shape_info.strides) {
-      args.append_ndim(in.strides());
-    }
-    if (shape_info.ndim) {
-      args.append<int32_t>(in.ndim());
-    }
-  }
-  for (auto& out : outputs) {
-    args.append(out);
-  }
-  for (auto& s : scalar_arguments_) {
-    if (std::holds_alternative<bool>(s)) {
-      args.append(std::get<bool>(s));
-    } else if (std::holds_alternative<int>(s)) {
-      args.append(std::get<int>(s));
-    } else if (std::holds_alternative<float>(s)) {
-      args.append(std::get<float>(s));
-    }
-  }
-
-  // Make the grid
-  const auto [tx, ty, tz] = threadgroup_;
-  const auto [gx, gy, gz] = grid_;
-  dim3 block(std::min(tx, gx), std::min(ty, gy), std::min(tz, gz));
-  dim3 grid((gx + tx - 1) / tx, (gy + ty - 1) / ty, (gz + tz - 1) / tz);
-
-  // Call the kernel
-  auto& encoder = cu::get_command_encoder(s);
-  for (const auto& in : checked_inputs) {
-    encoder.set_input_array(in);
-  }
-  for (const auto& out : outputs) {
-    encoder.set_output_array(out);
-  }
-  for (const auto& t : copies) {
-    encoder.add_temporary(t);
-  }
-  auto kernel =
-      mod.get_kernel(kernel_name, [smem = shared_memory_](CUfunction kernel) {
-        if (smem > 0 && smem > 48000) {
-          cuFuncSetAttribute(
-              kernel, CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES, smem);
-        }
-      });
-  encoder.add_kernel_node(kernel, grid, block, shared_memory_, args.args());
-}
-
-} // namespace mlx::core::fast

mlx/backend/cuda/device.cpp

@@ -29,18 +29,11 @@ void check_cudnn_error(const char* name, cudnnStatus_t err) {
 
 int cuda_graph_cache_size() {
   static int cache_size = []() {
-    return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 400);
+    return env::get_var("MLX_CUDA_GRAPH_CACHE_SIZE", 100);
   }();
   return cache_size;
 }
 
-bool use_cuda_graphs() {
-  static bool use_graphs = []() {
-    return env::get_var("MLX_USE_CUDA_GRAPHS", true);
-  }();
-  return use_graphs;
-}
-
 } // namespace
 
 Device::Device(int device) : device_(device) {
@@ -93,19 +86,14 @@ CommandEncoder& Device::get_command_encoder(Stream s) {
 
 CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
   enc.device().make_current();
-  if (!use_cuda_graphs()) {
-    return;
-  }
   CHECK_CUDA_ERROR(
       cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
 }
 
 CommandEncoder::CaptureContext::~CaptureContext() {
-  if (!use_cuda_graphs()) {
+  CHECK_CUDA_ERROR(cudaStreamEndCapture(enc.stream(), &graph));
-    return;
+  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
-  }
+      &graph, [](cudaGraph_t* p) { CHECK_CUDA_ERROR(cudaGraphDestroy(*p)); });
-
-  graph.end_capture(enc.stream());
   if (discard) {
     return;
   }
@@ -119,9 +107,6 @@ CommandEncoder::ConcurrentContext::ConcurrentContext(CommandEncoder& enc)
 
 CommandEncoder::ConcurrentContext::~ConcurrentContext() {
   enc.in_concurrent_ = false;
-  if (!use_cuda_graphs()) {
-    return;
-  }
 
   // Use an empty graph node for synchronization
   CommandEncoder::GraphNode empty{NULL, 'E', std::to_string(enc.node_count_++)};
@@ -200,28 +185,20 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
 }
 
 CommandEncoder::CommandEncoder(Device& d)
-    : device_(d),
+    : device_(d), stream_(d), graph_cache_(cuda_graph_cache_size()) {
-      stream_(d),
+  CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
-      graph_(d),
+}
-      graph_cache_(cuda_graph_cache_size()) {}
 
 void CommandEncoder::add_completed_handler(std::function<void()> task) {
   worker_.add_task(std::move(task));
 }
 
 void CommandEncoder::set_input_array(const array& arr) {
-  if (!use_cuda_graphs()) {
-    return;
-  }
   auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
   active_deps_.push_back(id);
 }
 
 void CommandEncoder::set_output_array(const array& arr) {
-  if (!use_cuda_graphs()) {
-    return;
-  }
-
   auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
   active_deps_.push_back(id);
   active_outputs_.push_back(id);
@@ -239,11 +216,6 @@ void CommandEncoder::add_kernel_node(
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
-  if (!use_cuda_graphs()) {
-    CHECK_CUDA_ERROR(cudaLaunchKernel(
-        func, grid_dim, block_dim, params, smem_bytes, stream()));
-    return;
-  }
   cudaKernelNodeParams kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDim = grid_dim;
@@ -259,22 +231,6 @@ void CommandEncoder::add_kernel_node(
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
-  if (!use_cuda_graphs()) {
-    CHECK_CUDA_ERROR(cuLaunchKernel(
-        func,
-        grid_dim.x,
-        grid_dim.y,
-        grid_dim.z,
-        block_dim.x,
-        block_dim.y,
-        block_dim.z,
-        smem_bytes,
-        stream(),
-        params,
-        nullptr));
-    return;
-  }
-
   CUDA_KERNEL_NODE_PARAMS kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDimX = grid_dim.x;
@@ -301,12 +257,6 @@ void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
 }
 
 void CommandEncoder::add_graph_node(cudaGraph_t child) {
-  if (!use_cuda_graphs()) {
-    CudaGraphExec graph_exec;
-    graph_exec.instantiate(child);
-    device_.make_current();
-    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream()));
-  }
   cudaGraphNode_t node;
   CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
   insert_graph_dependencies(GraphNode{node, 'G'});
@@ -320,13 +270,7 @@ void CommandEncoder::commit() {
   if (node_count_ > 0) {
     if (!from_nodes_.empty()) {
       CHECK_CUDA_ERROR(cudaGraphAddDependencies(
-          graph_,
+          graph_, from_nodes_.data(), to_nodes_.data(), from_nodes_.size()));
-          from_nodes_.data(),
-          to_nodes_.data(),
-#if CUDART_VERSION >= 13000
-          nullptr, // edgeData
-#endif // CUDART_VERSION >= 13000
-          from_nodes_.size()));
     }
 
     graph_key_ += ".";
@@ -367,7 +311,8 @@ void CommandEncoder::commit() {
     to_nodes_.clear();
     graph_key_.clear();
     node_map_.clear();
-    graph_ = CudaGraph(device_);
+    CHECK_CUDA_ERROR(cudaGraphDestroy(graph_));
+    CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
   }
 
   // Put completion handlers in a batch.

mlx/backend/cuda/device.h

@@ -21,7 +21,7 @@ class CommandEncoder {
   struct CaptureContext {
     CaptureContext(CommandEncoder& enc);
     ~CaptureContext();
-    CudaGraph graph;
+    cudaGraph_t graph;
     CommandEncoder& enc;
     bool discard{false};
   };
@@ -76,6 +76,9 @@ class CommandEncoder {
       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) {
@@ -98,9 +101,6 @@ class CommandEncoder {
   void synchronize();
 
  private:
-  void add_kernel_node(const cudaKernelNodeParams& params);
-  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
-
   struct GraphNode {
     cudaGraphNode_t node;
     // K = kernel
@@ -115,7 +115,7 @@ class CommandEncoder {
 
   Device& device_;
   CudaStream stream_;
-  CudaGraph graph_;
+  cudaGraph_t graph_;
   Worker worker_;
   char node_count_{0};
   char graph_node_count_{0};

mlx/backend/cuda/device/binary_ops.cuh

@@ -204,12 +204,6 @@ struct Power {
   __device__ T operator()(T base, T exp) {
     if constexpr (cuda::std::is_integral_v<T>) {
       T res = 1;
-      // Raising an integer to a negative power is undefined
-      if constexpr (cuda::std::is_signed_v<T>) {
-        if (exp < 0) {
-          return 0;
-        }
-      }
       while (exp) {
         if (exp & 1) {
           res *= base;

mlx/backend/cuda/device/cast_op.cuh

@@ -6,6 +6,7 @@
 
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
+#include <thrust/iterator/transform_iterator.h>
 
 namespace mlx::core::cu {
 
@@ -115,4 +116,15 @@ inline __host__ __device__ auto cast_to(SrcT x) {
   return CastOp<SrcT, DstT>{}(x);
 }
 
+// Return an iterator that cast the value to DstT using CastOp.
+template <typename DstT, typename Iterator>
+inline __host__ __device__ auto make_cast_iterator(Iterator it) {
+  using SrcT = typename cuda::std::iterator_traits<Iterator>::value_type;
+  if constexpr (std::is_same_v<SrcT, DstT>) {
+    return it;
+  } else {
+    return thrust::make_transform_iterator(it, CastOp<SrcT, DstT>{});
+  }
+}
+
 } // namespace mlx::core::cu

mlx/backend/cuda/device/utils.cuh

@@ -146,23 +146,6 @@ inline __device__ void store_vector(
   }
 }
 
-template <int N, typename T, typename SizeT>
-inline __device__ void store_vector(
-    T* ptr,
-    uint32_t offset,
-    const AlignedVector<T, N>& vec,
-    SizeT size,
-    int64_t stride) {
-  if (is_aligned<N>(ptr) && (offset + 1) * N <= size && stride == 1) {
-    auto* to = reinterpret_cast<AlignedVector<T, N>*>(ptr);
-    to[offset] = vec;
-  } else {
-    for (int i = 0; (offset * N + i) < size && i < N; ++i) {
-      ptr[stride * (offset * N + i)] = vec[i];
-    }
-  }
-}
-
 ///////////////////////////////////////////////////////////////////////////////
 // Type limits utils
 ///////////////////////////////////////////////////////////////////////////////

mlx/backend/cuda/distributed.cu

@@ -1,56 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/backend/gpu/copy.h"
-#include "mlx/distributed/primitives.h"
-#include "mlx/primitives.h"
-
-#include <cassert>
-
-namespace mlx::core::distributed {
-void AllReduce::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 1);
-  assert(outputs.size() == 1);
-
-  auto set_input_output =
-      [s = stream()](const array& in, array& out) -> std::pair<array, array> {
-    if (!in.flags().row_contiguous) {
-      copy_gpu(in, out, CopyType::General, s);
-      return {out, out};
-    } else if (in.is_donatable()) {
-      out.copy_shared_buffer(in);
-      return {in, out};
-    } else {
-      out.set_data(allocator::malloc(out.nbytes()));
-      return {in, out};
-    }
-  };
-
-  auto [input, output] = set_input_output(inputs[0], outputs[0]);
-
-  auto& encoder = cu::get_command_encoder(stream());
-  encoder.set_input_array(input);
-  encoder.set_output_array(output);
-
-  auto capture = encoder.capture_context();
-  auto& s = stream();
-
-  switch (reduce_type_) {
-    case Sum:
-      distributed::detail::all_sum(group(), input, output, s);
-      break;
-    case Max:
-      distributed::detail::all_max(group(), input, output, s);
-      break;
-    case Min:
-      distributed::detail::all_min(group(), input, output, s);
-      break;
-    default:
-      throw std::runtime_error(
-          "Only all reduce sum, max, and min are supported.");
-  }
-}
-} // namespace mlx::core::distributed

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -202,25 +202,6 @@ CublasGemm::~CublasGemm() {
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescDestroy(matmul_desc_));
 }
 
-void CublasGemm::set_out(
-    Dtype dtype,
-    bool transposed,
-    uint64_t rows,
-    uint64_t cols,
-    int64_t ld,
-    int32_t batch_count,
-    int64_t batch_stride) {
-  CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(out_desc_));
-  out_desc_ = create_matrix_layout(
-      dtype_to_cublas_type(dtype),
-      rows,
-      cols,
-      transposed,
-      ld,
-      batch_count,
-      batch_stride);
-}
-
 void CublasGemm::run(
     cu::CommandEncoder& encoder,
     array& out,

mlx/backend/cuda/gemms/cublas_gemm.h

@@ -44,17 +44,6 @@ class CublasGemm {
 
   ~CublasGemm();
 
-  // The output's descriptor is inferred from inputs by default, use this method
-  // for unusual output.
-  void set_out(
-      Dtype dtype,
-      bool transposed,
-      uint64_t rows,
-      uint64_t cols,
-      int64_t ld,
-      int32_t batch_count,
-      int64_t batch_stride);
-
   void run(
       cu::CommandEncoder& encoder,
       array& out,

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

@@ -94,7 +94,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
             large ? "int64_t" : "int32_t"));
       }
     }
-    return std::make_tuple(false, jit_source_gather, std::move(kernel_names));
+    return std::make_pair(jit_source_gather, std::move(kernel_names));
   });
 
   cu::KernelArgs args;
@@ -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(
@@ -189,7 +189,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
             large ? "int64_t" : "int32_t"));
       }
     }
-    return std::make_tuple(false, jit_source_scatter, std::move(kernel_names));
+    return std::make_pair(jit_source_scatter, std::move(kernel_names));
   });
 
   cu::KernelArgs args;
@@ -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(
@@ -268,8 +268,7 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
         }
       }
     }
-    return std::make_tuple(
+    return std::make_pair(jit_source_gather_axis, std::move(kernel_names));
-        false, jit_source_gather_axis, std::move(kernel_names));
   });
 
   size_t idx_size_pre = 1;
@@ -372,8 +371,7 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
         }
       }
     }
-    return std::make_tuple(
+    return std::make_pair(jit_source_scatter_axis, std::move(kernel_names));
-        false, jit_source_scatter_axis, std::move(kernel_names));
   });
 
   size_t idx_size_pre = 1;

mlx/backend/cuda/jit_module.cpp

@@ -67,12 +67,10 @@ const std::string& cccl_dir() {
       return path.string();
     }
     // Finally check the environment variable.
-    if (const char* env = std::getenv("MLX_CCCL_DIR"); env) {
+    path = std::getenv("MLX_CCCL_DIR");
-      path = env;
     if (!path.empty() && std::filesystem::exists(path)) {
       return path.string();
     }
-    }
     return std::string();
   }();
   return dir;
@@ -103,8 +101,8 @@ const std::filesystem::path& ptx_cache_dir() {
 bool read_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
-    std::string& ptx,
+    std::vector<char>* ptx,
-    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
+    std::vector<std::pair<std::string, std::string>>* ptx_kernels) {
   if (cache_dir.empty()) {
     return false;
   }
@@ -119,15 +117,15 @@ bool read_cached_ptx(
   if (!ptx_file.good()) {
     return false;
   }
-  ptx.resize(ptx_size);
+  ptx->resize(ptx_size);
-  ptx_file.read(ptx.data(), ptx_size);
+  ptx_file.read(ptx->data(), ptx_size);
 
   std::ifstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
   std::string line;
   while (std::getline(txt_file, line)) {
     auto tab = line.find('\t');
     if (tab != std::string::npos) {
-      ptx_kernels.emplace_back(line.substr(0, tab), line.substr(tab + 1));
+      ptx_kernels->emplace_back(line.substr(0, tab), line.substr(tab + 1));
     }
   }
   return true;
@@ -137,7 +135,7 @@ bool read_cached_ptx(
 void write_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
-    const std::string& ptx,
+    const std::vector<char>& ptx,
     const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
     const std::string& source_code) {
   if (cache_dir.empty()) {
@@ -219,18 +217,22 @@ constexpr const char* g_headers[] = {
     jit_source_utils,
 };
 
-void compile(
+} // namespace
+
+JitModule::JitModule(
     Device& device,
     const std::string& module_name,
-    const std::string& source,
+    const KernelBuilder& builder) {
-    const std::vector<std::string>& kernel_names,
+  // Check cache.
-    std::string& ptx,
+  std::vector<char> ptx;
-    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
+  std::vector<std::pair<std::string, std::string>> ptx_kernels;
-  // Create the program
+  if (!read_cached_ptx(ptx_cache_dir(), module_name, &ptx, &ptx_kernels)) {
+    // Create program.
+    auto [source_code, kernel_names] = builder();
     nvrtcProgram prog;
     CHECK_NVRTC_ERROR(nvrtcCreateProgram(
         &prog,
-      source.c_str(),
+        source_code.c_str(),
         (module_name + ".cu").c_str(),
         std::size(g_headers),
         g_headers,
@@ -284,20 +286,16 @@ void compile(
     } else {
       CHECK_NVRTC_ERROR(nvrtcGetPTXSize(prog, &ptx_size));
     }
-  ptx.resize(ptx_size);
+    ptx.resize(ptx_size, 0);
     if (use_sass) {
       CHECK_NVRTC_ERROR(nvrtcGetCUBIN(prog, ptx.data()));
     } else {
       CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
     }
+    write_cached_ptx(
+        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
   }
 
-void load_module(
-    const std::string& module_name,
-    const std::string& ptx,
-    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
-    CUmodule& module_,
-    std::unordered_map<std::string, std::pair<CUfunction, bool>>& kernels) {
   // Load module.
   char jit_log[4089] = {};
   CUjit_option options[] = {
@@ -314,69 +312,21 @@ void load_module(
   for (const auto& [name, mangled] : ptx_kernels) {
     CUfunction kernel;
     CHECK_CUDA_ERROR(cuModuleGetFunction(&kernel, module_, mangled.c_str()));
-    kernels[name] = std::make_pair(kernel, false);
+    kernels_[name] = kernel;
   }
 }
 
-} // namespace
-
-JitModule::JitModule(
-    Device& device,
-    const std::string& module_name,
-    const KernelBuilder& builder,
-    bool use_disk_cache) {
-  // Will hold the actual device executable source code and kernel names
-  std::string ptx;
-  std::vector<std::pair<std::string, std::string>> ptx_kernels;
-
-  // Try to load them from the file cache
-  if (!read_cached_ptx(ptx_cache_dir(), module_name, ptx, ptx_kernels)) {
-    auto [precompiled, source_code, kernel_names] = builder();
-
-    // Get the PTX or cubin
-    if (precompiled) {
-      ptx = std::move(source_code);
-      for (auto& name : kernel_names) {
-        ptx_kernels.emplace_back(name, name);
-      }
-    } else {
-      compile(device, module_name, source_code, kernel_names, ptx, ptx_kernels);
-    }
-
-    // If requested save them in the file cache for the next launch
-    if (use_disk_cache) {
-      write_cached_ptx(
-          ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
-    }
-  }
-
-  // Load the module
-  load_module(module_name, ptx, ptx_kernels, module_, kernels_);
-}
-
 JitModule::~JitModule() {
   CHECK_CUDA_ERROR(cuModuleUnload(module_));
 }
 
-CUfunction JitModule::get_kernel(
+CUfunction JitModule::get_kernel(const std::string& kernel_name) {
-    const std::string& kernel_name,
-    std::function<void(CUfunction)> configure_kernel) {
   auto it = kernels_.find(kernel_name);
   if (it == kernels_.end()) {
     throw std::runtime_error(
         fmt::format("There is no kernel named {}.", kernel_name));
   }
-
+  return it->second;
-  // If it is the first time we run this kernel then configure it. Do it only
-  // once!
-  if (!it->second.second) {
-    if (configure_kernel) {
-      configure_kernel(it->second.first);
-    }
-    it->second.second = true;
-  }
-
-  return it->second.first;
 }
 
 std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
@@ -387,12 +337,11 @@ std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,
-    const KernelBuilder& builder,
+    const KernelBuilder& builder) {
-    bool cache) {
   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, cache).first;
+    it = map.try_emplace(name, cu::device(device), name, builder).first;
   }
   return it->second;
 }

mlx/backend/cuda/jit_module.h

@@ -19,8 +19,7 @@ namespace mlx::core::cu {
 
 class Device;
 
-using KernelBuilderResult = std::tuple<
+using KernelBuilderResult = std::pair<
-    /* precompiled */ bool,
     /* source code */ std::string,
     /* kernel names */ std::vector<std::string>>;
 using KernelBuilder = std::function<KernelBuilderResult()>;
@@ -46,11 +45,6 @@ struct KernelArgs {
     append_ptr(std::get<SmallVector<T>>(storage_.back()).data());
   }
 
-  template <typename T>
-  void append(const std::vector<T>& vec) {
-    append(SmallVector<T>(vec.begin(), vec.end()));
-  }
-
   // Make sure the arg is copied to an array with size of NDIM.
   template <size_t NDIM = MAX_NDIM, typename T>
   void append_ndim(SmallVector<T> vec) {
@@ -69,16 +63,14 @@ struct KernelArgs {
  private:
   std::vector<void*> args_;
 
-  // The cuGraphAddKernelNode API requires passing pointers to arguments so
+  // The cuLaunchKernel API requires passing pointers to arguments so store
-  // store temporary values until the node is created.
+  // temporary values untill kernel is launched.
   using Arg = std::variant<
       std::monostate,
       CUdeviceptr,
-      bool,
       int32_t,
       uint32_t,
       int64_t,
-      float,
       SmallVector<const void*>,
       SmallVector<int32_t>,
       SmallVector<int64_t>>;
@@ -90,19 +82,16 @@ class JitModule {
   JitModule(
       Device& device,
       const std::string& module_name,
-      const KernelBuilder& builder,
+      const KernelBuilder& builder);
-      bool cache);
   ~JitModule();
 
   JitModule(const JitModule&) = delete;
   JitModule& operator=(const JitModule&) = delete;
-  CUfunction get_kernel(
+  CUfunction get_kernel(const std::string& kernel_name);
-      const std::string& kernel_name,
-      std::function<void(CUfunction)> configure_kernel = nullptr);
 
  private:
   CUmodule module_{nullptr};
-  std::unordered_map<std::string, std::pair<CUfunction, bool>> kernels_;
+  std::unordered_map<std::string, CUfunction> kernels_;
 };
 
 std::unordered_map<std::string, JitModule>& get_jit_module_cache();
@@ -110,7 +99,6 @@ std::unordered_map<std::string, JitModule>& get_jit_module_cache();
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,
-    const KernelBuilder& builder,
+    const KernelBuilder& builder);
-    bool use_disk_cache = true);
 
 } // namespace mlx::core::cu

mlx/backend/cuda/matmul.cpp

@@ -7,6 +7,8 @@
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/primitives.h"
 
+#include "mlx/backend/cuda/gemms/steel_gemm.h"
+
 #include <nvtx3/nvtx3.hpp>
 #include <numeric>
 
@@ -95,6 +97,24 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     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
   CublasGemm gemm(

mlx/backend/cuda/no_cuda.cpp

@@ -1,47 +1,11 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/cuda.h"
-#include "mlx/fast.h"
 
-namespace mlx::core {
+namespace mlx::core::cu {
-
-namespace cu {
 
 bool is_available() {
   return false;
 }
 
-} // namespace cu
+} // namespace mlx::core::cu
-
-namespace fast {
-
-CustomKernelFunction cuda_kernel(
-    const std::string&,
-    const std::vector<std::string>&,
-    const std::vector<std::string>&,
-    const std::string&,
-    const std::string&,
-    bool,
-    int) {
-  throw std::runtime_error("[cuda_kernel] No CUDA back-end.");
-}
-
-std::vector<array> precompiled_cuda_kernel(
-    const std::string&,
-    const std::string&,
-    const std::vector<array>&,
-    const std::vector<Shape>&,
-    const std::vector<Dtype>&,
-    const std::vector<ScalarArg>&,
-    std::tuple<int, int, int>,
-    std::tuple<int, int, int>,
-    int shared_memory,
-    std::optional<float> init_value,
-    bool ensure_row_contiguous,
-    StreamOrDevice) {
-  throw std::runtime_error("[cuda_kernel] No CUDA back-end.");
-}
-
-} // namespace fast
-
-} // namespace mlx::core

mlx/backend/cuda/primitives.cpp

@@ -24,6 +24,8 @@ namespace mlx::core {
   }
 
 NO_GPU(BlockMaskedMM)
+NO_GPU(DynamicSlice)
+NO_GPU(DynamicSliceUpdate)
 NO_GPU(FFT)
 NO_GPU(GatherMM)
 NO_GPU(GatherQMM)
@@ -39,7 +41,12 @@ 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)

mlx/backend/cuda/quantized/quantized.cpp

@@ -46,10 +46,10 @@ inline array ensure_row_contiguous_matrix(
 
 } // namespace
 
-void fast::Quantize::eval_gpu(
+void fast::AffineQuantize::eval_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
-  nvtx3::scoped_range r("Quantize::eval_gpu");
+  nvtx3::scoped_range r("AffineQuantize::eval_gpu");
   auto& s = stream();
   auto& d = cu::device(s.device);
   auto& enc = d.get_command_encoder(s);

mlx/backend/cuda/slicing.cpp

@@ -1,11 +1,8 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/common/slicing.h"
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/jit_module.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/gpu/slicing.h"
-#include "mlx/dtype_utils.h"
 
 #include <numeric>
 
@@ -30,7 +27,8 @@ void concatenate_gpu(
   flags.row_contiguous = false;
   flags.col_contiguous = false;
   flags.contiguous = false;
-  auto concurrent = cu::get_command_encoder(s).concurrent_context();
+  // TODO: Handle concurrent outputs:
+  // https://github.com/ml-explore/mlx/pull/2145#discussion_r2070753816
   for (int i = 0; i < inputs.size(); i++) {
     array out_slice(inputs[i].shape(), out.dtype(), nullptr, {});
     size_t data_offset = strides[axis] * sizes[i];
@@ -40,71 +38,4 @@ void concatenate_gpu(
   }
 }
 
-array compute_dynamic_offset(
-    const array& indices,
-    const Strides& strides,
-    const std::vector<int>& axes,
-    const Stream& s) {
-  Dtype dtype = indices.dtype();
-  int nidx = axes.size();
-
-  std::string module_name =
-      fmt::format("compute_dynamic_offset_{}_{}", dtype_to_string(dtype), nidx);
-  std::string kernel_name = fmt::format(
-      "mlx::core::cu::compute_dynamic_offset<{}, {}>",
-      dtype_to_cuda_type(dtype),
-      nidx);
-
-  cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
-    std::string source = R"(
-        #include "mlx/backend/cuda/device/utils.cuh"
-
-        namespace mlx::core::cu {
-
-        template <typename T, int NIDX>
-        __global__ void compute_dynamic_offset(
-            const T* indices,
-            int64_t* offset,
-            const __grid_constant__ Strides strides,
-            const __grid_constant__ cuda::std::array<int, NIDX> axes) {
-          int64_t acc = 0;
-          #pragma unroll
-          for (int i = 0; i < NIDX; ++i) {
-            acc += indices[i] * strides[axes[i]];
-          }
-          *offset = acc;
-        }
-
-        } // namespace mlx::core::cu
-    )";
-    return std::make_tuple(false, std::move(source), std::vector{kernel_name});
-  });
-
-  // Prepare output.
-  array offset({1}, int64, nullptr, {});
-  bool donate = indices.is_donatable() &&
-      (indices.data_size() * indices.itemsize()) >= offset.itemsize();
-  if (donate) {
-    offset.copy_shared_buffer(indices);
-  } else {
-    offset.set_data(allocator::malloc(offset.itemsize()));
-  }
-
-  auto& encoder = cu::get_command_encoder(s);
-  encoder.add_temporary(offset);
-  encoder.set_input_array(indices);
-  encoder.set_output_array(offset);
-
-  cu::KernelArgs args;
-  args.append(indices);
-  args.append(offset);
-  args.append_ndim(strides);
-  args.append(axes);
-
-  auto kernel = mod.get_kernel(kernel_name);
-  encoder.add_kernel_node(kernel, 1, 1, 0, args.args());
-
-  return offset;
-}
-
 } // namespace mlx::core

mlx/backend/cuda/sort.cu

@@ -1,5 +1,6 @@
 // Copyright © 2025 Apple Inc.
 
+#include "mlx/backend/common/utils.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/gpu/copy.h"
@@ -9,7 +10,7 @@
 #include <nvtx3/nvtx3.hpp>
 #include <thrust/device_ptr.h>
 #include <thrust/transform.h>
-#include <cub/device/device_segmented_radix_sort.cuh>
+#include <cub/device/device_segmented_sort.cuh>
 
 #include <cassert>
 
@@ -79,7 +80,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
         encoder.add_temporary(discard);
 
         size_t size;
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortPairs(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
             nullptr,
             size,
             in.data<Type>(),
@@ -90,8 +91,6 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
-            0,
-            sizeof(Type) * 8,
             stream));
 
         array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
@@ -106,7 +105,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             thrust::device_pointer_cast(indices.data<uint32_t>()),
             ModOp<uint32_t>{static_cast<uint32_t>(nsort)});
 
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortPairs(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortPairs(
             temp.data<void>(),
             size,
             in.data<Type>(),
@@ -117,12 +116,10 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
-            0,
-            sizeof(Type) * 8,
             stream));
       } else {
         size_t size;
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortKeys(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
             nullptr,
             size,
             in.data<Type>(),
@@ -131,8 +128,6 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
-            0,
-            sizeof(Type) * 8,
             stream));
 
         array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
@@ -140,7 +135,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
 
         // Start capturing after allocations
         auto capture = encoder.capture_context();
-        CHECK_CUDA_ERROR(cub::DeviceSegmentedRadixSort::SortKeys(
+        CHECK_CUDA_ERROR(cub::DeviceSegmentedSort::StableSortKeys(
             temp.data<void>(),
             size,
             in.data<Type>(),
@@ -149,8 +144,6 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             in.data_size() / nsort,
             offsets,
             offsets + 1,
-            0,
-            sizeof(Type) * 8,
             stream));
       }
     } else {

mlx/backend/cuda/steel/tiles.cuh

@@ -143,85 +143,87 @@ struct Tile16x16 {
   }
 };
 
-/**
+// /**
- * A simple container of multiple Tile16x16.
+//  * A simple container of multiple Tile16x16.
- *
+//  *
- * Provides utility functions for loading and manipulating collections of basic
+//  * Provides utility functions for loading and manipulating collections of
- * tiles.
+//  basic
- */
+//  * tiles.
-template <typename T, int ROWS_, int COLS_>
+//  */
-struct RegisterTile {
+// template <typename T, int ROWS_, int COLS_>
-  static constexpr int ROWS = ROWS_;
+// struct RegisterTile {
-  static constexpr int COLS = COLS_;
+//   static constexpr int ROWS = ROWS_;
-  static constexpr int TILES_X = COLS / 16;
+//   static constexpr int COLS = COLS_;
-  static constexpr int TILES_Y = ROWS / 16;
+//   static constexpr int TILES_X = COLS / 16;
+//   static constexpr int TILES_Y = ROWS / 16;
 
-  Tile16x16<T> data[TILES_X * TILES_Y];
+//   Tile16x16<T> data[TILES_X * TILES_Y];
 
-  __device__ inline void fill(T v) {
+//   __device__ inline void fill(T v) {
-    MLX_UNROLL
+//     MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
+//     for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
+//       MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
+//       for (int j = 0; j < TILES_X; j++) {
-        data[i * TILES_X + j].fill(v);
+//         data[i * TILES_X + j].fill(v);
-      }
+//       }
-    }
+//     }
-  }
+//   }
 
-  template <typename Tile>
+//   template <typename Tile>
-  __device__ __forceinline__ void
+//   __device__ __forceinline__ void
-  load(Tile& tile, uint32_t base_address, int row, int col) {
+//   load(Tile& tile, uint32_t base_address, int row, int col) {
-    MLX_UNROLL
+//     MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
+//     for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
+//       MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
+//       for (int j = 0; j < TILES_X; j++) {
-        data[i * TILES_X + j].load(
+//         data[i * TILES_X + j].load(
-            tile.loc(base_address, row + i * 16, col + j * 16));
+//             tile.loc(base_address, row + i * 16, col + j * 16));
-      }
+//       }
-    }
+//     }
-  }
+//   }
 
-  template <typename Tile, typename F>
+//   template <typename Tile, typename F>
-  __device__ __forceinline__ void
+//   __device__ __forceinline__ void
-  load(Tile& tile, F f, uint32_t base_address, int row, int col) {
+//   load(Tile& tile, F f, uint32_t base_address, int row, int col) {
-    MLX_UNROLL
+//     MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
+//     for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
+//       MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
+//       for (int j = 0; j < TILES_X; j++) {
-        f(data[i * TILES_X + j],
+//         f(data[i * TILES_X + j],
-          tile,
+//           tile,
-          base_address,
+//           base_address,
-          row + i * 16,
+//           row + i * 16,
-          col + j * 16);
+//           col + j * 16);
-      }
+//       }
-    }
+//     }
-  }
+//   }
 
-  template <typename U>
+//   template <typename U>
-  __device__ inline void store_global(U* x, int N, int row, int col) {
+//   __device__ inline void store_global(U* x, int N, int row, int col) {
-    MLX_UNROLL
+//     MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
+//     for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
+//       MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
+//       for (int j = 0; j < TILES_X; j++) {
-        data[i * TILES_X + j].store_global(
+//         data[i * TILES_X + j].store_global(
-            x + (row + i * 16) * N + col + j * 16, N);
+//             x + (row + i * 16) * N + col + j * 16, N);
-      }
+//       }
-    }
+//     }
-  }
+//   }
 
-  template <typename U>
+//   template <typename U>
-  __device__ inline void
+//   __device__ inline void
-  store_global_safe(U* x, int N, int row, int col, int max_rows) {
+//   store_global_safe(U* x, int N, int row, int col, int max_rows) {
-    MLX_UNROLL
+//     MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
+//     for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
+//       MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
+//       for (int j = 0; j < TILES_X; j++) {
-        data[i * TILES_X + j].store_global_safe(
+//         data[i * TILES_X + j].store_global_safe(
-            x + (row + i * 16) * N + col + j * 16, N, max_rows - row - i * 16);
+//             x + (row + i * 16) * N + col + j * 16, N, max_rows - row - i *
-      }
+//             16);
-    }
+//       }
-  }
+//     }
-};
+//   }
+// };
 
 /**
  * A simple container of multiple Tile16x16.

mlx/backend/cuda/ternary.cu

@@ -39,98 +39,52 @@ ternary_v(const bool* a, const T* b, const T* c, T* out, IdxT size) {
   }
 }
 
-template <typename Op, typename T, typename IdxT, int NDIM, int N_READS>
+template <typename Op, typename T, typename IdxT, int NDIM>
 __global__ void ternary_g_nd(
     const bool* a,
     const T* b,
     const T* c,
     T* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> c_strides) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[NDIM - 1];
-  auto a_stride_x = a_strides[NDIM - 1];
-  auto b_stride_x = b_strides[NDIM - 1];
-  auto c_stride_x = c_strides[NDIM - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [a_idx, b_idx, c_idx] = elem_to_loc_nd<NDIM>(
-      index_rest * shape_x,
+        index,
         shape.data(),
         a_strides.data(),
         b_strides.data(),
         c_strides.data());
-  auto a_vec =
+    out[index] = Op{}(a[a_idx], b[b_idx], c[c_idx]);
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, false);
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, T(0));
-  auto c_vec =
-      load_vector<N_READS>(c + c_idx, index_x, shape_x, c_stride_x, T(0));
-
-  AlignedVector<T, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(a_vec[i], b_vec[i], c_vec[i]);
   }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
-template <typename Op, typename T, typename IdxT, int N_READS>
+template <typename Op, typename T, typename IdxT>
 __global__ void ternary_g(
     const bool* a,
     const T* b,
     const T* c,
     T* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     const __grid_constant__ Strides c_strides,
     int ndim) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
-  if (index_rest >= size_rest) {
-    return;
-  }
-
-  auto shape_x = shape[ndim - 1];
-  auto a_stride_x = a_strides[ndim - 1];
-  auto b_stride_x = b_strides[ndim - 1];
-  auto c_stride_x = c_strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
     auto [a_idx, b_idx, c_idx] = elem_to_loc(
-      index_rest * shape_x,
+        index,
         shape.data(),
         a_strides.data(),
         b_strides.data(),
         c_strides.data(),
         ndim);
-  auto a_vec =
+    out[index] = Op{}(a[a_idx], b[b_idx], c[c_idx]);
-      load_vector<N_READS>(a + a_idx, index_x, shape_x, a_stride_x, false);
-  auto b_vec =
-      load_vector<N_READS>(b + b_idx, index_x, shape_x, b_stride_x, T(0));
-  auto c_vec =
-      load_vector<N_READS>(c + c_idx, index_x, shape_x, c_stride_x, T(0));
-
-  AlignedVector<T, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(a_vec[i], b_vec[i], c_vec[i]);
   }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 } // namespace cu
@@ -169,55 +123,36 @@ void ternary_op_gpu_inplace(
             auto& b_strides = strides[1];
             auto& c_strides = strides[2];
             int ndim = shape.size();
-            int work_per_thread = 1;
-            auto dim0 = ndim > 0 ? shape.back() : 1;
-            auto rest = out.size() / dim0;
-            if (dim0 >= 4) {
-              work_per_thread = 4;
-            }
-            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-            auto block_dims = get_block_dims(dim0, rest, 1);
-            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                auto kernel =
+                auto [num_blocks, block_dims] = get_launch_args(out, large());
-                    cu::ternary_g_nd<Op, DType, IdxT, dims_constant(), 1>;
-                if (work_per_thread == 4) {
-                  kernel =
-                      cu::ternary_g_nd<Op, DType, IdxT, dims_constant(), 4>;
-                }
                 encoder.add_kernel_node(
-                    kernel,
+                    cu::ternary_g_nd<Op, DType, IdxT, dims_constant()>,
-                    {num_blocks_x, num_blocks_y},
+                    num_blocks,
                     block_dims,
                     0,
                     a.data<bool>(),
                     b.data<DType>(),
                     c.data<DType>(),
                     out.data<DType>(),
-                    rest,
+                    out.size(),
                     const_param<dims_constant()>(shape),
                     const_param<dims_constant()>(a_strides),
                     const_param<dims_constant()>(b_strides),
                     const_param<dims_constant()>(c_strides));
               });
             } else {
-              auto kernel = cu::ternary_g<Op, DType, IdxT, 1>;
+              auto [num_blocks, block_dims] = get_launch_args(out, large());
-              if (work_per_thread == 4) {
-                kernel = cu::ternary_g<Op, DType, IdxT, 4>;
-              }
               encoder.add_kernel_node(
-                  kernel,
+                  cu::ternary_g<Op, DType, IdxT>,
-                  {num_blocks_x, num_blocks_y},
+                  num_blocks,
                   block_dims,
                   0,
                   a.data<bool>(),
                   b.data<DType>(),
                   c.data<DType>(),
                   out.data<DType>(),
-                  rest,
+                  out.data_size(),
                   const_param(shape),
                   const_param(a_strides),
                   const_param(b_strides),

mlx/backend/cuda/unary.cu

@@ -37,36 +37,19 @@ __global__ void unary_v(const In* in, Out* out, IdxT size) {
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT>
 __global__ void unary_g(
     const In* in,
     Out* out,
-    IdxT size_rest,
+    IdxT size,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides strides,
     int ndim) {
-  auto block = cg::this_thread_block();
+  IdxT index = cg::this_grid().thread_rank();
-  auto grid = cg::this_grid();
+  if (index < size) {
-  IdxT index_rest =
+    auto idx = elem_to_loc(index, shape.data(), strides.data(), ndim);
-      grid.block_index().y * block.dim_threads().y + block.thread_index().y;
+    out[index] = Op{}(in[idx]);
-  if (index_rest >= size_rest) {
-    return;
   }
-
-  auto shape_x = shape[ndim - 1];
-  auto stride_x = strides[ndim - 1];
-  IdxT index_x =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  auto idx =
-      elem_to_loc(index_rest * shape_x, shape.data(), strides.data(), ndim);
-  auto in_vec =
-      load_vector<N_READS>(in + idx, index_x, shape_x, stride_x, In(0));
-  AlignedVector<Out, N_READS> out_vec;
-#pragma unroll
-  for (int i = 0; i < N_READS; ++i) {
-    out_vec[i] = Op{}(in_vec[i]);
-  }
-  store_vector(out + shape_x * index_rest, index_x, out_vec, shape_x);
 }
 
 template <typename Op, typename In, typename Out>
@@ -144,7 +127,8 @@ void unary_op_gpu_inplace(
           using OutType = cuda_type_t<CTYPE_OUT>;
           if (contig) {
             using IdxT = std::conditional_t<large(), int64_t, uint32_t>;
-            constexpr int N_READS = 16 / sizeof(OutType);
+            // TODO: Choose optimized value based on type size.
+            constexpr int N_READS = 4;
             auto [num_blocks, block_dims] = get_launch_args(
                 out.data_size(), out.shape(), out.strides(), large, N_READS);
             encoder.add_kernel_node(
@@ -158,30 +142,18 @@ void unary_op_gpu_inplace(
           } else {
             using IdxT = std::conditional_t<large(), int64_t, int32_t>;
             auto [shape, strides] = collapse_contiguous_dims(in);
-            auto ndim = shape.size();
+            auto [num_blocks, block_dims] = get_launch_args(out, large);
-            int work_per_thread = 1;
-            auto kernel = cu::unary_g<Op, InType, OutType, IdxT, 1>;
-            auto dim0 = ndim > 0 ? shape.back() : 1;
-            auto rest = out.size() / dim0;
-            if (dim0 >= 4) {
-              kernel = cu::unary_g<Op, InType, OutType, IdxT, 4>;
-              work_per_thread = 4;
-            }
-            dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-            auto block_dims = get_block_dims(dim0, rest, 1);
-            uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-            uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
             encoder.add_kernel_node(
-                kernel,
+                cu::unary_g<Op, InType, OutType, IdxT>,
-                {num_blocks_x, num_blocks_y},
+                num_blocks,
                 block_dims,
                 0,
                 in.data<InType>(),
                 out.data<OutType>(),
-                rest,
+                out.data_size(),
                 const_param(shape),
                 const_param(strides),
-                ndim);
+                shape.size());
           }
         });
       } else {

mlx/backend/cuda/unary/CMakeLists.txt

@@ -1,34 +0,0 @@
-target_sources(
-  mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/abs.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arccos.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arccosh.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arcsin.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arcsinh.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctan.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctanh.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/bitwise_invert.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/ceil.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/conjugate.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/cos.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/cosh.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/erf.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/erf_inv.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/exp.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/expm1.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/floor.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/imag.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log1p.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_not.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/negative.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/real.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/round.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sigmoid.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sign.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sin.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sinh.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/sqrt.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/square.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tan.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tanh.cu)

mlx/backend/cuda/unary/abs.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Abs)
-} // namespace mlx::core

mlx/backend/cuda/unary/arccos.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(ArcCos)
-} // namespace mlx::core

mlx/backend/cuda/unary/arccosh.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(ArcCosh)
-} // namespace mlx::core

mlx/backend/cuda/unary/arcsin.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(ArcSin)
-} // namespace mlx::core

mlx/backend/cuda/unary/arcsinh.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(ArcSinh)
-} // namespace mlx::core

mlx/backend/cuda/unary/arctan.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(ArcTan)
-} // namespace mlx::core

mlx/backend/cuda/unary/arctanh.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(ArcTanh)
-} // namespace mlx::core

mlx/backend/cuda/unary/bitwise_invert.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(BitwiseInvert)
-} // namespace mlx::core

mlx/backend/cuda/unary/ceil.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Ceil)
-} // namespace mlx::core

mlx/backend/cuda/unary/conjugate.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Conjugate)
-} // namespace mlx::core

mlx/backend/cuda/unary/cos.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Cos)
-} // namespace mlx::core

mlx/backend/cuda/unary/cosh.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Cosh)
-} // namespace mlx::core

mlx/backend/cuda/unary/erf.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Erf)
-} // namespace mlx::core

mlx/backend/cuda/unary/erf_inv.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(ErfInv)
-} // namespace mlx::core

mlx/backend/cuda/unary/exp.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Exp)
-} // namespace mlx::core

mlx/backend/cuda/unary/expm1.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Expm1)
-} // namespace mlx::core

mlx/backend/cuda/unary/floor.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Floor)
-} // namespace mlx::core

mlx/backend/cuda/unary/imag.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Imag)
-} // namespace mlx::core

mlx/backend/cuda/unary/log.cu

@@ -1,21 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-void Log::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Log::eval_gpu");
-  auto& s = out.primitive().stream();
-  switch (base_) {
-    case Base::e:
-      unary_op_gpu<cu::Log>(inputs, out, name(), s);
-      break;
-    case Base::two:
-      unary_op_gpu<cu::Log2>(inputs, out, name(), s);
-      break;
-    case Base::ten:
-      unary_op_gpu<cu::Log10>(inputs, out, name(), s);
-      break;
-  }
-}
-} // namespace mlx::core

mlx/backend/cuda/unary/log1p.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Log1p)
-} // namespace mlx::core

mlx/backend/cuda/unary/logical_not.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(LogicalNot)
-} // namespace mlx::core

mlx/backend/cuda/unary/negative.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Negative)
-} // namespace mlx::core

mlx/backend/cuda/unary/real.cu

@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-UNARY_GPU(Real)
-} // namespace mlx::core

mlx/backend/cuda/unary/round.cu

@@ -1,18 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/unary/unary.cuh"
-
-namespace mlx::core {
-void Round::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Round::eval_gpu");
-  assert(inputs.size() == 1);
-  const auto& in = inputs[0];
-  auto& s = out.primitive().stream();
-  if (issubdtype(in.dtype(), inexact)) {
-    unary_op_gpu<cu::Round>(inputs, out, name(), s);
-  } else {
-    // No-op integer types
-    out.copy_shared_buffer(in);
-  }
-}
-} // namespace mlx::core