CMakeLists.txt update

* optionally load metallib from framework

* pre-commit

* adjust logic

.circleci/config.yml

@@ -18,16 +18,17 @@ jobs:
         type: boolean
         default: false
     macos:
-      xcode: "16.2.0"
-    resource_class: m2pro.medium
+      xcode: "26.0.0"
+    resource_class: m4pro.medium
     steps:
       - checkout
       - run:
           name: Install
           command: |
-            brew install python@3.9
+            xcodebuild -downloadComponent MetalToolchain
+            brew install python@3.10
             brew install doxygen
-            python3.9 -m venv env
+            python3.10 -m venv env
             source env/bin/activate
             pip install --upgrade pip
             pip install --upgrade cmake
@@ -89,7 +90,8 @@ jobs:
           command: |
             uv venv
             uv pip install cmake
-            uv pip install -e ".[dev]" -v
+            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
+              uv pip install -e ".[dev]" -v
       - run:
           name: Generate package stubs
           command: |
@@ -118,7 +120,7 @@ jobs:
     parameters:
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       macosx_deployment_target:
         type: string
         default: ""
@@ -126,18 +128,19 @@ jobs:
       xcode: << parameters.xcode_version >>
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.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:
           name: Install Python package
           command: |
-            uv venv --python 3.9
+            uv venv --python 3.10
             uv pip install \
               nanobind==2.4.0 \
               cmake \
@@ -196,7 +199,7 @@ jobs:
           name: Run Python tests with JIT
           command: |
             CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e .
+              uv pip install -e . -v
             LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
               METAL_DEBUG_ERROR_MODE=0 \
               uv run --no-project python -m xmlrunner discover \
@@ -222,15 +225,20 @@ 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
             curl -LsSf https://astral.sh/uv/install.sh | sh
+      - run:
+          name: Set CCache size
+          command: ccache --max-size 1G
       - run:
           name: Install Python package
           command: |
             uv venv
-            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+            uv pip install cmake
+            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
@@ -238,12 +246,23 @@ 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: |
             ccache --show-stats
             ccache --zero-stats
-            ccache --max-size 400MB
             ccache --cleanup
       - save_cache:
           key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
@@ -254,10 +273,10 @@ jobs:
     parameters:
       python_version:
         type: string
-        default: "3.9"
+        default: "3.10"
       xcode_version:
         type: string
-        default: "16.2.0"
+        default: "26.0.0"
       build_env:
         type: string
         default: ""
@@ -266,7 +285,7 @@ jobs:
         default: ""
     macos:
       xcode: << parameters.xcode_version >>
-    resource_class: m2pro.medium
+    resource_class: m4pro.medium
     environment:
       MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
     steps:
@@ -274,11 +293,15 @@ jobs:
       - run:
           name: Install dependencies
           command: |
-            brew install python@<< parameters.python_version >>
-            brew install openmpi
-            python<< parameters.python_version >> -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
+            xcodebuild -downloadComponent MetalToolchain
+            mkdir -p ~/miniconda3
+            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
+            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+            rm ~/miniconda3/miniconda.sh
+            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
@@ -288,29 +311,29 @@ jobs:
       - run:
           name: Install Python package
           command: |
-            source env/bin/activate
+            conda activate env
             env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
               pip install . -v
       - run:
           name: Generate package stubs
           command: |
-            source env/bin/activate
+            conda activate env
             pip install typing_extensions
             python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
-            source env/bin/activate
+            conda activate env
             python setup.py clean --all
             << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
       - when:
           condition:
-            equal: ["3.9", << parameters.python_version >>]
+            equal: ["3.10", << parameters.python_version >>]
           steps:
             - run:
                 name: Build common package
                 command: |
-                  source env/bin/activate
+                  conda activate env
                   python setup.py clean --all
                   << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
       - when:
@@ -319,7 +342,7 @@ jobs:
             - run:
                 name: Upload package
                 command: |
-                  source env/bin/activate
+                  conda activate env
                   twine upload dist/*
       - store_artifacts:
           path: dist/
@@ -328,7 +351,7 @@ jobs:
     parameters:
       python_version:
         type: string
-        default: "3.9"
+        default: "3.10"
       build_env:
         type: string
         default: ""
@@ -364,7 +387,7 @@ jobs:
             bash python/scripts/repair_linux.sh
       - when:
           condition:
-            equal: ["3.9", << parameters.python_version >>]
+            equal: ["3.10", << parameters.python_version >>]
           steps:
             - run:
                 name: Build common package
@@ -392,7 +415,7 @@ jobs:
         default: ""
     machine:
       image: ubuntu-2204:current
-      resource_class: large
+      resource_class: xlarge
     steps:
       - checkout
       - run:
@@ -439,7 +462,7 @@ workflows:
       - mac_build_and_test:
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test
       - cuda_build_and_test:
           matrix:
@@ -461,71 +484,10 @@ workflows:
               ignore: /.*/
           matrix:
             parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "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"
+              xcode_version: ["26.0.0"]
       - build_documentation:
           filters:
             tags:
@@ -541,7 +503,7 @@ workflows:
               ignore: /.*/
           matrix:
             parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
               build_env: ["PYPI_RELEASE=1"]
       - build_cuda_release:
           filters:
@@ -567,7 +529,7 @@ workflows:
           requires: [ hold ]
           matrix:
             parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
+              macosx_deployment_target: ["13.5", "15.0"]
       - linux_build_and_test:
           requires: [ hold ]
       - cuda_build_and_test:
@@ -584,59 +546,13 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              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"
+              xcode_version: ["26.0.0"]
       - build_linux_release:
           matrix:
             parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
       - build_cuda_release
 
   build_dev_release:
@@ -648,75 +564,14 @@ workflows:
       - build_release:
           matrix:
             parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
               macosx_deployment_target: ["13.5", "14.0", "15.0"]
               build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
+              xcode_version: ["26.0.0"]
       - build_linux_release:
           matrix:
             parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
               build_env: ["DEV_RELEASE=1"]
       - build_cuda_release:
           matrix:

ACKNOWLEDGMENTS.md

@@ -19,12 +19,17 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
-- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer, and the `ReLU²` activation function.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>
 
+# 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

@@ -20,12 +20,17 @@ project(
   LANGUAGES C CXX
   VERSION ${MLX_PROJECT_VERSION})
 
+if(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang")
+  add_compile_options(-Wall -Wextra)
+endif()
+
 # ----------------------------- Setup -----------------------------
 set(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # ----------------------------- Configuration -----------------------------
 option(MLX_BUILD_TESTS "Build tests for mlx" ON)
@@ -87,22 +92,21 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-if(MLX_BUILD_METAL)
-  set(METAL_LIB "-framework Metal")
-  set(FOUNDATION_LIB "-framework Foundation")
-  set(QUARTZ_LIB "-framework QuartzCore")
-endif()
-
 if(MLX_BUILD_CUDA)
   enable_language(CUDA)
 endif()
 
-if(MLX_BUILD_METAL AND NOT METAL_LIB)
-  message(STATUS "Metal not found. Unable to build GPU")
-  set(MLX_BUILD_METAL OFF)
-  set(MLX_METAL_DEBUG OFF)
-elseif(MLX_BUILD_METAL)
-  message(STATUS "Building METAL sources")
+if(MLX_BUILD_METAL)
+  find_library(METAL_LIB Metal)
+  find_library(FOUNDATION_LIB Foundation)
+  find_library(QUARTZ_LIB QuartzCore)
+  if(METAL_LIB)
+    message(STATUS "Metal found ${METAL_LIB}")
+  else()
+    message(
+      FATAL_ERROR
+        "Metal not found. Set MLX_BUILD_METAL=OFF to build without GPU")
+  endif()
 
   if(MLX_METAL_DEBUG)
     add_compile_definitions(MLX_METAL_DEBUG)
@@ -111,7 +115,8 @@ elseif(MLX_BUILD_METAL)
   # Throw an error if xcrun not found
   execute_process(
     COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
-    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
+    OUTPUT_VARIABLE MACOS_SDK_VERSION
+    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
 
   if(${MACOS_SDK_VERSION} LESS 14.0)
     message(
@@ -140,6 +145,12 @@ 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.
@@ -167,7 +178,7 @@ if(MLX_BUILD_CPU)
     message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
     set(MLX_BUILD_ACCELERATE ON)
   else()
-    message(STATUS "Accelerate or arm neon not found, using default backend.")
+    message(STATUS "Accelerate not found, using default backend.")
     set(MLX_BUILD_ACCELERATE OFF)
   endif()
 

README.md

@@ -2,7 +2,7 @@
 
 [**Quickstart**](#quickstart) | [**Installation**](#installation) |
 [**Documentation**](https://ml-explore.github.io/mlx/build/html/index.html) |
-[**Examples**](#examples) 
+[**Examples**](#examples)
 
 [![CircleCI](https://circleci.com/gh/ml-explore/mlx.svg?style=svg)](https://circleci.com/gh/ml-explore/mlx)
 
@@ -11,37 +11,37 @@ brought to you by Apple machine learning research.
 
 Some key features of MLX include:
 
- - **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
+- **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
    also has fully featured C++, [C](https://github.com/ml-explore/mlx-c), and
    [Swift](https://github.com/ml-explore/mlx-swift/) APIs, which closely mirror
    the Python API. MLX has higher-level packages like `mlx.nn` and
    `mlx.optimizers` with APIs that closely follow PyTorch to simplify building
    more complex models.
 
- - **Composable function transformations**: MLX supports composable function
-   transformations for automatic differentiation, automatic vectorization,
-   and computation graph optimization.
+- **Composable function transformations**: MLX supports composable function
+  transformations for automatic differentiation, automatic vectorization,
+  and computation graph optimization.
 
- - **Lazy computation**: Computations in MLX are lazy. Arrays are only
-   materialized when needed.
+- **Lazy computation**: Computations in MLX are lazy. Arrays are only
+  materialized when needed.
 
- - **Dynamic graph construction**: Computation graphs in MLX are constructed
-   dynamically. Changing the shapes of function arguments does not trigger
-   slow compilations, and debugging is simple and intuitive.
+- **Dynamic graph construction**: Computation graphs in MLX are constructed
+  dynamically. Changing the shapes of function arguments does not trigger
+  slow compilations, and debugging is simple and intuitive.
 
- - **Multi-device**: Operations can run on any of the supported devices
-   (currently the CPU and the GPU).
+- **Multi-device**: Operations can run on any of the supported devices
+  (currently the CPU and the GPU).
 
- - **Unified memory**: A notable difference from MLX and other frameworks
-   is the *unified memory model*. Arrays in MLX live in shared memory.
-   Operations on MLX arrays can be performed on any of the supported
-   device types without transferring data.
+- **Unified memory**: A notable difference from MLX and other frameworks
+  is the *unified memory model*. Arrays in MLX live in shared memory.
+  Operations on MLX arrays can be performed on any of the supported
+  device types without transferring data.
 
 MLX is designed by machine learning researchers for machine learning
 researchers. The framework is intended to be user-friendly, but still efficient
 to train and deploy models. The design of the framework itself is also
 conceptually simple. We intend to make it easy for researchers to extend and
-improve MLX with the goal of quickly exploring new ideas. 
+improve MLX with the goal of quickly exploring new ideas.
 
 The design of MLX is inspired by frameworks like
 [NumPy](https://numpy.org/doc/stable/index.html),
@@ -91,7 +91,7 @@ Checkout the
 [documentation](https://ml-explore.github.io/mlx/build/html/install.html#)
 for more information on building the C++ and Python APIs from source.
 
-## Contributing 
+## Contributing
 
 Check out the [contribution guidelines](https://github.com/ml-explore/mlx/tree/main/CONTRIBUTING.md) for more information
 on contributing to MLX. See the
@@ -110,7 +110,7 @@ Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. If you find
 MLX useful in your research and wish to cite it, please use the following
 BibTex entry:
 
-```
+```text
 @software{mlx2023,
   author = {Awni Hannun and Jagrit Digani and Angelos Katharopoulos and Ronan Collobert},
   title = {{MLX}: Efficient and flexible machine learning on Apple silicon},

benchmarks/python/blas/bench_gemm.py

@@ -142,9 +142,7 @@ def bench_shape(B, M, N, K, np_dtype, transpose="nn"):
     t_b = (0, 1, 2) if transpose[1] == "n" else (0, 2, 1)
 
     c_mlx = a_mx.transpose(t_a) @ b_mx.transpose(t_b)
-    c_npy = a_np.transpose(t_a).astype(np.float32) @ b_np.transpose(t_b).astype(
-        np.float32
-    )
+    c_npy = a_np.transpose(t_a).astype(np_dtype) @ b_np.transpose(t_b).astype(np_dtype)
 
     atol = 1e-5 if np_dtype == np.float32 else 1e-4
 
@@ -163,7 +161,7 @@ def get_gflop_count(B, M, N, K):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Run gemm benchmarks")
 
-    dtypes = ("float32", "float16")
+    dtypes = ("float32", "float16", "complex64")
     transposes = ("nn", "nt", "tn")
     shapes = (
         (16, 234, 768, 3072),
@@ -187,7 +185,7 @@ if __name__ == "__main__":
                 diff = gflops_mx / gflops_pt - 1.0
 
                 print(
-                    f"{B:3d}, {M:4d}, {N:4d}, {K:4d}, {dtype}, {transpose}, {gflops_pt:05.3f}, {gflops_mx:05.3f}, {100. * diff:+5.2f}%"
+                    f"{B:3d}, {M:4d}, {N:4d}, {K:4d}, {dtype}, {transpose}, {gflops_pt:05.3f}, {gflops_mx:05.3f}, {100.0 * diff:+5.2f}%"
                 )
                 if gflops_pt >= 2.0 * gflops_mx:
                     print("ATTENTION ^^^^^^^")

benchmarks/python/blas/bench_gemv.py

@@ -196,7 +196,7 @@ def bench_with_out_len(ax, out_vec_len, in_vector_lens, dtype, transpose):
 
 
 for transpose in (False, True):
-    for dtype in ("float32", "float16"):
+    for dtype in ("float32", "float16", "complex64"):
         fig, axs = plt.subplots(
             len(in_vec_sizes), 2, figsize=(8.5, 11), layout="constrained"
         )
@@ -215,7 +215,7 @@ for transpose in (False, True):
         fig.suptitle(f"{device_name}: {dtype} {op_name}")
         fig.savefig(
             os.path.join(
-                results_dir, f'{device_name.replace(" ", "_")}_{dtype}_{op_name}.pdf'
+                results_dir, f"{device_name.replace(' ', '_')}_{dtype}_{op_name}.pdf"
             )
         )
         plt.close(fig)

cmake/FindNCCL.cmake

@@ -0,0 +1,54 @@
+# 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

@@ -127,7 +127,8 @@ 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):
@@ -138,7 +139,6 @@ 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

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

docs/src/install.rst

@@ -16,7 +16,7 @@ silicon computer is
 To install from PyPI your system must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
-- Using a native Python >= 3.9
+- Using a native Python >= 3.10
 - macOS >= 13.5
 
 .. note::
@@ -39,7 +39,7 @@ requirements:
 - Nvidia driver >= 550.54.14
 - CUDA toolkit >= 12.0
 - Linux distribution with glibc >= 2.35
-- Python >= 3.9
+- Python >= 3.10
 
 
 CPU-only (Linux)
@@ -55,7 +55,7 @@ To install the CPU-only package from PyPi your system must meet the following
 requirements:
 
 - Linux distribution with glibc >= 2.35
-- Python >= 3.9
+- Python >= 3.10
 
 
 Troubleshooting
@@ -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 -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev libcudnn9-dev-cuda-12 -y
 
 
 When building either the Python or C++ APIs make sure to pass the cmake flag

docs/src/python/cuda.rst

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

docs/src/python/fast.rst

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

docs/src/python/nn/functions.rst

@@ -27,6 +27,7 @@ simple functions.
    mish
    prelu
    relu
+   relu2
    relu6
    selu
    sigmoid

docs/src/python/nn/layers.rst

@@ -50,6 +50,7 @@ Layers
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU2
    ReLU6
    RNN
    RoPE

docs/src/python/ops.rst

@@ -112,6 +112,7 @@ Operations
    max
    maximum
    mean
+   median
    meshgrid
    min
    minimum

docs/src/usage/compile.rst

@@ -130,8 +130,8 @@ Now make an array, and benchmark both functions:
 .. code-block:: python
 
   x = mx.random.uniform(shape=(32, 1000, 4096))
-  timeit(nn.gelu, x)
-  timeit(mx.compile(nn.gelu), x)
+  timeit(gelu, x)
+  timeit(mx.compile(gelu), x)
 
 On an M1 Max the times are 15.5 and 3.1 milliseconds. The compiled ``gelu`` is
 five times faster.
@@ -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), state
+      return mx.exp(z)
 
   fun(mx.array(1.0), mx.array(2.0))
   # Prints [array(3, dtype=float32)]

docs/src/usage/distributed.rst

@@ -184,7 +184,7 @@ almost identical to the example above:
 
     def step(model, x, y):
         loss, grads = loss_grad_fn(model, x, y)
-        grads = mlx.nn.average_gradients(grads) # <---- This line was added
+        grads = mx.nn.average_gradients(grads)  # <---- This line was added
         optimizer.update(model, grads)
         return loss
 

docs/src/usage/export.rst

@@ -164,11 +164,11 @@ to export a function which can be used for inputs with variable shapes:
 
 .. code-block:: python
 
-  mx.export_function("fun.mlxfn", mx.abs, mx.array(0.0), shapeless=True)
+  mx.export_function("fun.mlxfn", mx.abs, mx.array([0.0]), shapeless=True)
   imported_abs = mx.import_function("fun.mlxfn")
 
   # Ok
-  out, = imported_abs(mx.array(-1.0))
+  out, = imported_abs(mx.array([-1.0]))
 
   # Also ok
   out, = imported_abs(mx.array([-1.0, -2.0]))

docs/src/usage/indexing.rst

@@ -107,8 +107,20 @@ same array:
   >>> a
   array([1, 2, 0], dtype=int32)
 
+Note that unlike NumPy, slicing an array creates a copy, not a view. So
+mutating it does not mutate the original array:
 
-Note, unlike NumPy, updates to the same location are nondeterministic:
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> b = a[:]
+  >>> b[2] = 0
+  >>> b
+  array([1, 2, 0], dtype=int32)
+  >>> a
+  array([1, 2, 3], dtype=int32)
+
+Also unlike NumPy, updates to the same location are nondeterministic:
 
 .. code-block:: shell
 

examples/cpp/tutorial.cpp

@@ -14,14 +14,17 @@ void array_basics() {
   // Get the value out of it:
   auto s = x.item<float>();
   assert(s == 1.0);
+  (void)s;
 
   // Scalars have a size of 1:
-  size_t size = x.size();
+  int64_t size = x.size();
   assert(size == 1);
+  (void)size;
 
   // Scalars have 0 dimensions:
   int ndim = x.ndim();
   assert(ndim == 0);
+  (void)ndim;
 
   // The shape should be an empty vector:
   auto shape = x.shape();
@@ -30,6 +33,7 @@ void array_basics() {
   // The datatype should be float32:
   auto dtype = x.dtype();
   assert(dtype == mx::float32);
+  (void)dtype;
 
   // Specify the dtype when constructing the array:
   x = mx::array(1, mx::int32);

mlx/array.cpp

@@ -44,11 +44,11 @@ std::vector<array> array::make_arrays(
     const std::shared_ptr<Primitive>& primitive,
     const std::vector<array>& inputs) {
   std::vector<array> outputs;
-  for (size_t i = 0; i < shapes.size(); ++i) {
+  for (int i = 0; i < std::ssize(shapes); ++i) {
     outputs.emplace_back(std::move(shapes[i]), dtypes[i], primitive, inputs);
   }
   // For each node in |outputs|, its siblings are the other nodes.
-  for (size_t i = 0; i < outputs.size(); ++i) {
+  for (int i = 0; i < std::ssize(outputs); ++i) {
     auto siblings = outputs;
     siblings.erase(siblings.begin() + i);
     outputs[i].set_siblings(std::move(siblings), i);
@@ -145,8 +145,9 @@ void array::set_data(allocator::Buffer buffer, Deleter d) {
   array_desc_->data_size = size();
   array_desc_->flags.contiguous = true;
   array_desc_->flags.row_contiguous = true;
-  auto max_dim = std::max_element(shape().begin(), shape().end());
-  array_desc_->flags.col_contiguous = size() <= 1 || size() == *max_dim;
+  auto max_dim =
+      static_cast<int64_t>(*std::max_element(shape().begin(), shape().end()));
+  array_desc_->flags.col_contiguous = size() <= 1 || size() == max_dim;
 }
 
 void array::set_data(
@@ -192,7 +193,7 @@ array::~array() {
   }
 
   // Break circular reference for non-detached arrays with siblings
-  if (auto n = siblings().size(); n > 0) {
+  if (auto n = std::ssize(siblings()); n > 0) {
     bool do_detach = true;
     // If all siblings have siblings.size() references except
     // the one we are currently destroying (which has siblings.size() + 1)
@@ -241,8 +242,8 @@ array::ArrayDesc::ArrayDesc(
     std::vector<array> inputs)
     : shape(std::move(shape)),
       dtype(dtype),
-      status(Status::unscheduled),
       primitive(std::move(primitive)),
+      status(Status::unscheduled),
       inputs(std::move(inputs)) {
   init();
 }
@@ -274,7 +275,7 @@ array::ArrayDesc::~ArrayDesc() {
     ad.inputs.clear();
     for (auto& [_, a] : input_map) {
       bool is_deletable =
-          (a.array_desc_.use_count() <= a.siblings().size() + 1);
+          (a.array_desc_.use_count() <= std::ssize(a.siblings()) + 1);
       // An array with siblings is deletable only if all of its siblings
       // are deletable
       for (auto& s : a.siblings()) {
@@ -283,7 +284,7 @@ array::ArrayDesc::~ArrayDesc() {
         }
         int is_input = (input_map.find(s.id()) != input_map.end());
         is_deletable &=
-            s.array_desc_.use_count() <= a.siblings().size() + is_input;
+            s.array_desc_.use_count() <= std::ssize(a.siblings()) + is_input;
       }
       if (is_deletable) {
         for_deletion.push_back(std::move(a.array_desc_));

mlx/array.h

@@ -81,22 +81,22 @@ class array {
   }
 
   /** The size of the array's datatype in bytes. */
-  size_t itemsize() const {
+  int itemsize() const {
     return size_of(dtype());
   }
 
   /** The number of elements in the array. */
-  size_t size() const {
+  int64_t size() const {
     return array_desc_->size;
   }
 
   /** The number of bytes in the array. */
-  size_t nbytes() const {
+  int64_t nbytes() const {
     return size() * itemsize();
   }
 
   /** The number of dimensions of the array. */
-  size_t ndim() const {
+  int ndim() const {
     return array_desc_->shape.size();
   }
 
@@ -329,7 +329,7 @@ class array {
    * corresponding to ``arr[-1, -1, ...]``) then ``data_size = last - first``.
    * Note, ``data_size`` is in units of ``item_size`` (not bytes).
    **/
-  size_t data_size() const {
+  int64_t data_size() const {
     return array_desc_->data_size;
   }
 
@@ -340,7 +340,7 @@ class array {
     return array_desc_->data->buffer;
   }
 
-  size_t buffer_size() const {
+  int64_t buffer_size() const {
     return allocator::allocator().size(buffer());
   }
 
@@ -530,7 +530,7 @@ array::array(
     Shape shape,
     Dtype dtype /* = TypeToDtype<T>() */)
     : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
-  if (data.size() != size()) {
+  if (std::ssize(data) != size()) {
     throw std::invalid_argument(
         "Data size and provided shape mismatch in array construction.");
   }

mlx/backend/common/common.cpp

@@ -21,8 +21,8 @@ void AsStrided::eval(const std::vector<array>& inputs, array& out) {
 
   // Compute the flags given the shape and strides
   bool row_contiguous = true, col_contiguous = true;
-  size_t r = 1, c = 1;
-  for (int i = strides_.size() - 1, j = 0; i >= 0; i--, j++) {
+  int64_t r = 1, c = 1;
+  for (int i = std::ssize(strides_) - 1, j = 0; i >= 0; i--, j++) {
     row_contiguous &= (r == strides_[i]) || (shape_[i] == 1);
     col_contiguous &= (c == strides_[j]) || (shape_[j] == 1);
     r *= shape_[i];
@@ -60,7 +60,8 @@ void CustomTransforms::eval(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   assert(inputs.size() > outputs.size());
-  for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
+  for (int i = 0, j = std::ssize(inputs) - std::ssize(outputs);
+       i < std::ssize(outputs);
        i++, j++) {
     outputs[i].copy_shared_buffer(inputs[j]);
   }
@@ -70,7 +71,7 @@ void Depends::eval(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   assert(inputs.size() > outputs.size());
-  for (int i = 0; i < outputs.size(); i++) {
+  for (int i = 0; i < std::ssize(outputs); i++) {
     outputs[i].copy_shared_buffer(inputs[i]);
   }
 }
@@ -206,11 +207,11 @@ void Split::eval(
 
   auto compute_new_flags = [](const auto& shape,
                               const auto& strides,
-                              size_t in_data_size,
+                              int64_t in_data_size,
                               auto flags) {
-    size_t data_size = 1;
-    size_t f_stride = 1;
-    size_t b_stride = 1;
+    int64_t data_size = 1;
+    int64_t f_stride = 1;
+    int64_t b_stride = 1;
     flags.row_contiguous = true;
     flags.col_contiguous = true;
     for (int i = 0, ri = shape.size() - 1; ri >= 0; i++, ri--) {
@@ -240,7 +241,7 @@ void Split::eval(
 
   std::vector<int> indices(1, 0);
   indices.insert(indices.end(), indices_.begin(), indices_.end());
-  for (int i = 0; i < indices.size(); i++) {
+  for (int i = 0; i < std::ssize(indices); i++) {
     size_t offset = indices[i] * in.strides()[axis_];
     auto [new_flags, data_size] = compute_new_flags(
         outputs[i].shape(), in.strides(), in.data_size(), in.flags());
@@ -254,7 +255,7 @@ void Squeeze::eval(const std::vector<array>& inputs, array& out) {
   const auto& in = inputs[0];
   Strides strides;
   for (int i = 0, j = 0; i < in.ndim(); ++i) {
-    if (j < axes_.size() && i == axes_[j]) {
+    if (j < std::ssize(axes_) && i == axes_[j]) {
       j++;
     } else {
       strides.push_back(in.strides(i));
@@ -272,7 +273,7 @@ void Transpose::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   Strides out_strides(out.ndim());
   auto& in = inputs[0];
-  for (int ax = 0; ax < axes_.size(); ++ax) {
+  for (int ax = 0; ax < std::ssize(axes_); ++ax) {
     out_strides[ax] = in.strides()[axes_[ax]];
   }
 

mlx/backend/common/compiled.cpp

@@ -120,7 +120,7 @@ void compiled_allocate_outputs(
     Strides strides;
     size_t data_size;
     array::Flags flags;
-    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+    for (int i = 0; i < std::ssize(inputs) && o < std::ssize(outputs); ++i) {
       auto& in = inputs[i];
       // Conditions for donation
       // - Correct size
@@ -138,7 +138,7 @@ void compiled_allocate_outputs(
         data_size = in.data_size();
       }
     }
-    for (; o < outputs.size(); ++o) {
+    for (; o < std::ssize(outputs); ++o) {
       outputs[o].set_data(
           allocator::malloc(data_size * outputs[o].itemsize()),
           data_size,
@@ -147,7 +147,7 @@ void compiled_allocate_outputs(
     }
   } else {
     int o = 0;
-    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
+    for (int i = 0; i < std::ssize(inputs) && o < std::ssize(outputs); ++i) {
       auto& in = inputs[i];
       // Conditions for donation
       // - Row contiguous
@@ -162,7 +162,7 @@ void compiled_allocate_outputs(
         o++;
       }
     }
-    for (; o < outputs.size(); ++o) {
+    for (; o < std::ssize(outputs); ++o) {
       outputs[o].set_data(allocator::malloc(outputs[o].nbytes()));
     }
   }
@@ -193,7 +193,7 @@ std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(
 
     // Broadcast the inputs to the output shape.
     Strides xstrides;
-    size_t j = 0;
+    int j = 0;
     for (; j < shape.size() - x.ndim(); ++j) {
       if (shape[j] == 1) {
         xstrides.push_back(out.strides()[j]);
@@ -201,7 +201,7 @@ std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(
         xstrides.push_back(0);
       }
     }
-    for (size_t i = 0; i < x.ndim(); ++i, ++j) {
+    for (int i = 0; i < x.ndim(); ++i, ++j) {
       if (x.shape(i) == 1) {
         if (shape[j] == 1) {
           xstrides.push_back(out.strides()[j]);
@@ -224,13 +224,13 @@ bool compiled_use_large_index(
     const std::vector<array>& outputs,
     bool contiguous) {
   if (contiguous) {
-    size_t max_size = 0;
+    int64_t max_size = 0;
     for (const auto& in : inputs) {
       max_size = std::max(max_size, in.data_size());
     }
     return max_size > UINT32_MAX;
   } else {
-    size_t max_size = 0;
+    int64_t max_size = 0;
     for (const auto& o : outputs) {
       max_size = std::max(max_size, o.size());
     }

mlx/backend/common/load.cpp

@@ -27,7 +27,7 @@ void swap_endianness(uint8_t* data_bytes, size_t N) {
 
 namespace mlx::core {
 
-void Load::eval_cpu(const std::vector<array>& inputs, array& out) {
+void Load::eval_cpu(const std::vector<array>& /* inputs */, array& out) {
   out.set_data(allocator::malloc(out.nbytes()));
   auto read_task = [out_ptr = out.data<char>(),
                     size = out.size(),

mlx/backend/common/matmul.h

@@ -13,7 +13,7 @@ inline std::tuple<Shape, Strides, Strides> collapse_batches(
     const array& a,
     const array& b) {
   if (a.ndim() == 2) {
-    return {{1}, {0}, {0}};
+    return {Shape{1}, Strides{0}, Strides{0}};
   }
 
   Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
@@ -38,7 +38,7 @@ inline std::tuple<Shape, Strides, Strides> collapse_batches(
 inline std::tuple<Shape, Strides, Strides, Strides>
 collapse_batches(const array& a, const array& b, const array& c) {
   if (a.ndim() == 2) {
-    return {{1}, {0}, {0}, {0}};
+    return {Shape{1}, Strides{0}, Strides{0}, Strides{0}};
   }
 
   Shape A_bshape{a.shape().begin(), a.shape().end() - 2};

mlx/backend/common/reduce.cpp

@@ -28,7 +28,7 @@ std::pair<Shape, Strides> shapes_without_reduction_axes(
 
 ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
   // The data is all there and we are reducing over everything
-  if (x.size() == x.data_size() && axes.size() == x.ndim() &&
+  if (x.size() == x.data_size() && std::ssize(axes) == x.ndim() &&
       x.flags().contiguous) {
     return ContiguousAllReduce;
   }
@@ -38,7 +38,7 @@ ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
     // Merge consecutive axes
     Shape shape = {x.shape(axes[0])};
     Strides strides = {x.strides()[axes[0]]};
-    for (int i = 1; i < axes.size(); i++) {
+    for (int i = 1; i < std::ssize(axes); i++) {
       if (axes[i] - 1 == axes[i - 1] && x.shape(axes[i]) > 1) {
         shape.back() *= x.shape(axes[i]);
         strides.back() = x.strides()[axes[i]];

mlx/backend/common/slicing.cpp

@@ -24,8 +24,8 @@ std::tuple<int64_t, Strides> prepare_slice(
 void shared_buffer_slice(
     const array& in,
     const Strides& out_strides,
-    size_t data_offset,
-    size_t data_size,
+    int64_t data_offset,
+    int64_t data_size,
     array& out) {
   // Compute row/col contiguity
   auto [no_bsx_size, is_row_contiguous, is_col_contiguous] =
@@ -61,7 +61,7 @@ void slice(
   if (data_end < 0) {
     data_end += in.data_size();
   }
-  size_t data_size = (data_end - data_offset);
+  int64_t data_size = (data_end - data_offset);
   shared_buffer_slice(in, inp_strides, data_offset, data_size, out);
 }
 

mlx/backend/common/ternary.h

@@ -11,6 +11,8 @@ namespace mlx::core {
 enum class TernaryOpType {
   ScalarScalarScalar,
   VectorVectorVector,
+  VectorVectorScalar,
+  VectorScalarVector,
   General,
 };
 
@@ -25,6 +27,14 @@ get_ternary_op_type(const array& a, const array& b, const array& c) {
       (a.flags().col_contiguous && b.flags().col_contiguous &&
        c.flags().col_contiguous)) {
     topt = TernaryOpType::VectorVectorVector;
+  } else if (
+      b.data_size() == 1 && a.flags().row_contiguous &&
+      c.flags().row_contiguous) {
+    topt = TernaryOpType::VectorScalarVector;
+  } else if (
+      c.data_size() == 1 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    topt = TernaryOpType::VectorVectorScalar;
   } else {
     topt = TernaryOpType::General;
   }
@@ -59,6 +69,8 @@ inline void set_ternary_op_output_data(
             b.flags());
       }
       break;
+    case TernaryOpType::VectorVectorScalar:
+    case TernaryOpType::VectorScalarVector:
     case TernaryOpType::General:
       // Try to donate an input which is row_contiguous
       if (!((a.flags().row_contiguous && maybe_donate(a)) ||

mlx/backend/common/utils.cpp

@@ -28,7 +28,7 @@ std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
     if (shape[0] != 1) {
       to_collapse.push_back(0);
     }
-    size_t size = shape[0];
+    int64_t size = shape[0];
     for (int i = 1; i < shape.size(); i++) {
       bool contiguous = true;
       size *= shape[i];
@@ -64,7 +64,7 @@ std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
       current_shape *= shape[to_collapse[k]];
     }
     out_shape.push_back(current_shape);
-    for (int j = 0; j < strides.size(); j++) {
+    for (int j = 0; j < std::ssize(strides); j++) {
       const auto& st = strides[j];
       out_strides[j].push_back(st[to_collapse[k - 1]]);
     }
@@ -228,31 +228,4 @@ std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
       std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
 }
 
-array swapaxes_in_eval(const array& x, int axis1, int axis2) {
-  int ndim = x.ndim();
-  if (axis1 < 0) {
-    axis1 += ndim;
-  }
-  if (axis2 < 0) {
-    axis2 += ndim;
-  }
-
-  auto shape = x.shape();
-  std::swap(shape[axis1], shape[axis2]);
-  auto strides = x.strides();
-  std::swap(strides[axis1], strides[axis2]);
-
-  auto [data_size, row_contiguous, col_contiguous] =
-      check_contiguity(shape, strides);
-  bool contiguous = data_size == x.data_size();
-
-  array out(std::move(shape), x.dtype(), nullptr, {});
-  out.copy_shared_buffer(
-      x,
-      std::move(strides),
-      {contiguous, row_contiguous, col_contiguous},
-      x.data_size());
-  return out;
-}
-
 } // namespace mlx::core

mlx/backend/common/utils.h

@@ -162,7 +162,7 @@ struct ContiguousIterator {
 };
 
 inline auto check_contiguity(const Shape& shape, const Strides& strides) {
-  size_t no_broadcast_data_size = 1;
+  int64_t no_broadcast_data_size = 1;
   int64_t f_stride = 1;
   int64_t b_stride = 1;
   bool is_row_contiguous = true;
@@ -183,7 +183,7 @@ inline auto check_contiguity(const Shape& shape, const Strides& strides) {
 }
 
 inline bool is_donatable(const array& in, const array& out) {
-  constexpr size_t donation_extra = 16384;
+  constexpr int64_t donation_extra = 16384;
 
   return in.is_donatable() && in.itemsize() == out.itemsize() &&
       in.buffer_size() <= out.nbytes() + donation_extra;
@@ -196,9 +196,6 @@ 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/arange.h

@@ -10,7 +10,7 @@ namespace mlx::core {
 namespace {
 
 template <typename T>
-void arange(T start, T next, array& out, size_t size, Stream stream) {
+void arange(T start, T next, array& out, int64_t size, Stream stream) {
   auto ptr = out.data<T>();
   auto step_size = next - start;
   auto& encoder = cpu::get_command_encoder(stream);

mlx/backend/cpu/arg_reduce.cpp

@@ -19,12 +19,12 @@ void arg_reduce(const array& in, array& out, const OpT& op, int axis) {
   auto in_ptr = in.data<InT>();
   auto out_ptr = out.data<uint32_t>();
 
-  for (uint32_t i = 0; i < out.size(); ++i) {
+  for (int64_t i = 0; i < out.size(); ++i) {
     auto loc = elem_to_loc(i, shape, strides);
     auto local_in_ptr = in_ptr + loc;
     uint32_t ind_v = 0;
     InT v = (*local_in_ptr);
-    for (uint32_t j = 0; j < axis_size; ++j, local_in_ptr += axis_stride) {
+    for (int64_t j = 0; j < axis_size; ++j, local_in_ptr += axis_stride) {
       op(j, (*local_in_ptr), &ind_v, &v);
     }
     out_ptr[i] = ind_v;

mlx/backend/cpu/binary.cpp

@@ -17,7 +17,12 @@ namespace mlx::core {
 namespace {
 
 template <typename Op>
-void binary(const array& a, const array& b, array& out, Op op, Stream stream) {
+void binary(
+    const array& a,
+    const array& b,
+    array& out,
+    Op /* op */,
+    Stream stream) {
   auto bopt = get_binary_op_type(a, b);
   set_binary_op_output_data(a, b, out, bopt);
 
@@ -81,7 +86,7 @@ void comparison_op(
     const array& a,
     const array& b,
     array& out,
-    Op op,
+    Op /* op */,
     Stream stream) {
   auto bopt = get_binary_op_type(a, b);
   set_binary_op_output_data(a, b, out, bopt);
@@ -146,7 +151,7 @@ void binary_float(
     const array& a,
     const array& b,
     array& out,
-    Op op,
+    Op /* op */,
     Stream stream) {
   auto bopt = get_binary_op_type(a, b);
   set_binary_op_output_data(a, b, out, bopt);
@@ -187,7 +192,7 @@ void binary_int(
     const array& a,
     const array& b,
     array& out,
-    Op op,
+    Op /* op */,
     Stream stream) {
   auto bopt = get_binary_op_type(a, b);
   set_binary_op_output_data(a, b, out, bopt);

mlx/backend/cpu/binary_two.h

@@ -99,7 +99,7 @@ void binary_op_dispatch_dims(
   ContiguousIterator a_it(shape, a_strides, ndim - 2);
   ContiguousIterator b_it(shape, b_strides, ndim - 2);
   auto stride = out_strides[ndim - 3];
-  for (size_t elem = 0; elem < a.size(); elem += stride) {
+  for (int64_t elem = 0; elem < std::ssize(a); elem += stride) {
     binary_op_dims<T, U, Op, 2>(
         a_ptr + a_it.loc,
         b_ptr + b_it.loc,
@@ -137,21 +137,21 @@ void binary_op(
   if (bopt == BinaryOpType::ScalarScalar) {
     std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
   } else if (bopt == BinaryOpType::ScalarVector) {
-    for (size_t i = 0; i < b.data_size(); ++i) {
+    for (int64_t i = 0; i < b.data_size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
       b_ptr++;
     }
   } else if (bopt == BinaryOpType::VectorScalar) {
-    for (size_t i = 0; i < a.data_size(); ++i) {
+    for (int64_t i = 0; i < a.data_size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;
       a_ptr++;
     }
   } else { // VectorVector
-    for (size_t i = 0; i < a.size(); ++i) {
+    for (int64_t i = 0; i < a.size(); ++i) {
       std::tie(*out_a_ptr, *out_b_ptr) = op(*a_ptr, *b_ptr);
       out_a_ptr++;
       out_b_ptr++;

mlx/backend/cpu/cholesky.cpp

@@ -33,8 +33,8 @@ void cholesky_impl(const array& a, array& factor, bool upper, Stream stream) {
                     N = a.shape(-1),
                     size = a.size()]() mutable {
     char uplo = (upper) ? 'L' : 'U';
-    size_t num_matrices = size / (N * N);
-    for (int i = 0; i < num_matrices; i++) {
+    int64_t num_matrices = size / (N * N);
+    for (int64_t i = 0; i < num_matrices; i++) {
       // Compute Cholesky factorization.
       int info;
       potrf<T>(

mlx/backend/cpu/compiled.cpp

@@ -15,6 +15,7 @@
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
+#include "mlx/version.h"
 
 namespace mlx::core {
 
@@ -48,7 +49,7 @@ static CompilerCache& cache() {
 // GPU compile is always available if the GPU is available and since we are in
 // this file CPU compile is also available.
 namespace detail {
-bool compile_available_for_device(const Device& device) {
+bool compile_available_for_device(const Device& /* device */) {
   return true;
 }
 
@@ -94,7 +95,11 @@ void* compile(
     kernel_file_name = kernel_name;
   }
 
-  auto output_dir = std::filesystem::temp_directory_path();
+  auto output_dir =
+      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();
@@ -157,11 +162,13 @@ inline void build_kernel(
 #endif
 
   // Start the kernel
-  os << "void " << kernel_name << "(void** args) {" << std::endl;
+  os << "void " << kernel_name
+     << "(int* shape, int64_t** strides, void** args) {" << std::endl;
 
   // Add the input arguments
   int cnt = 0;
-  for (size_t i = 0; i < inputs.size(); ++i) {
+  int strides_index = 1;
+  for (int i = 0; i < std::ssize(inputs); ++i) {
     // Skip constants from the input list
     if (is_constant(i)) {
       continue;
@@ -175,8 +182,8 @@ inline void build_kernel(
        << "];" << std::endl;
     // Scalars and contiguous need no strides
     if (!is_scalar(x) && !contiguous) {
-      os << "  const size_t* " << xname << "_strides = (size_t*)args[" << cnt++
-         << "];" << std::endl;
+      os << "  const int64_t* " << xname << "_strides = strides["
+         << strides_index++ << "];" << std::endl;
     }
   }
 
@@ -186,10 +193,8 @@ inline void build_kernel(
     os << "  " << tstr << "* " << namer.get_name(x) << " = (" << tstr
        << "*)args[" << cnt++ << "];" << std::endl;
   }
-  // Add output strides and shape to extract the indices.
-  if (!contiguous) {
-    os << "  const int* shape = (int*)args[" << cnt++ << "];" << std::endl;
-  } else {
+  // Add output size
+  if (contiguous) {
     os << "  const size_t size = (size_t)args[" << cnt++ << "];" << std::endl;
   }
 
@@ -233,7 +238,7 @@ inline void build_kernel(
     } else {
       os << x.primitive().name();
       os << "()(";
-      for (int i = 0; i < x.inputs().size() - 1; i++) {
+      for (int i = 0; i < std::ssize(x.inputs()) - 1; i++) {
         os << "tmp_" << namer.get_name(x.inputs()[i]) << ", ";
       }
       os << "tmp_" << namer.get_name(x.inputs().back()) << ");" << std::endl;
@@ -288,17 +293,8 @@ 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;
@@ -306,9 +302,6 @@ 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
@@ -343,16 +336,20 @@ void Compiled::eval_cpu(
     args.push_back(x.data<void>());
     encoder.set_output_array(x);
   }
-  if (!contiguous) {
-    args.push_back((void*)shape.data());
-  } else {
+  if (contiguous) {
     args.push_back((void*)outputs[0].data_size());
   }
-  auto fun = (void (*)(void**))fn_ptr;
+  auto fun = reinterpret_cast<void (*)(int*, int64_t**, void**)>(fn_ptr);
   encoder.dispatch([fun,
                     args = std::move(args),
                     strides = std::move(strides),
-                    shape = std::move(shape)]() mutable { fun(args.data()); });
+                    shape = std::move(shape)]() mutable {
+    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/conv.cpp

@@ -860,7 +860,7 @@ void explicit_gemm_conv_1D_cpu(
     const std::vector<int>& padding_lo,
     const std::vector<int>& padding_hi,
     const std::vector<int>& wt_strides,
-    const std::vector<int>& wt_dilation,
+    const std::vector<int>& /* wt_dilation */,
     Stream stream) {
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
   const int iH = in.shape(1); // Input spatial dim
@@ -996,131 +996,6 @@ void explicit_gemm_conv_1D_cpu(
   encoder.add_temporaries(std::move(temps));
 }
 
-void explicit_gemm_conv_2D_cpu(
-    const array& in,
-    const array& wt,
-    array out,
-    const std::vector<int>& padding_lo,
-    const std::vector<int>& padding_hi,
-    const std::vector<int>& wt_strides,
-    const std::vector<int>& wt_dilation,
-    Stream stream) {
-  const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
-  const int iH = in.shape(1); // Input spatial dim
-  const int iW = in.shape(2); // Input spatial dim
-  const int oH = out.shape(1); // Output spatial dim
-  const int oW = out.shape(2); // Output spatial dim
-  const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(3); // In channels
-  const int wH = wt.shape(1); // Weight spatial dim
-  const int wW = wt.shape(2); // Weight spatial dim
-
-  auto conv_dtype = out.dtype();
-  auto& encoder = cpu::get_command_encoder(stream);
-
-  // Pad input
-  Shape padded_shape = {
-      N,
-      iH + padding_lo[0] + padding_hi[0],
-      iW + padding_lo[1] + padding_hi[1],
-      C};
-  array in_padded(padded_shape, conv_dtype, nullptr, {});
-
-  // Fill with zeros
-  std::vector<array> temps;
-  temps.push_back(array(0, conv_dtype));
-  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
-
-  // Pick input slice from padded
-  size_t data_offset = padding_lo[0] * in_padded.strides()[1] +
-      padding_lo[1] * in_padded.strides()[2];
-  array in_padded_slice(in.shape(), in_padded.dtype(), nullptr, {});
-  in_padded_slice.copy_shared_buffer(
-      in_padded,
-      in_padded.strides(),
-      in_padded.flags(),
-      in_padded_slice.size(),
-      data_offset);
-  temps.push_back(in_padded_slice);
-
-  // Copy input values into the slice
-  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
-
-  // Make strided view
-  Shape strided_shape = {N, oH, oW, wH, wW, C};
-
-  Strides strided_strides = {
-      in_padded.strides()[0],
-      in_padded.strides()[1] * wt_strides[0],
-      in_padded.strides()[2] * wt_strides[1],
-      in_padded.strides()[1],
-      in_padded.strides()[2],
-      in_padded.strides()[3]};
-  auto flags = in_padded.flags();
-
-  array in_strided_view(strided_shape, in_padded.dtype(), nullptr, {});
-  in_strided_view.copy_shared_buffer(
-      in_padded, strided_strides, flags, in_strided_view.size(), 0);
-
-  // Materialize strided view
-  Shape strided_reshape = {N * oH * oW, wH * wW * C};
-  array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
-  temps.push_back(in_strided);
-
-  // Check wt dtype and prepare
-  auto gemm_wt = wt;
-  auto gemm_out = out;
-
-  if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
-    auto ctype =
-        wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-    gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy_cpu(wt, gemm_wt, ctype, stream);
-    temps.push_back(gemm_wt);
-  }
-
-  if (out.dtype() != float32) {
-    gemm_out = array(out.shape(), float32, nullptr, {});
-    gemm_out.set_data(allocator::malloc(gemm_out.nbytes()));
-    temps.push_back(gemm_out);
-  }
-
-  encoder.set_input_array(in_strided);
-  encoder.set_input_array(gemm_wt);
-  encoder.set_output_array(gemm_out);
-
-  encoder.dispatch([in_strided_ptr = in_strided.data<float>(),
-                    gemm_wt_ptr = gemm_wt.data<float>(),
-                    gemm_out_ptr = gemm_out.data<float>(),
-                    strided_reshape = std::move(strided_reshape),
-                    O]() {
-    // Perform gemm
-    cblas_sgemm(
-        CblasRowMajor,
-        CblasNoTrans, // no trans A
-        CblasTrans, // transB
-        strided_reshape[0], // M
-        O, // N
-        strided_reshape[1], // K
-        1.0f, // alpha
-        in_strided_ptr,
-        strided_reshape[1], // lda
-        gemm_wt_ptr,
-        strided_reshape[1], // ldb
-        0.0f, // beta
-        gemm_out_ptr,
-        O // ldc
-    );
-  });
-
-  // Copy results if needed
-  if (out.dtype() != float32) {
-    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
-  }
-  encoder.add_temporaries(std::move(temps));
-}
-
 void explicit_gemm_conv_ND_cpu(
     const array& in,
     const array& wt,
@@ -1128,7 +1003,7 @@ void explicit_gemm_conv_ND_cpu(
     const std::vector<int>& padding_lo,
     const std::vector<int>& padding_hi,
     const std::vector<int>& wt_strides,
-    const std::vector<int>& wt_dilation,
+    const std::vector<int>& /* wt_dilation */,
     const bool flip,
     Stream stream) {
   const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
@@ -1148,7 +1023,7 @@ void explicit_gemm_conv_ND_cpu(
   // Pad input
   Shape padded_shape(in.shape().size());
   padded_shape.front() = N;
-  for (size_t i = 0; i < iDim.size(); i++) {
+  for (int i = 0; i < iDim.size(); i++) {
     padded_shape[i + 1] = iDim[i] + padding_lo[i] + padding_hi[i];
   }
   padded_shape.back() = C;
@@ -1179,20 +1054,20 @@ void explicit_gemm_conv_ND_cpu(
   // Make strided view
   Shape strided_shape(oDim.size() + wDim.size() + 2);
   strided_shape.front() = N;
-  for (size_t i = 0; i < oDim.size(); i++) {
+  for (int i = 0; i < oDim.size(); i++) {
     strided_shape[i + 1] = oDim[i];
   }
-  for (size_t i = 0; i < wDim.size(); i++) {
+  for (int i = 0; i < wDim.size(); i++) {
     strided_shape[i + 1 + oDim.size()] = wDim[i];
   }
   strided_shape.back() = C;
 
   Strides strided_strides(in.shape().size() * 2 - 2);
   strided_strides[0] = in_padded.strides()[0];
-  for (size_t i = 0; i < wt_strides.size(); i++) {
+  for (int i = 0; i < std::ssize(wt_strides); i++) {
     strided_strides[i + 1] = in_padded.strides()[i + 1] * wt_strides[i];
   }
-  for (size_t i = 1; i < in_padded.strides().size(); i++) {
+  for (int i = 1; i < std::ssize(in_padded.strides()); i++) {
     strided_strides[i + wt_strides.size()] = in_padded.strides()[i];
   }
 

mlx/backend/cpu/distributed.cpp

@@ -90,6 +90,7 @@ void Recv::eval_cpu(
     std::vector<array>& outputs) {
   assert(inputs.size() == 0);
   assert(outputs.size() == 1);
+  (void)inputs;
 
   outputs[0].set_data(allocator::malloc(outputs[0].nbytes()));
   distributed::detail::recv(group(), outputs[0], src_, stream());

mlx/backend/cpu/eig.cpp

@@ -46,7 +46,6 @@ void eig_impl(
     int info;
     {
       T work;
-      int iwork;
       geev<T>(
           &jobl,
           &jobr,
@@ -71,7 +70,7 @@ void eig_impl(
     auto eig_tmp = static_cast<T*>(eig_tmp_data.buffer.raw_ptr());
     auto vec_tmp = static_cast<T*>(vec_tmp_data.buffer.raw_ptr());
     auto work_buf = array::Data{allocator::malloc(sizeof(T) * lwork)};
-    for (size_t i = 0; i < size / (N * N); ++i) {
+    for (int64_t i = 0; i < size / (N * N); ++i) {
       geev<T>(
           &jobl,
           &jobr,

mlx/backend/cpu/eigh.cpp

@@ -165,7 +165,7 @@ void eigh_impl(
     EighWork<T> work(jobz, uplo, N);
 
     // Work loop
-    for (size_t i = 0; i < size / (N * N); ++i) {
+    for (int64_t i = 0; i < size / (N * N); ++i) {
       work.run(vec_ptr, eig_ptr);
       vec_ptr += N * N;
       eig_ptr += N;

mlx/backend/cpu/encoder.h

@@ -20,8 +20,8 @@ struct CommandEncoder {
   CommandEncoder(CommandEncoder&&) = delete;
   CommandEncoder& operator=(CommandEncoder&&) = delete;
 
-  void set_input_array(const array& a) {}
-  void set_output_array(array& a) {}
+  void set_input_array(const array& /* a */) {}
+  void set_output_array(array& /* a */) {}
 
   // Hold onto a temporary until any already scheduled tasks which use it as
   // an input are complete.

mlx/backend/cpu/gemm.h

@@ -12,12 +12,12 @@ void matmul(
     T* out,
     bool a_transposed,
     bool b_transposed,
-    size_t lda,
-    size_t ldb,
-    size_t ldc,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
     float alpha,
     float beta,
-    size_t batch_size,
+    int64_t batch_size,
     const Shape& a_shape,
     const Strides& a_strides,
     const Shape& b_shape,

mlx/backend/cpu/gemms/bnns.cpp

@@ -1,5 +1,4 @@
 // Copyright © 2023-2024 Apple Inc.
-
 #include <Accelerate/Accelerate.h>
 
 #include "mlx/array.h"
@@ -35,7 +34,7 @@ void matmul_bnns(
     bool b_transposed,
     size_t lda,
     size_t ldb,
-    size_t ldc,
+    size_t /* ldc */,
     float alpha,
     float beta,
     size_t batch_size,
@@ -49,9 +48,15 @@ void matmul_bnns(
   size_t K = a_shape[ndim - 1];
 
   BNNSDataType bnns_dtype = to_bnns_dtype<T>();
-
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+  if (beta != 1.0 && beta != 0.0) {
+    // scale the output
+    for (size_t i = 0; i < batch_size * M * N; ++i) {
+      out[i] *= beta;
+    }
+    beta = 1.0;
+  }
   const BNNSLayerParametersBroadcastMatMul gemm_params{
       /* float alpha = */ alpha,
       /* float beta = */ beta,
@@ -122,7 +127,7 @@ void matmul_bnns(
   auto bnns_filter =
       BNNSFilterCreateLayerBroadcastMatMul(&gemm_params, nullptr);
 
-  for (int i = 0; i < batch_size; ++i) {
+  for (size_t i = 0; i < batch_size; ++i) {
     BNNSFilterApplyTwoInput(
         bnns_filter,
         reinterpret_cast<const uint8_t*>(
@@ -143,12 +148,12 @@ void matmul<float16_t>(
     float16_t* out,
     bool a_transposed,
     bool b_transposed,
-    size_t lda,
-    size_t ldb,
-    size_t ldc,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
     float alpha,
     float beta,
-    size_t batch_size,
+    int64_t batch_size,
     const Shape& a_shape,
     const Strides& a_strides,
     const Shape& b_shape,
@@ -178,12 +183,12 @@ void matmul<bfloat16_t>(
     bfloat16_t* out,
     bool a_transposed,
     bool b_transposed,
-    size_t lda,
-    size_t ldb,
-    size_t ldc,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
     float alpha,
     float beta,
-    size_t batch_size,
+    int64_t batch_size,
     const Shape& a_shape,
     const Strides& a_strides,
     const Shape& b_shape,

mlx/backend/cpu/gemms/cblas.cpp

@@ -13,20 +13,20 @@ void matmul<float>(
     float* out,
     bool a_transposed,
     bool b_transposed,
-    size_t lda,
-    size_t ldb,
-    size_t ldc,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
     float alpha,
     float beta,
-    size_t batch_size,
+    int64_t batch_size,
     const Shape& a_shape,
     const Strides& a_strides,
     const Shape& b_shape,
     const Strides& b_strides) {
   auto ndim = a_shape.size();
-  size_t M = a_shape[ndim - 2];
-  size_t N = b_shape[ndim - 1];
-  size_t K = a_shape[ndim - 1];
+  int64_t M = a_shape[ndim - 2];
+  int64_t N = b_shape[ndim - 1];
+  int64_t K = a_shape[ndim - 1];
 
   for (int i = 0; i < batch_size; ++i) {
     cblas_sgemm(
@@ -54,20 +54,20 @@ void matmul<double>(
     double* out,
     bool a_transposed,
     bool b_transposed,
-    size_t lda,
-    size_t ldb,
-    size_t ldc,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
     float alpha,
     float beta,
-    size_t batch_size,
+    int64_t batch_size,
     const Shape& a_shape,
     const Strides& a_strides,
     const Shape& b_shape,
     const Strides& b_strides) {
   auto ndim = a_shape.size();
-  size_t M = a_shape[ndim - 2];
-  size_t N = b_shape[ndim - 1];
-  size_t K = a_shape[ndim - 1];
+  int64_t M = a_shape[ndim - 2];
+  int64_t N = b_shape[ndim - 1];
+  int64_t K = a_shape[ndim - 1];
 
   for (int i = 0; i < batch_size; ++i) {
     cblas_dgemm(
@@ -88,4 +88,47 @@ void matmul<double>(
   }
 }
 
+template <>
+void matmul<complex64_t>(
+    const complex64_t* a,
+    const complex64_t* b,
+    complex64_t* out,
+    bool a_transposed,
+    bool b_transposed,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    float alpha,
+    float beta,
+    int64_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  int64_t M = a_shape[ndim - 2];
+  int64_t N = b_shape[ndim - 1];
+  int64_t K = a_shape[ndim - 1];
+  auto calpha = static_cast<complex64_t>(alpha);
+  auto cbeta = static_cast<complex64_t>(beta);
+
+  for (int i = 0; i < batch_size; ++i) {
+    cblas_cgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        &calpha,
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
+        lda,
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
+        ldb,
+        &cbeta,
+        out + M * N * i,
+        ldc);
+  }
+}
+
 } // namespace mlx::core

mlx/backend/cpu/hadamard.cpp

@@ -11,9 +11,9 @@ namespace mlx::core {
 
 // n = 2^k component
 template <typename T>
-void hadamard_n(T* out, int n, int m, float scale, size_t size) {
+void hadamard_n(T* out, int n, int /* m */, float scale, int64_t size) {
   for (int b = 0; b < size / n; b++) {
-    size_t loc = b * n;
+    int64_t loc = b * n;
     T* data_ptr = out + loc;
     int h = 1;
     int n_over_2 = n / 2;
@@ -37,7 +37,7 @@ void hadamard_n(T* out, int n, int m, float scale, size_t size) {
 
 // m component
 template <typename T>
-void hadamard_m(T* out, int n, int m, float scale, size_t size) {
+void hadamard_m(T* out, int n, int m, float scale, int64_t size) {
   auto h_matrices = hadamard_matrices();
   auto& matrix = h_matrices[m];
   auto start = 1;
@@ -45,7 +45,7 @@ void hadamard_m(T* out, int n, int m, float scale, size_t size) {
   std::vector<bool> hmat_vec;
   while (end != std::string_view::npos) {
     auto row = matrix.substr(start, end - start);
-    for (int i = 0; i < row.length(); i++) {
+    for (int i = 0; i < std::ssize(row); i++) {
       hmat_vec.push_back(row[i] == '+');
     }
     start = end + 1;
@@ -53,7 +53,7 @@ void hadamard_m(T* out, int n, int m, float scale, size_t size) {
   }
 
   for (int b = 0; b < size / m / n; b++) {
-    size_t loc = b * n * m;
+    int64_t loc = b * n * m;
     T* data_ptr = out + loc;
     for (int i = 0; i < n; i++) {
       std::vector<float> out(m);

mlx/backend/cpu/indexing.cpp

@@ -78,7 +78,7 @@ void gather(
     can_copy = true;
 
     // Ignore leading 1s
-    int i = 0;
+    int64_t i = 0;
     for (; i < slice_sizes.size() && slice_sizes[i] == 1; ++i)
       ;
 
@@ -91,7 +91,7 @@ void gather(
     can_copy = true;
 
     // Ignore trailing 1s
-    int i = slice_sizes.size() - 1;
+    int64_t i = slice_sizes.size() - 1;
     for (; i >= 0 && slice_sizes[i] == 1; --i)
       ;
 
@@ -101,11 +101,11 @@ void gather(
       can_copy = (src.shape(i) == slice_sizes[i]);
     }
   }
-  size_t slice_size = 1;
+  int64_t slice_size = 1;
   for (auto s : slice_sizes) {
     slice_size *= s;
   }
-  size_t ind_size = slice_size == 0 ? 0 : out.size() / slice_size;
+  int64_t ind_size = slice_size == 0 ? 0 : out.size() / slice_size;
   const T* src_ptr = src.data<T>();
   T* dst_ptr = out.data<T>();
 
@@ -115,10 +115,10 @@ void gather(
     src_it = ContiguousIterator(slice_sizes, src.strides(), src.ndim());
   }
 
-  size_t out_idx = 0;
-  for (int idx = 0; idx < ind_size; idx++) {
-    size_t src_idx = 0;
-    for (int ii = 0; ii < inds.size(); ++ii) {
+  int64_t out_idx = 0;
+  for (int64_t idx = 0; idx < ind_size; idx++) {
+    int64_t src_idx = 0;
+    for (int ii = 0; ii < std::ssize(inds); ++ii) {
       auto ax = axes[ii];
       auto idx_loc = its[ii].loc;
       its[ii].step();
@@ -134,7 +134,7 @@ void gather(
           src_ptr + src_idx, src_ptr + src_idx + slice_size, dst_ptr + out_idx);
       out_idx += slice_size;
     } else {
-      for (int jj = 0; jj < slice_size; jj++) {
+      for (int64_t jj = 0; jj < slice_size; jj++) {
         dst_ptr[out_idx++] = src_ptr[src_idx + src_it.loc];
         src_it.step();
       }
@@ -403,11 +403,11 @@ void scatter(
     const std::vector<int>& axes) {
   int nind = inds.size();
   auto inds_ndim = updates.ndim() - out.ndim();
-  size_t n_updates = nind ? inds[0].size() : 1;
+  int64_t n_updates = nind ? inds[0].size() : 1;
 
   Shape update_shape(
       updates.shape().begin() + inds_ndim, updates.shape().end());
-  size_t update_size = 1;
+  int64_t update_size = 1;
   for (auto us : update_shape) {
     update_size *= us;
   }
@@ -418,9 +418,9 @@ void scatter(
 
   auto out_ptr = out.data<InT>();
   auto upd_ptr = updates.data<InT>();
-  for (int i = 0; i < n_updates; ++i) {
-    size_t out_offset = 0;
-    for (int j = 0; j < inds.size(); ++j) {
+  for (int64_t i = 0; i < n_updates; ++i) {
+    int64_t out_offset = 0;
+    for (int j = 0; j < std::ssize(inds); ++j) {
       auto ax = axes[j];
       auto idx_loc = its[j].loc;
       its[j].step();
@@ -429,7 +429,7 @@ void scatter(
       out_offset += (idx_val * out.strides()[ax]);
     }
     update_it.seek(i * update_size);
-    for (int j = 0; j < update_size; ++j) {
+    for (int64_t j = 0; j < update_size; ++j) {
       OpT{}(upd_ptr[update_it.loc], out_ptr + out_offset + out_it.loc);
       update_it.step();
       out_it.step();

mlx/backend/cpu/inverse.cpp

@@ -122,7 +122,7 @@ void inverse_impl(
       stream);
 
   const int N = a.shape(-1);
-  const size_t num_matrices = a.size() / (N * N);
+  const int64_t num_matrices = a.size() / (N * N);
 
   auto& encoder = cpu::get_command_encoder(stream);
   encoder.set_output_array(inv);
@@ -130,13 +130,13 @@ void inverse_impl(
   auto inv_ptr = inv.data<T>();
   if (tri) {
     encoder.dispatch([inv_ptr, N, num_matrices, upper]() {
-      for (int i = 0; i < num_matrices; i++) {
+      for (int64_t i = 0; i < num_matrices; i++) {
         tri_inv<T>(inv_ptr + N * N * i, N, upper);
       }
     });
   } else {
     encoder.dispatch([inv_ptr, N, num_matrices]() {
-      for (int i = 0; i < num_matrices; i++) {
+      for (int64_t i = 0; i < num_matrices; i++) {
         general_inv<T>(inv_ptr + N * N * i, N);
       }
     });

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(gesvdx)
+INSTANTIATE_LAPACK_REAL(gesdd)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)

mlx/backend/cpu/masked_mm.cpp

@@ -25,7 +25,7 @@ inline void mask_matrix(
     const int64_t Y_data_str,
     const int64_t X_mask_str,
     const int64_t Y_mask_str,
-    const size_t mask_offset) {
+    const int64_t mask_offset) {
   int tX = (X + block_size - 1) / block_size;
   int tY = (Y + block_size - 1) / block_size;
 
@@ -61,13 +61,13 @@ inline void segmented_mm(
     T* out,
     bool a_transposed,
     bool b_transposed,
-    size_t lda,
-    size_t ldb,
+    int64_t lda,
+    int64_t ldb,
     const Shape& a_shape,
     const Strides& a_strides,
     const Shape& b_shape,
     const Strides& b_strides,
-    size_t num_segments,
+    int64_t num_segments,
     const Shape& segments_shape,
     const Strides& segments_strides) {
   int ndim = a_shape.size();
@@ -149,9 +149,9 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto [b_transposed, ldb, b, b_copied] =
       check_transpose(b_pre, has_op_mask, inputs.back().dtype() != bool_);
 
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
+  int64_t M = a.shape(-2);
+  int64_t N = b.shape(-1);
+  int64_t K = a.shape(-1);
 
   if (M == 0 || N == 0) {
     return;
@@ -172,8 +172,8 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
                        int batch_idx,
                        int X,
                        int Y,
-                       size_t X_data_str,
-                       size_t Y_data_str,
+                       int64_t X_data_str,
+                       int64_t Y_data_str,
                        const Shape& mask_shape,
                        const Strides& mask_strides,
                        bool is_bool) {
@@ -215,18 +215,18 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   encoder.set_input_array(a);
   encoder.set_input_array(b);
-  const void* a_mask_ptr;
-  const void* b_mask_ptr;
-  const void* out_mask_ptr;
+  const void* a_mask_ptr = nullptr;
+  const void* b_mask_ptr = nullptr;
+  const void* out_mask_ptr = nullptr;
   Shape a_mask_shape;
   Shape b_mask_shape;
   Shape out_mask_shape;
   Strides a_mask_strides;
   Strides b_mask_strides;
   Strides out_mask_strides;
-  bool a_mask_bool;
-  bool b_mask_bool;
-  bool out_mask_bool;
+  bool a_mask_bool = false;
+  bool b_mask_bool = false;
+  bool out_mask_bool = false;
   if (has_op_mask) {
     auto& a_mask = inputs[inputs.size() - 2];
     auto& b_mask = inputs[inputs.size() - 1];
@@ -253,7 +253,7 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto a_ptr = a.data<float>();
   auto b_ptr = b.data<float>();
   auto out_ptr = out.data<float>();
-  size_t num_matrices = out.size() / (M * size_t(N));
+  int64_t num_matrices = out.size() / (M * int64_t(N));
   auto ldc = out.shape(-1);
 
   encoder.dispatch([a_ptr,
@@ -394,9 +394,9 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto [a_transposed, lda, a] = check_transpose(a_pre);
   auto [b_transposed, ldb, b] = check_transpose(b_pre);
 
-  size_t M = a.shape(-2);
-  size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
+  int64_t M = a.shape(-2);
+  int64_t N = b.shape(-1);
+  int64_t K = a.shape(-1);
 
   if (M == 0 || N == 0) {
     return;
@@ -413,7 +413,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   // Get batch dims
   auto batch_size_out = out.size() / (M * N);
-  size_t matrix_stride_out = M * N;
+  int64_t matrix_stride_out = M * N;
 
   auto get_batch_dims = [](const auto& v) {
     return decltype(v){v.begin(), v.end() - 2};
@@ -423,7 +423,6 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& rhs_indices = inputs[3];
 
   auto batch_shape = get_batch_dims(out.shape());
-  int batch_ndim = batch_shape.size();
 
   auto batch_shape_A = get_batch_dims(a.shape());
   auto batch_strides_A = get_batch_dims(a.strides());

mlx/backend/cpu/matmul.cpp

@@ -91,7 +91,6 @@ void matmul_general(
   auto [b_transposed, ldb, b] = check_transpose(b_pre);
   size_t M = a.shape(-2);
   size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
   if (M == 0 || N == 0) {
     return;
   }
@@ -108,6 +107,9 @@ void matmul_general(
   } else if (out.dtype() == float64) {
     matmul_dispatch<double>(
         a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
+  } else if (out.dtype() == complex64) {
+    matmul_dispatch<complex64_t>(
+        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else {
     throw std::runtime_error("[Matmul::eval_cpu] Invalid type.");
   }
@@ -128,10 +130,6 @@ void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
 }
 
 void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[AddMM::eval_cpu] Currently only supports float32.");
-  }
   if (out.size() == 0) {
     out.set_data(allocator::malloc(out.nbytes()));
     return;

mlx/backend/cpu/primitives.cpp

@@ -48,7 +48,7 @@ static std::pair<array, bool> compute_dynamic_offset(
   auto compute_offset =
       [strides, axes, offset = offset.data<int64_t>()](const auto* indices) {
         int64_t offset_ = 0;
-        for (int i = 0; i < axes.size(); ++i) {
+        for (int i = 0; i < std::ssize(axes); ++i) {
           offset_ += indices[i] * strides[axes[i]];
         }
         offset[0] = offset_;
@@ -124,6 +124,7 @@ void Transpose::eval_cpu(const std::vector<array>& inputs, array& out) {
 
 void Arange::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 0);
+  (void)inputs;
   out.set_data(allocator::malloc(out.nbytes()));
   switch (out.dtype()) {
     case bool_:
@@ -193,9 +194,9 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
   flags.row_contiguous = false;
   flags.col_contiguous = false;
   flags.contiguous = false;
-  for (int i = 0; i < inputs.size(); i++) {
+  for (int i = 0; i < std::ssize(inputs); i++) {
     array out_slice(inputs[i].shape(), out.dtype(), nullptr, {});
-    size_t data_offset = strides[axis_] * sizes[i];
+    int64_t data_offset = strides[axis_] * sizes[i];
     out_slice.copy_shared_buffer(
         out, strides, flags, out_slice.size(), data_offset);
     copy_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
@@ -205,7 +206,7 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
 void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  constexpr size_t extra_bytes = 16384;
+  constexpr int64_t extra_bytes = 16384;
   if (in.buffer_size() <= out.nbytes() + extra_bytes &&
       (in.flags().row_contiguous ||
        (allow_col_major_ && in.flags().col_contiguous))) {
@@ -254,8 +255,8 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
   copy_cpu(val, out, CopyType::Scalar, stream());
 
   // Find offset for start of input values
-  size_t data_offset = 0;
-  for (int i = 0; i < axes_.size(); i++) {
+  int64_t data_offset = 0;
+  for (int i = 0; i < std::ssize(axes_); i++) {
     auto ax = axes_[i] < 0 ? out.ndim() + axes_[i] : axes_[i];
     data_offset += out.strides()[ax] * low_pad_size_[i];
   }
@@ -274,10 +275,10 @@ void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
   // keys has shape (N1, ..., NK, 2)
   // out has shape (N1, ..., NK, M1, M2, ...)
   auto& keys = inputs[0];
-  size_t num_keys = keys.size() / 2;
+  int64_t num_keys = keys.size() / 2;
 
-  size_t elems_per_key = out.size() / num_keys;
-  size_t bytes_per_key = out.itemsize() * elems_per_key;
+  int64_t elems_per_key = out.size() / num_keys;
+  int64_t bytes_per_key = out.itemsize() * elems_per_key;
   out.set_data(allocator::malloc(out.nbytes()));
 
   auto kptr = inputs[0].data<uint32_t>();
@@ -291,8 +292,8 @@ void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
                     num_keys,
                     kshape = keys.shape(),
                     kstrides = keys.strides()]() mutable {
-    size_t out_skip = (bytes_per_key + 4 - 1) / 4;
-    auto half_size = out_skip / 2;
+    int64_t out_skip = (bytes_per_key + 4 - 1) / 4;
+    uintptr_t half_size = out_skip / 2;
     bool even = out_skip % 2 == 0;
     for (int i = 0; i < num_keys; ++i, cptr += bytes_per_key) {
       auto ptr = reinterpret_cast<uint32_t*>(cptr);

mlx/backend/cpu/qrf.cpp

@@ -13,7 +13,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
   const int M = a.shape(-2);
   const int N = a.shape(-1);
   const int lda = M;
-  size_t num_matrices = a.size() / (M * N);
+  int64_t num_matrices = a.size() / (M * N);
 
   // Copy A to inplace input and make it col-contiguous
   array in(a.shape(), a.dtype(), nullptr, {});
@@ -54,7 +54,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
     auto work = allocator::malloc(sizeof(T) * lwork);
 
     // Loop over matrices
-    for (int i = 0; i < num_matrices; ++i) {
+    for (int64_t i = 0; i < num_matrices; ++i) {
       // Solve
       geqrf<T>(
           &M,
@@ -68,7 +68,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
     }
     allocator::free(work);
 
-    for (int i = 0; i < num_matrices; ++i) {
+    for (int64_t i = 0; i < num_matrices; ++i) {
       /// num_reflectors x N
       for (int j = 0; j < num_reflectors; ++j) {
         for (int k = 0; k < j; ++k) {
@@ -97,7 +97,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
     work = allocator::malloc(sizeof(T) * lwork);
 
     // Loop over matrices
-    for (int i = 0; i < num_matrices; ++i) {
+    for (int64_t i = 0; i < num_matrices; ++i) {
       // Compute Q
       orgqr<T>(
           &M,
@@ -111,7 +111,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
           &info);
     }
 
-    for (int i = 0; i < num_matrices; ++i) {
+    for (int64_t i = 0; i < num_matrices; ++i) {
       // M x num_reflectors
       for (int j = 0; j < M; ++j) {
         for (int k = 0; k < num_reflectors; ++k) {

mlx/backend/cpu/quantized.cpp

@@ -1,7 +1,5 @@
 // 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"
@@ -13,6 +11,35 @@ 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);
 }
@@ -407,6 +434,229 @@ 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 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 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,
@@ -513,115 +763,198 @@ 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];
 
-  std::vector<array> temps;
-  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      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_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),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous(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),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      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& biases_pre = inputs[3];
-  auto& lhs_indices = inputs[4];
-  auto& rhs_indices = inputs[5];
+  auto& lhs_indices = inputs[inputs.size() - 2];
+  auto& rhs_indices = inputs[inputs.size() - 1];
 
-  std::vector<array> temps;
+  auto& encoder = cpu::get_command_encoder(stream());
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &temps](const array& arr) {
+                                          &encoder](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 {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      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);
-  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),
-                    biases = array::unsafe_weak_copy(biases),
-                    lhs_indices = array::unsafe_weak_copy(lhs_indices),
-                    rhs_indices = array::unsafe_weak_copy(rhs_indices),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _bs_qmm_dispatch(
-        out,
-        x,
-        w,
-        scales,
-        biases,
-        lhs_indices,
-        rhs_indices,
-        group_size_,
-        bits_,
-        transpose_);
-  });
+  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),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _bs_qmm_dispatch(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          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),
+                      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>
@@ -705,7 +1038,7 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::AffineQuantize::eval_cpu(
+void fast::Quantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto ensure_row_contiguous = [s = stream()](const array& arr) {
@@ -764,7 +1097,7 @@ void fast::AffineQuantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
     }
   });
 }

mlx/backend/cpu/simd/accelerate_simd.h

@@ -9,7 +9,7 @@
 
 #include "mlx/backend/cpu/simd/base_simd.h"
 
-// There seems to be a bug in sims/base.h
+// There seems to be a bug in simd/base_simd.h
 // __XROS_2_0 is not defined, the expression evaluates
 // to true instead of false setting the SIMD library
 // higher than it should be even on macOS < 15
@@ -234,6 +234,7 @@ 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) {
@@ -251,9 +252,13 @@ 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;
-    while (any(exp)) {
-      res = select(exp & 1, res * base, res);
-      base = select(exp, base * base, base);
+    // Raising an integer to a negative power is undefined
+    if (any(exp < static_cast<T>(0))) {
+      return 0;
+    }
+    while (any(exp > static_cast<T>(0))) {
+      res = select((exp & 1) != 0, res * base, res);
+      base = select(exp > static_cast<T>(0), base * base, base);
       exp = exp >> 1;
     }
     return res;

mlx/backend/cpu/simd/math.h

@@ -79,7 +79,8 @@ Simd<T, N> sincos(Simd<T, N> in) {
 
   // Get the polynom selection mask. There is one polynom for 0 <= x <= Pi/4
   // and another one for Pi/4<x<=Pi/2. Both branches will be computed.
-  auto poly_mask = (emm2 & 2) != 0;
+  auto poly_mask =
+      (emm2 & static_cast<uint32_t>(2)) != static_cast<uint32_t>(0);
 
   // The magic pass: "Extended precision modular arithmetic"
   // x = ((x - y * DP1) - y * DP2) - y * DP3
@@ -87,8 +88,8 @@ Simd<T, N> sincos(Simd<T, N> in) {
   x = fma(y, Simd<float, N>(-2.4187564849853515625e-4f), x);
   x = fma(y, Simd<float, N>(-3.77489497744594108e-8f), x);
 
-  sign_mask_sin = sign_mask_sin ^ ((emm2 & 4) != 0);
-  auto sign_mask_cos = ((emm2 - 2) & 4) != 0;
+  sign_mask_sin = sign_mask_sin ^ ((emm2 & 4) != static_cast<uint32_t>(0));
+  auto sign_mask_cos = ((emm2 - 2) & 4) != static_cast<uint32_t>(0);
 
   // Evaluate the first polynom  (0 <= x <= Pi/4) in y1,
   // and the second polynom      (Pi/4 <= x <= 0) in y2

mlx/backend/cpu/sort.cpp

@@ -15,6 +15,18 @@ namespace mlx::core {
 
 namespace {
 
+// NaN-aware comparator that places NaNs at the end
+template <typename T>
+bool nan_aware_less(T a, T b) {
+  if constexpr (std::is_floating_point_v<T> || std::is_same_v<T, complex64_t>) {
+    if (std::isnan(a))
+      return false;
+    if (std::isnan(b))
+      return true;
+  }
+  return a < b;
+}
+
 template <typename T>
 struct StridedIterator {
   using iterator_category = std::random_access_iterator_tag;
@@ -27,7 +39,7 @@ struct StridedIterator {
   StridedIterator() = default;
 
   explicit StridedIterator(T* ptr, int64_t stride, difference_type offset = 0)
-      : ptr_(ptr + offset * stride), stride_(stride) {}
+      : stride_(stride), ptr_(ptr + offset * stride) {}
 
   explicit StridedIterator(array& arr, int axis, difference_type offset = 0)
       : StridedIterator(arr.data<T>(), arr.strides()[axis], offset) {}
@@ -108,8 +120,8 @@ template <typename T>
 void sort(array& out, int axis) {
   // Get axis, shape and stride info
   axis = axis < 0 ? axis + out.ndim() : axis;
-  size_t in_size = out.size();
-  size_t n_rows = in_size / out.shape(axis);
+  int64_t in_size = out.size();
+  int64_t n_rows = in_size / out.shape(axis);
 
   auto remaining_shape = out.shape();
   remaining_shape.erase(remaining_shape.begin() + axis);
@@ -124,13 +136,13 @@ void sort(array& out, int axis) {
   ContiguousIterator src_it(
       remaining_shape, remaining_strides, remaining_shape.size());
   auto out_ptr = out.data<T>();
-  for (int i = 0; i < n_rows; i++) {
+  for (int64_t i = 0; i < n_rows; i++) {
     T* data_ptr = out_ptr + src_it.loc;
 
     StridedIterator st(data_ptr, axis_stride, 0);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::stable_sort(st, ed);
+    std::stable_sort(st, ed, nan_aware_less<T>);
     src_it.step();
   }
 }
@@ -139,7 +151,7 @@ template <typename T, typename IdxT = uint32_t>
 void argsort(const array& in, array& out, int axis) {
   // Get axis, shape and stride info
   axis = axis < 0 ? axis + in.ndim() : axis;
-  size_t n_rows = in.size() / in.shape(axis);
+  int64_t n_rows = in.size() / in.shape(axis);
 
   auto in_remaining_shape = in.shape();
   in_remaining_shape.erase(in_remaining_shape.begin() + axis);
@@ -164,7 +176,7 @@ void argsort(const array& in, array& out, int axis) {
       out_remaining_shape, out_remaining_strides, out_remaining_shape.size());
   auto in_ptr = in.data<T>();
   auto out_ptr = out.data<IdxT>();
-  for (int i = 0; i < n_rows; i++) {
+  for (int64_t i = 0; i < n_rows; i++) {
     const T* data_ptr = in_ptr + in_it.loc;
     IdxT* idx_ptr = out_ptr + out_it.loc;
 
@@ -184,6 +196,15 @@ void argsort(const array& in, array& out, int axis) {
     std::stable_sort(st, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
+
+      // Handle NaNs (place them at the end)
+      if (std::is_floating_point<T>::value) {
+        if (std::isnan(v1))
+          return false;
+        if (std::isnan(v2))
+          return true;
+      }
+
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }
@@ -193,8 +214,8 @@ template <typename T>
 void partition(array& out, int axis, int kth) {
   // Get axis, shape and stride info
   axis = axis < 0 ? axis + out.ndim() : axis;
-  size_t in_size = out.size();
-  size_t n_rows = in_size / out.shape(axis);
+  int64_t in_size = out.size();
+  int64_t n_rows = in_size / out.shape(axis);
 
   auto remaining_shape = out.shape();
   remaining_shape.erase(remaining_shape.begin() + axis);
@@ -211,7 +232,7 @@ void partition(array& out, int axis, int kth) {
   ContiguousIterator src_it(
       remaining_shape, remaining_strides, remaining_shape.size());
   auto out_ptr = out.data<T>();
-  for (int i = 0; i < n_rows; i++) {
+  for (int64_t i = 0; i < n_rows; i++) {
     T* data_ptr = out_ptr + src_it.loc;
     src_it.step();
 
@@ -219,7 +240,7 @@ void partition(array& out, int axis, int kth) {
     StridedIterator md(data_ptr, axis_stride, kth);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::nth_element(st, md, ed);
+    std::nth_element(st, md, ed, nan_aware_less<T>);
   }
 }
 
@@ -227,7 +248,7 @@ template <typename T, typename IdxT = uint32_t>
 void argpartition(const array& in, array& out, int axis, int kth) {
   // Get axis, shape and stride info
   axis = axis < 0 ? axis + in.ndim() : axis;
-  size_t n_rows = in.size() / in.shape(axis);
+  int64_t n_rows = in.size() / in.shape(axis);
 
   auto in_remaining_shape = in.shape();
   in_remaining_shape.erase(in_remaining_shape.begin() + axis);
@@ -256,7 +277,7 @@ void argpartition(const array& in, array& out, int axis, int kth) {
   auto in_ptr = in.data<T>();
   auto out_ptr = out.data<IdxT>();
 
-  for (int i = 0; i < n_rows; i++) {
+  for (int64_t i = 0; i < n_rows; i++) {
     const T* data_ptr = in_ptr + in_it.loc;
     IdxT* idx_ptr = out_ptr + out_it.loc;
     in_it.step();
@@ -276,6 +297,15 @@ void argpartition(const array& in, array& out, int axis, int kth) {
     std::nth_element(st, md, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
+
+      // Handle NaNs (place them at the end)
+      if (std::is_floating_point<T>::value) {
+        if (std::isnan(v1))
+          return false;
+        if (std::isnan(v2))
+          return true;
+      }
+
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }

mlx/backend/cpu/svd.cpp

@@ -27,7 +27,7 @@ void svd_impl(
   const int N = a.shape(-1);
   const int K = std::min(M, N);
 
-  size_t num_matrices = a.size() / (M * N);
+  int64_t num_matrices = a.size() / (M * N);
 
   // lapack clobbers the input, so we have to make a copy.
   array in(a.shape(), a.dtype(), nullptr, {});
@@ -81,40 +81,26 @@ void svd_impl(
     // Vᵀ of shape N x N. (M x M in lapack).
     const int ldvt = M;
 
-    auto job_u = (u_ptr) ? "V" : "N";
-    auto job_vt = (u_ptr) ? "V" : "N";
-    static constexpr auto range = "A";
+    auto jobz = (u_ptr) ? "A" : "N";
 
-    // Will contain the number of singular values after the call has returned.
-    int ns = 0;
     T workspace_dimension = 0;
 
     // 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) * 12 * K)};
+    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
 
     static const int lwork_query = -1;
 
-    static const int ignored_int = 0;
-    static const T ignored_float = 0;
-
     int info;
 
     // Compute workspace size.
-    gesvdx<T>(
-        /* jobu = */ job_u,
-        /* jobvt = */ job_vt,
-        /* range = */ range,
+    gesdd<T>(
+        /* jobz = */ jobz,
         // 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,
@@ -135,21 +121,14 @@ void svd_impl(
     auto scratch = array::Data{allocator::malloc(sizeof(T) * lwork)};
 
     // Loop over matrices.
-    for (int i = 0; i < num_matrices; i++) {
-      gesvdx<T>(
-          /* jobu = */ job_u,
-          /* jobvt = */ job_vt,
-          /* range = */ range,
+    for (int64_t i = 0; i < num_matrices; i++) {
+      gesdd<T>(
+          /* jobz = */ jobz,
           // 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,
@@ -167,13 +146,6 @@ 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);
@@ -181,10 +153,10 @@ void svd_impl(
 
 template <typename T>
 void compute_svd(
-    const array& a,
-    bool compute_uv,
-    std::vector<array>& outputs,
-    Stream stream) {}
+    const array& /* a */,
+    bool /* compute_uv */,
+    std::vector<array>& /* outputs */,
+    Stream /* stream */) {}
 
 void SVD::eval_cpu(
     const std::vector<array>& inputs,

mlx/backend/cpu/ternary.h

@@ -136,7 +136,7 @@ void ternary_op(
   if (topt == TernaryOpType::ScalarScalarScalar) {
     *out_ptr = op(*a_ptr, *b_ptr, *c_ptr);
   } else if (topt == TernaryOpType::VectorVectorVector) {
-    for (size_t i = 0; i < out.size(); ++i) {
+    for (int64_t i = 0; i < out.size(); ++i) {
       *out_ptr = op(*a_ptr, *b_ptr, *c_ptr);
       a_ptr++;
       b_ptr++;

mlx/backend/cpu/unary.h

@@ -10,8 +10,8 @@
 namespace mlx::core {
 
 template <typename T, typename U = T, typename Op>
-void unary_op(const T* a, U* out, size_t shape, size_t stride) {
-  for (size_t i = 0; i < shape; i += 1) {
+void unary_op(const T* a, U* out, int64_t shape, int64_t stride) {
+  for (int64_t i = 0; i < shape; i += 1) {
     out[i] = Op{}(*a);
     a += stride;
   }
@@ -38,14 +38,14 @@ void unary_op(const array& a, array& out, Op) {
       src++;
     }
   } else {
-    size_t shape = ndim > 0 ? a.shape().back() : 1;
-    size_t stride = ndim > 0 ? a.strides().back() : 1;
+    int64_t shape = ndim > 0 ? a.shape().back() : 1;
+    int64_t stride = ndim > 0 ? a.strides().back() : 1;
     if (ndim <= 1) {
       unary_op<T, U, Op>(src, dst, shape, stride);
       return;
     }
     auto it = ContiguousIterator(a.shape(), a.strides(), ndim - 1);
-    for (size_t elem = 0; elem < a.size(); elem += shape) {
+    for (int64_t elem = 0; elem < a.size(); elem += shape) {
       unary_op<T, U, Op>(src + it.loc, dst + elem, shape, stride);
       it.step();
     }

mlx/backend/cpu/unary_ops.h

@@ -77,7 +77,8 @@ struct Real {
 struct Sigmoid {
   template <int N, typename T>
   Simd<T, N> operator()(Simd<T, N> x) {
-    return 1.0f / (1.0f + simd::exp(-x));
+    auto y = 1.0f / (1.0f + simd::exp(simd::abs(x)));
+    return simd::select(x < Simd<T, N>{0}, y, Simd<T, N>{1} - y);
   }
   SINGLE()
 };

mlx/backend/cuda/CMakeLists.txt

@@ -8,7 +8,6 @@ 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
@@ -17,14 +16,18 @@ 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
@@ -46,12 +49,14 @@ 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)
@@ -149,7 +154,7 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 FetchContent_Declare(
   cudnn
   GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
-  GIT_TAG v1.12.1
+  GIT_TAG v1.14.0
   GIT_SHALLOW TRUE
   EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)
@@ -165,6 +170,10 @@ target_link_libraries(mlx PRIVATE CUDNN::cudnn_all)
 # Suppress nvcc warnings on MLX headers.
 target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
                                    --diag_suppress=997>)
+# Supress warnings: note: parameter passing for argument of type
+# ‘std::pair<float, float>’ when C++17 is enabled changed to match C++14 in GCC
+# 10.1
+target_compile_options(mlx PRIVATE -Wno-psabi)
 
 # Install CCCL headers for JIT.
 install(DIRECTORY ${cccl_SOURCE_DIR}/include/cuda

mlx/backend/cuda/allocator.cpp

@@ -30,8 +30,20 @@ SmallSizePool::SmallSizePool() {
   next_free_ = buffer_;
 
   CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
-  CHECK_CUDA_ERROR(
-      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, 0));
+
+  int device_count = 0;
+  CHECK_CUDA_ERROR(cudaGetDeviceCount(&device_count));
+  for (int i = 0; i < device_count; ++i) {
+#if CUDART_VERSION >= 13000
+    cudaMemLocation loc;
+    loc.type = cudaMemLocationTypeDevice;
+    loc.id = i;
+#else
+    int loc = i;
+#endif // CUDART_VERSION >= 13000
+    CHECK_CUDA_ERROR(
+        cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetAccessedBy, loc));
+  }
 
   auto curr = next_free_;
   for (size_t i = 1; i < num_blocks; ++i) {
@@ -79,7 +91,7 @@ CudaAllocator::CudaAllocator()
   // TODO: Set memory limit for multi-device.
   size_t free, total;
   CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
-  memory_limit_ = total * 0.8;
+  memory_limit_ = total * 0.95;
   max_pool_size_ = memory_limit_;
 }
 

mlx/backend/cuda/arange.cu

@@ -6,23 +6,33 @@
 #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 {
 
-template <typename T>
-struct Arange {
-  const T start;
-  const T step;
+namespace cg = cooperative_groups;
 
-  __device__ T operator()(uint32_t i) const {
-    return start + i * step;
+template <typename T, typename IdxT, int N_WRITES>
+__global__ void arange(T* out, IdxT size, T start, T 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
 
@@ -36,19 +46,23 @@ 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>;
-    CTYPE step =
-        static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
-    thrust::transform(
-        cu::thrust_policy(encoder.stream()),
-        thrust::counting_iterator<uint32_t>(0),
-        thrust::counting_iterator<uint32_t>(out.data_size()),
-        thrust::device_pointer_cast(out.data<OutType>()),
-        cu::Arange<OutType>{
-            static_cast<OutType>(start_), static_cast<OutType>(step)});
+    constexpr int N_WRITES = 16 / sizeof(OutType);
+    dispatch_bool(out.data_size() > INT32_MAX, [&](auto large) {
+      using IdxT = std::conditional_t<large(), int64_t, int32_t>;
+      auto [num_blocks, block_dims] = get_launch_args(out, large(), N_WRITES);
+      encoder.add_kernel_node(
+          cu::arange<OutType, IdxT, N_WRITES>,
+          num_blocks,
+          block_dims,
+          0,
+          out.data<OutType>(),
+          out.data_size(),
+          static_cast<CTYPE>(start_),
+          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));
+    });
   });
 }
 

mlx/backend/cuda/binary/CMakeLists.txt

@@ -0,0 +1,21 @@
+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

@@ -0,0 +1,7 @@
+// 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

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

mlx/backend/cuda/binary/binary.cuh

@@ -99,39 +99,89 @@ __global__ void binary_vv(const In* a, const In* b, Out* out, IdxT size) {
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+template <
+    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,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
-        index, shape.data(), a_strides.data(), b_strides.data());
-    out[index] = Op{}(a[a_idx], b[b_idx]);
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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());
+  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, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_g(
     const In* a,
     const In* b,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc(
-        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
-    out[index] = Op{}(a[a_idx], b[b_idx]);
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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,
+      shape.data(),
+      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>
@@ -209,39 +259,61 @@ 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 [num_blocks, block_dims] =
-                        get_launch_args(out, large());
+                    auto kernel = cu::binary_g_nd<
+                        Op,
+                        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(
-                        cu::binary_g_nd<
-                            Op,
-                            InType,
-                            OutType,
-                            IdxT,
-                            dims_constant()>,
-                        num_blocks,
+                        kernel,
+                        {num_blocks_x, num_blocks_y},
                         block_dims,
                         0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out.data<OutType>(),
-                        out.size(),
+                        rest,
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto [num_blocks, block_dims] = get_launch_args(out, large());
+                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT, 1>;
+                  if (work_per_thread == 4) {
+                    kernel = cu::binary_g<Op, InType, OutType, IdxT, 4>;
+                  }
                   encoder.add_kernel_node(
-                      cu::binary_g<Op, InType, OutType, IdxT>,
-                      num_blocks,
+                      kernel,
+                      {num_blocks_x, num_blocks_y},
                       block_dims,
                       0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out.data<OutType>(),
-                      out.size(),
+                      rest,
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),
@@ -304,54 +376,4 @@ 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/bitwise_binary.cu

@@ -0,0 +1,27 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,15 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -0,0 +1,7 @@
+// 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

@@ -127,45 +127,99 @@ binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
   }
 }
 
-template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+template <
+    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,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
     const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
-        index, shape.data(), a_strides.data(), b_strides.data());
-    auto out = Op{}(a[a_idx], b[b_idx]);
-    out_a[index] = out[0];
-    out_b[index] = out[1];
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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());
+  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, typename IdxT>
+template <typename Op, typename In, typename Out, typename IdxT, int N_READS>
 __global__ void binary_two_g(
     const In* a,
     const In* b,
     Out* out_a,
     Out* out_b,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides a_strides,
     const __grid_constant__ Strides b_strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [a_idx, b_idx] = elem_to_loc(
-        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
-    auto out = Op{}(a[a_idx], b[b_idx]);
-    out_a[index] = out[0];
-    out_b[index] = out[1];
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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,
+      shape.data(),
+      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_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>
@@ -225,42 +279,64 @@ 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 [num_blocks, block_dims] =
-                        get_launch_args(out_a, large());
+                    auto kernel = cu::binary_two_g_nd<
+                        Op,
+                        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(
-                        cu::binary_two_g_nd<
-                            Op,
-                            InType,
-                            OutType,
-                            IdxT,
-                            dims_constant()>,
-                        num_blocks,
+                        kernel,
+                        {num_blocks_x, num_blocks_y},
                         block_dims,
                         0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out_a.data<OutType>(),
                         out_b.data<OutType>(),
-                        out_a.size(),
+                        rest,
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
                         const_param<dims_constant()>(b_strides));
                   });
                 } else {
-                  auto [num_blocks, block_dims] =
-                      get_launch_args(out_a, large());
+                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 1>;
+                  if (work_per_thread == 4) {
+                    kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 4>;
+                  }
                   encoder.add_kernel_node(
-                      cu::binary_two_g<Op, InType, OutType, IdxT>,
-                      num_blocks,
+                      kernel,
+                      {num_blocks_x, num_blocks_y},
                       block_dims,
                       0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out_a.data<OutType>(),
                       out_b.data<OutType>(),
-                      out_a.size(),
+                      rest,
                       const_param(shape),
                       const_param(a_strides),
                       const_param(b_strides),

mlx/backend/cuda/compiled.cpp

@@ -267,7 +267,8 @@ void Compiled::eval_gpu(
       }
     }
 
-    return std::make_pair(std::move(builder.os), std::move(kernel_names));
+    return std::make_tuple(
+        false, std::move(builder.os), std::move(kernel_names));
   });
 
   // Collapse contiguous dims to route to a faster kernel if possible. Also
@@ -331,9 +332,9 @@ void Compiled::eval_gpu(
     encoder.set_output_array(out);
   }
 
-  auto kernel = mod.get_kernel(kernel_name);
+  auto [kernel, max_block_dims] = mod.get_kernel_and_dims(kernel_name);
   auto [num_blocks, block_dims] =
-      get_launch_args(outputs[0], large, work_per_thread);
+      get_launch_args(outputs[0], large, work_per_thread, max_block_dims);
   encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 

mlx/backend/cuda/conv.cpp

@@ -1,18 +1,12 @@
 // 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>
@@ -21,9 +15,6 @@ 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 \
@@ -31,6 +22,9 @@ 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;
@@ -50,203 +44,13 @@ struct ConvCacheKey {
 auto& conv_cache() {
   static LRUBytesKeyCache<
       ConvCacheKey,
-      std::pair<cudnnBackendDescriptorType_t, cudnn_frontend::ExecutionPlan>>
-      cache(/* capacity */ 128);
+      std::pair<
+          cudnnBackendDescriptorType_t,
+          std::optional<cudnn_frontend::ExecutionPlan>>>
+      cache("MLX_CUDA_CONV_CACHE_SIZE", /* default_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,
@@ -291,7 +95,7 @@ auto get_conv_op_settings(
   }
 }
 
-std::optional<cudnn_frontend::OperationGraph> build_op_graph(
+std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
     cu::CommandEncoder& encoder,
     cudnnBackendDescriptorType_t backend_type,
     Dtype dtype,
@@ -317,9 +121,9 @@ std::optional<cudnn_frontend::OperationGraph> build_op_graph(
                          .build();
 
     auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .setxDesc(build_tensor('x', x))
-                  .setwDesc(build_tensor('w', w))
-                  .setyDesc(build_tensor('y', y))
+                  .setxDesc(build_cudnn_tensor_nchw('x', x))
+                  .setwDesc(build_cudnn_tensor_nchw('w', w))
+                  .setyDesc(build_cudnn_tensor_nchw('y', y))
                   .setcDesc(conv_desc)
                   .build();
 
@@ -336,6 +140,42 @@ std::optional<cudnn_frontend::OperationGraph> build_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.
@@ -345,13 +185,14 @@ 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 = swapaxes_in_eval(wt, 0, -1);
+    wt = group_transpose(wt, groups, 0, 0, -1, s);
   } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    in = swapaxes_in_eval(in, 0, -1);
+    in = group_transpose(in, groups, -1, 0, -1, s);
     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.
@@ -444,12 +285,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   ConvCacheKey cache_key{
       encoder.device().cuda_device(),
       dtype_to_cudnn_type(dtype),
-      fixed_vector(in.shape()),
-      fixed_vector(wt.shape()),
-      fixed_vector(kernel_strides_),
-      fixed_vector(padding_lo_),
-      fixed_vector(padding_hi_),
-      fixed_vector(kernel_dilation_),
+      vector_key(in.shape()),
+      vector_key(wt.shape()),
+      vector_key(kernel_strides_),
+      vector_key(padding_lo_),
+      vector_key(padding_hi_),
+      vector_key(kernel_dilation_),
       groups_,
       flip_,
       get_alignment(in),
@@ -457,11 +298,29 @@ 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;
-    std::tie(in, wt, out) = prepare_args(encoder, backend_type, in, wt, out, s);
-    register_args(encoder, backend_type, in, wt, out, out_);
-    auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-    if (!execute_plan(encoder, plan, x, w, y)) {
-      throw std::runtime_error("[conv] Cached plan failed to execute.");
+    if (plan) {
+      // 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)) {
+        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;
   }
@@ -490,7 +349,7 @@ 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, s);
+        prepare_args(encoder, try_backend, in, wt, out, groups_, 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,
@@ -502,7 +361,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_hi_,
         kernel_dilation_,
         input_dilation_);
-    op_graph = build_op_graph(
+    op_graph = build_conv_op_graph(
         encoder,
         try_backend,
         dtype,
@@ -521,26 +380,38 @@ 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) {
+    // Find a plan for the graph and execute it.
+    auto plan = find_cudnn_plan_from_op_graph(
+        encoder.device().cudnn_handle(), backend_type, dtype, *op_graph);
+    if (plan) {
+      // Setup inputs and outputs.
+      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)) {
+        conv_cache().emplace(
+            cache_key, std::make_pair(backend_type, std::move(*plan)));
+        return;
+      }
+    }
   }
 
-  // Get ready to execute the graph.
-  register_args(encoder, backend_type, in, wt, out, out_);
-
-  // Try to run plans based on heuristics.
-  auto configs = get_engine_configs(backend_type, dtype, *op_graph);
-  auto tag = op_graph->getTag();
-  auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-    return;
-  }
-  // Then try fallback plans.
-  configs = get_engine_configs(backend_type, dtype, *op_graph);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-    return;
-  }
-  throw std::runtime_error("[conv] Unable to find a working engine.");
+  // 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

@@ -0,0 +1,126 @@
+// 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

@@ -0,0 +1,217 @@
+// 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

@@ -0,0 +1,231 @@
+// 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,
-    const std::optional<array>& dynamic_offset_in,
-    const std::optional<array>& dynamic_offset_out) {
+    std::optional<array> dynamic_offset_in,
+    std::optional<array> dynamic_offset_out) {
   if (out.size() == 0) {
     return;
   }
@@ -44,6 +44,16 @@ 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,
@@ -54,8 +64,8 @@ void copy_gpu_inplace(
             shape_collapsed,
             strides_vec[0],
             strides_vec[1],
-            dynamic_offset_in ? *dynamic_offset_in : array(0, int64),
-            dynamic_offset_out ? *dynamic_offset_out : array(0, int64));
+            *dynamic_offset_in,
+            *dynamic_offset_out);
       } else {
         copy_general(
             encoder,

mlx/backend/cuda/copy/copy_general.cu

@@ -10,37 +10,80 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM>
+template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
 __global__ void copy_gg_nd(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     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) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc_nd<NDIM>(
-        index, shape.data(), strides_in.data(), strides_out.data());
-    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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,
+      shape.data(),
+      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>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_gg(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
     const __grid_constant__ Strides strides_in,
     const __grid_constant__ Strides strides_out,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    auto [idx_in, idx_out] = elem_to_loc(
-        index, shape.data(), strides_in.data(), strides_out.data(), ndim);
-    out[idx_out] = CastOp<In, Out>{}(in[idx_in]);
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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,
+      shape.data(),
+      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
@@ -69,33 +112,52 @@ 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 [num_blocks, block_dims] =
-                    get_launch_args(data_size, shape, out.strides(), large());
+                auto kernel =
+                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 1>;
+                if (work_per_thread == 4) {
+                  kernel =
+                      cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant(), 4>;
+                }
                 encoder.add_kernel_node(
-                    cu::copy_gg_nd<InType, OutType, IdxT, ndim_constant()>,
-                    num_blocks,
+                    kernel,
+                    {num_blocks_x, num_blocks_y},
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    data_size,
+                    rest,
                     const_param<ndim_constant()>(shape),
                     const_param<ndim_constant()>(strides_in),
                     const_param<ndim_constant()>(strides_out));
               });
             } else { // ndim >= 4
-              auto [num_blocks, block_dims] =
-                  get_launch_args(data_size, shape, out.strides(), large());
+              auto kernel = cu::copy_gg<InType, OutType, IdxT, 1>;
+              if (work_per_thread == 4) {
+                kernel = cu::copy_gg<InType, OutType, IdxT, 4>;
+              }
               encoder.add_kernel_node(
-                  cu::copy_gg<InType, OutType, IdxT>,
-                  num_blocks,
+                  kernel,
+                  {num_blocks_x, num_blocks_y},
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  data_size,
+                  rest,
                   const_param(shape),
                   const_param(strides_in),
                   const_param(strides_out),

mlx/backend/cuda/copy/copy_general_input.cu

@@ -10,33 +10,67 @@ namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <typename In, typename Out, typename IdxT, int NDIM>
+template <typename In, typename Out, typename IdxT, int NDIM, int N_READS>
 __global__ void copy_g_nd(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
-    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides_in) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    IdxT idx_in = elem_to_loc_nd<NDIM>(index, shape.data(), strides_in.data());
-    out[index] = CastOp<In, Out>{}(in[idx_in]);
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> strides) {
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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 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>
+template <typename In, typename Out, typename IdxT, int N_READS>
 __global__ void copy_g(
     const In* in,
     Out* out,
-    IdxT size,
+    IdxT size_rest,
     const __grid_constant__ Shape shape,
-    const __grid_constant__ Strides strides_in,
+    const __grid_constant__ Strides strides,
     int ndim) {
-  IdxT index = cg::this_grid().thread_rank();
-  if (index < size) {
-    IdxT idx_in = elem_to_loc(index, shape.data(), strides_in.data(), ndim);
-    out[index] = CastOp<In, Out>{}(in[idx_in]);
+  auto block = cg::this_thread_block();
+  auto grid = cg::this_grid();
+  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 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
@@ -61,30 +95,49 @@ 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 [num_blocks, block_dims] = get_launch_args(out, large());
+                auto kernel =
+                    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(
-                    cu::copy_g_nd<InType, OutType, IdxT, dims_constant()>,
-                    num_blocks,
+                    kernel,
+                    {num_blocks_x, num_blocks_y},
                     block_dims,
                     0,
                     in_ptr,
                     out_ptr,
-                    out.size(),
+                    rest,
                     const_param<dims_constant()>(shape),
                     const_param<dims_constant()>(strides_in));
               });
             } else { // ndim >= 4
-              auto [num_blocks, block_dims] = get_launch_args(out, large());
+              auto kernel = cu::copy_g<InType, OutType, IdxT, 1>;
+              if (work_per_thread == 4) {
+                kernel = cu::copy_g<InType, OutType, IdxT, 4>;
+              }
               encoder.add_kernel_node(
-                  cu::copy_g<InType, OutType, IdxT>,
-                  num_blocks,
+                  kernel,
+                  {num_blocks_x, num_blocks_y},
                   block_dims,
                   0,
                   in_ptr,
                   out_ptr,
-                  out.size(),
+                  rest,
                   const_param(shape),
                   const_param(strides_in),
                   ndim);

mlx/backend/cuda/cudnn_utils.cpp

@@ -0,0 +1,275 @@
+// 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);
+  if (engine_configs.empty()) {
+    return std::nullopt;
+  }
+  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

@@ -0,0 +1,164 @@
+// 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

@@ -0,0 +1,379 @@
+// 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

@@ -14,10 +14,6 @@ namespace mlx::core::cu {
 
 namespace {
 
-// Can be tuned with MLX_MAX_OPS_PER_BUFFER
-// This should be less than 255
-constexpr int default_max_nodes_per_graph = 20;
-
 #define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
 
 void check_cudnn_error(const char* name, cudnnStatus_t err) {
@@ -27,11 +23,11 @@ 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", 100);
+bool use_cuda_graphs() {
+  static bool use_graphs = []() {
+    return env::get_var("MLX_USE_CUDA_GRAPHS", true);
   }();
-  return cache_size;
+  return use_graphs;
 }
 
 } // namespace
@@ -68,8 +64,8 @@ Device::~Device() {
 
 void Device::make_current() {
   // We need to set/get current CUDA device very frequently, cache it to reduce
-  // actual calls of CUDA APIs. This function assumes single-thread in host.
-  static int current = 0;
+  // actual calls of CUDA APIs.
+  static thread_local int current = 0;
   if (current != device_) {
     CHECK_CUDA_ERROR(cudaSetDevice(device_));
     current = device_;
@@ -86,14 +82,20 @@ 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() {
-  CHECK_CUDA_ERROR(cudaStreamEndCapture(enc.stream(), &graph));
-  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
-      &graph, [](cudaGraph_t* p) { CHECK_CUDA_ERROR(cudaGraphDestroy(*p)); });
+  if (!use_cuda_graphs()) {
+    enc.node_count_++;
+    return;
+  }
+
+  graph.end_capture(enc.stream());
   if (discard) {
     return;
   }
@@ -107,6 +109,9 @@ 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_++)};
@@ -185,37 +190,46 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
 }
 
 CommandEncoder::CommandEncoder(Device& d)
-    : device_(d), stream_(d), graph_cache_(cuda_graph_cache_size()) {
-  CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
-}
+    : device_(d),
+      stream_(d),
+      graph_(d),
+      worker_(d),
+      graph_cache_("MLX_CUDA_GRAPH_CACHE_SIZE", /* default_capacity */ 400) {}
 
 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);
 }
 
-void CommandEncoder::maybe_commit() {
-  if (node_count_ >= env::max_ops_per_buffer(default_max_nodes_per_graph)) {
-    commit();
-  }
-}
-
 void CommandEncoder::add_kernel_node(
     void* func,
     dim3 grid_dim,
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
+  if (!use_cuda_graphs()) {
+    node_count_++;
+    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;
@@ -231,6 +245,23 @@ void CommandEncoder::add_kernel_node(
     dim3 block_dim,
     uint32_t smem_bytes,
     void** params) {
+  if (!use_cuda_graphs()) {
+    node_count_++;
+    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;
@@ -257,20 +288,38 @@ void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
 }
 
 void CommandEncoder::add_graph_node(cudaGraph_t child) {
+  if (!use_cuda_graphs()) {
+    node_count_++;
+    CudaGraphExec graph_exec;
+    graph_exec.instantiate(child);
+    device_.make_current();
+    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream()));
+    return;
+  }
   cudaGraphNode_t node;
   CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
   insert_graph_dependencies(GraphNode{node, 'G'});
 }
 
+int CommandEncoder::get_num_ops() {
+  return node_count_;
+}
+
 void CommandEncoder::commit() {
   nvtx3::scoped_range r("CommandEncoder::commit");
   if (!temporaries_.empty()) {
     add_completed_handler([temporaries = std::move(temporaries_)]() {});
   }
-  if (node_count_ > 0) {
+  if (use_cuda_graphs() && node_count_ > 0) {
     if (!from_nodes_.empty()) {
       CHECK_CUDA_ERROR(cudaGraphAddDependencies(
-          graph_, from_nodes_.data(), to_nodes_.data(), from_nodes_.size()));
+          graph_,
+          from_nodes_.data(),
+          to_nodes_.data(),
+#if CUDART_VERSION >= 13000
+          nullptr, // edgeData
+#endif // CUDART_VERSION >= 13000
+          from_nodes_.size()));
     }
 
     graph_key_ += ".";
@@ -304,19 +353,18 @@ void CommandEncoder::commit() {
     CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
 
     // Reset state
-    node_count_ = 0;
     graph_node_count_ = 0;
     empty_node_count_ = 0;
     from_nodes_.clear();
     to_nodes_.clear();
     graph_key_.clear();
     node_map_.clear();
-    CHECK_CUDA_ERROR(cudaGraphDestroy(graph_));
-    CHECK_CUDA_ERROR(cudaGraphCreate(&graph_, 0));
+    graph_ = CudaGraph(device_);
   }
 
   // Put completion handlers in a batch.
   worker_.commit(stream_);
+  node_count_ = 0;
 }
 
 void CommandEncoder::synchronize() {

mlx/backend/cuda/device.h

@@ -21,7 +21,7 @@ class CommandEncoder {
   struct CaptureContext {
     CaptureContext(CommandEncoder& enc);
     ~CaptureContext();
-    cudaGraph_t graph;
+    CudaGraph graph;
     CommandEncoder& enc;
     bool discard{false};
   };
@@ -76,9 +76,6 @@ 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) {
@@ -86,7 +83,7 @@ class CommandEncoder {
   }
 
   void add_completed_handler(std::function<void()> task);
-  void maybe_commit();
+  int get_num_ops();
   void commit();
 
   Device& device() {
@@ -101,6 +98,9 @@ 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_t graph_;
+  CudaGraph graph_;
   Worker worker_;
   char node_count_{0};
   char graph_node_count_{0};
@@ -140,7 +140,7 @@ class Device {
   Device(const Device&) = delete;
   Device& operator=(const Device&) = delete;
 
-  // Make this device the current cuda device, required by some cuda calls.
+  // Make this device the current cuda device, this method is thread-safe.
   void make_current();
 
   CommandEncoder& get_command_encoder(Stream s);

mlx/backend/cuda/device/binary_ops.cuh

@@ -204,6 +204,12 @@ 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;