Install uv

There is no uvx
Revert "Save uv cache"
2025-12-16 01:49:05 +08:00 · 2025-07-30 15:27:00 +09:00 · 2025-07-30 15:20:40 +09:00 · 2025-07-30 15:17:00 +09:00 · 2025-07-30 15:15:32 +09:00 · 2025-07-30 15:02:49 +09:00
379 changed files with 5278 additions and 17429 deletions
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -18,17 +18,16 @@ jobs:
        type: boolean
        default: false
    macos:
-      xcode: "26.0.0"
-    resource_class: m4pro.medium
+      xcode: "16.2.0"
+    resource_class: m2pro.medium
    steps:
      - checkout
      - run:
          name: Install
          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            brew install python@3.10
+            brew install python@3.9
            brew install doxygen
-            python3.10 -m venv env
+            python3.9 -m venv env
            source env/bin/activate
            pip install --upgrade pip
            pip install --upgrade cmake
@@ -73,8 +72,10 @@ jobs:
      - run:
          name: Run style checks
          command: |
-            pip install pre-commit
-            pre-commit run --all
+            curl -LsSf https://astral.sh/uv/install.sh | sh
+            uv venv
+            uv pip install pre-commit
+            uv run --no-project pre-commit run --all
            if ! git diff --quiet; then echo 'Style checks failed, please install pre-commit and run pre-commit run --all and push the change'; exit 1; fi
      - run:
          name: Install dependencies
@@ -84,14 +85,11 @@ jobs:
            sudo apt-get update
            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-            curl -LsSf https://astral.sh/uv/install.sh | sh
      - run:
          name: Install Python package
          command: |
-            uv venv
            uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e ".[dev]" -v
+            uv pip install -e ".[dev]"
      - run:
          name: Generate package stubs
          command: |
@@ -120,7 +118,7 @@ jobs:
    parameters:
      xcode_version:
        type: string
-        default: "26.0.0"
+        default: "16.2.0"
      macosx_deployment_target:
        type: string
        default: ""
@@ -128,19 +126,15 @@ jobs:
      xcode: << parameters.xcode_version >>
    environment:
      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m4pro.medium
+    resource_class: m2pro.medium
    steps:
      - checkout
      - run:
          name: Install dependencies
          command: |
-            xcodebuild -downloadComponent MetalToolchain
            HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
              brew install openmpi uv
-      - run:
-          name: Install Python package
-          command: |
-            uv venv --python 3.10
+            uv venv --python 3.9
            uv pip install \
              nanobind==2.4.0 \
              cmake \
@@ -148,8 +142,11 @@ jobs:
              torch \
              tensorflow \
              unittest-xml-reporting
+      - run:
+          name: Install Python package
+          command: |
            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e . -v
+              uv pip install -e .
      - run:
          name: Generate package stubs
          command: |
@@ -167,11 +164,9 @@ jobs:
      - run:
          name: Build example extension
          command: |
-            source .venv/bin/activate
            cd examples/extensions
            uv pip install -r requirements.txt
-            uv run --no-project setup.py build_ext --inplace
-            uv run --no-project python test.py
+            uv run --no-project setup.py build_ext -j8
      - store_test_results:
          path: test-results
      - run:
@@ -199,7 +194,7 @@ jobs:
          name: Run Python tests with JIT
          command: |
            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e . -v
+              uv pip install -e .
            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
              METAL_DEBUG_ERROR_MODE=0 \
              uv run --no-project python -m xmlrunner discover \
@@ -225,44 +220,28 @@ 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
-            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
+            CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              uv pip install -e ".[dev]"
      - run:
          name: Run Python tests
          command: |
            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 }}
@@ -273,10 +252,10 @@ jobs:
    parameters:
      python_version:
        type: string
-        default: "3.10"
+        default: "3.9"
      xcode_version:
        type: string
-        default: "26.0.0"
+        default: "16.2.0"
      build_env:
        type: string
        default: ""
@@ -285,7 +264,7 @@ jobs:
        default: ""
    macos:
      xcode: << parameters.xcode_version >>
-    resource_class: m4pro.medium
+    resource_class: m2pro.medium
    environment:
      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
    steps:
@@ -293,15 +272,11 @@ jobs:
      - run:
          name: Install dependencies
          command: |
-            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
+            brew install python@<< parameters.python_version >>
+            brew install openmpi
+            python<< parameters.python_version >> -m venv env
+            source env/bin/activate
+            pip install --upgrade pip
            pip install --upgrade cmake
            pip install nanobind==2.4.0
            pip install --upgrade setuptools
@@ -311,29 +286,29 @@ jobs:
      - run:
          name: Install Python package
          command: |
-            conda activate env
+            source env/bin/activate
            env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
              pip install . -v
      - run:
          name: Generate package stubs
          command: |
-            conda activate env
+            source env/bin/activate
            pip install typing_extensions
            python setup.py generate_stubs
      - run:
          name: Build Python package
          command: |
-            conda activate env
+            source env/bin/activate
            python setup.py clean --all
            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
      - when:
          condition:
-            equal: ["3.10", << parameters.python_version >>]
+            equal: ["3.9", << parameters.python_version >>]
          steps:
            - run:
                name: Build common package
                command: |
-                  conda activate env
+                  source env/bin/activate
                  python setup.py clean --all
                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
      - when:
@@ -342,7 +317,7 @@ jobs:
            - run:
                name: Upload package
                command: |
-                  conda activate env
+                  source env/bin/activate
                  twine upload dist/*
      - store_artifacts:
          path: dist/
@@ -351,7 +326,7 @@ jobs:
    parameters:
      python_version:
        type: string
-        default: "3.10"
+        default: "3.9"
      build_env:
        type: string
        default: ""
@@ -387,7 +362,7 @@ jobs:
            bash python/scripts/repair_linux.sh
      - when:
          condition:
-            equal: ["3.10", << parameters.python_version >>]
+            equal: ["3.9", << parameters.python_version >>]
          steps:
            - run:
                name: Build common package
@@ -415,7 +390,7 @@ jobs:
        default: ""
    machine:
      image: ubuntu-2204:current
-      resource_class: xlarge
+      resource_class: large
    steps:
      - checkout
      - run:
@@ -462,7 +437,7 @@ workflows:
      - mac_build_and_test:
          matrix:
            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
      - linux_build_and_test
      - cuda_build_and_test:
          matrix:
@@ -484,10 +459,71 @@ workflows:
              ignore: /.*/
          matrix:
            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
              macosx_deployment_target: ["13.5", "14.0", "15.0"]
              build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["26.0.0"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "PYPI_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "PYPI_RELEASE=1"
      - build_documentation:
          filters:
            tags:
@@ -503,7 +539,7 @@ workflows:
              ignore: /.*/
          matrix:
            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
              build_env: ["PYPI_RELEASE=1"]
      - build_cuda_release:
          filters:
@@ -529,7 +565,7 @@ workflows:
          requires: [ hold ]
          matrix:
            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
+              macosx_deployment_target: ["13.5", "14.0"]
      - linux_build_and_test:
          requires: [ hold ]
      - cuda_build_and_test:
@@ -546,13 +582,59 @@ workflows:
      - build_release:
          matrix:
            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["26.0.0"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
      - build_linux_release:
          matrix:
            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
      - build_cuda_release

  build_dev_release:
@@ -564,14 +646,75 @@ workflows:
      - build_release:
          matrix:
            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
              macosx_deployment_target: ["13.5", "14.0", "15.0"]
              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["26.0.0"]
+              xcode_version: ["16.2.0", "15.0.0"]
+            exclude:
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "13.5"
+                xcode_version: "16.2.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "14.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.9"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.10"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.11"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.12"
+                build_env: "DEV_RELEASE=1"
+              - macosx_deployment_target: "15.0"
+                xcode_version: "15.0.0"
+                python_version: "3.13"
+                build_env: "DEV_RELEASE=1"
      - build_linux_release:
          matrix:
            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
              build_env: ["DEV_RELEASE=1"]
      - build_cuda_release:
          matrix:
--- a/ACKNOWLEDGMENTS.md
+++ b/ACKNOWLEDGMENTS.md
@@ -19,17 +19,12 @@ 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, and the `ReLU²` activation function.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.

 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
  <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>

-# 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
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -20,17 +20,12 @@ 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 20)
+set(CMAKE_CXX_STANDARD 17)
 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)
@@ -48,7 +43,6 @@ option(MLX_BUILD_BLAS_FROM_SOURCE "Build OpenBLAS from source code" OFF)
 option(MLX_METAL_JIT "Use JIT compilation for Metal kernels" OFF)
 option(MLX_USE_CCACHE "Use CCache for compilation cache when available" ON)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
-option(USE_SYSTEM_FMT "Use system's provided fmt library" OFF)

 # --------------------- Processor tests -------------------------
 message(
@@ -92,21 +86,22 @@ 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)
-  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_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_METAL_DEBUG)
    add_compile_definitions(MLX_METAL_DEBUG)
@@ -115,8 +110,7 @@ if(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
-    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
+    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)

  if(${MACOS_SDK_VERSION} LESS 14.0)
    message(
@@ -145,12 +139,6 @@ if(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.
@@ -178,7 +166,7 @@ if(MLX_BUILD_CPU)
    message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
    set(MLX_BUILD_ACCELERATE ON)
  else()
-    message(STATUS "Accelerate not found, using default backend.")
+    message(STATUS "Accelerate or arm neon not found, using default backend.")
    set(MLX_BUILD_ACCELERATE OFF)
  endif()

@@ -254,16 +242,12 @@ target_include_directories(
 # Do not add mlx_EXPORTS define for shared library.
 set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")

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

 if(MLX_BUILD_PYTHON_BINDINGS)
--- a/README.md
+++ b/README.md
@@ -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),
@@ -78,20 +78,20 @@ pip install mlx
 To install the CUDA backend on Linux, run:

 ```bash
-pip install mlx[cuda]
+pip install "mlx[cuda]"
 ```

 To install a CPU-only Linux package, run:

 ```bash
-pip install mlx[cpu]
+pip install "mlx[cpu]"
 ```

 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},
--- a/benchmarks/python/blas/bench_gemm.py
+++ b/benchmarks/python/blas/bench_gemm.py
@@ -142,7 +142,9 @@ 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_dtype) @ b_np.transpose(t_b).astype(np_dtype)
+    c_npy = a_np.transpose(t_a).astype(np.float32) @ b_np.transpose(t_b).astype(
+        np.float32
+    )

    atol = 1e-5 if np_dtype == np.float32 else 1e-4

@@ -161,7 +163,7 @@ def get_gflop_count(B, M, N, K):
 if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Run gemm benchmarks")

-    dtypes = ("float32", "float16", "complex64")
+    dtypes = ("float32", "float16")
    transposes = ("nn", "nt", "tn")
    shapes = (
        (16, 234, 768, 3072),
@@ -185,7 +187,7 @@ 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}, {100.0 * diff:+5.2f}%"
+                    f"{B:3d}, {M:4d}, {N:4d}, {K:4d}, {dtype}, {transpose}, {gflops_pt:05.3f}, {gflops_mx:05.3f}, {100. * diff:+5.2f}%"
                )
                if gflops_pt >= 2.0 * gflops_mx:
                    print("ATTENTION ^^^^^^^")
--- a/benchmarks/python/blas/bench_gemv.py
+++ b/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", "complex64"):
+    for dtype in ("float32", "float16"):
        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)
--- a/cmake/FindNCCL.cmake
+++ b/cmake/FindNCCL.cmake
@@ -1,54 +0,0 @@
-# FindNCCL.cmake This module finds the NVIDIA NCCL library and its include
-# directories.
-
-set(NCCL_ROOT_DIR
-    $ENV{NCCL_ROOT_DIR}
-    CACHE PATH "Folder contains NVIDIA NCCL")
-
-find_path(
-  NCCL_INCLUDE_DIRS
-  NAMES nccl.h
-  HINTS ${NCCL_INCLUDE_DIR} ${NCCL_ROOT_DIR} ${NCCL_ROOT_DIR}/include
-        ${CUDA_TOOLKIT_ROOT_DIR}/include)
-
-if($ENV{USE_STATIC_NCCL})
-  message(
-    STATUS "USE_STATIC_NCCL detected. Linking against static NCCL library")
-  set(NCCL_LIBNAME "libnccl_static.a")
-else()
-  set(NCCL_LIBNAME "nccl")
-endif()
-
-find_library(
-  NCCL_LIBRARIES
-  NAMES ${NCCL_LIBNAME}
-  HINTS ${NCCL_LIB_DIR}
-        ${NCCL_ROOT_DIR}
-        ${NCCL_ROOT_DIR}/lib
-        ${NCCL_ROOT_DIR}/lib/x86_64-linux-gnu
-        ${NCCL_ROOT_DIR}/lib64
-        ${CUDA_TOOLKIT_ROOT_DIR}/lib
-        ${CUDA_TOOLKIT_ROOT_DIR}/lib64)
-
-include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS
-                                  NCCL_LIBRARIES)
-
-if(NCCL_FOUND)
-  set(NCCL_HEADER_FILE "${NCCL_INCLUDE_DIRS}/nccl.h")
-  message(
-    STATUS "Determining NCCL version from the header file: ${NCCL_HEADER_FILE}")
-  file(
-    STRINGS ${NCCL_HEADER_FILE} NCCL_MAJOR_VERSION_DEFINED
-    REGEX "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+[0-9]+.*$"
-    LIMIT_COUNT 1)
-  if(NCCL_MAJOR_VERSION_DEFINED)
-    string(REGEX REPLACE "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+" ""
-                         NCCL_MAJOR_VERSION ${NCCL_MAJOR_VERSION_DEFINED})
-    message(STATUS "NCCL_MAJOR_VERSION: ${NCCL_MAJOR_VERSION}")
-  endif()
-  message(
-    STATUS
-      "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
-  mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
-endif()
--- a/docs/requirements.txt
+++ b/docs/requirements.txt
@@ -1,5 +1,4 @@
 sphinx
 breathe
 sphinx-book-theme
-sphinx-copybutton
 mlx
--- a/docs/src/conf.py
+++ b/docs/src/conf.py
@@ -18,7 +18,6 @@ release = version
 # -- General configuration ---------------------------------------------------

 extensions = [
-    "sphinx_copybutton",
    "sphinx.ext.autodoc",
    "sphinx.ext.autosummary",
    "sphinx.ext.intersphinx",
--- a/docs/src/dev/custom_metal_kernels.rst
+++ b/docs/src/dev/custom_metal_kernels.rst
@@ -127,8 +127,7 @@ relying on a copy from ``ensure_row_contiguous``:
      name="myexp_strided",
      input_names=["inp"],
      output_names=["out"],
-      source=source,
-      ensure_row_contiguous=False,
+      source=source
  )

  def exp_elementwise(a: mx.array):
@@ -139,6 +138,7 @@ relying on a copy from ``ensure_row_contiguous``:
          threadgroup=(256, 1, 1),
          output_shapes=[a.shape],
          output_dtypes=[a.dtype],
+          ensure_row_contiguous=False,
      )
      return outputs[0]

--- a/docs/src/dev/extensions.rst
+++ b/docs/src/dev/extensions.rst
@@ -394,14 +394,14 @@ below.
        out.set_data(allocator::malloc(out.nbytes()));

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

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

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

        // Prepare to encode kernel
        auto& compute_encoder = d.get_command_encoder(s.index);
--- a/docs/src/index.rst
+++ b/docs/src/index.rst
@@ -70,7 +70,6 @@ are the CPU and GPU.
   python/fft
   python/linalg
   python/metal
-   python/cuda
   python/memory_management
   python/nn
   python/optimizers
--- a/docs/src/install.rst
+++ b/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 >= 3.10
+- Using a native Python >= 3.9
 - macOS >= 13.5

 .. note::
@@ -30,7 +30,7 @@ MLX has a CUDA backend which you can install with:

 .. code-block:: shell

-    pip install mlx[cuda]
+    pip install "mlx[cuda]"

 To install the CUDA package from PyPi your system must meet the following
 requirements:
@@ -39,7 +39,7 @@ requirements:
 - Nvidia driver >= 550.54.14
 - CUDA toolkit >= 12.0
 - Linux distribution with glibc >= 2.35
- Python >= 3.10
+- Python >= 3.9


 CPU-only (Linux)
@@ -49,13 +49,13 @@ For a CPU-only version of MLX that runs on Linux use:

 .. code-block:: shell

-    pip install mlx[cpu]
+    pip install "mlx[cpu]"

 To install the CPU-only package from PyPi your system must meet the following
 requirements:

 - Linux distribution with glibc >= 2.35
- Python >= 3.10
+- Python >= 3.9


 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 libcudnn9-dev-cuda-12 -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y


 When building either the Python or C++ APIs make sure to pass the cmake flag
--- a/docs/src/python/cuda.rst
+++ b/docs/src/python/cuda.rst
@@ -1,9 +0,0 @@
-CUDA
-=====
-
-.. currentmodule:: mlx.core.cuda
-
-.. autosummary::
-  :toctree: _autosummary
-
-  is_available
--- a/docs/src/python/fast.rst
+++ b/docs/src/python/fast.rst
@@ -13,4 +13,3 @@ Fast
  rope
  scaled_dot_product_attention
  metal_kernel
-  cuda_kernel
--- a/docs/src/python/nn/functions.rst
+++ b/docs/src/python/nn/functions.rst
@@ -27,7 +27,6 @@ simple functions.
   mish
   prelu
   relu
-   relu2
   relu6
   selu
   sigmoid
--- a/docs/src/python/nn/layers.rst
+++ b/docs/src/python/nn/layers.rst
@@ -50,7 +50,6 @@ Layers
   QuantizedLinear
   RMSNorm
   ReLU
-   ReLU2
   ReLU6
   RNN
   RoPE
--- a/docs/src/python/ops.rst
+++ b/docs/src/python/ops.rst
@@ -112,7 +112,6 @@ Operations
   max
   maximum
   mean
-   median
   meshgrid
   min
   minimum
--- a/docs/src/python/optimizers.rst
+++ b/docs/src/python/optimizers.rst
@@ -51,14 +51,14 @@ the saved state. Here's a simple example:
   optimizer.update(model, grads)

   # Save the state
-   state = tree_flatten(optimizer.state, destination={})
-   mx.save_safetensors("optimizer.safetensors", state)
+   state = tree_flatten(optimizer.state)
+   mx.save_safetensors("optimizer.safetensors", dict(state))

   # Later on, for example when loading from a checkpoint,
   # recreate the optimizer and load the state
   optimizer = optim.Adam(learning_rate=1e-2)

-   state = tree_unflatten(mx.load("optimizer.safetensors"))
+   state = tree_unflatten(list(mx.load("optimizer.safetensors").items()))
   optimizer.state = state

 Note, not every optimizer configuation parameter is saved in the state. For
--- a/docs/src/usage/compile.rst
+++ b/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(gelu, x)
-  timeit(mx.compile(gelu), x)
+  timeit(nn.gelu, x)
+  timeit(mx.compile(nn.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)
+      return mx.exp(z), state

  fun(mx.array(1.0), mx.array(2.0))
  # Prints [array(3, dtype=float32)]
--- a/docs/src/usage/distributed.rst
+++ b/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 = mx.nn.average_gradients(grads)  # <---- This line was added
+        grads = mlx.nn.average_gradients(grads) # <---- This line was added
        optimizer.update(model, grads)
        return loss

--- a/docs/src/usage/export.rst
+++ b/docs/src/usage/export.rst
@@ -7,17 +7,17 @@ Exporting Functions

 MLX has an API to export and import functions to and from a file. This lets you
 run computations written in one MLX front-end (e.g. Python) in another MLX
-front-end (e.g. C++).
+front-end (e.g. C++). 

 This guide walks through the basics of the MLX export API with some examples.
 To see the full list of functions check-out the :ref:`API documentation
 <export>`.

-Basics of Exporting
+Basics of Exporting 
 -------------------

 Let's start with a simple example:
-
+ 
 .. code-block:: python

  def fun(x, y):
@@ -67,7 +67,7 @@ specified as variable positional arguments or as a tuple of arrays:

  x = mx.array(1.0)
  y = mx.array(1.0)
-
+   
  # Both arguments to fun are positional
  mx.export_function("add.mlxfn", fun, x, y)

@@ -133,7 +133,7 @@ parameters are also saved to the ``model.mlxfn`` file.
   For enclosed arrays inside an exported function, be extra careful to ensure
   they are evaluated. The computation graph that gets exported will include
   the computation that produces enclosed inputs.
-
+  
   If the above example was missing ``mx.eval(model.parameters()``, the
   exported function would include the random initialization of the
   :obj:`mlx.nn.Module` parameters.
@@ -150,8 +150,8 @@ parameters, pass them as inputs to the ``call`` wrapper:
     # Set the model's parameters to the input parameters
     model.update(tree_unflatten(list(params.items())))
     return model(x)
-
-   params = tree_flatten(model.parameters(), destination={})
+ 
+   params = dict(tree_flatten(model.parameters()))
   mx.export_function("model.mlxfn", call, (mx.zeros(4),), params)


@@ -164,13 +164,13 @@ 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]))
-
-  # Also ok
+  out, = imported_abs(mx.array(-1.0))
+  
+  # Also ok 
  out, = imported_abs(mx.array([-1.0, -2.0]))

 With ``shapeless=False`` (which is the default), the second call to
@@ -197,7 +197,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
  def fun(x, y=None):
      constant = mx.array(3.0)
      if y is not None:
-        x += y
+        x += y 
      return x + constant

  with mx.exporter("fun.mlxfn", fun) as exporter:
@@ -215,7 +215,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
  print(out)

 In the above example the function constant data, (i.e. ``constant``), is only
-saved once.
+saved once. 

 Transformations with Imported Functions
 ---------------------------------------
@@ -238,7 +238,7 @@ on imported functions just like regular Python functions:
  # Prints: array(1, dtype=float32)
  print(dfdx(x))

-  # Compile the imported function
+  # Compile the imported function 
  mx.compile(imported_fun)
  # Prints: array(0, dtype=float32)
  print(compiled_fun(x)[0])
@@ -275,7 +275,7 @@ Import and run the function in C++ with only a few lines of code:
  // Prints: array(2, dtype=float32)
  std::cout << outputs[0] << std::endl;

-Imported functions can be transformed in C++ just like in Python. Use
+Imported functions can be transformed in C++ just like in Python. Use 
 ``std::vector<mx::array>`` for positional arguments and ``std::map<std::string,
 mx::array>`` for keyword arguments when calling imported functions in C++.

--- a/docs/src/usage/indexing.rst
+++ b/docs/src/usage/indexing.rst
@@ -107,20 +107,8 @@ 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:

-.. 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:
+Note, unlike NumPy, updates to the same location are nondeterministic:

 .. code-block:: shell

--- a/examples/cpp/tutorial.cpp
+++ b/examples/cpp/tutorial.cpp
@@ -14,17 +14,14 @@ 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:
-  int64_t size = x.size();
+  size_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();
@@ -33,7 +30,6 @@ 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);
--- a/examples/extensions/axpby/axpby.cpp
+++ b/examples/extensions/axpby/axpby.cpp
@@ -1,6 +1,5 @@
 // Copyright © 2023-2025 Apple Inc.

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

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

 namespace my_ext {

-// A helper function to find the location of the current binary on disk.
-// The Metal library ("mlx_ext.mtllib"), should be in the same directory.
-std::string current_binary_dir() {
-  static std::string binary_dir = []() {
-    Dl_info info;
-    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
-      throw std::runtime_error("Unable to get current binary dir.");
-    }
-    return std::filesystem::path(info.dli_fname).parent_path().string();
-  }();
-  return binary_dir;
-}
-
 ///////////////////////////////////////////////////////////////////////////////
 // Operation Implementation
 ///////////////////////////////////////////////////////////////////////////////
@@ -181,15 +167,16 @@ void Axpby::eval_gpu(
  }

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

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

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

  // Prepare to encode kernel
  auto& compute_encoder = d.get_command_encoder(s.index);
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,4 +1,4 @@
 setuptools>=42
 cmake>=3.25
 mlx>=0.21.0
-nanobind==2.4.0
+nanobind==2.2.0
--- a/examples/extensions/test.py
+++ b/examples/extensions/test.py
@@ -3,10 +3,8 @@ from mlx_sample_extensions import axpby

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

-print(f"c shape: {c_cpu.shape}")
-print(f"c dtype: {c_cpu.dtype}")
-print(f"c_cpu correct: {mx.all(c_cpu == 6.0).item()}")
-print(f"c_gpu correct: {mx.all(c_gpu == 6.0).item()}")
+print(f"c shape: {c.shape}")
+print(f"c dtype: {c.dtype}")
+print(f"c correct: {mx.all(c == 6.0).item()}")
--- a/mlx/array.cpp
+++ b/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 (int i = 0; i < std::ssize(shapes); ++i) {
+  for (size_t i = 0; i < shapes.size(); ++i) {
    outputs.emplace_back(std::move(shapes[i]), dtypes[i], primitive, inputs);
  }
  // For each node in |outputs|, its siblings are the other nodes.
-  for (int i = 0; i < std::ssize(outputs); ++i) {
+  for (size_t i = 0; i < outputs.size(); ++i) {
    auto siblings = outputs;
    siblings.erase(siblings.begin() + i);
    outputs[i].set_siblings(std::move(siblings), i);
@@ -145,9 +145,8 @@ 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 =
-      static_cast<int64_t>(*std::max_element(shape().begin(), shape().end()));
-  array_desc_->flags.col_contiguous = size() <= 1 || size() == max_dim;
+  auto max_dim = std::max_element(shape().begin(), shape().end());
+  array_desc_->flags.col_contiguous = size() <= 1 || size() == *max_dim;
 }

 void array::set_data(
@@ -193,7 +192,7 @@ array::~array() {
  }

  // Break circular reference for non-detached arrays with siblings
-  if (auto n = std::ssize(siblings()); n > 0) {
+  if (auto n = siblings().size(); 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)
@@ -242,8 +241,8 @@ array::ArrayDesc::ArrayDesc(
    std::vector<array> inputs)
    : shape(std::move(shape)),
      dtype(dtype),
-      primitive(std::move(primitive)),
      status(Status::unscheduled),
+      primitive(std::move(primitive)),
      inputs(std::move(inputs)) {
  init();
 }
@@ -275,7 +274,7 @@ array::ArrayDesc::~ArrayDesc() {
    ad.inputs.clear();
    for (auto& [_, a] : input_map) {
      bool is_deletable =
-          (a.array_desc_.use_count() <= std::ssize(a.siblings()) + 1);
+          (a.array_desc_.use_count() <= a.siblings().size() + 1);
      // An array with siblings is deletable only if all of its siblings
      // are deletable
      for (auto& s : a.siblings()) {
@@ -284,7 +283,7 @@ array::ArrayDesc::~ArrayDesc() {
        }
        int is_input = (input_map.find(s.id()) != input_map.end());
        is_deletable &=
-            s.array_desc_.use_count() <= std::ssize(a.siblings()) + is_input;
+            s.array_desc_.use_count() <= a.siblings().size() + is_input;
      }
      if (is_deletable) {
        for_deletion.push_back(std::move(a.array_desc_));
--- a/mlx/array.h
+++ b/mlx/array.h
@@ -10,7 +10,6 @@
 #include "mlx/allocator.h"
 #include "mlx/dtype.h"
 #include "mlx/event.h"
-#include "mlx/small_vector.h"

 namespace mlx::core {

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

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

 class array {
  /* An array is really a node in a graph. It contains a shared ArrayDesc
@@ -81,22 +80,22 @@ class array {
  }

  /** The size of the array's datatype in bytes. */
-  int itemsize() const {
+  size_t itemsize() const {
    return size_of(dtype());
  }

  /** The number of elements in the array. */
-  int64_t size() const {
+  size_t size() const {
    return array_desc_->size;
  }

  /** The number of bytes in the array. */
-  int64_t nbytes() const {
+  size_t nbytes() const {
    return size() * itemsize();
  }

  /** The number of dimensions of the array. */
-  int ndim() const {
+  size_t ndim() const {
    return array_desc_->shape.size();
  }

@@ -329,7 +328,7 @@ class array {
   * corresponding to ``arr[-1, -1, ...]``) then ``data_size = last - first``.
   * Note, ``data_size`` is in units of ``item_size`` (not bytes).
   **/
-  int64_t data_size() const {
+  size_t data_size() const {
    return array_desc_->data_size;
  }

@@ -340,7 +339,7 @@ class array {
    return array_desc_->data->buffer;
  }

-  int64_t buffer_size() const {
+  size_t buffer_size() const {
    return allocator::allocator().size(buffer());
  }

@@ -530,7 +529,7 @@ array::array(
    Shape shape,
    Dtype dtype /* = TypeToDtype<T>() */)
    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
-  if (std::ssize(data) != size()) {
+  if (data.size() != size()) {
    throw std::invalid_argument(
        "Data size and provided shape mismatch in array construction.");
  }
--- a/mlx/backend/common/common.cpp
+++ b/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;
-  int64_t r = 1, c = 1;
-  for (int i = std::ssize(strides_) - 1, j = 0; i >= 0; i--, j++) {
+  size_t r = 1, c = 1;
+  for (int i = strides_.size() - 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];
@@ -60,8 +60,7 @@ void CustomTransforms::eval(
    const std::vector<array>& inputs,
    std::vector<array>& outputs) {
  assert(inputs.size() > outputs.size());
-  for (int i = 0, j = std::ssize(inputs) - std::ssize(outputs);
-       i < std::ssize(outputs);
+  for (int i = 0, j = inputs.size() - outputs.size(); i < outputs.size();
       i++, j++) {
    outputs[i].copy_shared_buffer(inputs[j]);
  }
@@ -71,7 +70,7 @@ void Depends::eval(
    const std::vector<array>& inputs,
    std::vector<array>& outputs) {
  assert(inputs.size() > outputs.size());
-  for (int i = 0; i < std::ssize(outputs); i++) {
+  for (int i = 0; i < outputs.size(); i++) {
    outputs[i].copy_shared_buffer(inputs[i]);
  }
 }
@@ -207,11 +206,11 @@ void Split::eval(

  auto compute_new_flags = [](const auto& shape,
                              const auto& strides,
-                              int64_t in_data_size,
+                              size_t in_data_size,
                              auto flags) {
-    int64_t data_size = 1;
-    int64_t f_stride = 1;
-    int64_t b_stride = 1;
+    size_t data_size = 1;
+    size_t f_stride = 1;
+    size_t b_stride = 1;
    flags.row_contiguous = true;
    flags.col_contiguous = true;
    for (int i = 0, ri = shape.size() - 1; ri >= 0; i++, ri--) {
@@ -241,7 +240,7 @@ void Split::eval(

  std::vector<int> indices(1, 0);
  indices.insert(indices.end(), indices_.begin(), indices_.end());
-  for (int i = 0; i < std::ssize(indices); i++) {
+  for (int i = 0; i < indices.size(); 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());
@@ -255,7 +254,7 @@ 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 < std::ssize(axes_) && i == axes_[j]) {
+    if (j < axes_.size() && i == axes_[j]) {
      j++;
    } else {
      strides.push_back(in.strides(i));
@@ -273,7 +272,7 @@ 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 < std::ssize(axes_); ++ax) {
+  for (int ax = 0; ax < axes_.size(); ++ax) {
    out_strides[ax] = in.strides()[axes_[ax]];
  }

--- a/mlx/backend/common/compiled.cpp
+++ b/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 < std::ssize(inputs) && o < std::ssize(outputs); ++i) {
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++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 < std::ssize(outputs); ++o) {
+    for (; o < outputs.size(); ++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 < std::ssize(inputs) && o < std::ssize(outputs); ++i) {
+    for (int i = 0; i < inputs.size() && o < outputs.size(); ++i) {
      auto& in = inputs[i];
      // Conditions for donation
      // - Row contiguous
@@ -162,7 +162,7 @@ void compiled_allocate_outputs(
        o++;
      }
    }
-    for (; o < std::ssize(outputs); ++o) {
+    for (; o < outputs.size(); ++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;
-    int j = 0;
+    size_t 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 (int i = 0; i < x.ndim(); ++i, ++j) {
+    for (size_t 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) {
-    int64_t max_size = 0;
+    size_t max_size = 0;
    for (const auto& in : inputs) {
      max_size = std::max(max_size, in.data_size());
    }
    return max_size > UINT32_MAX;
  } else {
-    int64_t max_size = 0;
+    size_t max_size = 0;
    for (const auto& o : outputs) {
      max_size = std::max(max_size, o.size());
    }
--- a/mlx/backend/common/load.cpp
+++ b/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(),
--- a/mlx/backend/common/matmul.h
+++ b/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 {Shape{1}, Strides{0}, Strides{0}};
+    return {{1}, {0}, {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 {Shape{1}, Strides{0}, Strides{0}, Strides{0}};
+    return {{1}, {0}, {0}, {0}};
  }

  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
--- a/mlx/backend/common/reduce.cpp
+++ b/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() && std::ssize(axes) == x.ndim() &&
+  if (x.size() == x.data_size() && axes.size() == 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 < std::ssize(axes); i++) {
+    for (int i = 1; i < axes.size(); 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]];
--- a/mlx/backend/common/slicing.cpp
+++ b/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,
-    int64_t data_offset,
-    int64_t data_size,
+    size_t data_offset,
+    size_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();
  }
-  int64_t data_size = (data_end - data_offset);
+  size_t data_size = (data_end - data_offset);
  shared_buffer_slice(in, inp_strides, data_offset, data_size, out);
 }

--- a/mlx/backend/common/ternary.h
+++ b/mlx/backend/common/ternary.h
@@ -11,8 +11,6 @@ namespace mlx::core {
 enum class TernaryOpType {
  ScalarScalarScalar,
  VectorVectorVector,
-  VectorVectorScalar,
-  VectorScalarVector,
  General,
 };

@@ -27,14 +25,6 @@ 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;
  }
@@ -69,8 +59,6 @@ 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)) ||
--- a/mlx/backend/common/utils.cpp
+++ b/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);
    }
-    int64_t size = shape[0];
+    size_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 < std::ssize(strides); j++) {
+    for (int j = 0; j < strides.size(); j++) {
      const auto& st = strides[j];
      out_strides[j].push_back(st[to_collapse[k - 1]]);
    }
--- a/mlx/backend/common/utils.h
+++ b/mlx/backend/common/utils.h
@@ -162,7 +162,7 @@ struct ContiguousIterator {
 };

 inline auto check_contiguity(const Shape& shape, const Strides& strides) {
-  int64_t no_broadcast_data_size = 1;
+  size_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 int64_t donation_extra = 16384;
+  constexpr size_t donation_extra = 16384;

  return in.is_donatable() && in.itemsize() == out.itemsize() &&
      in.buffer_size() <= out.nbytes() + donation_extra;
@@ -197,7 +197,7 @@ void shared_buffer_reshape(
    array& out);

 template <typename T>
-inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
+inline std::vector<T> remove_index(std::vector<T> vec, size_t index) {
  vec.erase(std::next(vec.begin(), index));
  return vec;
 }
--- a/mlx/backend/cpu/arange.h
+++ b/mlx/backend/cpu/arange.h
@@ -10,7 +10,7 @@ namespace mlx::core {
 namespace {

 template <typename T>
-void arange(T start, T next, array& out, int64_t size, Stream stream) {
+void arange(T start, T next, array& out, size_t size, Stream stream) {
  auto ptr = out.data<T>();
  auto step_size = next - start;
  auto& encoder = cpu::get_command_encoder(stream);
--- a/mlx/backend/cpu/arg_reduce.cpp
+++ b/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 (int64_t i = 0; i < out.size(); ++i) {
+  for (uint32_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 (int64_t j = 0; j < axis_size; ++j, local_in_ptr += axis_stride) {
+    for (uint32_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;
--- a/mlx/backend/cpu/binary.cpp
+++ b/mlx/backend/cpu/binary.cpp
@@ -17,12 +17,7 @@ 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);

@@ -86,7 +81,7 @@ 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);
@@ -151,7 +146,7 @@ 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);
@@ -192,7 +187,7 @@ 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);
--- a/mlx/backend/cpu/binary_two.h
+++ b/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 (int64_t elem = 0; elem < std::ssize(a); elem += stride) {
+  for (size_t elem = 0; elem < a.size(); 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 (int64_t i = 0; i < b.data_size(); ++i) {
+    for (size_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 (int64_t i = 0; i < a.data_size(); ++i) {
+    for (size_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 (int64_t i = 0; i < a.size(); ++i) {
+    for (size_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++;
--- a/mlx/backend/cpu/cholesky.cpp
+++ b/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';
-    int64_t num_matrices = size / (N * N);
-    for (int64_t i = 0; i < num_matrices; i++) {
+    size_t num_matrices = size / (N * N);
+    for (int i = 0; i < num_matrices; i++) {
      // Compute Cholesky factorization.
      int info;
      potrf<T>(
--- a/mlx/backend/cpu/compiled.cpp
+++ b/mlx/backend/cpu/compiled.cpp
@@ -15,7 +15,6 @@
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
-#include "mlx/version.h"

 namespace mlx::core {

@@ -49,7 +48,7 @@ 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;
 }

@@ -95,11 +94,7 @@ void* compile(
    kernel_file_name = kernel_name;
  }

-  auto output_dir =
-      std::filesystem::temp_directory_path() / "mlx" / version() / "cpu";
-  if (!std::filesystem::exists(output_dir)) {
-    std::filesystem::create_directories(output_dir);
-  }
+  auto output_dir = std::filesystem::temp_directory_path();

  std::string shared_lib_name = "lib" + kernel_file_name + ".so";
  auto shared_lib_path = (output_dir / shared_lib_name).string();
@@ -162,13 +157,11 @@ inline void build_kernel(
 #endif

  // Start the kernel
-  os << "void " << kernel_name
-     << "(int* shape, int64_t** strides, void** args) {" << std::endl;
+  os << "void " << kernel_name << "(void** args) {" << std::endl;

  // Add the input arguments
  int cnt = 0;
-  int strides_index = 1;
-  for (int i = 0; i < std::ssize(inputs); ++i) {
+  for (size_t i = 0; i < inputs.size(); ++i) {
    // Skip constants from the input list
    if (is_constant(i)) {
      continue;
@@ -182,8 +175,8 @@ inline void build_kernel(
       << "];" << std::endl;
    // Scalars and contiguous need no strides
    if (!is_scalar(x) && !contiguous) {
-      os << "  const int64_t* " << xname << "_strides = strides["
-         << strides_index++ << "];" << std::endl;
+      os << "  const size_t* " << xname << "_strides = (size_t*)args[" << cnt++
+         << "];" << std::endl;
    }
  }

@@ -193,8 +186,10 @@ inline void build_kernel(
    os << "  " << tstr << "* " << namer.get_name(x) << " = (" << tstr
       << "*)args[" << cnt++ << "];" << std::endl;
  }
-  // Add output size
-  if (contiguous) {
+  // Add output strides and shape to extract the indices.
+  if (!contiguous) {
+    os << "  const int* shape = (int*)args[" << cnt++ << "];" << std::endl;
+  } else {
    os << "  const size_t size = (size_t)args[" << cnt++ << "];" << std::endl;
  }

@@ -238,7 +233,7 @@ inline void build_kernel(
    } else {
      os << x.primitive().name();
      os << "()(";
-      for (int i = 0; i < std::ssize(x.inputs()) - 1; i++) {
+      for (int i = 0; i < x.inputs().size() - 1; i++) {
        os << "tmp_" << namer.get_name(x.inputs()[i]) << ", ";
      }
      os << "tmp_" << namer.get_name(x.inputs().back()) << ");" << std::endl;
@@ -295,6 +290,7 @@ void Compiled::eval_cpu(

  // Collect function input arguments.
  std::vector<void*> args;
+  int strides_index = 1;
  for (size_t i = 0; i < inputs.size(); ++i) {
    if (is_constant_(i)) {
      continue;
@@ -302,6 +298,9 @@ void Compiled::eval_cpu(
    const auto& x = inputs[i];
    encoder.set_input_array(x);
    args.push_back((void*)x.data<void>());
+    if (!contiguous && !is_scalar(x)) {
+      args.push_back(strides[strides_index++].data());
+    }
  }

  // Get the kernel name from the lib
@@ -336,20 +335,16 @@ void Compiled::eval_cpu(
    args.push_back(x.data<void>());
    encoder.set_output_array(x);
  }
-  if (contiguous) {
+  if (!contiguous) {
+    args.push_back((void*)shape.data());
+  } else {
    args.push_back((void*)outputs[0].data_size());
  }
-  auto fun = reinterpret_cast<void (*)(int*, int64_t**, void**)>(fn_ptr);
+  auto fun = (void (*)(void**))fn_ptr;
  encoder.dispatch([fun,
                    args = std::move(args),
                    strides = std::move(strides),
-                    shape = std::move(shape)]() mutable {
-    SmallVector<int64_t*> strides_ptrs;
-    for (auto& s : strides) {
-      strides_ptrs.push_back(s.data());
-    }
-    fun(shape.data(), strides_ptrs.data(), args.data());
-  });
+                    shape = std::move(shape)]() mutable { fun(args.data()); });
 }

 } // namespace mlx::core
--- a/mlx/backend/cpu/conv.cpp
+++ b/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
@@ -996,6 +996,131 @@ 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,
@@ -1003,7 +1128,7 @@ 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)
@@ -1023,7 +1148,7 @@ void explicit_gemm_conv_ND_cpu(
  // Pad input
  Shape padded_shape(in.shape().size());
  padded_shape.front() = N;
-  for (int i = 0; i < iDim.size(); i++) {
+  for (size_t i = 0; i < iDim.size(); i++) {
    padded_shape[i + 1] = iDim[i] + padding_lo[i] + padding_hi[i];
  }
  padded_shape.back() = C;
@@ -1054,20 +1179,20 @@ void explicit_gemm_conv_ND_cpu(
  // Make strided view
  Shape strided_shape(oDim.size() + wDim.size() + 2);
  strided_shape.front() = N;
-  for (int i = 0; i < oDim.size(); i++) {
+  for (size_t i = 0; i < oDim.size(); i++) {
    strided_shape[i + 1] = oDim[i];
  }
-  for (int i = 0; i < wDim.size(); i++) {
+  for (size_t 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 (int i = 0; i < std::ssize(wt_strides); i++) {
+  for (size_t i = 0; i < wt_strides.size(); i++) {
    strided_strides[i + 1] = in_padded.strides()[i + 1] * wt_strides[i];
  }
-  for (int i = 1; i < std::ssize(in_padded.strides()); i++) {
+  for (size_t i = 1; i < in_padded.strides().size(); i++) {
    strided_strides[i + wt_strides.size()] = in_padded.strides()[i];
  }

--- a/mlx/backend/cpu/distributed.cpp
+++ b/mlx/backend/cpu/distributed.cpp
@@ -90,7 +90,6 @@ 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());
--- a/mlx/backend/cpu/eig.cpp
+++ b/mlx/backend/cpu/eig.cpp
@@ -46,6 +46,7 @@ void eig_impl(
    int info;
    {
      T work;
+      int iwork;
      geev<T>(
          &jobl,
          &jobr,
@@ -70,7 +71,7 @@ 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 (int64_t i = 0; i < size / (N * N); ++i) {
+    for (size_t i = 0; i < size / (N * N); ++i) {
      geev<T>(
          &jobl,
          &jobr,
--- a/mlx/backend/cpu/eigh.cpp
+++ b/mlx/backend/cpu/eigh.cpp
@@ -165,7 +165,7 @@ void eigh_impl(
    EighWork<T> work(jobz, uplo, N);

    // Work loop
-    for (int64_t i = 0; i < size / (N * N); ++i) {
+    for (size_t i = 0; i < size / (N * N); ++i) {
      work.run(vec_ptr, eig_ptr);
      vec_ptr += N * N;
      eig_ptr += N;
--- a/mlx/backend/cpu/encoder.h
+++ b/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.
--- a/mlx/backend/cpu/gemm.h
+++ b/mlx/backend/cpu/gemm.h
@@ -12,12 +12,12 @@ void matmul(
    T* out,
    bool a_transposed,
    bool b_transposed,
-    int64_t lda,
-    int64_t ldb,
-    int64_t ldc,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
    float alpha,
    float beta,
-    int64_t batch_size,
+    size_t batch_size,
    const Shape& a_shape,
    const Strides& a_strides,
    const Shape& b_shape,
--- a/mlx/backend/cpu/gemms/bnns.cpp
+++ b/mlx/backend/cpu/gemms/bnns.cpp
@@ -1,4 +1,5 @@
 // Copyright © 2023-2024 Apple Inc.
+
 #include <Accelerate/Accelerate.h>

 #include "mlx/array.h"
@@ -34,7 +35,7 @@ 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,
@@ -48,15 +49,9 @@ 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,
@@ -127,7 +122,7 @@ void matmul_bnns(
  auto bnns_filter =
      BNNSFilterCreateLayerBroadcastMatMul(&gemm_params, nullptr);

-  for (size_t i = 0; i < batch_size; ++i) {
+  for (int i = 0; i < batch_size; ++i) {
    BNNSFilterApplyTwoInput(
        bnns_filter,
        reinterpret_cast<const uint8_t*>(
@@ -148,12 +143,12 @@ void matmul<float16_t>(
    float16_t* out,
    bool a_transposed,
    bool b_transposed,
-    int64_t lda,
-    int64_t ldb,
-    int64_t ldc,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
    float alpha,
    float beta,
-    int64_t batch_size,
+    size_t batch_size,
    const Shape& a_shape,
    const Strides& a_strides,
    const Shape& b_shape,
@@ -183,12 +178,12 @@ void matmul<bfloat16_t>(
    bfloat16_t* out,
    bool a_transposed,
    bool b_transposed,
-    int64_t lda,
-    int64_t ldb,
-    int64_t ldc,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
    float alpha,
    float beta,
-    int64_t batch_size,
+    size_t batch_size,
    const Shape& a_shape,
    const Strides& a_strides,
    const Shape& b_shape,
--- a/mlx/backend/cpu/gemms/cblas.cpp
+++ b/mlx/backend/cpu/gemms/cblas.cpp
@@ -13,20 +13,20 @@ void matmul<float>(
    float* out,
    bool a_transposed,
    bool b_transposed,
-    int64_t lda,
-    int64_t ldb,
-    int64_t ldc,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
    float alpha,
    float beta,
-    int64_t batch_size,
+    size_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];
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_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,
-    int64_t lda,
-    int64_t ldb,
-    int64_t ldc,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
    float alpha,
    float beta,
-    int64_t batch_size,
+    size_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];
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];

  for (int i = 0; i < batch_size; ++i) {
    cblas_dgemm(
@@ -88,47 +88,4 @@ 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
--- a/mlx/backend/cpu/hadamard.cpp
+++ b/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, int64_t size) {
+void hadamard_n(T* out, int n, int m, float scale, size_t size) {
  for (int b = 0; b < size / n; b++) {
-    int64_t loc = b * n;
+    size_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, int64_t size) {

 // m component
 template <typename T>
-void hadamard_m(T* out, int n, int m, float scale, int64_t size) {
+void hadamard_m(T* out, int n, int m, float scale, size_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, int64_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 < std::ssize(row); i++) {
+    for (int i = 0; i < row.length(); 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, int64_t size) {
  }

  for (int b = 0; b < size / m / n; b++) {
-    int64_t loc = b * n * m;
+    size_t loc = b * n * m;
    T* data_ptr = out + loc;
    for (int i = 0; i < n; i++) {
      std::vector<float> out(m);
--- a/mlx/backend/cpu/indexing.cpp
+++ b/mlx/backend/cpu/indexing.cpp
@@ -78,7 +78,7 @@ void gather(
    can_copy = true;

    // Ignore leading 1s
-    int64_t i = 0;
+    int i = 0;
    for (; i < slice_sizes.size() && slice_sizes[i] == 1; ++i)
      ;

@@ -91,7 +91,7 @@ void gather(
    can_copy = true;

    // Ignore trailing 1s
-    int64_t i = slice_sizes.size() - 1;
+    int 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]);
    }
  }
-  int64_t slice_size = 1;
+  size_t slice_size = 1;
  for (auto s : slice_sizes) {
    slice_size *= s;
  }
-  int64_t ind_size = slice_size == 0 ? 0 : out.size() / slice_size;
+  size_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());
  }

-  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) {
+  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) {
      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 (int64_t jj = 0; jj < slice_size; jj++) {
+      for (int 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();
-  int64_t n_updates = nind ? inds[0].size() : 1;
+  size_t n_updates = nind ? inds[0].size() : 1;

  Shape update_shape(
      updates.shape().begin() + inds_ndim, updates.shape().end());
-  int64_t update_size = 1;
+  size_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 (int64_t i = 0; i < n_updates; ++i) {
-    int64_t out_offset = 0;
-    for (int j = 0; j < std::ssize(inds); ++j) {
+  for (int i = 0; i < n_updates; ++i) {
+    size_t out_offset = 0;
+    for (int j = 0; j < inds.size(); ++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 (int64_t j = 0; j < update_size; ++j) {
+    for (int 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();
--- a/mlx/backend/cpu/inverse.cpp
+++ b/mlx/backend/cpu/inverse.cpp
@@ -122,7 +122,7 @@ void inverse_impl(
      stream);

  const int N = a.shape(-1);
-  const int64_t num_matrices = a.size() / (N * N);
+  const size_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 (int64_t i = 0; i < num_matrices; i++) {
+      for (int 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 (int64_t i = 0; i < num_matrices; i++) {
+      for (int i = 0; i < num_matrices; i++) {
        general_inv<T>(inv_ptr + N * N * i, N);
      }
    });
--- a/mlx/backend/cpu/jit_compiler.cpp
+++ b/mlx/backend/cpu/jit_compiler.cpp
@@ -2,7 +2,6 @@

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

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

--- a/mlx/backend/cpu/lapack.h
+++ b/mlx/backend/cpu/lapack.h
@@ -47,7 +47,7 @@ INSTANTIATE_LAPACK_REAL(orgqr)
 INSTANTIATE_LAPACK_REAL(syevd)
 INSTANTIATE_LAPACK_REAL(geev)
 INSTANTIATE_LAPACK_REAL(potrf)
-INSTANTIATE_LAPACK_REAL(gesdd)
+INSTANTIATE_LAPACK_REAL(gesvdx)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)
--- a/mlx/backend/cpu/masked_mm.cpp
+++ b/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 int64_t mask_offset) {
+    const size_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,
-    int64_t lda,
-    int64_t ldb,
+    size_t lda,
+    size_t ldb,
    const Shape& a_shape,
    const Strides& a_strides,
    const Shape& b_shape,
    const Strides& b_strides,
-    int64_t num_segments,
+    size_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_);

-  int64_t M = a.shape(-2);
-  int64_t N = b.shape(-1);
-  int64_t K = a.shape(-1);
+  size_t M = a.shape(-2);
+  size_t N = b.shape(-1);
+  size_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,
-                       int64_t X_data_str,
-                       int64_t Y_data_str,
+                       size_t X_data_str,
+                       size_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 = nullptr;
-  const void* b_mask_ptr = nullptr;
-  const void* out_mask_ptr = nullptr;
+  const void* a_mask_ptr;
+  const void* b_mask_ptr;
+  const void* out_mask_ptr;
  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 = false;
-  bool b_mask_bool = false;
-  bool out_mask_bool = false;
+  bool a_mask_bool;
+  bool b_mask_bool;
+  bool out_mask_bool;
  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>();
-  int64_t num_matrices = out.size() / (M * int64_t(N));
+  size_t num_matrices = out.size() / (M * size_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);

-  int64_t M = a.shape(-2);
-  int64_t N = b.shape(-1);
-  int64_t K = a.shape(-1);
+  size_t M = a.shape(-2);
+  size_t N = b.shape(-1);
+  size_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);
-  int64_t matrix_stride_out = M * N;
+  size_t matrix_stride_out = M * N;

  auto get_batch_dims = [](const auto& v) {
    return decltype(v){v.begin(), v.end() - 2};
@@ -423,6 +423,7 @@ 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());
--- a/mlx/backend/cpu/matmul.cpp
+++ b/mlx/backend/cpu/matmul.cpp
@@ -91,6 +91,7 @@ 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;
  }
@@ -107,9 +108,6 @@ 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.");
  }
@@ -130,6 +128,10 @@ 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;
--- a/mlx/backend/cpu/primitives.cpp
+++ b/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 < std::ssize(axes); ++i) {
+        for (int i = 0; i < axes.size(); ++i) {
          offset_ += indices[i] * strides[axes[i]];
        }
        offset[0] = offset_;
@@ -124,7 +124,6 @@ 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_:
@@ -194,9 +193,9 @@ 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 < std::ssize(inputs); i++) {
+  for (int i = 0; i < inputs.size(); i++) {
    array out_slice(inputs[i].shape(), out.dtype(), nullptr, {});
-    int64_t data_offset = strides[axis_] * sizes[i];
+    size_t data_offset = strides[axis_] * sizes[i];
    out_slice.copy_shared_buffer(
        out, strides, flags, out_slice.size(), data_offset);
    copy_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
@@ -206,7 +205,7 @@ 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 int64_t extra_bytes = 16384;
+  constexpr size_t extra_bytes = 16384;
  if (in.buffer_size() <= out.nbytes() + extra_bytes &&
      (in.flags().row_contiguous ||
       (allow_col_major_ && in.flags().col_contiguous))) {
@@ -255,8 +254,8 @@ 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
-  int64_t data_offset = 0;
-  for (int i = 0; i < std::ssize(axes_); i++) {
+  size_t data_offset = 0;
+  for (int i = 0; i < axes_.size(); i++) {
    auto ax = axes_[i] < 0 ? out.ndim() + axes_[i] : axes_[i];
    data_offset += out.strides()[ax] * low_pad_size_[i];
  }
@@ -275,10 +274,10 @@ 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];
-  int64_t num_keys = keys.size() / 2;
+  size_t num_keys = keys.size() / 2;

-  int64_t elems_per_key = out.size() / num_keys;
-  int64_t bytes_per_key = out.itemsize() * elems_per_key;
+  size_t elems_per_key = out.size() / num_keys;
+  size_t bytes_per_key = out.itemsize() * elems_per_key;
  out.set_data(allocator::malloc(out.nbytes()));

  auto kptr = inputs[0].data<uint32_t>();
@@ -292,8 +291,8 @@ void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
                    num_keys,
                    kshape = keys.shape(),
                    kstrides = keys.strides()]() mutable {
-    int64_t out_skip = (bytes_per_key + 4 - 1) / 4;
-    uintptr_t half_size = out_skip / 2;
+    size_t out_skip = (bytes_per_key + 4 - 1) / 4;
+    auto 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);
--- a/mlx/backend/cpu/qrf.cpp
+++ b/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;
-  int64_t num_matrices = a.size() / (M * N);
+  size_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 (int64_t i = 0; i < num_matrices; ++i) {
+    for (int 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 (int64_t i = 0; i < num_matrices; ++i) {
+    for (int 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 (int64_t i = 0; i < num_matrices; ++i) {
+    for (int 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 (int64_t i = 0; i < num_matrices; ++i) {
+    for (int 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) {
--- a/mlx/backend/cpu/quantized.cpp
+++ b/mlx/backend/cpu/quantized.cpp
@@ -1,5 +1,7 @@
 // Copyright © 2023 Apple Inc.

+#include <cassert>
+
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
@@ -11,35 +13,6 @@ namespace mlx::core {

 namespace {

-const static float MXFP4_LUT[16] = {
-    +0.0f,
-    +0.5f,
-    +1.0f,
-    +1.5f,
-    +2.0f,
-    +3.0f,
-    +4.0f,
-    +6.0f,
-    -0.0f,
-    -0.5f,
-    -1.0f,
-    -1.5f,
-    -2.0f,
-    -3.0f,
-    -4.0f,
-    -6.0f};
-
-template <typename T>
-static inline T dequantize_scale(uint8_t s) {
-  using FOrI = union {
-    bfloat16_t f;
-    uint16_t i;
-  };
-  FOrI out;
-  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
-  return static_cast<T>(out.f);
-}
-
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
  return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -434,229 +407,6 @@ void _qmm_dispatch(
  }
 }

-template <typename T>
-void mxfp4_qmm(
-    T* result,
-    const T* x,
-    const uint32_t* w,
-    const uint8_t* scales,
-    int M,
-    int N,
-    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = get_pack_factor(4, 8);
-  constexpr int 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,
@@ -763,198 +513,115 @@ void _bs_qmm_dispatch(
  }
 }

-template <typename T>
-void mxfp4_bs_qmm_dispatch_typed(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& lhs_indices,
-    const array& rhs_indices,
-    bool transposed_w) {
-  int K = x.shape(-1);
-  int M = x.shape(-2);
-  int N = out.shape(-1);
-
-  int w_els = w.shape(-1) * w.shape(-2);
-  int g_els = scales.shape(-1) * scales.shape(-2);
-
-  auto out_ptr = out.data<T>();
-  auto x_ptr = x.data<T>();
-  auto w_ptr = w.data<uint32_t>();
-  auto scales_ptr = scales.data<uint8_t>();
-  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
-  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
-
-  for (int i = 0; i < lhs_indices.size(); i++) {
-    int x_idx = lhs_indices_ptr[elem_to_loc(
-        i, lhs_indices.shape(), lhs_indices.strides())];
-    int w_idx = rhs_indices_ptr[elem_to_loc(
-        i, rhs_indices.shape(), rhs_indices.strides())];
-    mxfp4_qmm_dispatch_transpose<T>(
-        out_ptr + i * M * N,
-        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
-        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
-        scales_ptr +
-            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
-        M,
-        N,
-        K,
-        transposed_w);
-  }
-}
-
-void mxfp4_bs_qmm_dispatch(
-    array& out,
-    const array& x,
-    const array& w,
-    const array& scales,
-    const array& lhs_indices,
-    const array& rhs_indices,
-    bool transposed_w) {
-  switch (x.dtype()) {
-    case float32:
-      mxfp4_bs_qmm_dispatch_typed<float>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    case float16:
-      mxfp4_bs_qmm_dispatch_typed<float16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    case bfloat16:
-      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
-      break;
-    default:
-      throw std::invalid_argument(
-          "[quantized_matmul] only floating types are supported");
-  }
-}
-
 } // namespace

 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 4);
+
  auto& x_pre = inputs[0];
  auto& w_pre = inputs[1];
  auto& scales_pre = inputs[2];
+  auto& biases_pre = inputs[3];

-  auto& encoder = cpu::get_command_encoder(stream());
-  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
+  std::vector<array> temps;
+  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
    if (arr.flags().row_contiguous) {
      return arr;
    } else {
-      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, arr_cpy, CopyType::General, s);
-      encoder.add_temporary(arr_cpy);
-      return arr_cpy;
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy_cpu(arr, temps.back(), CopyType::General, s);
+      return temps.back();
    }
  };

  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);
-  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_);
-    });
-  }
+  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_);
+  });
 }

 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 6);
+
  auto& x_pre = inputs[0];
  auto& w_pre = inputs[1];
  auto& scales_pre = inputs[2];
-  auto& lhs_indices = inputs[inputs.size() - 2];
-  auto& rhs_indices = inputs[inputs.size() - 1];
+  auto& biases_pre = inputs[3];
+  auto& lhs_indices = inputs[4];
+  auto& rhs_indices = inputs[5];

-  auto& encoder = cpu::get_command_encoder(stream());
+  std::vector<array> temps;
  auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &encoder](const array& arr) {
+                                          &temps](const array& arr) {
    auto stride_0 = arr.strides()[arr.ndim() - 2];
    auto stride_1 = arr.strides()[arr.ndim() - 1];
    if (stride_0 == arr.shape(-1) && stride_1 == 1) {
      return arr;
    } else {
-      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, arr_cpy, CopyType::General, s);
-      encoder.add_temporary(arr_cpy);
-      return arr_cpy;
+      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
+      copy_cpu(arr, temps.back(), CopyType::General, s);
+      return temps.back();
    }
  };

  auto x = ensure_row_contiguous_last_dims(x_pre);
  auto w = ensure_row_contiguous_last_dims(w_pre);
  auto scales = ensure_row_contiguous_last_dims(scales_pre);
+  auto biases = ensure_row_contiguous_last_dims(biases_pre);

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

+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.add_temporaries(std::move(temps));
  encoder.set_input_array(x);
  encoder.set_input_array(w);
  encoder.set_input_array(scales);
+  encoder.set_input_array(biases);
  encoder.set_input_array(lhs_indices);
  encoder.set_input_array(rhs_indices);
  encoder.set_output_array(out);
-  if (mode_ == QuantizationMode::Affine) {
-    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
-    encoder.set_input_array(biases);
-    encoder.dispatch([out = array::unsafe_weak_copy(out),
-                      x = array::unsafe_weak_copy(x),
-                      w = array::unsafe_weak_copy(w),
-                      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_);
-    });
-  }
+  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_);
+  });
 }

 template <typename T, typename U>
@@ -1038,7 +705,7 @@ void dispatch_quantize(
      w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }

-void fast::Quantize::eval_cpu(
+void fast::AffineQuantize::eval_cpu(
    const std::vector<array>& inputs,
    std::vector<array>& outputs) {
  auto ensure_row_contiguous = [s = stream()](const array& arr) {
@@ -1097,7 +764,7 @@ void fast::Quantize::eval_cpu(
      }
    } else {
      throw std::runtime_error(
-          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
+          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
    }
  });
 }
--- a/mlx/backend/cpu/reduce.cpp
+++ b/mlx/backend/cpu/reduce.cpp
@@ -491,27 +491,19 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
        switch (in.dtype()) {
          case bool_:
          case uint8:
-            reduce_dispatch_sum_prod<uint8_t>(in, out, reduce_type_, axes_);
-            break;
-          case uint16:
-            reduce_dispatch_sum_prod<uint16_t>(in, out, reduce_type_, axes_);
-            break;
-          case uint32:
-            reduce_dispatch_sum_prod<uint32_t>(in, out, reduce_type_, axes_);
-            break;
-          case uint64:
-            reduce_dispatch_sum_prod<uint64_t>(in, out, reduce_type_, axes_);
-            break;
          case int8:
            reduce_dispatch_sum_prod<int8_t>(in, out, reduce_type_, axes_);
            break;
          case int16:
+          case uint16:
            reduce_dispatch_sum_prod<int16_t>(in, out, reduce_type_, axes_);
            break;
          case int32:
+          case uint32:
            reduce_dispatch_sum_prod<int32_t>(in, out, reduce_type_, axes_);
            break;
          case int64:
+          case uint64:
            reduce_dispatch_sum_prod<int64_t>(in, out, reduce_type_, axes_);
            break;
          case float16:
--- a/mlx/backend/cpu/simd/accelerate_simd.h
+++ b/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 simd/base_simd.h
+// There seems to be a bug in sims/base.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
@@ -234,7 +234,6 @@ Simd<T, N> remainder(Simd<T, N> a, Simd<T, N> b) {

 template <typename MaskT, typename T1, typename T2, int N>
 Simd<T1, N> select(Simd<MaskT, N> mask, Simd<T1, N> x, Simd<T2, N> y) {
-  static_assert(std::is_same_v<MaskT, bool>);
  if constexpr (sizeof(T1) == 1) {
    return asd::bitselect(y.value, x.value, asd::convert<char>(mask.value));
  } else if constexpr (sizeof(T1) == 2) {
@@ -252,13 +251,9 @@ Simd<T, N> pow(Simd<T, N> base, Simd<T, N> exp) {
    return asd::pow(base.value, exp.value);
  } else {
    Simd<T, N> res = 1;
-    // Raising an integer to a negative power is undefined
-    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);
+    while (any(exp)) {
+      res = select(exp & 1, res * base, res);
+      base = select(exp, base * base, base);
      exp = exp >> 1;
    }
    return res;
--- a/mlx/backend/cpu/simd/math.h
+++ b/mlx/backend/cpu/simd/math.h
@@ -79,8 +79,7 @@ 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 & static_cast<uint32_t>(2)) != static_cast<uint32_t>(0);
+  auto poly_mask = (emm2 & 2) != 0;

  // The magic pass: "Extended precision modular arithmetic"
  // x = ((x - y * DP1) - y * DP2) - y * DP3
@@ -88,8 +87,8 @@ 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) != static_cast<uint32_t>(0));
-  auto sign_mask_cos = ((emm2 - 2) & 4) != static_cast<uint32_t>(0);
+  sign_mask_sin = sign_mask_sin ^ ((emm2 & 4) != 0);
+  auto sign_mask_cos = ((emm2 - 2) & 4) != 0;

  // Evaluate the first polynom  (0 <= x <= Pi/4) in y1,
  // and the second polynom      (Pi/4 <= x <= 0) in y2
--- a/mlx/backend/cpu/sort.cpp
+++ b/mlx/backend/cpu/sort.cpp
@@ -8,25 +8,13 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
-#include "mlx/dtype_utils.h"
+
 #include "mlx/primitives.h"

 namespace mlx::core {

 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;
@@ -39,7 +27,7 @@ struct StridedIterator {
  StridedIterator() = default;

  explicit StridedIterator(T* ptr, int64_t stride, difference_type offset = 0)
-      : stride_(stride), ptr_(ptr + offset * stride) {}
+      : ptr_(ptr + offset * stride), stride_(stride) {}

  explicit StridedIterator(array& arr, int axis, difference_type offset = 0)
      : StridedIterator(arr.data<T>(), arr.strides()[axis], offset) {}
@@ -120,8 +108,8 @@ template <typename T>
 void sort(array& out, int axis) {
  // Get axis, shape and stride info
  axis = axis < 0 ? axis + out.ndim() : axis;
-  int64_t in_size = out.size();
-  int64_t n_rows = in_size / out.shape(axis);
+  size_t in_size = out.size();
+  size_t n_rows = in_size / out.shape(axis);

  auto remaining_shape = out.shape();
  remaining_shape.erase(remaining_shape.begin() + axis);
@@ -136,13 +124,13 @@ void sort(array& out, int axis) {
  ContiguousIterator src_it(
      remaining_shape, remaining_strides, remaining_shape.size());
  auto out_ptr = out.data<T>();
-  for (int64_t i = 0; i < n_rows; i++) {
+  for (int 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, nan_aware_less<T>);
+    std::stable_sort(st, ed);
    src_it.step();
  }
 }
@@ -151,7 +139,7 @@ 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;
-  int64_t n_rows = in.size() / in.shape(axis);
+  size_t n_rows = in.size() / in.shape(axis);

  auto in_remaining_shape = in.shape();
  in_remaining_shape.erase(in_remaining_shape.begin() + axis);
@@ -176,7 +164,7 @@ 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 (int64_t i = 0; i < n_rows; i++) {
+  for (int i = 0; i < n_rows; i++) {
    const T* data_ptr = in_ptr + in_it.loc;
    IdxT* idx_ptr = out_ptr + out_it.loc;

@@ -196,15 +184,6 @@ 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);
    });
  }
@@ -214,8 +193,8 @@ template <typename T>
 void partition(array& out, int axis, int kth) {
  // Get axis, shape and stride info
  axis = axis < 0 ? axis + out.ndim() : axis;
-  int64_t in_size = out.size();
-  int64_t n_rows = in_size / out.shape(axis);
+  size_t in_size = out.size();
+  size_t n_rows = in_size / out.shape(axis);

  auto remaining_shape = out.shape();
  remaining_shape.erase(remaining_shape.begin() + axis);
@@ -232,7 +211,7 @@ 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 (int64_t i = 0; i < n_rows; i++) {
+  for (int i = 0; i < n_rows; i++) {
    T* data_ptr = out_ptr + src_it.loc;
    src_it.step();

@@ -240,7 +219,7 @@ 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, nan_aware_less<T>);
+    std::nth_element(st, md, ed);
  }
 }

@@ -248,7 +227,7 @@ 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;
-  int64_t n_rows = in.size() / in.shape(axis);
+  size_t n_rows = in.size() / in.shape(axis);

  auto in_remaining_shape = in.shape();
  in_remaining_shape.erase(in_remaining_shape.begin() + axis);
@@ -277,7 +256,7 @@ void argpartition(const array& in, array& out, int axis, int kth) {
  auto in_ptr = in.data<T>();
  auto out_ptr = out.data<IdxT>();

-  for (int64_t i = 0; i < n_rows; i++) {
+  for (int 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();
@@ -297,15 +276,6 @@ 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);
    });
  }
@@ -363,24 +333,47 @@ void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 1);
  auto& in = inputs[0];

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

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

 void ArgPartition::eval_cpu(const std::vector<array>& inputs, array& out) {
--- a/mlx/backend/cpu/svd.cpp
+++ 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);

-  int64_t num_matrices = a.size() / (M * N);
+  size_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, {});
@@ -81,26 +81,40 @@ void svd_impl(
    // Vᵀ of shape N x N. (M x M in lapack).
    const int ldvt = M;

-    auto jobz = (u_ptr) ? "A" : "N";
+    auto job_u = (u_ptr) ? "V" : "N";
+    auto job_vt = (u_ptr) ? "V" : "N";
+    static constexpr auto range = "A";

+    // Will contain the number of singular values after the call has returned.
+    int ns = 0;
    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) * 8 * K)};
+    auto iwork = array::Data{allocator::malloc(sizeof(int) * 12 * K)};

    static const int lwork_query = -1;

+    static const int ignored_int = 0;
+    static const T ignored_float = 0;
+
    int info;

    // Compute workspace size.
-    gesdd<T>(
-        /* jobz = */ jobz,
+    gesvdx<T>(
+        /* jobu = */ job_u,
+        /* jobvt = */ job_vt,
+        /* range = */ range,
        // M and N are swapped since lapack expects column-major.
        /* m = */ &N,
        /* n = */ &M,
        /* a = */ nullptr,
        /* lda = */ &lda,
+        /* vl = */ &ignored_float,
+        /* vu = */ &ignored_float,
+        /* il = */ &ignored_int,
+        /* iu = */ &ignored_int,
+        /* ns = */ &ns,
        /* s = */ nullptr,
        /* u = */ nullptr,
        /* ldu = */ &ldu,
@@ -121,14 +135,21 @@ void svd_impl(
    auto scratch = array::Data{allocator::malloc(sizeof(T) * lwork)};

    // Loop over matrices.
-    for (int64_t i = 0; i < num_matrices; i++) {
-      gesdd<T>(
-          /* jobz = */ jobz,
+    for (int i = 0; i < num_matrices; i++) {
+      gesvdx<T>(
+          /* jobu = */ job_u,
+          /* jobvt = */ job_vt,
+          /* range = */ range,
          // M and N are swapped since lapack expects column-major.
          /* m = */ &N,
          /* n = */ &M,
          /* a = */ in_ptr + M * N * i,
          /* lda = */ &lda,
+          /* vl = */ &ignored_float,
+          /* vu = */ &ignored_float,
+          /* il = */ &ignored_int,
+          /* iu = */ &ignored_int,
+          /* ns = */ &ns,
          /* s = */ s_ptr + K * i,
          // According to the identity above, lapack will write Vᵀᵀ as U.
          /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
@@ -146,6 +167,13 @@ void svd_impl(
        ss << "svd_impl: sgesvdx_ failed with code " << info;
        throw std::runtime_error(ss.str());
      }
+
+      if (ns != K) {
+        std::stringstream ss;
+        ss << "svd_impl: expected " << K << " singular values, but " << ns
+           << " were computed.";
+        throw std::runtime_error(ss.str());
+      }
    }
  });
  encoder.add_temporary(in);
@@ -153,10 +181,10 @@ 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,
--- a/mlx/backend/cpu/ternary.h
+++ b/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 (int64_t i = 0; i < out.size(); ++i) {
+    for (size_t i = 0; i < out.size(); ++i) {
      *out_ptr = op(*a_ptr, *b_ptr, *c_ptr);
      a_ptr++;
      b_ptr++;
--- a/mlx/backend/cpu/unary.h
+++ b/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, int64_t shape, int64_t stride) {
-  for (int64_t i = 0; i < shape; i += 1) {
+void unary_op(const T* a, U* out, size_t shape, size_t stride) {
+  for (size_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 {
-    int64_t shape = ndim > 0 ? a.shape().back() : 1;
-    int64_t stride = ndim > 0 ? a.strides().back() : 1;
+    size_t shape = ndim > 0 ? a.shape().back() : 1;
+    size_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 (int64_t elem = 0; elem < a.size(); elem += shape) {
+    for (size_t elem = 0; elem < a.size(); elem += shape) {
      unary_op<T, U, Op>(src + it.loc, dst + elem, shape, stride);
      it.step();
    }
--- a/mlx/backend/cpu/unary_ops.h
+++ b/mlx/backend/cpu/unary_ops.h
@@ -77,8 +77,7 @@ struct Real {
 struct Sigmoid {
  template <int N, typename T>
  Simd<T, N> operator()(Simd<T, N> 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);
+    return 1.0f / (1.0f + simd::exp(-x));
  }
  SINGLE()
 };
--- a/mlx/backend/cuda/CMakeLists.txt
+++ b/mlx/backend/cuda/CMakeLists.txt
@@ -8,6 +8,7 @@ target_sources(
  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/arange.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cu
@@ -16,13 +17,8 @@ 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
@@ -43,26 +39,22 @@ target_sources(
          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
          ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/softmax.cu
          ${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}/quantized.cu
          ${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)
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_batched_gemm_12_9.cu)
 else()
  target_sources(
-    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_0.cpp)
+    mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_batched_gemm_12_0.cpp)
 endif()

 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
@@ -154,7 +146,7 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 FetchContent_Declare(
  cudnn
  GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
-  GIT_TAG v1.14.0
+  GIT_TAG v1.12.1
  GIT_SHALLOW TRUE
  EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)
@@ -170,10 +162,6 @@ 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
--- a/mlx/backend/cuda/allocator.cpp
+++ b/mlx/backend/cuda/allocator.cpp
@@ -30,20 +30,8 @@ SmallSizePool::SmallSizePool() {
  next_free_ = buffer_;

  CHECK_CUDA_ERROR(cudaMallocManaged(&data_, small_pool_size));
-
-  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));
-  }
+  CHECK_CUDA_ERROR(
+      cudaMemAdvise(data_, small_pool_size, cudaMemAdviseSetReadMostly, 0));

  auto curr = next_free_;
  for (size_t i = 1; i < num_blocks; ++i) {
@@ -91,7 +79,7 @@ CudaAllocator::CudaAllocator()
  // TODO: Set memory limit for multi-device.
  size_t free, total;
  CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
-  memory_limit_ = total * 0.95;
+  memory_limit_ = total * 0.8;
  max_pool_size_ = memory_limit_;
 }

--- a/mlx/backend/cuda/arange.cu
+++ b/mlx/backend/cuda/arange.cu
@@ -6,33 +6,23 @@
 #include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"

-#include <cooperative_groups.h>
 #include <nvtx3/nvtx3.hpp>
+#include <thrust/device_ptr.h>
+#include <thrust/transform.h>

 namespace mlx::core {

 namespace cu {

-namespace cg = cooperative_groups;
+template <typename T>
+struct Arange {
+  const T start;
+  const T step;

-template <typename T, typename IdxT, int N_WRITES>
-__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);
+  __device__ T operator()(uint32_t i) const {
+    return start + i * step;
  }
-}
+};

 } // namespace cu

@@ -46,23 +36,19 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
  auto& encoder = cu::get_command_encoder(stream());
  encoder.set_output_array(out);

+  auto capture = encoder.capture_context();
  dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
    using CTYPE = MLX_GET_TYPE(type_tag);
    using OutType = cuda_type_t<CTYPE>;
-    constexpr int N_WRITES = 16 / sizeof(OutType);
-    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_));
-    });
+    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)});
  });
 }

--- a/mlx/backend/cuda/arg_reduce.cu
+++ b/mlx/backend/cuda/arg_reduce.cu
@@ -44,11 +44,8 @@ struct ArgMin {
  }

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

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

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

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

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

@@ -171,7 +166,6 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
          kernel,
          num_blocks,
          block_dim(),
-          0,
          in.data<T>(),
          out.data<uint32_t>(),
          out.size(),
--- a/mlx/backend/cuda/binary/binary.cuh
+++ b/mlx/backend/cuda/binary/binary.cuh
@@ -99,89 +99,39 @@ __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,
-    int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
 __global__ void binary_g_nd(
    const In* a,
    const In* b,
    Out* out,
-    IdxT size_rest,
+    IdxT size,
    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  auto block = cg::this_thread_block();
-  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;
+  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 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, int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT>
 __global__ void binary_g(
    const In* a,
    const In* b,
    Out* out,
-    IdxT size_rest,
+    IdxT size,
    const __grid_constant__ Shape shape,
    const __grid_constant__ Strides a_strides,
    const __grid_constant__ Strides b_strides,
    int ndim) {
-  auto block = cg::this_thread_block();
-  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;
+  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 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>
@@ -259,61 +209,37 @@ void binary_op_gpu_inplace(
                auto& a_strides = strides[0];
                auto& b_strides = strides[1];
                int ndim = shape.size();
-                int work_per_thread = 1;
-                auto dim0 = ndim > 0 ? shape.back() : 1;
-                auto rest = out.size() / dim0;
-                if (dim0 >= 4) {
-                  work_per_thread = 4;
-                }
-                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-                auto block_dims = get_block_dims(dim0, rest, 1);
-                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
                if (ndim <= 3) {
                  dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::binary_g_nd<
-                        Op,
-                        InType,
-                        OutType,
-                        IdxT,
-                        dims_constant(),
-                        1>;
-                    if (work_per_thread == 4) {
-                      kernel = cu::binary_g_nd<
-                          Op,
-                          InType,
-                          OutType,
-                          IdxT,
-                          dims_constant(),
-                          4>;
-                    }
+                    auto [num_blocks, block_dims] =
+                        get_launch_args(out, large());
                    encoder.add_kernel_node(
-                        kernel,
-                        {num_blocks_x, num_blocks_y},
+                        cu::binary_g_nd<
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
+                        num_blocks,
                        block_dims,
-                        0,
                        a.data<InType>(),
                        b.data<InType>(),
                        out.data<OutType>(),
-                        rest,
+                        out.size(),
                        const_param<dims_constant()>(shape),
                        const_param<dims_constant()>(a_strides),
                        const_param<dims_constant()>(b_strides));
                  });
                } else {
-                  auto kernel = cu::binary_g<Op, InType, OutType, IdxT, 1>;
-                  if (work_per_thread == 4) {
-                    kernel = cu::binary_g<Op, InType, OutType, IdxT, 4>;
-                  }
+                  auto [num_blocks, block_dims] = get_launch_args(out, large());
                  encoder.add_kernel_node(
-                      kernel,
-                      {num_blocks_x, num_blocks_y},
+                      cu::binary_g<Op, InType, OutType, IdxT>,
+                      num_blocks,
                      block_dims,
-                      0,
                      a.data<InType>(),
                      b.data<InType>(),
                      out.data<OutType>(),
-                      rest,
+                      out.size(),
                      const_param(shape),
                      const_param(a_strides),
                      const_param(b_strides),
@@ -338,7 +264,6 @@ void binary_op_gpu_inplace(
                kernel,
                num_blocks,
                block_dims,
-                0,
                a.data<InType>(),
                b.data<InType>(),
                out.data<OutType>(),
@@ -376,4 +301,54 @@ void binary_op_gpu(
    binary_op_gpu<cu::func>(inputs, out, name(), s);                  \
  }

+BINARY_GPU(Add)
+BINARY_GPU(ArcTan2)
+BINARY_GPU(Divide)
+BINARY_GPU(Remainder)
+BINARY_GPU(Greater)
+BINARY_GPU(GreaterEqual)
+BINARY_GPU(Less)
+BINARY_GPU(LessEqual)
+BINARY_GPU(LogicalAnd)
+BINARY_GPU(LogicalOr)
+BINARY_GPU(LogAddExp)
+BINARY_GPU(Maximum)
+BINARY_GPU(Minimum)
+BINARY_GPU(Multiply)
+BINARY_GPU(NotEqual)
+BINARY_GPU(Power)
+BINARY_GPU(Subtract)
+
+void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Equal::eval_gpu");
+  auto& s = out.primitive().stream();
+  if (equal_nan_) {
+    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
+  } else {
+    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
+  }
+}
+
+void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
+  auto& s = out.primitive().stream();
+  switch (op_) {
+    case BitwiseBinary::And:
+      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Or:
+      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::Xor:
+      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::LeftShift:
+      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
+      break;
+    case BitwiseBinary::RightShift:
+      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
+      break;
+  }
+}
+
 } // namespace mlx::core
--- a/mlx/backend/cuda/binary/CMakeLists.txt
+++ b/mlx/backend/cuda/binary/CMakeLists.txt
@@ -1,21 +0,0 @@
-target_sources(
-  mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/add.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arctan2.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/bitwise_binary.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/divide.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/greater_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/less_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_and.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/logical_or.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/log_add_exp.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/minimum.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/maximum.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/multiply.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/power.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/remainder.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/not_equal.cu
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/subtract.cu)
--- a/mlx/backend/cuda/binary/add.cu
+++ b/mlx/backend/cuda/binary/add.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Add)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/arctan2.cu
+++ b/mlx/backend/cuda/binary/arctan2.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(ArcTan2)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/bitwise_binary.cu
+++ b/mlx/backend/cuda/binary/bitwise_binary.cu
@@ -1,27 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-void BitwiseBinary::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("BitwiseBinary::eval_gpu");
-  auto& s = out.primitive().stream();
-  switch (op_) {
-    case BitwiseBinary::And:
-      binary_op_gpu<cu::BitwiseAnd>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Or:
-      binary_op_gpu<cu::BitwiseOr>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::Xor:
-      binary_op_gpu<cu::BitwiseXor>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::LeftShift:
-      binary_op_gpu<cu::LeftShift>(inputs, out, name(), s);
-      break;
-    case BitwiseBinary::RightShift:
-      binary_op_gpu<cu::RightShift>(inputs, out, name(), s);
-      break;
-  }
-}
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/divide.cu
+++ b/mlx/backend/cuda/binary/divide.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Divide)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/equal.cu
+++ b/mlx/backend/cuda/binary/equal.cu
@@ -1,15 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-void Equal::eval_gpu(const std::vector<array>& inputs, array& out) {
-  nvtx3::scoped_range r("Equal::eval_gpu");
-  auto& s = out.primitive().stream();
-  if (equal_nan_) {
-    binary_op_gpu<cu::NaNEqual>(inputs, out, name(), s);
-  } else {
-    binary_op_gpu<cu::Equal>(inputs, out, name(), s);
-  }
-}
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/greater.cu
+++ b/mlx/backend/cuda/binary/greater.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Greater)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/greater_equal.cu
+++ b/mlx/backend/cuda/binary/greater_equal.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(GreaterEqual)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/less.cu
+++ b/mlx/backend/cuda/binary/less.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Less)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/less_equal.cu
+++ b/mlx/backend/cuda/binary/less_equal.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LessEqual)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/log_add_exp.cu
+++ b/mlx/backend/cuda/binary/log_add_exp.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogAddExp)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/logical_and.cu
+++ b/mlx/backend/cuda/binary/logical_and.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogicalAnd)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/logical_or.cu
+++ b/mlx/backend/cuda/binary/logical_or.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(LogicalOr)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/maximum.cu
+++ b/mlx/backend/cuda/binary/maximum.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Maximum)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/minimum.cu
+++ b/mlx/backend/cuda/binary/minimum.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Minimum)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/multiply.cu
+++ b/mlx/backend/cuda/binary/multiply.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Multiply)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/not_equal.cu
+++ b/mlx/backend/cuda/binary/not_equal.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(NotEqual)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/power.cu
+++ b/mlx/backend/cuda/binary/power.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Power)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/remainder.cu
+++ b/mlx/backend/cuda/binary/remainder.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Remainder)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary/subtract.cu
+++ b/mlx/backend/cuda/binary/subtract.cu
@@ -1,7 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/binary/binary.cuh"
-
-namespace mlx::core {
-BINARY_GPU(Subtract)
-} // namespace mlx::core
--- a/mlx/backend/cuda/binary_two.cu
+++ b/mlx/backend/cuda/binary_two.cu
@@ -127,99 +127,45 @@ binary_two_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
  }
 }

-template <
-    typename Op,
-    typename In,
-    typename Out,
-    typename IdxT,
-    int NDIM,
-    int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
 __global__ void binary_two_g_nd(
    const In* a,
    const In* b,
    Out* out_a,
    Out* out_b,
-    IdxT size_rest,
+    IdxT size,
    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
-  auto block = cg::this_thread_block();
-  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;
+  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 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, int N_READS>
+template <typename Op, typename In, typename Out, typename IdxT>
 __global__ void binary_two_g(
    const In* a,
    const In* b,
    Out* out_a,
    Out* out_b,
-    IdxT size_rest,
+    IdxT size,
    const __grid_constant__ Shape shape,
    const __grid_constant__ Strides a_strides,
    const __grid_constant__ Strides b_strides,
    int ndim) {
-  auto block = cg::this_thread_block();
-  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;
+  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 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>
@@ -279,64 +225,40 @@ void binary_two_op_gpu_inplace(
                auto& a_strides = strides[0];
                auto& b_strides = strides[1];
                int ndim = shape.size();
-                int work_per_thread = 1;
-                auto dim0 = ndim > 0 ? shape.back() : 1;
-                auto rest = out_a.size() / dim0;
-                if (dim0 >= 4) {
-                  work_per_thread = 4;
-                }
-                dim0 = (dim0 + work_per_thread - 1) / work_per_thread;
-                auto block_dims = get_block_dims(dim0, rest, 1);
-                uint32_t num_blocks_x = cuda::ceil_div(dim0, block_dims.x);
-                uint32_t num_blocks_y = cuda::ceil_div(rest, block_dims.y);
-
                if (ndim <= 3) {
                  dispatch_1_2_3(ndim, [&](auto dims_constant) {
-                    auto kernel = cu::binary_two_g_nd<
-                        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>;
-                    }
+                    auto [num_blocks, block_dims] =
+                        get_launch_args(out_a, large());
                    encoder.add_kernel_node(
-                        kernel,
-                        {num_blocks_x, num_blocks_y},
+                        cu::binary_two_g_nd<
+                            Op,
+                            InType,
+                            OutType,
+                            IdxT,
+                            dims_constant()>,
+                        num_blocks,
                        block_dims,
-                        0,
                        a.data<InType>(),
                        b.data<InType>(),
                        out_a.data<OutType>(),
                        out_b.data<OutType>(),
-                        rest,
+                        out_a.size(),
                        const_param<dims_constant()>(shape),
                        const_param<dims_constant()>(a_strides),
                        const_param<dims_constant()>(b_strides));
                  });
                } else {
-                  auto kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 1>;
-                  if (work_per_thread == 4) {
-                    kernel = cu::binary_two_g<Op, InType, OutType, IdxT, 4>;
-                  }
+                  auto [num_blocks, block_dims] =
+                      get_launch_args(out_a, large());
                  encoder.add_kernel_node(
-                      kernel,
-                      {num_blocks_x, num_blocks_y},
+                      cu::binary_two_g<Op, InType, OutType, IdxT>,
+                      num_blocks,
                      block_dims,
-                      0,
                      a.data<InType>(),
                      b.data<InType>(),
                      out_a.data<OutType>(),
                      out_b.data<OutType>(),
-                      rest,
+                      out_a.size(),
                      const_param(shape),
                      const_param(a_strides),
                      const_param(b_strides),
@@ -365,7 +287,6 @@ void binary_two_op_gpu_inplace(
                kernel,
                num_blocks,
                block_dims,
-                0,
                a.data<InType>(),
                b.data<InType>(),
                out_a.data<OutType>(),
--- a/mlx/backend/cuda/compiled.cpp
+++ b/mlx/backend/cuda/compiled.cpp
@@ -104,41 +104,10 @@ struct FusedKernelBuilder {
          "  }\n";
    }

-    // Vectorized read loop
-    if (contiguous) {
-      for (size_t i = 0; i < inputs.size(); ++i) {
-        const auto& x = inputs[i];
-        if (is_scalar(x) || is_constant(i)) {
-          continue;
-        }
-        const std::string& xname = namer.get_name(x);
-        std::string type = dtype_to_cuda_type(x.dtype());
-        os += fmt::format(
-            "  auto vec_{0} = load_vector<work_per_thread, {1}>({0} + index, 0, size - index, 0);\n",
-            xname,
-            type);
-      }
-    }
-
-    // Create some space for the outputs
-    for (const auto& x : outputs) {
-      const std::string& xname = namer.get_name(x);
-      std::string type = dtype_to_cuda_type(x.dtype());
-      os += fmt::format(
-          "  AlignedVector<{}, work_per_thread> vec_{};\n", type, xname);
-    }
-
    // Work loop
-    if (!contiguous) {
-      os +=
-          "\n"
-          "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
-    } else {
-      os +=
-          "\n"
-          "  #pragma unroll\n"
-          "  for (int i = 0; i < work_per_thread; i++) {\n";
-    }
+    os +=
+        "\n"
+        "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";

    // Read inputs.
    for (size_t i = 0; i < inputs.size(); ++i) {
@@ -153,7 +122,7 @@ struct FusedKernelBuilder {
      } else if (is_scalar(x)) {
        value = fmt::format("{}[0]", xname);
      } else if (contiguous) {
-        value = fmt::format("vec_{}[i]", xname);
+        value = fmt::format("{}[index]", xname);
      } else {
        value = fmt::format("{}[{}_idx]", xname, xname);
      }
@@ -181,30 +150,25 @@ struct FusedKernelBuilder {

    // Write output.
    for (const auto& x : outputs) {
-      os += fmt::format("    vec_{0}[i] = tmp_{0};\n", namer.get_name(x));
+      os += fmt::format("    {0}[index] = tmp_{0};\n", namer.get_name(x));
    }

    // End of work loop
+    os +=
+        "\n"
+        "    index++;\n";
    if (!contiguous) {
-      os += "\n";
      for (size_t i = 0; i < inputs.size(); ++i) {
        const auto& x = inputs[i];
        const std::string& xname = namer.get_name(x);
        if (is_scalar(x) || is_constant(i)) {
          continue;
        }
-        os += fmt::format("    {0}_idx += {0}_strides[NDIM - 1];\n", xname);
+        os += "    " + xname + "_idx += " + xname + "_strides[NDIM - 1];\n";
      }
    }
    os += "  }\n";

-    // Store the output to global memory
-    for (const auto& x : outputs) {
-      os += fmt::format(
-          "  store_vector({0} + index, 0, vec_{0}, size - index);\n",
-          namer.get_name(x));
-    }
-
    os += "}\n";
  }
 };
@@ -228,15 +192,6 @@ void Compiled::eval_gpu(
  nvtx3::scoped_range r("Compiled::eval_gpu");
  auto& s = stream();

-  // Determine the work per thread for the vectorized reads/writes. We take it
-  // as 16 over the max itemsize for the outputs. Another heuristic could be
-  // over the max itemsize of all arrays.
-  int max_size = 1;
-  for (const auto& x : outputs) {
-    max_size = (max_size > x.itemsize()) ? max_size : x.itemsize();
-  }
-  int work_per_thread = 16 / max_size;
-
  cu::JitModule& mod = cu::get_jit_module(s.device, lib_name(), [&]() {
    // Build source code.
    cu::FusedKernelBuilder builder{
@@ -250,25 +205,29 @@ void Compiled::eval_gpu(
    builder.os += "\n} // namespace mlx::core::cu\n";
    // Build kernel names.
    std::vector<std::string> kernel_names;
-    kernel_names.push_back(fmt::format(
-        "mlx::core::cu::{}_contiguous<uint32_t, {}>",
-        lib_name(),
-        work_per_thread));
-    kernel_names.push_back(fmt::format(
-        "mlx::core::cu::{}_contiguous<int64_t, {}>",
-        lib_name(),
-        work_per_thread));
-    for (auto wpt : std::array<int, 2>{1, work_per_thread}) {
+    for (auto work_per_thread : std::array<int, 2>{1, 4}) {
+      kernel_names.push_back(fmt::format(
+          "mlx::core::cu::{}_contiguous<uint32_t, {}>",
+          lib_name(),
+          work_per_thread));
+      kernel_names.push_back(fmt::format(
+          "mlx::core::cu::{}_contiguous<int64_t, {}>",
+          lib_name(),
+          work_per_thread));
      for (int i = 1; i <= MAX_NDIM; ++i) {
        kernel_names.push_back(fmt::format(
-            "mlx::core::cu::{}_strided<{}, uint32_t, {}>", lib_name(), i, wpt));
+            "mlx::core::cu::{}_strided<{}, uint32_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
        kernel_names.push_back(fmt::format(
-            "mlx::core::cu::{}_strided<{}, int64_t, {}>", lib_name(), i, wpt));
+            "mlx::core::cu::{}_strided<{}, int64_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
      }
    }
-
-    return std::make_tuple(
-        false, std::move(builder.os), std::move(kernel_names));
+    return std::make_pair(std::move(builder.os), std::move(kernel_names));
  });

  // Collapse contiguous dims to route to a faster kernel if possible. Also
@@ -310,6 +269,7 @@ void Compiled::eval_gpu(
  }

  // Choose work per thread
+  int work_per_thread = 4;
  if (!contiguous && shape.back() % work_per_thread != 0) {
    work_per_thread = 1;
  }
@@ -332,10 +292,10 @@ void Compiled::eval_gpu(
    encoder.set_output_array(out);
  }

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

 } // namespace mlx::core
--- a/mlx/backend/cuda/conv.cpp
+++ b/mlx/backend/cuda/conv.cpp
@@ -1,214 +1,257 @@
 // 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>
+#include <numeric>

 namespace mlx::core {

 namespace {

-// Alias for better readability.
-#define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
-#define CONV_BACKWARD_INPUT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DATA_DESCRIPTOR
-#define CONV_BACKWARD_WEIGHT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
-
-// Custom placeholder representing fallback kernel.
-#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
+// Not all engines support it so can not use this API now.
+#define MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API 0

 struct ConvCacheKey {
  int device_id;
-  cudnnDataType_t cudnn_dtype;
+  cudnnBackendDescriptorType_t backend_type;
+  cudnnDataType_t cudnn_type;
  std::array<int, MAX_NDIM> input_shape;
-  std::array<int, MAX_NDIM> weight_shape;
-  std::array<int, MAX_NDIM> stride;
+  std::array<int, MAX_NDIM> filter_shape;
  std::array<int, MAX_NDIM> padding_lo;
  std::array<int, MAX_NDIM> padding_hi;
+  std::array<int, MAX_NDIM> stride;
  std::array<int, MAX_NDIM> dilation;
  int groups;
-  bool flip;
  uint8_t input_alignment;
-  uint8_t weight_alignment;
+  uint8_t filter_alignment;
  uint8_t output_alignment;
 };

 auto& conv_cache() {
-  static LRUBytesKeyCache<
-      ConvCacheKey,
-      std::pair<
-          cudnnBackendDescriptorType_t,
-          std::optional<cudnn_frontend::ExecutionPlan>>>
-      cache("MLX_CUDA_CONV_CACHE_SIZE", /* default_capacity */ 128);
+  static LRUBytesKeyCache<ConvCacheKey, cudnn_frontend::ExecutionPlan> cache(
+      /* capacity */ 128);
  return cache;
 }

-auto get_conv_op_settings(
-    cudnnBackendDescriptorType_t backend_type,
-    array& x,
-    array& w,
-    array& y,
-    const std::vector<int>& kernel_strides,
-    const std::vector<int>& padding_lo_,
-    const std::vector<int>& padding_hi_,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation) {
-  auto padding_lo = convert_vector<int64_t>(padding_lo_);
-  auto padding_hi = convert_vector<int64_t>(padding_hi_);
+template <typename T, typename U>
+inline std::vector<T> convert_vector(const std::vector<U>& vec) {
+  return std::vector<T>(vec.begin(), vec.end());
+}

-  if (backend_type == CONV_BACKWARD_INPUT) {
-    for (int i = 0; i < padding_lo.size(); ++i) {
-      int wt_size = 1 + kernel_dilation[i] * (w.shape(1 + i) - 1);
-      padding_lo[i] = wt_size - padding_lo[i] - 1;
-      int in_size = 1 + kernel_strides[i] * (x.shape(1 + i) - 1);
-      int out_size = 1 + input_dilation[i] * (y.shape(1 + i) - 1);
-      padding_hi[i] = out_size - in_size + padding_hi[i];
-    }
-    return std::make_tuple(
-        convert_vector<int64_t>(input_dilation),
-        std::move(padding_lo),
-        std::move(padding_hi),
-        convert_vector<int64_t>(kernel_dilation));
+template <typename T>
+inline std::array<T, MAX_NDIM> fixed_vector(const std::vector<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;
+}

-  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    padding_hi = padding_lo;
-    return std::make_tuple(
-        convert_vector<int64_t>(kernel_dilation),
-        std::move(padding_lo),
-        std::move(padding_hi),
-        convert_vector<int64_t>(kernel_strides));
+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(shape, strides);
+}

-  } else {
-    return std::make_tuple(
-        convert_vector<int64_t>(kernel_strides),
-        std::move(padding_lo),
-        std::move(padding_hi),
-        convert_vector<int64_t>(kernel_dilation));
+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)));
  }
 }

-std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
-    cu::CommandEncoder& encoder,
+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,
-    array& x,
-    array& w,
-    array& y,
-    const SmallVector<int64_t>& stride,
-    const SmallVector<int64_t>& padding_lo,
-    const SmallVector<int64_t>& padding_hi,
-    const SmallVector<int64_t>& dilation) {
-  try {
-    auto compute_dtype = (dtype == float16 || dtype == bfloat16)
-        ? CUDNN_DATA_FLOAT
-        : dtype_to_cudnn_type(dtype);
-    auto conv_desc = cudnn_frontend::ConvDescBuilder()
-                         .setDataType(compute_dtype)
-                         .setMathMode(CUDNN_CROSS_CORRELATION)
-                         .setNDims(stride.size())
-                         .setStrides(stride.size(), stride.data())
-                         .setPrePadding(padding_lo.size(), padding_lo.data())
-                         .setPostPadding(padding_hi.size(), padding_hi.data())
-                         .setDilation(dilation.size(), dilation.data())
+    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,
+    const array& in,
+    const array& wt,
+    array& out) {
+  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] = {
+      const_cast<void*>(in.data<void>()),
+      const_cast<void*>(wt.data<void>()),
+      out.data<void>(),
+  };
+
+  auto variantPack = cudnn_frontend::VariantPackBuilder()
+                         .setWorkspacePointer(workspace.data<void>())
+                         .setDataPointers(3, data_ptrs)
+                         .setUids(3, uids)
                         .build();

-    auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .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();
+  auto handle = encoder.device().cudnn_handle();
+  cudnnSetStream(handle, encoder.stream());

-    std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
-    return cudnn_frontend::OperationGraphBuilder()
-        .setHandle(encoder.device().cudnn_handle())
-        .setOperationGraph(ops.size(), ops.data())
-        .build();
-  } catch (cudnn_frontend::cudnnException& error) {
-    if (error.getCudnnStatus() != CUDNN_STATUS_BAD_PARAM) {
-      throw;
-    }
-    return std::nullopt;
+#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;
 }

-// 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.
-std::tuple<array, array, array> prepare_args(
+bool try_engines(
    cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
-    array in,
-    array wt,
-    array out,
-    int groups,
-    Stream s) {
-  // Transpose the args depending on the backend type.
-  // TODO: Handle groups.
-  if (backend_type == CONV_BACKWARD_INPUT) {
-    wt = group_transpose(wt, groups, 0, 0, -1, s);
-  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    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.
-    Shape shape(out.shape());
-    std::swap(shape.front(), shape.back());
-    Strides strides(shape.size(), 1);
-    for (int i = shape.size() - 2; i >= 0; --i) {
-      strides[i] = shape[i + 1] * strides[i + 1];
+    cudnn_frontend::EngineConfigList& configs,
+    const ConvCacheKey& cache_key,
+    const std::string& op_graph_tag,
+    const array& in,
+    const array& wt,
+    array& out) {
+  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, in, wt, out)) {
+        conv_cache().emplace(cache_key, std::move(plan));
+        return true;
+      }
+    } catch (cudnn_frontend::cudnnException&) {
    }
-    array intermediate(std::move(shape), out.dtype(), nullptr, {});
-    intermediate.copy_shared_buffer(
-        out, std::move(strides), {true, true, false}, out.data_size());
-    out = intermediate;
  }
+  return false;
+}
+
+} // namespace
+
+void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Convolution::eval_gpu");
+  if (out.size() == 0) {
+    return;
+  }
+
+  assert(inputs.size() == 2);
+  array in = inputs[0];
+  array wt = inputs[1];
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);

  // cuDNN requires contiguous input.
+  // TODO: Handle NCHW format specially.
  if (!in.flags().row_contiguous) {
    in = contiguous_copy_gpu(in, s);
    encoder.add_temporary(in);
@@ -218,200 +261,80 @@ std::tuple<array, array, array> prepare_args(
    encoder.add_temporary(wt);
  }

-  return {std::move(in), std::move(wt), std::move(out)};
-}
-
-// Get the x/w/y args from the in/wt/out args depending on backend type.
-inline std::tuple<array&, array&, array&> dispatch_args(
-    cudnnBackendDescriptorType_t backend_type,
-    array& in,
-    array& wt,
-    array& out) {
-  switch (backend_type) {
-    case CONV_BACKWARD_INPUT:
-      return {out, wt, in};
-    case CONV_BACKWARD_WEIGHT:
-      return {in, out, wt};
-    default:
-      return {in, wt, out};
-  }
-}
-
-// Register inputs and outputs before actually running conv op. Can only be
-// called once per eval_gpu.
-void register_args(
-    cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
-    array& in,
-    array& wt,
-    array& intermediate_out,
-    array& final_out) {
  encoder.set_input_array(in);
  encoder.set_input_array(wt);
-  encoder.set_output_array(final_out);
+  encoder.set_output_array(out);

-  if (backend_type == CONV_BACKWARD_WEIGHT) {
-    // Turn |out| into a strided array, which will have C_in and C_out swapped
-    // in vjp and the final |grad_weight| will then be contiguous.
-    Strides strides = intermediate_out.strides();
-    std::swap(strides.front(), strides.back());
-    final_out.copy_shared_buffer(
-        intermediate_out,
-        std::move(strides),
-        {false, false, false},
-        intermediate_out.data_size());
-  }
-}
-
-} // namespace
-
-void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
-  nvtx3::scoped_range r("Convolution::eval_gpu");
-  if (out_.size() == 0) {
-    return;
-  }
-
-  assert(inputs.size() == 2);
-  array in = inputs[0];
-  array wt = inputs[1];
-  array out = out_;
-  out.set_data(allocator::malloc(out.nbytes()));
-  Dtype dtype = out.dtype();
-
-  auto& s = stream();
-  auto& encoder = cu::get_command_encoder(s);
+  auto backend_type = CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR;
+  auto cudnn_type = dtype_to_cudnn_type(in.dtype());

  // Search cache.
  ConvCacheKey cache_key{
      encoder.device().cuda_device(),
-      dtype_to_cudnn_type(dtype),
-      vector_key(in.shape()),
-      vector_key(wt.shape()),
-      vector_key(kernel_strides_),
-      vector_key(padding_lo_),
-      vector_key(padding_hi_),
-      vector_key(kernel_dilation_),
+      backend_type,
+      cudnn_type,
+      fixed_vector(in.shape()),
+      fixed_vector(wt.shape()),
+      fixed_vector(padding_lo_),
+      fixed_vector(padding_hi_),
+      fixed_vector(kernel_strides_),
+      fixed_vector(kernel_dilation_),
      groups_,
-      flip_,
      get_alignment(in),
      get_alignment(wt),
      get_alignment(out)};
  if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
-    auto& [backend_type, plan] = it->second;
-    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);
+    if (!execute_plan(encoder, it->second, in, wt, out)) {
+      throw std::runtime_error("Cached convolution plan failed to execute.");
    }
    return;
  }

-  // There is no reliable way to deduce the proper cuDNN backend for the
-  // convolution, so we make a best guess and then try.
-  SmallVector<cudnnBackendDescriptorType_t, 2> try_backends;
-  if (flip_) {
-    // When weight is flipped, we assume it is backward input convolution.
-    try_backends.push_back(CONV_BACKWARD_INPUT);
-  } else {
-    // Otherwise it could be backward weight convolution or forward convolution,
-    // mathematically there is no difference so we have to use heuristics.
-    // Empirically backward convolutions have large kernel dimensions, and
-    // usually have |in| and |wt| transposed.
-    if (!in.flags().row_contiguous && !wt.flags().row_contiguous &&
-        wt.shape(2) > out.shape(2)) {
-      try_backends = {CONV_BACKWARD_WEIGHT, CONV_FORWARD};
-    } else {
-      try_backends = {CONV_FORWARD, CONV_BACKWARD_WEIGHT};
-    }
+  // Build operation graph.
+  auto compute_data_type = (in.dtype() == float16 || in.dtype() == bfloat16)
+      ? CUDNN_DATA_FLOAT
+      : cudnn_type;
+
+  auto stride = convert_vector<int64_t>(kernel_strides_);
+  auto padding_lo = convert_vector<int64_t>(padding_lo_);
+  auto padding_hi = convert_vector<int64_t>(padding_hi_);
+  auto dilation = convert_vector<int64_t>(kernel_dilation_);
+
+  auto conv_desc = cudnn_frontend::ConvDescBuilder()
+                       .setDataType(compute_data_type)
+                       .setMathMode(CUDNN_CROSS_CORRELATION)
+                       .setNDims(stride.size())
+                       .setStrides(stride.size(), stride.data())
+                       .setPrePadding(padding_lo.size(), padding_lo.data())
+                       .setPostPadding(padding_hi.size(), padding_hi.data())
+                       .setDilation(dilation.size(), dilation.data())
+                       .build();
+
+  auto op = cudnn_frontend::OperationBuilder(backend_type)
+                .setxDesc(build_tensor('x', in))
+                .setwDesc(build_tensor('w', wt))
+                .setyDesc(build_tensor('y', out))
+                .setcDesc(conv_desc)
+                .build();
+
+  std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
+  auto op_graph = cudnn_frontend::OperationGraphBuilder()
+                      .setHandle(encoder.device().cudnn_handle())
+                      .setOperationGraph(ops.size(), ops.data())
+                      .build();
+
+  // Try to run plans based on heuristics.
+  auto configs = get_engine_configs(backend_type, in.dtype(), op_graph);
+  auto op_graph_tag = op_graph.getTag();
+  if (try_engines(encoder, configs, cache_key, op_graph_tag, in, wt, out)) {
+    return;
  }
-
-  // Try to build op graph.
-  cudnnBackendDescriptorType_t backend_type;
-  std::optional<cudnn_frontend::OperationGraph> op_graph;
-  for (auto try_backend : try_backends) {
-    auto [in_copy, wt_copy, out_copy] =
-        prepare_args(encoder, try_backend, in, wt, out, 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,
-        x,
-        w,
-        y,
-        kernel_strides_,
-        padding_lo_,
-        padding_hi_,
-        kernel_dilation_,
-        input_dilation_);
-    op_graph = build_conv_op_graph(
-        encoder,
-        try_backend,
-        dtype,
-        x,
-        w,
-        y,
-        stride,
-        padding_lo,
-        padding_hi,
-        dilation);
-    if (op_graph) {
-      backend_type = try_backend;
-      in = std::move(in_copy);
-      wt = std::move(wt_copy);
-      out = std::move(out_copy);
-      break;
-    }
+  // Then try fallback plans.
+  configs = get_engine_configs(backend_type, in.dtype(), op_graph);
+  if (try_engines(encoder, configs, cache_key, op_graph_tag, in, wt, out)) {
+    return;
  }
-
-  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;
-      }
-    }
-  }
-
-  // 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));
+  throw std::runtime_error("Unable to find an engine for convolution.");
 }

 } // namespace mlx::core
--- a/mlx/backend/cuda/conv/conv.h
+++ b/mlx/backend/cuda/conv/conv.h
@@ -1,126 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/gpu/copy.h"
-
-namespace mlx::core {
-
-template <int NDIM>
-struct ConvParams {
-  int N; // Batch size
-  int C; // In channels
-  int O; // Out channels
-  int strides[NDIM];
-  int padding[NDIM];
-  int kernel_dilation[NDIM];
-  int input_dilation[NDIM];
-  int groups;
-  bool flip;
-  int in_spatial_dims[NDIM];
-  int wt_spatial_dims[NDIM];
-  int out_spatial_dims[NDIM];
-  int64_t in_strides[NDIM + 2];
-
-  ConvParams(
-      const array& in,
-      const array& wt,
-      const array& out,
-      const std::vector<int>& strides,
-      const std::vector<int>& padding,
-      const std::vector<int>& kernel_dilation,
-      const std::vector<int>& input_dilation,
-      int groups,
-      bool flip)
-      : N(in.shape(0)),
-        C(in.shape(-1)),
-        O(wt.shape(0)),
-        groups(groups),
-        flip(flip) {
-    std::copy_n(strides.begin(), NDIM, this->strides);
-    std::copy_n(padding.begin(), NDIM, this->padding);
-    std::copy_n(kernel_dilation.begin(), NDIM, this->kernel_dilation);
-    std::copy_n(input_dilation.begin(), NDIM, this->input_dilation);
-    std::copy_n(in.shape().begin() + 1, NDIM, this->in_spatial_dims);
-    std::copy_n(wt.shape().begin() + 1, NDIM, this->wt_spatial_dims);
-    std::copy_n(out.shape().begin() + 1, NDIM, this->out_spatial_dims);
-    std::copy_n(in.strides().begin(), NDIM + 2, this->in_strides);
-  }
-};
-
-void gemm_grouped_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s);
-
-void gemm_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    bool flip,
-    Stream s);
-
-inline void gemm_conv(
-    cu::CommandEncoder& encoder,
-    array in,
-    array wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    int groups,
-    bool flip,
-    Stream s) {
-  if (!in.flags().row_contiguous) {
-    in = contiguous_copy_gpu(in, s);
-    encoder.add_temporary(in);
-  }
-  if (!wt.flags().row_contiguous) {
-    wt = contiguous_copy_gpu(wt, s);
-    encoder.add_temporary(wt);
-  }
-
-  if (groups == 1) {
-    gemm_conv(
-        encoder,
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        flip,
-        s);
-  } else {
-    gemm_grouped_conv(
-        encoder,
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        groups,
-        flip,
-        s);
-  }
-}
-
-} // namespace mlx::core
--- a/mlx/backend/cuda/conv/gemm_conv.cu
+++ b/mlx/backend/cuda/conv/gemm_conv.cu
@@ -1,217 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/conv/conv.h"
-#include "mlx/backend/cuda/gemms/cublas_gemm.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/dtype_utils.h"
-
-#include <cooperative_groups.h>
-
-namespace mlx::core {
-
-namespace cu {
-
-namespace cg = cooperative_groups;
-
-template <typename T, int NDIM>
-__global__ void naive_unfold_nd(
-    const T* in,
-    T* out,
-    int filter_size,
-    int out_pixels,
-    const __grid_constant__ ConvParams<NDIM> params) {
-  auto block = cg::this_thread_block();
-  auto tid = block.group_index();
-  auto lid = block.thread_index();
-
-  int index_batch = tid.z / out_pixels; // [0, N)
-  int index_out_spatial = tid.z % out_pixels; // [0, H_out * W_out)
-  int index_wt_spatial =
-      tid.x * block.dim_threads().x + lid.x; // [0, H_wt * W_wt)
-
-  if (index_wt_spatial >= filter_size / params.C) {
-    return;
-  }
-
-  in += tid.y; // [0, C)
-  out += tid.z * filter_size + index_wt_spatial * params.C + tid.y;
-
-  bool valid = index_batch < params.N;
-
-  // Get the coordinates in input.
-  int index_in[NDIM] = {};
-#pragma unroll
-  for (int i = NDIM - 1; i >= 0; --i) {
-    int index_out = index_out_spatial % params.out_spatial_dims[i];
-    int index_wt = index_wt_spatial % params.wt_spatial_dims[i];
-
-    if (params.flip) {
-      index_wt = params.wt_spatial_dims[i] - index_wt - 1;
-    }
-
-    int index = index_out * params.strides[i] - params.padding[i] +
-        index_wt * params.kernel_dilation[i];
-    int index_max =
-        1 + params.input_dilation[i] * (params.in_spatial_dims[i] - 1);
-
-    valid &= (index >= 0) && (index < index_max) &&
-        (index % params.input_dilation[i] == 0);
-
-    index_in[i] = index / params.input_dilation[i];
-
-    index_out_spatial /= params.out_spatial_dims[i];
-    index_wt_spatial /= params.wt_spatial_dims[i];
-  }
-
-  if (valid) {
-    int in_offset = index_batch * params.in_strides[0];
-#pragma unroll
-    for (int i = 0; i < NDIM; ++i) {
-      in_offset += index_in[i] * params.in_strides[i + 1];
-    }
-    *out = in[in_offset];
-  } else {
-    *out = T{0};
-  }
-}
-
-} // namespace cu
-
-template <int NDIM>
-array unfold_inputs_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    int mat_M,
-    int mat_K,
-    int mat_N,
-    ConvParams<NDIM>& params) {
-  array unfolded({mat_M, mat_K}, in.dtype(), nullptr, {});
-  unfolded.set_data(allocator::malloc(unfolded.nbytes()));
-  encoder.add_temporary(unfolded);
-
-  int filter_size = params.C;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    filter_size *= params.wt_spatial_dims[i];
-  }
-
-  int out_pixels = 1;
-#pragma unroll
-  for (int i = 0; i < NDIM; ++i) {
-    out_pixels *= params.out_spatial_dims[i];
-  }
-
-  int wt_spatial_size = mat_K / params.C;
-  dim3 block_dims;
-  block_dims.x = std::min(std::max(wt_spatial_size, 32), 1024);
-  dim3 num_blocks;
-  num_blocks.x = cuda::ceil_div(wt_spatial_size, block_dims.x);
-  num_blocks.y = params.C;
-  num_blocks.z = mat_M;
-
-  encoder.set_input_array(in);
-  encoder.set_output_array(unfolded);
-  dispatch_float_types(in.dtype(), "unfold", [&](auto type_tag) {
-    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-    encoder.add_kernel_node(
-        cu::naive_unfold_nd<DataType, NDIM>,
-        num_blocks,
-        block_dims,
-        0,
-        in.data<DataType>(),
-        unfolded.data<DataType>(),
-        filter_size,
-        out_pixels,
-        params);
-  });
-
-  return unfolded;
-}
-
-template <int NDIM>
-void gemm_conv_nd(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    ConvParams<NDIM>& params,
-    Stream s) {
-  // Get gemm shapes.
-  int mat_M = out.size() / params.O; // N * H_out * W_out
-  int mat_K = wt.size() / params.O; // C * H_wt * W_wt
-  int mat_N = params.O; // O
-
-  // Unfold input to (N * H_out * W_out, C * H_wt * W_wt) for gemm.
-  array in_unfolded =
-      unfold_inputs_nd<NDIM>(encoder, in, mat_M, mat_K, mat_N, params);
-
-  // Reshape weight to (C * H_wt * W_wt, O) for gemm.
-  array wt_reshaped({mat_K, mat_N}, wt.dtype(), nullptr, {});
-  wt_reshaped.copy_shared_buffer(
-      wt,
-      {1, mat_K},
-      {false, false, /* col_contiguous */ true},
-      wt.data_size());
-
-  // Single batch.
-  Shape batch_shape{1};
-  Strides a_batch_strides{0};
-  Strides b_batch_strides{0};
-
-  // Run matmul.
-  CublasGemm gemm(
-      encoder.device(),
-      in.dtype(),
-      false, // a_transposed
-      mat_M, // a_rows
-      mat_K, // a_cols
-      mat_K, // lda
-      true, // b_transposed
-      mat_K, // b_rows
-      mat_N, // b_cols
-      mat_K, // ldb
-      batch_shape.back(),
-      a_batch_strides.back(),
-      b_batch_strides.back());
-  gemm.run(
-      encoder,
-      out,
-      in_unfolded,
-      wt_reshaped,
-      batch_shape,
-      a_batch_strides,
-      b_batch_strides);
-}
-
-void gemm_conv(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    const array& wt,
-    array& out,
-    const std::vector<int>& strides,
-    const std::vector<int>& padding,
-    const std::vector<int>& kernel_dilation,
-    const std::vector<int>& input_dilation,
-    bool flip,
-    Stream s) {
-  int conv_ndim = in.ndim() - 2;
-  if (conv_ndim < 1 || conv_ndim > 3) {
-    throw std::runtime_error(
-        fmt::format("[conv] Unsupported gemm_conv for {}D conv.", conv_ndim));
-  }
-  dispatch_1_2_3(conv_ndim, [&](auto ndim_constant) {
-    ConvParams<ndim_constant()> params(
-        in,
-        wt,
-        out,
-        strides,
-        padding,
-        kernel_dilation,
-        input_dilation,
-        1, // groups
-        flip);
-    gemm_conv_nd<ndim_constant()>(encoder, in, wt, out, params, s);
-  });
-}
-
-} // namespace mlx::core
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Cheng	4ba778c08a	Install uv	2025-07-30 15:27:00 +09:00
Cheng	893c55459e	There is no uvx	2025-07-30 15:20:40 +09:00
Cheng	8fb222f4a7	Revert "Save uv cache" This reverts commit `aaa7b822cc`.	2025-07-30 15:17:00 +09:00
Cheng	0bb6108fdf	uv is installed by default on linux	2025-07-30 15:15:32 +09:00
Cheng	acc09ed3a7	Use uv in mac	2025-07-30 15:02:49 +09:00
Cheng	aaa7b822cc	Save uv cache	2025-07-30 14:42:52 +09:00
Cheng	48e0eba372	Use uv in cuda build	2025-07-30 14:20:21 +09:00
Cheng	8d9bc0c421	Use uv to install packages in linux_build_and_test	2025-07-30 13:19:14 +09:00
Cheng	51da73c855	Remove unused apt installs	2025-07-30 13:11:38 +09:00