fix quantized vjp for mxfp4 (#2555)

* faster contiguous gather for indices in the first axis

* work per thread > 1

* angelos suggestion for scales / biases

.circleci/config.yml

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

ACKNOWLEDGMENTS.md

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

CMakeLists.txt

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

cmake/FindNCCL.cmake

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

docs/src/dev/custom_metal_kernels.rst

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

docs/src/index.rst

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

docs/src/install.rst

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

docs/src/python/cuda.rst

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

docs/src/python/fast.rst

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

docs/src/usage/compile.rst

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

mlx/backend/common/utils.cpp

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

mlx/backend/common/utils.h

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

mlx/backend/cpu/compiled.cpp

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

mlx/backend/cpu/lapack.h

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

mlx/backend/cpu/quantized.cpp

@@ -1,7 +1,5 @@
 // Copyright © 2023 Apple Inc.
 
-#include <cassert>
-
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
@@ -13,6 +11,35 @@ namespace mlx::core {
 
 namespace {
 
+const static float MXFP4_LUT[16] = {
+    +0.0f,
+    +0.5f,
+    +1.0f,
+    +1.5f,
+    +2.0f,
+    +3.0f,
+    +4.0f,
+    +6.0f,
+    -0.0f,
+    -0.5f,
+    -1.0f,
+    -1.5f,
+    -2.0f,
+    -3.0f,
+    -4.0f,
+    -6.0f};
+
+template <typename T>
+static inline T dequantize_scale(uint8_t s) {
+  using FOrI = union {
+    bfloat16_t f;
+    uint16_t i;
+  };
+  FOrI out;
+  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
+  return static_cast<T>(out.f);
+}
+
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
   return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -407,6 +434,231 @@ void _qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_qmm(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(4);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    std::fill(result, result + N, 0);
+
+    for (int k = 0; k < K; k++) {
+      T* result_local = result;
+      T xi = *x++;
+
+      for (int n = 0; n < N; n += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+        for (int ng = 0; ng < packs_in_group; ng++) {
+          uint8_t wi = *w_local++;
+#pragma clang loop unroll(full)
+          for (int p = 0; p < pack_factor; p++) {
+            (*result_local++) +=
+                xi * scale * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+      }
+    }
+
+    result += N;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int bytes_per_pack = get_bytes_per_pack(4);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      const T* x_local = x;
+      T sum = 0;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        T gsum = 0;
+        for (int kw = 0; kw < packs_in_group; kw++) {
+          uint8_t wi = *w_local++;
+#pragma clang loop unroll(full)
+          for (int p = 0; p < pack_factor; p++) {
+            gsum += (*x_local++) * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+        sum += scale * gsum;
+      }
+      *result = sum;
+      result++;
+    }
+
+    x += K;
+  }
+}
+
+template <int S>
+simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
+  if constexpr (S == 8) {
+    constexpr std::array<uint32_t, 8> shifts_ = {{0, 4, 8, 12, 16, 20, 24, 28}};
+    auto shifts(*(simd::Simd<uint32_t, S>*)&shifts_);
+    auto wi = simd::Simd<uint32_t, S>(*w);
+    wi = wi >> shifts;
+    wi = wi & 0xf;
+    simd::Simd<float, S> w_out;
+    for (int i = 0; i < S; ++i) {
+      w_out[i] = MXFP4_LUT[wi[i]];
+    }
+    return w_out;
+  } else {
+    // Appease compiler.. but should never get here
+    throw std::runtime_error("Unsupported combination for simd qmm.");
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t_simd(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = 32 / 4;
+  constexpr int packs_in_group = group_size / pack_factor;
+  constexpr int S = simd::max_size<T>;
+  static_assert(
+      S % pack_factor == 0, "SIMD size must be divisible by pack factor");
+  constexpr int packs_per_simd = S / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint32_t* w_local = w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      simd::Simd<float, S> acc(0);
+      auto x_local = x;
+      for (int k = 0; k < K; k += group_size) {
+        T scale = dequantize_scale<T>(*scales_local++);
+
+        simd::Simd<float, S> g_acc(0);
+        for (int kw = 0; kw < packs_in_group; kw += packs_per_simd) {
+          // Extract bits
+          auto wf = mxfp4_extract_bits_simd<S>(w_local);
+          w_local += packs_per_simd;
+          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
+          g_acc = g_acc + x_simd * wf;
+          x_local += S;
+        }
+        acc = acc + scale * g_acc;
+      }
+
+      *result = T(simd::sum(acc));
+      result++;
+    }
+    x += K;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_transpose(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K,
+    bool transposed_w) {
+  if (transposed_w) {
+    // the simd size must be a multiple of the number of elements per word
+    if constexpr (simd::max_size<T> % 8 == 0) {
+      mxfp4_qmm_t_simd<T>(result, x, w, scales, M, N, K);
+    } else {
+      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
+    }
+  } else {
+    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.ndim() > 1 ? x.shape(-2) : 1;
+  int N = out.shape(-1);
+  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
+  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
+  int batch_size = x.size() / (K * M);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  for (int i = 0; i < batch_size; i++) {
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
+        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case bfloat16:
+      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float16:
+      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float32:
+      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 template <typename T>
 void _bs_qmm_dispatch_typed(
     array& out,
@@ -513,115 +765,198 @@ void _bs_qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_bs_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.shape(-2);
+  int N = out.shape(-1);
+
+  int w_els = w.shape(-1) * w.shape(-2);
+  int g_els = scales.shape(-1) * scales.shape(-2);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
+  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
+
+  for (int i = 0; i < lhs_indices.size(); i++) {
+    int x_idx = lhs_indices_ptr[elem_to_loc(
+        i, lhs_indices.shape(), lhs_indices.strides())];
+    int w_idx = rhs_indices_ptr[elem_to_loc(
+        i, rhs_indices.shape(), rhs_indices.strides())];
+    mxfp4_qmm_dispatch_transpose<T>(
+        out_ptr + i * M * N,
+        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
+        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
+        scales_ptr +
+            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case float32:
+      mxfp4_bs_qmm_dispatch_typed<float>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case float16:
+      mxfp4_bs_qmm_dispatch_typed<float16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case bfloat16:
+      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 } // namespace
 
 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 4);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
 
-  std::vector<array> temps;
-  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
     if (arr.flags().row_contiguous) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous(x_pre);
   auto w = ensure_row_contiguous(w_pre);
   auto scales = ensure_row_contiguous(scales_pre);
-  auto biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
+    });
+  }
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 6);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
-  auto& lhs_indices = inputs[4];
-  auto& rhs_indices = inputs[5];
+  auto& lhs_indices = inputs[inputs.size() - 2];
+  auto& rhs_indices = inputs[inputs.size() - 1];
 
-  std::vector<array> temps;
+  auto& encoder = cpu::get_command_encoder(stream());
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &temps](const array& arr) {
+                                          &encoder](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy_cpu(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
+      copy_cpu(arr, arr_cpy, CopyType::General, s);
+      encoder.add_temporary(arr_cpy);
+      return arr_cpy;
     }
   };
 
   auto x = ensure_row_contiguous_last_dims(x_pre);
   auto w = ensure_row_contiguous_last_dims(w_pre);
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
-  auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_input_array(lhs_indices);
   encoder.set_input_array(rhs_indices);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    lhs_indices = array::unsafe_weak_copy(lhs_indices),
-                    rhs_indices = array::unsafe_weak_copy(rhs_indices),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _bs_qmm_dispatch(
-        out,
-        x,
-        w,
-        scales,
-        biases,
-        lhs_indices,
-        rhs_indices,
-        group_size_,
-        bits_,
-        transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _bs_qmm_dispatch(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          group_size_,
+          bits_,
+          transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_bs_qmm_dispatch(
+          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
+    });
+  }
 }
 
 template <typename T, typename U>
@@ -705,7 +1040,7 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::AffineQuantize::eval_cpu(
+void fast::Quantize::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto ensure_row_contiguous = [s = stream()](const array& arr) {
@@ -764,7 +1099,7 @@ void fast::AffineQuantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
     }
   });
 }

mlx/backend/cpu/simd/accelerate_simd.h

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

mlx/backend/cpu/svd.cpp

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

mlx/backend/cuda/CMakeLists.txt

@@ -16,8 +16,13 @@ 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
@@ -149,7 +154,7 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 FetchContent_Declare(
   cudnn
   GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
-  GIT_TAG v1.12.1
+  GIT_TAG v1.14.0
   GIT_SHALLOW TRUE
   EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)

mlx/backend/cuda/allocator.cpp

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

mlx/backend/cuda/arange.cu

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

mlx/backend/cuda/compiled.cpp

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

mlx/backend/cuda/conv.cpp

@@ -1,18 +1,12 @@
 // Copyright © 2025 Apple Inc.
 
+#include "mlx/backend/cuda/conv/conv.h"
+#include "mlx/backend/cuda/cudnn_utils.h"
 #include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/config.h"
 #include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/gpu/copy.h"
-#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
-// cudnn_frontend.h redefines this macro.
-#undef CHECK_CUDA_ERROR
-
-#include <cudnn_frontend.h>
-#include <cudnn_frontend_find_plan.h>
-#include <fmt/format.h>
 #include <nvtx3/nvtx3.hpp>
 
 #include <cassert>
@@ -21,9 +15,6 @@ namespace mlx::core {
 
 namespace {
 
-// Not all engines support it so can not use this API now.
-#define MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API 0
-
 // Alias for better readability.
 #define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
 #define CONV_BACKWARD_INPUT \
@@ -31,6 +22,9 @@ namespace {
 #define CONV_BACKWARD_WEIGHT \
   CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
 
+// Custom placeholder representing fallback kernel.
+#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
+
 struct ConvCacheKey {
   int device_id;
   cudnnDataType_t cudnn_dtype;
@@ -50,203 +44,13 @@ struct ConvCacheKey {
 auto& conv_cache() {
   static LRUBytesKeyCache<
       ConvCacheKey,
-      std::pair<cudnnBackendDescriptorType_t, cudnn_frontend::ExecutionPlan>>
+      std::pair<
+          cudnnBackendDescriptorType_t,
+          std::optional<cudnn_frontend::ExecutionPlan>>>
       cache(/* capacity */ 128);
   return cache;
 }
 
-template <typename T, typename Vec>
-inline SmallVector<T> convert_vector(const Vec& vec) {
-  return SmallVector<T>(vec.begin(), vec.end());
-}
-
-template <typename T, template <typename U> class Vec>
-inline std::array<T, MAX_NDIM> fixed_vector(const Vec<T>& vec) {
-  if (vec.size() > MAX_NDIM) {
-    throw std::runtime_error(
-        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
-  }
-  std::array<T, MAX_NDIM> result = {};
-  std::copy_n(vec.begin(), vec.size(), result.begin());
-  return result;
-}
-
-auto nhwc_to_nchw(const array& x) {
-  auto shape = convert_vector<int64_t>(x.shape());
-  shape.insert(shape.begin() + 1, shape.back());
-  shape.erase(shape.end() - 1);
-  auto strides = convert_vector<int64_t>(x.strides());
-  strides.insert(strides.begin() + 1, strides.back());
-  strides.erase(strides.end() - 1);
-  return std::make_tuple(std::move(shape), std::move(strides));
-}
-
-inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
-  switch (dtype) {
-    case int8:
-      return CUDNN_DATA_INT8;
-    case int32:
-      return CUDNN_DATA_INT32;
-    case uint8:
-      return CUDNN_DATA_UINT8;
-    case float16:
-      return CUDNN_DATA_HALF;
-    case bfloat16:
-      return CUDNN_DATA_BFLOAT16;
-    case float32:
-      return CUDNN_DATA_FLOAT;
-    case float64:
-      return CUDNN_DATA_DOUBLE;
-    default:
-      throw std::runtime_error(fmt::format(
-          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
-  }
-}
-
-inline uint8_t get_alignment(const array& x) {
-  uint8_t alignment = 1;
-  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
-  for (; alignment < 32; alignment *= 2) {
-    if (address % (alignment * 2)) {
-      return alignment;
-    }
-  }
-  return alignment;
-}
-
-inline cudnn_frontend::Tensor build_tensor(int64_t id, const array& x) {
-  auto [shape, strides] = nhwc_to_nchw(x);
-  return cudnn_frontend::TensorBuilder()
-      .setDim(shape.size(), shape.data())
-      .setStrides(strides.size(), strides.data())
-      .setId(id)
-      .setAlignment(get_alignment(x))
-      .setDataType(dtype_to_cudnn_type(x.dtype()))
-      .build();
-}
-
-cudnn_frontend::EngineConfigList get_engine_configs(
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph,
-    bool use_fallback = false) {
-  cudnn_frontend::GeneratorSource source;
-  if (use_fallback) {
-    source = [&backend_type](cudnn_frontend::OperationGraph& op_graph) {
-      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
-                          .setOperationGraph(op_graph)
-                          .setOperation(backend_type)
-                          .build();
-      return fallback.getFallbackList();
-    };
-  } else {
-    source = [](cudnn_frontend::OperationGraph& op_graph) {
-      auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
-                            .setOperationGraph(op_graph)
-                            .setHeurMode(CUDNN_HEUR_MODE_A)
-                            .build();
-      return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
-    };
-  }
-
-  cudnn_frontend::EngineConfigGenerator generator(1, &source);
-  auto configs = generator.generate_engine_config(op_graph);
-
-  cudnn_frontend::EngineConfigList filtered_configs;
-  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
-    if (cudnn_frontend::hasNumericalNote<
-            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
-      return true;
-    }
-    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
-        dtype == float32 && !env::enable_tf32()) {
-      return true;
-    }
-    return false;
-  });
-  return filtered_configs;
-}
-
-bool execute_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    array& x,
-    array& w,
-    array& y) {
-  int workspace_size = plan.getWorkspaceSize();
-  array workspace(allocator::malloc(workspace_size), {workspace_size}, uint8);
-
-  int64_t uids[3] = {'x', 'w', 'y'};
-  void* data_ptrs[3] = {
-      x.data<void>(),
-      w.data<void>(),
-      y.data<void>(),
-  };
-
-  auto variantPack = cudnn_frontend::VariantPackBuilder()
-                         .setWorkspacePointer(workspace.data<void>())
-                         .setDataPointers(3, data_ptrs)
-                         .setUids(3, uids)
-                         .build();
-
-  auto handle = encoder.device().cudnn_handle();
-  cudnnSetStream(handle, encoder.stream());
-
-#if CUDNN_VERSION >= 90500 && MLX_USE_CUDNN_NATIVE_CUDA_GRAPH_API
-  cudaGraph_t graph;
-  cudaGraphCreate(&graph, 0);
-  std::unique_ptr<cudaGraph_t, void (*)(cudaGraph_t*)> graph_freer(
-      &graph, [](cudaGraph_t* p) { cudaGraphDestroy(*p); });
-  if (cudnnBackendPopulateCudaGraph(
-          handle, plan.get_raw_desc(), variantPack.get_raw_desc(), graph) !=
-      CUDNN_STATUS_SUCCESS) {
-    return false;
-  }
-  encoder.add_graph_node(graph);
-#else
-  auto capture = encoder.capture_context();
-  if (cudnnBackendExecute(
-          handle, plan.get_raw_desc(), variantPack.get_raw_desc()) !=
-      CUDNN_STATUS_SUCCESS) {
-    // Discard the captured graph when failed.
-    capture.discard = true;
-    return false;
-  }
-#endif
-
-  encoder.add_temporary(workspace);
-  return true;
-}
-
-bool try_engines(
-    cu::CommandEncoder& encoder,
-    const ConvCacheKey& cache_key,
-    cudnnBackendDescriptorType_t backend_type,
-    cudnn_frontend::EngineConfigList& configs,
-    const std::string& op_graph_tag,
-    array& x,
-    array& w,
-    array& y) {
-  for (auto& config : configs) {
-    try {
-      auto plan = cudnn_frontend::ExecutionPlanBuilder()
-                      .setHandle(encoder.device().cudnn_handle())
-                      .setEngineConfig(config, op_graph_tag)
-                      .build();
-      if (execute_plan(encoder, plan, x, w, y)) {
-        conv_cache().emplace(
-            cache_key, std::make_pair(backend_type, std::move(plan)));
-        return true;
-      }
-    } catch (cudnn_frontend::cudnnException& error) {
-      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
-        throw;
-      }
-    }
-  }
-  return false;
-}
-
 auto get_conv_op_settings(
     cudnnBackendDescriptorType_t backend_type,
     array& x,
@@ -291,7 +95,7 @@ auto get_conv_op_settings(
   }
 }
 
-std::optional<cudnn_frontend::OperationGraph> build_op_graph(
+std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
     cu::CommandEncoder& encoder,
     cudnnBackendDescriptorType_t backend_type,
     Dtype dtype,
@@ -317,9 +121,9 @@ std::optional<cudnn_frontend::OperationGraph> build_op_graph(
                          .build();
 
     auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .setxDesc(build_tensor('x', x))
-                  .setwDesc(build_tensor('w', w))
-                  .setyDesc(build_tensor('y', y))
+                  .setxDesc(build_cudnn_tensor_nchw('x', x))
+                  .setwDesc(build_cudnn_tensor_nchw('w', w))
+                  .setyDesc(build_cudnn_tensor_nchw('y', y))
                   .setcDesc(conv_desc)
                   .build();
 
@@ -336,6 +140,42 @@ std::optional<cudnn_frontend::OperationGraph> build_op_graph(
   }
 }
 
+// Transpose from (C_out, H, W, C_in / groups) to (C_in, H, W, C_out / groups).
+array group_transpose(
+    const array& x,
+    int groups,
+    int group_dim,
+    int axis1,
+    int axis2,
+    Stream s) {
+  if (groups == 1) {
+    return swapaxes_in_eval(x, axis1, axis2);
+  }
+  int ndim = x.ndim();
+  if (group_dim < 0) {
+    group_dim += ndim;
+  }
+  if (axis1 < 0) {
+    axis1 += ndim;
+  }
+  if (axis2 < 0) {
+    axis2 += ndim;
+  }
+  if (group_dim <= axis1) {
+    axis1 += 1;
+  }
+  if (group_dim <= axis2) {
+    axis2 += 1;
+  }
+  auto shape = x.shape();
+  shape.insert(shape.begin() + group_dim, groups);
+  shape[group_dim + 1] = shape[group_dim + 1] / groups;
+  array x_trans = reshape_in_eval(x, std::move(shape), s);
+  x_trans = swapaxes_in_eval(x_trans, axis1, axis2);
+  x_trans = flatten_in_eval(x_trans, group_dim, group_dim + 1, s);
+  return x_trans;
+}
+
 // Do necessary transposes and copies to prepare the inputs and outputs for
 // building the cuDNN conv op. It is safe to be called multiple times in one
 // eval_gpu, with cost of possible redundant copies.
@@ -345,13 +185,14 @@ std::tuple<array, array, array> prepare_args(
     array in,
     array wt,
     array out,
+    int groups,
     Stream s) {
   // Transpose the args depending on the backend type.
   // TODO: Handle groups.
   if (backend_type == CONV_BACKWARD_INPUT) {
-    wt = swapaxes_in_eval(wt, 0, -1);
+    wt = group_transpose(wt, groups, 0, 0, -1, s);
   } else if (backend_type == CONV_BACKWARD_WEIGHT) {
-    in = swapaxes_in_eval(in, 0, -1);
+    in = group_transpose(in, groups, -1, 0, -1, s);
     wt = swapaxes_in_eval(wt, 0, -1);
     // Create a contiguous array that shares the data with |out|, but with dim
     // C_in and C_out swapped.
@@ -444,12 +285,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   ConvCacheKey cache_key{
       encoder.device().cuda_device(),
       dtype_to_cudnn_type(dtype),
-      fixed_vector(in.shape()),
-      fixed_vector(wt.shape()),
-      fixed_vector(kernel_strides_),
-      fixed_vector(padding_lo_),
-      fixed_vector(padding_hi_),
-      fixed_vector(kernel_dilation_),
+      vector_key(in.shape()),
+      vector_key(wt.shape()),
+      vector_key(kernel_strides_),
+      vector_key(padding_lo_),
+      vector_key(padding_hi_),
+      vector_key(kernel_dilation_),
       groups_,
       flip_,
       get_alignment(in),
@@ -457,11 +298,29 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       get_alignment(out)};
   if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
     auto& [backend_type, plan] = it->second;
-    std::tie(in, wt, out) = prepare_args(encoder, backend_type, in, wt, out, s);
-    register_args(encoder, backend_type, in, wt, out, out_);
-    auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-    if (!execute_plan(encoder, plan, x, w, y)) {
-      throw std::runtime_error("[conv] Cached plan failed to execute.");
+    if (plan) {
+      // Run cached plan.
+      std::tie(in, wt, out) =
+          prepare_args(encoder, backend_type, in, wt, out, groups_, s);
+      register_args(encoder, backend_type, in, wt, out, out_);
+      auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
+      if (!encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+        throw std::runtime_error("[conv] Cached plan failed to execute.");
+      }
+    } else {
+      // Run fallback kernel.
+      gemm_conv(
+          encoder,
+          in,
+          wt,
+          out,
+          kernel_strides_,
+          padding_lo_,
+          kernel_dilation_,
+          input_dilation_,
+          groups_,
+          flip_,
+          s);
     }
     return;
   }
@@ -490,7 +349,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   std::optional<cudnn_frontend::OperationGraph> op_graph;
   for (auto try_backend : try_backends) {
     auto [in_copy, wt_copy, out_copy] =
-        prepare_args(encoder, try_backend, in, wt, out, s);
+        prepare_args(encoder, try_backend, in, wt, out, groups_, s);
     auto [x, w, y] = dispatch_args(try_backend, in_copy, wt_copy, out_copy);
     auto [stride, padding_lo, padding_hi, dilation] = get_conv_op_settings(
         try_backend,
@@ -502,7 +361,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_hi_,
         kernel_dilation_,
         input_dilation_);
-    op_graph = build_op_graph(
+    op_graph = build_conv_op_graph(
         encoder,
         try_backend,
         dtype,
@@ -521,26 +380,39 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       break;
     }
   }
-  if (!op_graph) {
-    throw std::runtime_error("[conv] Can not build op graph.");
+
+  if (op_graph) {
+    // Setup inputs and outputs.
+    register_args(encoder, backend_type, in, wt, out, out_);
+
+    // 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) {
+      throw std::runtime_error("[conv] Unable to find an execution plan.");
+    }
+    auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
+    if (encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
+      conv_cache().emplace(
+          cache_key, std::make_pair(backend_type, std::move(*plan)));
+      return;
+    }
   }
 
-  // Get ready to execute the graph.
-  register_args(encoder, backend_type, in, wt, out, out_);
-
-  // Try to run plans based on heuristics.
-  auto configs = get_engine_configs(backend_type, dtype, *op_graph);
-  auto tag = op_graph->getTag();
-  auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-    return;
-  }
-  // Then try fallback plans.
-  configs = get_engine_configs(backend_type, dtype, *op_graph);
-  if (try_engines(encoder, cache_key, backend_type, configs, tag, x, w, y)) {
-    return;
-  }
-  throw std::runtime_error("[conv] Unable to find a working engine.");
+  // Use fallback kernel for settings not supported by cuDNN.
+  gemm_conv(
+      encoder,
+      in,
+      wt,
+      out,
+      kernel_strides_,
+      padding_lo_,
+      kernel_dilation_,
+      input_dilation_,
+      groups_,
+      flip_,
+      s);
+  conv_cache().emplace(cache_key, std::make_pair(CONV_FALLBACK, std::nullopt));
 }
 
 } // namespace mlx::core

mlx/backend/cuda/conv/conv.h

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

mlx/backend/cuda/conv/gemm_conv.cu

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

mlx/backend/cuda/conv/gemm_grouped_conv.cu

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

mlx/backend/cuda/copy.cu

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

mlx/backend/cuda/cudnn_utils.cpp

@@ -0,0 +1,272 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/cudnn_utils.h"
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core {
+
+namespace {
+
+// Create a cudnn tensor descriptor.
+template <typename Vec>
+inline cudnn_frontend::Tensor build_cudnn_tensor(
+    int64_t id,
+    const array& x,
+    const Vec& shape,
+    const Vec& strides) {
+  return cudnn_frontend::TensorBuilder()
+      .setDim(shape.size(), shape.data())
+      .setStrides(strides.size(), strides.data())
+      .setId(id)
+      .setAlignment(get_alignment(x))
+      .setDataType(dtype_to_cudnn_type(x.dtype()))
+      .build();
+}
+
+// In MLX a singleton dim (shape[dim] == 1) can have any stride, but in cuDNN
+// whether a tensor is contiguous is determined with:
+// shape[dim] == shape[dim + 1] * strides[dim + 1]
+// So a contiguous array with singleton dims in MLX may be mistakenly treated
+// as strided in cuDNN, and we work around it by normalizing the strides.
+Strides normalized_strides(const array& x) {
+  if (!x.flags().row_contiguous || x.ndim() < 2) {
+    return x.strides();
+  }
+  Strides strides = x.strides();
+  for (int i = x.ndim() - 2; i >= 0; --i) {
+    if (x.shape(i) == 1) {
+      strides[i] = x.shape(i + 1) * strides[i + 1];
+    }
+  }
+  return strides;
+}
+
+// Return the shape and strides after transposing from NHWC to NCHW.
+auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
+  assert(shape.size() >= 3);
+  shape.insert(shape.begin() + 1, shape.back());
+  shape.erase(shape.end() - 1);
+  strides.insert(strides.begin() + 1, strides.back());
+  strides.erase(strides.end() - 1);
+  return std::make_tuple(std::move(shape), std::move(strides));
+}
+
+inline auto nhwc_to_nchw(const array& x) {
+  return nhwc_to_nchw(
+      convert_vector<int64_t>(x.shape()), normalized_strides(x));
+}
+
+// Return available engines for a |op_graph|.
+cudnn_frontend::EngineConfigList get_cudnn_engine_configs(
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph,
+    bool use_fallback = true) {
+  SmallVector<cudnn_frontend::GeneratorSource, 2> sources;
+  sources.push_back([](auto& op_graph) {
+    auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
+                          .setOperationGraph(op_graph)
+                          .setHeurMode(CUDNN_HEUR_MODE_A)
+                          .build();
+    return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
+  });
+  if (use_fallback) {
+    sources.push_back([&backend_type](auto& op_graph) {
+      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
+                          .setOperationGraph(op_graph)
+                          .setOperation(backend_type)
+                          .build();
+      return fallback.getFallbackList();
+    });
+  }
+
+  auto configs =
+      cudnn_frontend::EngineConfigGenerator(sources.size(), sources.data())
+          .generate_engine_config(op_graph);
+
+  cudnn_frontend::EngineConfigList filtered_configs;
+  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
+    if (cudnn_frontend::hasNumericalNote<
+            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
+      return true;
+    }
+    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
+        dtype == float32 && !env::enable_tf32()) {
+      return true;
+    }
+    return false;
+  });
+  return filtered_configs;
+}
+
+// Take |engine_configs| and |op_graph| and find a working execution plans
+// from them.
+std::optional<cudnn_frontend::ExecutionPlan>
+find_cudnn_plan_from_engine_configs(
+    cudnnHandle_t handle,
+    const cudnn_frontend::EngineConfigList& engine_configs,
+    const cudnn_frontend::OperationGraph& op_graph) {
+  auto op_graph_tag = op_graph.getTag();
+  for (const auto& config : engine_configs) {
+    try {
+      return cudnn_frontend::ExecutionPlanBuilder()
+          .setHandle(handle)
+          .setEngineConfig(config, op_graph_tag)
+          .build();
+    } catch (cudnn_frontend::cudnnException& error) {
+      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
+        throw;
+      }
+    }
+  }
+  return std::nullopt;
+}
+
+// Prepare workspace and args to execute plan.
+template <typename F>
+bool prepare_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs,
+    F&& execute) {
+  int workspace_size = plan.getWorkspaceSize();
+  array workspace(
+      workspace_size > 0 ? allocator::malloc(workspace_size)
+                         : allocator::Buffer(nullptr),
+      {workspace_size},
+      uint8);
+
+  auto args = cudnn_frontend::VariantPackBuilder()
+                  .setWorkspacePointer(workspace.data<void>())
+                  .setDataPointers(num_args, data_ptrs)
+                  .setUids(num_args, uids)
+                  .build();
+
+  auto handle = encoder.device().cudnn_handle();
+  cudnnSetStream(handle, encoder.stream());
+
+  if (!execute(handle, plan.get_raw_desc(), args.get_raw_desc())) {
+    return false;
+  }
+
+  encoder.add_temporary(workspace);
+  return true;
+}
+
+} // namespace
+
+cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  return build_cudnn_tensor(id, x, shape, normalized_strides(x));
+}
+
+cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x) {
+  auto [shape, strides] = nhwc_to_nchw(x);
+  return build_cudnn_tensor(id, x, shape, strides);
+}
+
+cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x) {
+  if (x.ndim() == 0) {
+    SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
+    return build_cudnn_tensor(id, x, scalar_dims, scalar_dims);
+  }
+  if (x.ndim() == 1) {
+    int64_t s = x.shape(0);
+    SmallVector<int64_t, 4> shape = {1, x.shape(0), 1, 1};
+    SmallVector<int64_t, 4> strides = {s, 1, s, s};
+    return build_cudnn_tensor(id, x, shape, strides);
+  }
+  if (x.ndim() == 2) {
+    int64_t s =
+        x.flags().row_contiguous ? x.shape(1) * x.strides(1) : x.strides(0);
+    SmallVector<int64_t, 4> shape = {x.shape(0), x.shape(1), 1, 1};
+    SmallVector<int64_t, 4> strides = {s, x.strides(1), s, s};
+    return build_cudnn_tensor(id, x, shape, strides);
+  }
+  if (x.ndim() == 3 || x.ndim() == 4) {
+    return build_cudnn_tensor_nchw(id, x);
+  }
+  throw std::runtime_error(
+      fmt::format("Unsupported array with {} dims.", x.ndim()));
+}
+
+cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype) {
+  SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
+  return cudnn_frontend::TensorBuilder()
+      .setDim(scalar_dims.size(), scalar_dims.data())
+      .setStrides(scalar_dims.size(), scalar_dims.data())
+      .setId(id)
+      .setAlignment(16)
+      .setDataType(dtype_to_cudnn_type(dtype))
+      .setByValue(true)
+      .build();
+}
+
+std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
+    cudnnHandle_t handle,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph) {
+  auto engine_configs = get_cudnn_engine_configs(backend_type, dtype, op_graph);
+  return find_cudnn_plan_from_engine_configs(handle, engine_configs, op_graph);
+}
+
+bool encode_cudnn_plan_with_capturing(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs) {
+  return prepare_cudnn_plan(
+      encoder,
+      plan,
+      num_args,
+      uids,
+      data_ptrs,
+      [&](auto handle, auto plan, auto args) {
+        auto capture = encoder.capture_context();
+        if (cudnnBackendExecute(handle, plan, args) != CUDNN_STATUS_SUCCESS) {
+          // Discard the captured graph when failed.
+          capture.discard = true;
+          return false;
+        }
+        return true;
+      });
+}
+
+#if CUDNN_VERSION >= 90500
+bool encode_cudnn_plan_with_graph_api(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs) {
+  return prepare_cudnn_plan(
+      encoder,
+      plan,
+      num_args,
+      uids,
+      data_ptrs,
+      [&](auto handle, auto plan, auto args) {
+        if (!graph) {
+          graph = CudaGraph(encoder.device());
+          if (cudnnBackendPopulateCudaGraph(handle, plan, args, graph) !=
+              CUDNN_STATUS_SUCCESS) {
+            return false;
+          }
+        } else {
+          if (cudnnBackendUpdateCudaGraph(handle, plan, args, graph) !=
+              CUDNN_STATUS_SUCCESS) {
+            return false;
+          }
+        }
+        encoder.add_graph_node(graph);
+        return true;
+      });
+}
+#endif
+
+} // namespace mlx::core

mlx/backend/cuda/cudnn_utils.h

@@ -0,0 +1,164 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/cuda/device/config.h"
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/dtype_utils.h"
+
+#include <cudnn_frontend.h>
+#include <cudnn_frontend_find_plan.h>
+#include <fmt/format.h>
+
+#include <algorithm>
+#include <array>
+
+namespace mlx::core {
+
+namespace cu {
+class CommandEncoder;
+}
+
+// Return pointer alignment of |x|'s data.
+inline uint8_t get_alignment(const array& x) {
+  uint8_t alignment = 1;
+  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
+  for (; alignment < 32; alignment *= 2) {
+    if (address % (alignment * 2)) {
+      return alignment;
+    }
+  }
+  return alignment;
+}
+
+// Convert the type of elements in |vec| to |T|.
+template <typename T, typename Vec>
+inline SmallVector<T> convert_vector(const Vec& vec) {
+  return SmallVector<T>(vec.begin(), vec.end());
+}
+
+// Return an array that can be used as map key for |vec| with size <= MAX_NDIM.
+//
+// There are 2 differences from the const_param util from kernel_utils.cuh:
+// 1. The rest of array is filled with 0.
+// 2. This util can be used in .cpp files.
+template <typename T, template <typename U> class Vec>
+inline std::array<T, MAX_NDIM> vector_key(const Vec<T>& vec) {
+  if (vec.size() > MAX_NDIM) {
+    throw std::runtime_error(
+        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
+  }
+  std::array<T, MAX_NDIM> result = {};
+  std::copy_n(vec.begin(), vec.size(), result.begin());
+  return result;
+}
+
+// Helpers used by get_data_ptrs to get pointers.
+inline void* get_data_ptr(const array& arr) {
+  return const_cast<void*>(arr.data<void>());
+}
+
+template <typename T, typename = std::enable_if_t<std::is_scalar_v<T>>>
+inline void* get_data_ptr(T& scalar) {
+  return &scalar;
+}
+
+// Return an array filled with data pointers of args.
+template <typename... Args>
+inline std::array<void*, sizeof...(Args)> get_data_ptrs(Args&... args) {
+  return {get_data_ptr(args)...};
+}
+
+// Map dtype to cudnn data type.
+inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return CUDNN_DATA_INT8;
+    case int32:
+      return CUDNN_DATA_INT32;
+    case uint8:
+      return CUDNN_DATA_UINT8;
+    case float16:
+      return CUDNN_DATA_HALF;
+    case bfloat16:
+      return CUDNN_DATA_BFLOAT16;
+    case float32:
+      return CUDNN_DATA_FLOAT;
+    case float64:
+      return CUDNN_DATA_DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+  }
+}
+
+// Create a tensor descriptor from |x|.
+cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x);
+
+// Create a tensor descriptor from |x|, and transpose from NHWC to NCHW.
+cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x);
+
+// Create a tensor descriptor from |x|, make sure it is 4D, and transpose it
+// from NHWC to NCHW.
+cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x);
+
+// Create a 4D scalar tensor descriptor, which is passed by value.
+cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype);
+
+// Find a working plan for |op_graph|.
+std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
+    cudnnHandle_t handle,
+    cudnnBackendDescriptorType_t backend_type,
+    Dtype dtype,
+    cudnn_frontend::OperationGraph& op_graph);
+
+// Encode the plan to command buffer by capturing.
+bool encode_cudnn_plan_with_capturing(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs);
+
+#if CUDNN_VERSION >= 90500
+// Encode the plan to command buffer by using native graph api of cudnn. If the
+// |graph| is empty it will be populated, otherwise it will be updated.
+bool encode_cudnn_plan_with_graph_api(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    int num_args,
+    const int64_t* uids,
+    void** data_ptrs);
+#endif
+
+// Helpers to make calls like encode_cudnn_plan(..., {'x', 'y', 'z'}, x, y, z).
+template <typename... Args>
+bool encode_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    std::initializer_list<int64_t> uids,
+    Args&... args) {
+  assert(uids.size() == sizeof...(args));
+  auto data_ptrs = get_data_ptrs(args...);
+  return encode_cudnn_plan_with_capturing(
+      encoder, plan, uids.size(), uids.begin(), data_ptrs.data());
+}
+
+#if CUDNN_VERSION >= 90500
+template <typename... Args>
+bool encode_cudnn_plan(
+    cu::CommandEncoder& encoder,
+    cudnn_frontend::ExecutionPlan& plan,
+    CudaGraph& graph,
+    std::initializer_list<int64_t> uids,
+    Args&... args) {
+  assert(uids.size() == sizeof...(args));
+  auto data_ptrs = get_data_ptrs(args...);
+  return encode_cudnn_plan_with_graph_api(
+      encoder, plan, graph, uids.size(), uids.begin(), data_ptrs.data());
+}
+#endif
+
+} // namespace mlx::core

mlx/backend/cuda/custom_kernel.cpp

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

mlx/backend/cuda/device.cpp

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

mlx/backend/cuda/device.h

@@ -21,7 +21,7 @@ class CommandEncoder {
   struct CaptureContext {
     CaptureContext(CommandEncoder& enc);
     ~CaptureContext();
-    cudaGraph_t graph;
+    CudaGraph graph;
     CommandEncoder& enc;
     bool discard{false};
   };
@@ -76,9 +76,6 @@ class CommandEncoder {
       uint32_t smem_bytes,
       void** params);
 
-  // Low-level graph helpers.
-  void add_kernel_node(const cudaKernelNodeParams& params);
-  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
   void add_graph_node(cudaGraph_t child);
 
   void add_temporary(const array& arr) {
@@ -101,6 +98,9 @@ class CommandEncoder {
   void synchronize();
 
  private:
+  void add_kernel_node(const cudaKernelNodeParams& params);
+  void add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params);
+
   struct GraphNode {
     cudaGraphNode_t node;
     // K = kernel
@@ -115,7 +115,7 @@ class CommandEncoder {
 
   Device& device_;
   CudaStream stream_;
-  cudaGraph_t graph_;
+  CudaGraph graph_;
   Worker worker_;
   char node_count_{0};
   char graph_node_count_{0};

mlx/backend/cuda/device/binary_ops.cuh

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

mlx/backend/cuda/device/cast_op.cuh

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

mlx/backend/cuda/distributed.cu

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

mlx/backend/cuda/gemms/cublas_gemm.cpp

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

mlx/backend/cuda/gemms/cublas_gemm.h

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

mlx/backend/cuda/indexing.cpp

@@ -94,7 +94,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
             large ? "int64_t" : "int32_t"));
       }
     }
-    return std::make_pair(jit_source_gather, std::move(kernel_names));
+    return std::make_tuple(false, jit_source_gather, std::move(kernel_names));
   });
 
   cu::KernelArgs args;
@@ -110,7 +110,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append<int32_t>(src.ndim());
   args.append_ndim(slice_sizes_);
   args.append(slice_size);
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
@@ -189,7 +189,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
             large ? "int64_t" : "int32_t"));
       }
     }
-    return std::make_pair(jit_source_scatter, std::move(kernel_names));
+    return std::make_tuple(false, jit_source_scatter, std::move(kernel_names));
   });
 
   cu::KernelArgs args;
@@ -211,7 +211,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   args.append_ndim(out.shape());
   args.append_ndim(out.strides());
   args.append<int32_t>(out.ndim());
-  args.append(SmallVector<int32_t>(axes_.begin(), axes_.end()));
+  args.append(axes_);
   append_indices_arg(args, inputs, nidx, idx_ndim);
 
   std::string kernel_name = fmt::format(
@@ -268,7 +268,8 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
         }
       }
     }
-    return std::make_pair(jit_source_gather_axis, std::move(kernel_names));
+    return std::make_tuple(
+        false, jit_source_gather_axis, std::move(kernel_names));
   });
 
   size_t idx_size_pre = 1;
@@ -371,7 +372,8 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
         }
       }
     }
-    return std::make_pair(jit_source_scatter_axis, std::move(kernel_names));
+    return std::make_tuple(
+        false, jit_source_scatter_axis, std::move(kernel_names));
   });
 
   size_t idx_size_pre = 1;

mlx/backend/cuda/jit_module.cpp

@@ -67,9 +67,11 @@ const std::string& cccl_dir() {
       return path.string();
     }
     // Finally check the environment variable.
-    path = std::getenv("MLX_CCCL_DIR");
-    if (!path.empty() && std::filesystem::exists(path)) {
-      return path.string();
+    if (const char* env = std::getenv("MLX_CCCL_DIR"); env) {
+      path = env;
+      if (!path.empty() && std::filesystem::exists(path)) {
+        return path.string();
+      }
     }
     return std::string();
   }();
@@ -101,8 +103,8 @@ const std::filesystem::path& ptx_cache_dir() {
 bool read_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
-    std::vector<char>* ptx,
-    std::vector<std::pair<std::string, std::string>>* ptx_kernels) {
+    std::string& ptx,
+    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
   if (cache_dir.empty()) {
     return false;
   }
@@ -117,15 +119,15 @@ bool read_cached_ptx(
   if (!ptx_file.good()) {
     return false;
   }
-  ptx->resize(ptx_size);
-  ptx_file.read(ptx->data(), ptx_size);
+  ptx.resize(ptx_size);
+  ptx_file.read(ptx.data(), ptx_size);
 
   std::ifstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
   std::string line;
   while (std::getline(txt_file, line)) {
     auto tab = line.find('\t');
     if (tab != std::string::npos) {
-      ptx_kernels->emplace_back(line.substr(0, tab), line.substr(tab + 1));
+      ptx_kernels.emplace_back(line.substr(0, tab), line.substr(tab + 1));
     }
   }
   return true;
@@ -135,7 +137,7 @@ bool read_cached_ptx(
 void write_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
-    const std::vector<char>& ptx,
+    const std::string& ptx,
     const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
     const std::string& source_code) {
   if (cache_dir.empty()) {
@@ -217,85 +219,85 @@ constexpr const char* g_headers[] = {
     jit_source_utils,
 };
 
-} // namespace
-
-JitModule::JitModule(
+void compile(
     Device& device,
     const std::string& module_name,
-    const KernelBuilder& builder) {
-  // Check cache.
-  std::vector<char> ptx;
-  std::vector<std::pair<std::string, std::string>> ptx_kernels;
-  if (!read_cached_ptx(ptx_cache_dir(), module_name, &ptx, &ptx_kernels)) {
-    // Create program.
-    auto [source_code, kernel_names] = builder();
-    nvrtcProgram prog;
-    CHECK_NVRTC_ERROR(nvrtcCreateProgram(
-        &prog,
-        source_code.c_str(),
-        (module_name + ".cu").c_str(),
-        std::size(g_headers),
-        g_headers,
-        g_include_names));
-    std::unique_ptr<nvrtcProgram, void (*)(nvrtcProgram*)> prog_freer(
-        &prog,
-        [](nvrtcProgram* p) { CHECK_NVRTC_ERROR(nvrtcDestroyProgram(p)); });
-    for (const auto& name : kernel_names) {
-      CHECK_NVRTC_ERROR(nvrtcAddNameExpression(prog, name.c_str()));
-    }
-
-    // Compile program.
-    std::vector<const char*> args;
-    bool use_sass = compiler_supports_device_sass(device);
-    std::string compute = fmt::format(
-        "--gpu-architecture={}_{}{}",
-        use_sass ? "sm" : "compute",
-        device.compute_capability_major(),
-        device.compute_capability_minor());
-    args.push_back(compute.c_str());
-    std::string cccl_include = cccl_dir();
-    if (!cccl_include.empty()) {
-      cccl_include = fmt::format("--include-path={}", cccl_include);
-      args.push_back(cccl_include.c_str());
-    }
-    std::string cuda_include =
-        fmt::format("--include-path={}/include", cuda_home());
-    args.push_back(cuda_include.c_str());
-    nvrtcResult compile_result =
-        nvrtcCompileProgram(prog, args.size(), args.data());
-    if (compile_result != NVRTC_SUCCESS) {
-      size_t log_size;
-      CHECK_NVRTC_ERROR(nvrtcGetProgramLogSize(prog, &log_size));
-      std::vector<char> log(log_size + 1, 0);
-      CHECK_NVRTC_ERROR(nvrtcGetProgramLog(prog, log.data()));
-      throw std::runtime_error(
-          fmt::format("Failed to compile kernel: {}.", log.data()));
-    }
-
-    // Get mangled names of kernel names.
-    for (const auto& name : kernel_names) {
-      const char* mangled;
-      CHECK_NVRTC_ERROR(nvrtcGetLoweredName(prog, name.c_str(), &mangled));
-      ptx_kernels.emplace_back(name, mangled);
-    }
-
-    // Get ptx data.
-    size_t ptx_size;
-    if (use_sass) {
-      CHECK_NVRTC_ERROR(nvrtcGetCUBINSize(prog, &ptx_size));
-    } else {
-      CHECK_NVRTC_ERROR(nvrtcGetPTXSize(prog, &ptx_size));
-    }
-    ptx.resize(ptx_size, 0);
-    if (use_sass) {
-      CHECK_NVRTC_ERROR(nvrtcGetCUBIN(prog, ptx.data()));
-    } else {
-      CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
-    }
-    write_cached_ptx(
-        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
+    const std::string& source,
+    const std::vector<std::string>& kernel_names,
+    std::string& ptx,
+    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
+  // Create the program
+  nvrtcProgram prog;
+  CHECK_NVRTC_ERROR(nvrtcCreateProgram(
+      &prog,
+      source.c_str(),
+      (module_name + ".cu").c_str(),
+      std::size(g_headers),
+      g_headers,
+      g_include_names));
+  std::unique_ptr<nvrtcProgram, void (*)(nvrtcProgram*)> prog_freer(
+      &prog,
+      [](nvrtcProgram* p) { CHECK_NVRTC_ERROR(nvrtcDestroyProgram(p)); });
+  for (const auto& name : kernel_names) {
+    CHECK_NVRTC_ERROR(nvrtcAddNameExpression(prog, name.c_str()));
   }
 
+  // Compile program.
+  std::vector<const char*> args;
+  bool use_sass = compiler_supports_device_sass(device);
+  std::string compute = fmt::format(
+      "--gpu-architecture={}_{}{}",
+      use_sass ? "sm" : "compute",
+      device.compute_capability_major(),
+      device.compute_capability_minor());
+  args.push_back(compute.c_str());
+  std::string cccl_include = cccl_dir();
+  if (!cccl_include.empty()) {
+    cccl_include = fmt::format("--include-path={}", cccl_include);
+    args.push_back(cccl_include.c_str());
+  }
+  std::string cuda_include =
+      fmt::format("--include-path={}/include", cuda_home());
+  args.push_back(cuda_include.c_str());
+  nvrtcResult compile_result =
+      nvrtcCompileProgram(prog, args.size(), args.data());
+  if (compile_result != NVRTC_SUCCESS) {
+    size_t log_size;
+    CHECK_NVRTC_ERROR(nvrtcGetProgramLogSize(prog, &log_size));
+    std::vector<char> log(log_size + 1, 0);
+    CHECK_NVRTC_ERROR(nvrtcGetProgramLog(prog, log.data()));
+    throw std::runtime_error(
+        fmt::format("Failed to compile kernel: {}.", log.data()));
+  }
+
+  // Get mangled names of kernel names.
+  for (const auto& name : kernel_names) {
+    const char* mangled;
+    CHECK_NVRTC_ERROR(nvrtcGetLoweredName(prog, name.c_str(), &mangled));
+    ptx_kernels.emplace_back(name, mangled);
+  }
+
+  // Get ptx data.
+  size_t ptx_size;
+  if (use_sass) {
+    CHECK_NVRTC_ERROR(nvrtcGetCUBINSize(prog, &ptx_size));
+  } else {
+    CHECK_NVRTC_ERROR(nvrtcGetPTXSize(prog, &ptx_size));
+  }
+  ptx.resize(ptx_size);
+  if (use_sass) {
+    CHECK_NVRTC_ERROR(nvrtcGetCUBIN(prog, ptx.data()));
+  } else {
+    CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
+  }
+}
+
+void load_module(
+    const std::string& module_name,
+    const std::string& ptx,
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
+    CUmodule& module_,
+    std::unordered_map<std::string, std::pair<CUfunction, bool>>& kernels) {
   // Load module.
   char jit_log[4089] = {};
   CUjit_option options[] = {
@@ -312,21 +314,69 @@ JitModule::JitModule(
   for (const auto& [name, mangled] : ptx_kernels) {
     CUfunction kernel;
     CHECK_CUDA_ERROR(cuModuleGetFunction(&kernel, module_, mangled.c_str()));
-    kernels_[name] = kernel;
+    kernels[name] = std::make_pair(kernel, false);
   }
 }
 
+} // namespace
+
+JitModule::JitModule(
+    Device& device,
+    const std::string& module_name,
+    const KernelBuilder& builder,
+    bool use_disk_cache) {
+  // Will hold the actual device executable source code and kernel names
+  std::string ptx;
+  std::vector<std::pair<std::string, std::string>> ptx_kernels;
+
+  // Try to load them from the file cache
+  if (!read_cached_ptx(ptx_cache_dir(), module_name, ptx, ptx_kernels)) {
+    auto [precompiled, source_code, kernel_names] = builder();
+
+    // Get the PTX or cubin
+    if (precompiled) {
+      ptx = std::move(source_code);
+      for (auto& name : kernel_names) {
+        ptx_kernels.emplace_back(name, name);
+      }
+    } else {
+      compile(device, module_name, source_code, kernel_names, ptx, ptx_kernels);
+    }
+
+    // If requested save them in the file cache for the next launch
+    if (use_disk_cache) {
+      write_cached_ptx(
+          ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
+    }
+  }
+
+  // Load the module
+  load_module(module_name, ptx, ptx_kernels, module_, kernels_);
+}
+
 JitModule::~JitModule() {
   CHECK_CUDA_ERROR(cuModuleUnload(module_));
 }
 
-CUfunction JitModule::get_kernel(const std::string& kernel_name) {
+CUfunction JitModule::get_kernel(
+    const std::string& kernel_name,
+    std::function<void(CUfunction)> configure_kernel) {
   auto it = kernels_.find(kernel_name);
   if (it == kernels_.end()) {
     throw std::runtime_error(
         fmt::format("There is no kernel named {}.", kernel_name));
   }
-  return it->second;
+
+  // If it is the first time we run this kernel then configure it. Do it only
+  // once!
+  if (!it->second.second) {
+    if (configure_kernel) {
+      configure_kernel(it->second.first);
+    }
+    it->second.second = true;
+  }
+
+  return it->second.first;
 }
 
 std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
@@ -337,11 +387,12 @@ std::unordered_map<std::string, JitModule>& get_jit_module_cache() {
 JitModule& get_jit_module(
     const mlx::core::Device& device,
     const std::string& name,
-    const KernelBuilder& builder) {
+    const KernelBuilder& builder,
+    bool cache) {
   auto& map = get_jit_module_cache();
   auto it = map.find(name);
   if (it == map.end()) {
-    it = map.try_emplace(name, cu::device(device), name, builder).first;
+    it = map.try_emplace(name, cu::device(device), name, builder, cache).first;
   }
   return it->second;
 }

mlx/backend/cuda/jit_module.h

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

mlx/backend/cuda/no_cuda.cpp

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

mlx/backend/cuda/primitives.cpp

@@ -24,8 +24,6 @@ namespace mlx::core {
   }
 
 NO_GPU(BlockMaskedMM)
-NO_GPU(DynamicSlice)
-NO_GPU(DynamicSliceUpdate)
 NO_GPU(FFT)
 NO_GPU(GatherMM)
 NO_GPU(GatherQMM)
@@ -41,12 +39,7 @@ NO_GPU(Cholesky)
 NO_GPU_MULTI(Eig)
 NO_GPU_MULTI(Eigh)
 
-namespace fast {
-NO_GPU_MULTI(CustomKernel)
-} // namespace fast
-
 namespace distributed {
-NO_GPU_MULTI(AllReduce)
 NO_GPU_MULTI(AllGather)
 NO_GPU_MULTI(Send)
 NO_GPU_MULTI(Recv)

mlx/backend/cuda/quantized/quantized.cpp

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

mlx/backend/cuda/slicing.cpp

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

mlx/backend/cuda/sort.cu

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

mlx/backend/cuda/utils.cpp

@@ -8,36 +8,6 @@
 
 namespace mlx::core {
 
-CudaStream::CudaStream(cu::Device& device) {
-  device.make_current();
-  CHECK_CUDA_ERROR(cudaStreamCreateWithFlags(&stream_, cudaStreamNonBlocking));
-}
-
-CudaStream::~CudaStream() {
-  CHECK_CUDA_ERROR(cudaStreamDestroy(stream_));
-}
-
-CudaGraphExec::CudaGraphExec(cudaGraphExec_t handle) : handle_(handle) {}
-
-CudaGraphExec::CudaGraphExec(CudaGraphExec&& other) : handle_(other.handle_) {
-  other.handle_ = nullptr;
-};
-
-CudaGraphExec::~CudaGraphExec() {
-  reset();
-}
-
-void CudaGraphExec::instantiate(cudaGraph_t graph) {
-  CHECK_CUDA_ERROR(cudaGraphInstantiate(&handle_, graph, nullptr, nullptr, 0));
-}
-
-void CudaGraphExec::reset() {
-  if (handle_ != nullptr) {
-    CHECK_CUDA_ERROR(cudaGraphExecDestroy(handle_));
-    handle_ = nullptr;
-  }
-}
-
 void check_cublas_error(const char* name, cublasStatus_t err) {
   if (err != CUBLAS_STATUS_SUCCESS) {
     // TODO: Use cublasGetStatusString when it is widely available.
@@ -96,4 +66,24 @@ const char* dtype_to_cuda_type(const Dtype& dtype) {
   }
 }
 
+CudaGraph::CudaGraph(cu::Device& device) {
+  device.make_current();
+  CHECK_CUDA_ERROR(cudaGraphCreate(&handle_, 0));
+}
+
+void CudaGraph::end_capture(cudaStream_t stream) {
+  assert(handle_ == nullptr);
+  CHECK_CUDA_ERROR(cudaStreamEndCapture(stream, &handle_));
+}
+
+void CudaGraphExec::instantiate(cudaGraph_t graph) {
+  assert(handle_ == nullptr);
+  CHECK_CUDA_ERROR(cudaGraphInstantiate(&handle_, graph, nullptr, nullptr, 0));
+}
+
+CudaStream::CudaStream(cu::Device& device) {
+  device.make_current();
+  CHECK_CUDA_ERROR(cudaStreamCreateWithFlags(&handle_, cudaStreamNonBlocking));
+}
+
 } // namespace mlx::core

mlx/backend/cuda/utils.h

@@ -16,44 +16,6 @@ class Device;
 
 struct Dtype;
 
-// Cuda stream managed with RAII.
-class CudaStream {
- public:
-  explicit CudaStream(cu::Device& device);
-  ~CudaStream();
-
-  CudaStream(const CudaStream&) = delete;
-  CudaStream& operator=(const CudaStream&) = delete;
-
-  operator cudaStream_t() const {
-    return stream_;
-  }
-
- private:
-  cudaStream_t stream_;
-};
-
-// Move-able RAII handle of cudaGraphExec_t.
-class CudaGraphExec {
- public:
-  CudaGraphExec(cudaGraphExec_t handle = nullptr);
-  CudaGraphExec(CudaGraphExec&& other);
-  ~CudaGraphExec();
-
-  CudaGraphExec(const CudaGraphExec&) = delete;
-  CudaGraphExec& operator=(const CudaGraphExec&) = delete;
-
-  void instantiate(cudaGraph_t graph);
-  void reset();
-
-  operator cudaGraphExec_t() const {
-    return handle_;
-  }
-
- private:
-  cudaGraphExec_t handle_;
-};
-
 // Throw exception if the cuda API does not succeed.
 void check_cublas_error(const char* name, cublasStatus_t err);
 void check_cuda_error(const char* name, cudaError_t err);
@@ -66,4 +28,62 @@ void check_cuda_error(const char* name, CUresult err);
 // Convert Dtype to CUDA C++ types.
 const char* dtype_to_cuda_type(const Dtype& dtype);
 
+// Base class for RAII managed CUDA resources.
+template <typename Handle, cudaError_t (*Destroy)(Handle)>
+class CudaHandle {
+ public:
+  CudaHandle(Handle handle = nullptr) : handle_(handle) {}
+
+  CudaHandle(CudaHandle&& other) : handle_(other.handle_) {
+    assert(this != &other);
+    other.handle_ = nullptr;
+  }
+
+  ~CudaHandle() {
+    reset();
+  }
+
+  CudaHandle(const CudaHandle&) = delete;
+  CudaHandle& operator=(const CudaHandle&) = delete;
+
+  CudaHandle& operator=(CudaHandle&& other) {
+    assert(this != &other);
+    reset();
+    std::swap(handle_, other.handle_);
+    return *this;
+  }
+
+  void reset() {
+    if (handle_ != nullptr) {
+      CHECK_CUDA_ERROR(Destroy(handle_));
+      handle_ = nullptr;
+    }
+  }
+
+  operator Handle() const {
+    return handle_;
+  }
+
+ protected:
+  Handle handle_;
+};
+
+// Wrappers of CUDA resources.
+class CudaGraph : public CudaHandle<cudaGraph_t, cudaGraphDestroy> {
+ public:
+  using CudaHandle::CudaHandle;
+  explicit CudaGraph(cu::Device& device);
+  void end_capture(cudaStream_t stream);
+};
+
+class CudaGraphExec : public CudaHandle<cudaGraphExec_t, cudaGraphExecDestroy> {
+ public:
+  void instantiate(cudaGraph_t graph);
+};
+
+class CudaStream : public CudaHandle<cudaStream_t, cudaStreamDestroy> {
+ public:
+  explicit CudaStream(cu::Device& device);
+};
+
 } // namespace mlx::core

mlx/backend/gpu/copy.cpp

@@ -52,4 +52,70 @@ array contiguous_copy_gpu(const array& arr, const Stream& s) {
   return arr_copy;
 }
 
+void reshape_gpu(const array& in, array& out, Stream s) {
+  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
+  if (copy_necessary) {
+    out.set_data(allocator::malloc(out.nbytes()));
+    copy_gpu_inplace(
+        in,
+        out,
+        in.shape(),
+        in.strides(),
+        make_contiguous_strides(in.shape()),
+        0,
+        0,
+        CopyType::General,
+        s);
+  } else {
+    shared_buffer_reshape(in, out_strides, out);
+  }
+}
+
+array flatten_in_eval(const array& x, int start_axis, int end_axis, Stream s) {
+  int ndim = x.ndim();
+  if (start_axis < 0) {
+    start_axis += ndim;
+  }
+  if (end_axis < 0) {
+    end_axis += ndim;
+  }
+  start_axis = std::max(0, start_axis);
+  end_axis = std::min(ndim - 1, end_axis);
+
+  return reshape_in_eval(x, Flatten::output_shape(x, start_axis, end_axis), s);
+}
+
+array reshape_in_eval(const array& x, Shape shape, Stream s) {
+  array out(std::move(shape), x.dtype(), nullptr, {});
+  reshape_gpu(x, out, s);
+  return out;
+}
+
+array swapaxes_in_eval(const array& x, int axis1, int axis2) {
+  int ndim = x.ndim();
+  if (axis1 < 0) {
+    axis1 += ndim;
+  }
+  if (axis2 < 0) {
+    axis2 += ndim;
+  }
+
+  auto shape = x.shape();
+  std::swap(shape[axis1], shape[axis2]);
+  auto strides = x.strides();
+  std::swap(strides[axis1], strides[axis2]);
+
+  auto [data_size, row_contiguous, col_contiguous] =
+      check_contiguity(shape, strides);
+  bool contiguous = data_size == x.data_size();
+
+  array out(std::move(shape), x.dtype(), nullptr, {});
+  out.copy_shared_buffer(
+      x,
+      std::move(strides),
+      {contiguous, row_contiguous, col_contiguous},
+      x.data_size());
+  return out;
+}
+
 } // namespace mlx::core

mlx/backend/gpu/copy.h

@@ -20,8 +20,8 @@ void copy_gpu_inplace(
     int64_t o_offset,
     CopyType ctype,
     const Stream& s,
-    const std::optional<array>& dynamic_i_offset = std::nullopt,
-    const std::optional<array>& dynamic_o_offset = std::nullopt);
+    std::optional<array> dynamic_i_offset = std::nullopt,
+    std::optional<array> dynamic_o_offset = std::nullopt);
 
 void copy_gpu(const array& src, array& out, CopyType ctype, const Stream& s);
 void copy_gpu(const array& src, array& out, CopyType ctype);
@@ -46,4 +46,12 @@ void fill_gpu(const array& val, array& out, const Stream& s);
 // Return a contiguous array with same shape that copies the data of |arr|.
 array contiguous_copy_gpu(const array& arr, const Stream& s);
 
+// Copy data from |in| and transpose to |out|'s shape.
+void reshape_gpu(const array& in, array& out, Stream s);
+
+// Like the normal ops but safe to call in eval_gpu.
+array flatten_in_eval(const array& x, int start_axis, int end_axis, Stream s);
+array reshape_in_eval(const array& x, Shape shape, Stream s);
+array swapaxes_in_eval(const array& x, int axis1, int axis2);
+
 } // namespace mlx::core

mlx/backend/gpu/primitives.cpp

@@ -20,29 +20,6 @@
 
 namespace mlx::core {
 
-namespace {
-
-void reshape(const array& in, array& out, Stream s) {
-  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
-  if (copy_necessary) {
-    out.set_data(allocator::malloc(out.nbytes()));
-    copy_gpu_inplace(
-        in,
-        out,
-        in.shape(),
-        in.strides(),
-        make_contiguous_strides(in.shape()),
-        0,
-        0,
-        CopyType::General,
-        s);
-  } else {
-    shared_buffer_reshape(in, out_strides, out);
-  }
-}
-
-} // namespace
-
 void AsStrided::eval_gpu(const std::vector<array>& inputs, array& out) {
   MLX_PROFILER_RANGE("AsStrided::eval_gpu");
   eval(inputs, out);
@@ -103,6 +80,74 @@ void Depends::eval_gpu(
   eval(inputs, outputs);
 }
 
+void DynamicSlice::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("DynamicSlice::eval_gpu");
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  auto& in = inputs[0];
+  auto& start = inputs[1];
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto s = stream();
+  auto in_offset = compute_dynamic_offset(start, in.strides(), axes_, s);
+  copy_gpu_inplace(
+      /* const array& src = */ in,
+      /* array& dst = */ out,
+      /* const Shape& data_shape = */ out.shape(),
+      /* const Strides& i_strides = */ in.strides(),
+      /* const Strides& o_strides = */ out.strides(),
+      /* int64_t i_offset = */ 0,
+      /* int64_t o_offset = */ 0,
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      /* const Stream& s = */ s,
+      /* std::optional<array> dynamic_i_offset = */ std::move(in_offset),
+      /* std::optional<array> dynamic_o_offset = */ std::nullopt);
+}
+
+void DynamicSliceUpdate::eval_gpu(
+    const std::vector<array>& inputs,
+    array& out) {
+  MLX_PROFILER_RANGE("DynamicSliceUpdate::eval_gpu");
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  auto& in = inputs[0];
+  auto& upd = inputs[1];
+  auto& start_indices = inputs[2];
+
+  if (upd.size() == 0) {
+    out.copy_shared_buffer(in);
+    return;
+  }
+
+  // Copy or donate input to output
+  auto s = stream();
+  auto ctype = in.flags().contiguous && in.size() == in.data_size()
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_gpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, s);
+
+  auto out_offset =
+      compute_dynamic_offset(start_indices, out.strides(), axes_, s);
+  copy_gpu_inplace(
+      /* const array& src = */ upd,
+      /* array& dst = */ out,
+      /* const Shape& data_shape = */ upd.shape(),
+      /* const Strides& i_strides = */ upd.strides(),
+      /* const Strides& o_strides = */ out.strides(),
+      /* int64_t i_offset = */ 0,
+      /* int64_t o_offset = */ 0,
+      /* CopyType ctype = */ CopyType::GeneralGeneral,
+      /* const Stream& s = */ s,
+      /* std::optional<array> dynamic_i_offset = */ std::nullopt,
+      /* std::optional<array> dynamic_o_offset = */ std::move(out_offset));
+}
+
 void ExpandDims::eval_gpu(const std::vector<array>& inputs, array& out) {
   MLX_PROFILER_RANGE("ExpandDims::eval_gpu");
   eval(inputs, out);
@@ -124,7 +169,7 @@ void Full::eval_gpu(const std::vector<array>& inputs, array& out) {
 
 void Flatten::eval_gpu(const std::vector<array>& inputs, array& out) {
   MLX_PROFILER_RANGE("Flatten::eval_gpu");
-  reshape(inputs[0], out, stream());
+  reshape_gpu(inputs[0], out, stream());
 }
 
 void NumberOfElements::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -150,7 +195,7 @@ void Pad::eval_gpu(const std::vector<array>& inputs, array& out) {
 
 void Reshape::eval_gpu(const std::vector<array>& inputs, array& out) {
   MLX_PROFILER_RANGE("Reshape::eval_gpu");
-  reshape(inputs[0], out, stream());
+  reshape_gpu(inputs[0], out, stream());
 }
 
 void Split::eval_gpu(
@@ -224,7 +269,7 @@ void Transpose::eval_gpu(const std::vector<array>& inputs, array& out) {
 
 void Unflatten::eval_gpu(const std::vector<array>& inputs, array& out) {
   MLX_PROFILER_RANGE("Unflatten::eval_gpu");
-  reshape(inputs[0], out, stream());
+  reshape_gpu(inputs[0], out, stream());
 }
 
 void View::eval_gpu(const std::vector<array>& inputs, array& out) {

mlx/backend/gpu/slicing.h

@@ -27,4 +27,10 @@ void pad_gpu(
     const Shape& low_pad_size,
     const Stream& s);
 
+array compute_dynamic_offset(
+    const array& indices,
+    const Strides& strides,
+    const std::vector<int>& axes,
+    const Stream& s);
+
 } // namespace mlx::core

mlx/backend/metal/CMakeLists.txt

@@ -33,10 +33,11 @@ make_jit_source(unary_ops kernels/erf.h kernels/expm1f.h)
 make_jit_source(binary_ops)
 make_jit_source(ternary_ops)
 make_jit_source(reduce_utils kernels/atomic.h kernels/reduction/ops.h)
-make_jit_source(scatter kernels/indexing.h)
-make_jit_source(gather kernels/indexing.h)
-make_jit_source(gather_axis)
-make_jit_source(scatter_axis)
+make_jit_source(indexing/scatter kernels/indexing/indexing.h)
+make_jit_source(indexing/gather kernels/indexing/indexing.h)
+make_jit_source(indexing/gather_front kernels/indexing/indexing.h)
+make_jit_source(indexing/gather_axis)
+make_jit_source(indexing/scatter_axis)
 make_jit_source(hadamard)
 
 if(MLX_METAL_JIT)
@@ -77,7 +78,10 @@ if(MLX_METAL_JIT)
   make_jit_source(steel/conv/kernels/steel_conv)
   make_jit_source(steel/conv/kernels/steel_conv_general kernels/steel/defines.h
                   kernels/steel/conv/loaders/loader_general.h)
-  make_jit_source(quantized)
+
+  make_jit_source(quantized_utils)
+  make_jit_source(quantized kernels/quantized_utils.h)
+  make_jit_source(fp4_quantized kernels/quantized_utils.h)
   make_jit_source(gemv_masked)
 else()
   target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/nojit_kernels.cpp)

mlx/backend/metal/copy.cpp

@@ -20,8 +20,8 @@ void copy_gpu_inplace(
     int64_t out_offset,
     CopyType ctype,
     const Stream& s,
-    const std::optional<array>& dynamic_i_offset /* = std::nullopt */,
-    const std::optional<array>& dynamic_o_offset /* = std::nullopt */) {
+    std::optional<array> dynamic_i_offset /* = std::nullopt */,
+    std::optional<array> dynamic_o_offset /* = std::nullopt */) {
   if (out.size() == 0) {
     return;
   }

mlx/backend/metal/custom_kernel.cpp

@@ -172,7 +172,7 @@ std::string write_template(
   return template_def.str();
 }
 
-MetalKernelFunction metal_kernel(
+CustomKernelFunction metal_kernel(
     const std::string& name,
     const std::vector<std::string>& input_names,
     const std::vector<std::string>& output_names,
@@ -316,7 +316,10 @@ MetalKernelFunction metal_kernel(
             threadgroup,
             shape_infos,
             ensure_row_contiguous,
-            init_value),
+            init_value,
+            std::vector<ScalarArg>{},
+            false,
+            0),
         std::move(inputs));
   };
 }

mlx/backend/metal/device.h

@@ -60,22 +60,12 @@ struct CommandEncoder {
     enc_->updateFence(fence);
   }
 
-  template <typename T>
-  void set_vector_bytes(const SmallVector<T>& vec, size_t nelems, int idx) {
-    enc_->setBytes(vec.data(), nelems * sizeof(T), idx);
+  template <typename Vec, typename = std::enable_if_t<is_vector_v<Vec>>>
+  void set_vector_bytes(const Vec& vec, size_t nelems, int idx) {
+    enc_->setBytes(vec.data(), nelems * sizeof(typename Vec::value_type), idx);
   }
-  template <typename T>
-  void set_vector_bytes(const SmallVector<T>& vec, int idx) {
-    return set_vector_bytes(vec, vec.size(), idx);
-  }
-
-  // TODO: Code is duplicated but they should be deleted soon.
-  template <typename T>
-  void set_vector_bytes(const std::vector<T>& vec, size_t nelems, int idx) {
-    enc_->setBytes(vec.data(), nelems * sizeof(T), idx);
-  }
-  template <typename T>
-  void set_vector_bytes(const std::vector<T>& vec, int idx) {
+  template <typename Vec, typename = std::enable_if_t<is_vector_v<Vec>>>
+  void set_vector_bytes(const Vec& vec, int idx) {
     return set_vector_bytes(vec, vec.size(), idx);
   }
 

mlx/backend/metal/indexing.cpp

@@ -52,8 +52,10 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& s = stream();
   auto& d = metal::device(s.device);
 
-  int idx_ndim = nidx ? inputs[1].ndim() : 0;
-  size_t ndim = src.ndim();
+  size_t slice_size = 1;
+  for (auto s : slice_sizes_) {
+    slice_size *= s;
+  }
 
   bool large_index = nidx && inputs[1].size() > INT32_MAX;
   bool large_src = src.size() > INT32_MAX;
@@ -61,6 +63,55 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   bool large = large_index || large_src || large_out;
 
   std::string idx_type_name = nidx ? type_to_name(inputs[1]) : "";
+
+  if (src.flags().row_contiguous && nidx == 1 && axes_[0] == 0 &&
+      inputs[1].flags().row_contiguous && slice_size == src.strides()[0]) {
+    int work_per_thread = (slice_size > 8 && src.dtype().size() < 4) ? 2 : 1;
+    auto& indices = inputs[1];
+    std::string kernel_name = fmt::format(
+        "gather_front{0}_{1}_{2}_{3}",
+        type_to_name(out),
+        idx_type_name,
+        large ? "int64_t" : "int",
+        work_per_thread);
+    std::string lib_name = kernel_name;
+
+    auto lib = d.get_library(lib_name, [&]() {
+      std::string kernel_source = metal::utils();
+      kernel_source += metal::gather_front();
+      kernel_source += get_template_definition(
+          kernel_name,
+          "gather_front",
+          get_type_string(out.dtype()),
+          get_type_string(indices.dtype()),
+          large ? "int64_t" : "int",
+          work_per_thread);
+
+      return kernel_source;
+    });
+
+    auto& compute_encoder = d.get_command_encoder(s.index);
+    auto kernel = d.get_kernel(kernel_name, lib);
+    compute_encoder.set_compute_pipeline_state(kernel);
+
+    size_t dim_x = (slice_size + work_per_thread - 1) / work_per_thread;
+    size_t dim_y = indices.size();
+    auto group_dims = get_block_dims(dim_x, dim_y, 1);
+    MTL::Size grid_dims = MTL::Size(dim_x, dim_y, 1);
+
+    compute_encoder.set_input_array(src, 0);
+    compute_encoder.set_input_array(indices, 1);
+    compute_encoder.set_output_array(out, 2);
+    compute_encoder.set_bytes(slice_size, 3);
+    compute_encoder.set_bytes(src.shape(0), 4);
+    compute_encoder.dispatch_threads(grid_dims, group_dims);
+
+    return;
+  }
+
+  int idx_ndim = nidx ? inputs[1].ndim() : 0;
+  size_t ndim = src.ndim();
+
   std::string kernel_name = fmt::format(
       "gather{0}{1}_{2}_{3}_{4}",
       type_to_name(out),
@@ -96,11 +147,6 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto kernel = d.get_kernel(kernel_name, lib);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  size_t slice_size = 1;
-  for (auto s : slice_sizes_) {
-    slice_size *= s;
-  }
-
   // Launch 3D grid of threads
   // First two dimensions for the indices, the last one for the slice
   size_t dim0 = 1;

mlx/backend/metal/jit/includes.h

@@ -19,9 +19,12 @@ const char* binary_two();
 const char* copy();
 const char* fft();
 const char* gather_axis();
+const char* gather_front();
 const char* hadamard();
 const char* logsumexp();
+const char* quantized_utils();
 const char* quantized();
+const char* fp4_quantized();
 const char* ternary();
 const char* scan();
 const char* scatter_axis();

mlx/backend/metal/jit_kernels.cpp

@@ -804,13 +804,19 @@ MTL::ComputePipelineState* get_fft_kernel(
 MTL::ComputePipelineState* get_quantized_kernel(
     metal::Device& d,
     const std::string& kernel_name,
-    const std::string& template_def) {
+    const std::string& template_def,
+    const std::string& mode) {
   const auto& lib_name = kernel_name;
   auto lib = d.get_library(lib_name, [&]() {
-    std::ostringstream kernel_source;
-    kernel_source << metal::utils() << metal::gemm() << metal::quantized()
-                  << template_def;
-    return kernel_source.str();
+    std::string kernel_source;
+    concatenate(
+        kernel_source,
+        metal::utils(),
+        metal::gemm(),
+        metal::quantized_utils(),
+        (mode == "affine") ? metal::quantized() : metal::fp4_quantized(),
+        template_def);
+    return kernel_source;
   });
   return d.get_kernel(kernel_name, lib);
 }
@@ -823,6 +829,7 @@ MTL::ComputePipelineState* get_gather_qmm_kernel(
     const array& x,
     int group_size,
     int bits,
+    const std::string& mode,
     int bm,
     int bn,
     int bk,
@@ -833,22 +840,40 @@ MTL::ComputePipelineState* get_gather_qmm_kernel(
   auto lib = d.get_library(lib_name, [&]() {
     std::string kernel_source;
     concatenate(
-        kernel_source,
-        metal::utils(),
-        metal::gemm(),
-        metal::quantized(),
-        get_template_definition(
-            lib_name,
-            "gather_qmm_rhs",
-            get_type_string(x.dtype()),
-            group_size,
-            bits,
-            bm,
-            bn,
-            bk,
-            wm,
-            wn,
-            transpose));
+        kernel_source, metal::utils(), metal::quantized_utils(), metal::gemm());
+    if (mode == "affine") {
+      concatenate(
+          kernel_source,
+          metal::quantized(),
+          get_template_definition(
+              lib_name,
+              mode + "_gather_qmm_rhs",
+              get_type_string(x.dtype()),
+              group_size,
+              bits,
+              bm,
+              bn,
+              bk,
+              wm,
+              wn,
+              transpose));
+    } else {
+      concatenate(
+          kernel_source,
+          metal::fp4_quantized(),
+          get_template_definition(
+              lib_name,
+              mode + "_gather_qmm_rhs",
+              get_type_string(x.dtype()),
+              group_size,
+              "uint8_t",
+              bm,
+              bn,
+              bk,
+              wm,
+              wn,
+              transpose));
+    }
     return kernel_source;
   });
   return d.get_kernel(kernel_name, lib, hash_name, func_consts);

mlx/backend/metal/kernels.h

@@ -238,7 +238,8 @@ MTL::ComputePipelineState* get_fft_kernel(
 MTL::ComputePipelineState* get_quantized_kernel(
     metal::Device& d,
     const std::string& kernel_name,
-    const std::string& template_def);
+    const std::string& template_def,
+    const std::string& mode);
 
 MTL::ComputePipelineState* get_gather_qmm_kernel(
     metal::Device& d,
@@ -248,6 +249,7 @@ MTL::ComputePipelineState* get_gather_qmm_kernel(
     const array& x,
     int group_size,
     int bits,
+    const std::string& mode,
     int bm,
     int bn,
     int bk,

mlx/backend/metal/kernels/CMakeLists.txt

@@ -108,7 +108,8 @@ if(NOT MLX_METAL_JIT)
     reduction/reduce_all.h
     reduction/reduce_col.h
     reduction/reduce_row.h)
-  build_kernel(quantized quantized.h ${STEEL_HEADERS})
+  build_kernel(quantized quantized.h quantized_utils.h ${STEEL_HEADERS})
+  build_kernel(fp4_quantized fp4_quantized.h quantized_utils.h ${STEEL_HEADERS})
   build_kernel(scan scan.h)
   build_kernel(softmax softmax.h)
   build_kernel(logsumexp logsumexp.h)

mlx/backend/metal/kernels/binary_ops.h

@@ -223,6 +223,11 @@ struct Power {
   template <typename T>
   metal::enable_if_t<metal::is_integral_v<T>, T> operator()(T base, T exp) {
     T res = 1;
+    // Undefined to raise integer to negative power
+    if (exp < 0) {
+      return 0;
+    }
+
     while (exp) {
       if (exp & 1) {
         res *= base;

mlx/backend/metal/kernels/conv.metal

@@ -166,115 +166,6 @@ instantiate_naive_unfold_nd_dims(float32, float);
 instantiate_naive_unfold_nd_dims(float16, half);
 instantiate_naive_unfold_nd_dims(bfloat16, bfloat16_t);
 
-///////////////////////////////////////////////////////////////////////////////
-/// Slow and naive conv2d kernels
-///////////////////////////////////////////////////////////////////////////////
-
-template <
-    typename T,
-    const int BM, /* Threadgroup rows (in threads) */
-    const int BN, /* Threadgroup cols (in threads) */
-    const int TM, /* Thread rows (in elements) */
-    const int TN, /* Thread cols (in elements) */
-    const int BC = 16>
-[[kernel]] void naive_conv_2d(
-    const device T* in [[buffer(0)]],
-    const device T* wt [[buffer(1)]],
-    device T* out [[buffer(2)]],
-    const constant MLXConvParams<2>& params [[buffer(3)]],
-    uint3 tid [[threadgroup_position_in_grid]],
-    uint3 lid [[thread_position_in_threadgroup]],
-    uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]]) {
-  (void)simd_gid;
-  (void)simd_lid;
-
-  out += tid.z * params.out_strides[0];
-  in += tid.z * params.in_strides[0];
-
-  int out_o = tid.y * BN * TN + lid.y * TN;
-  int out_hw = tid.x * BM * TM + lid.x * TM;
-
-  int out_h[TM];
-  int out_w[TN];
-
-  for (int m = 0; m < TM; ++m) {
-    int mm = (out_hw + m);
-    out_h[m] = mm / params.oS[1];
-    out_w[m] = mm % params.oS[1];
-  }
-
-  T in_local[TM];
-  T wt_local[TN];
-  T out_local[TM * TN] = {T(0)};
-
-  for (int h = 0; h < params.wS[0]; ++h) {
-    for (int w = 0; w < params.wS[1]; ++w) {
-      for (int c = 0; c < params.C; ++c) {
-        // Local in
-        for (int m = 0; m < TM; m++) {
-          int i = out_h[m] * params.str[0] - params.pad[0] + h * params.kdil[0];
-          int j = out_w[m] * params.str[1] - params.pad[1] + w * params.kdil[1];
-
-          bool valid = i >= 0 && i < params.iS[0] && j >= 0 && j < params.iS[1];
-          in_local[m] = valid
-              ? in[i * params.in_strides[1] + j * params.in_strides[2] + c]
-              : T(0);
-        }
-
-        // Load weight
-        for (int n = 0; n < TN; ++n) {
-          int o = out_o + n;
-          wt_local[n] = o < params.O
-              ? wt[o * params.wt_strides[0] + h * params.wt_strides[1] +
-                   w * params.wt_strides[2] + c]
-              : T(0);
-        }
-
-        // Accumulate
-        for (int m = 0; m < TM; ++m) {
-          for (int n = 0; n < TN; ++n) {
-            out_local[m * TN + n] += in_local[m] * wt_local[n];
-          }
-        }
-      }
-    }
-  }
-
-  for (int m = 0; m < TM; ++m) {
-    for (int n = 0; n < TN; ++n) {
-      if (out_h[m] < params.oS[0] && out_w[m] < params.oS[1] &&
-          (out_o + n) < params.O)
-        out[out_h[m] * params.out_strides[1] +
-            out_w[m] * params.out_strides[2] + out_o + n] =
-            out_local[m * TN + n];
-    }
-  }
-}
-
-// Instantiations
-
-#define instantiate_naive_conv_2d(name, itype, bm, bn, tm, tn)              \
-  template [[host_name("naive_conv_2d_" #name "_bm" #bm "_bn" #bn "_tm" #tm \
-                       "_tn" #tn)]] [[kernel]] void                         \
-  naive_conv_2d<itype, bm, bn, tm, tn>(                                     \
-      const device itype* in [[buffer(0)]],                                 \
-      const device itype* wt [[buffer(1)]],                                 \
-      device itype* out [[buffer(2)]],                                      \
-      const constant MLXConvParams<2>& params [[buffer(3)]],                \
-      uint3 tid [[threadgroup_position_in_grid]],                           \
-      uint3 lid [[thread_position_in_threadgroup]],                         \
-      uint simd_gid [[simdgroup_index_in_threadgroup]],                     \
-      uint simd_lid [[thread_index_in_simdgroup]]);
-
-#define instantiate_naive_conv_2d_blocks(name, itype) \
-  instantiate_naive_conv_2d(name, itype, 16, 8, 4, 4) \
-      instantiate_naive_conv_2d(name, itype, 16, 8, 2, 4)
-
-instantiate_naive_conv_2d_blocks(float32, float);
-instantiate_naive_conv_2d_blocks(float16, half);
-instantiate_naive_conv_2d_blocks(bfloat16, bfloat16_t);
-
 ///////////////////////////////////////////////////////////////////////////////
 /// Depthwise convolution kernels
 ///////////////////////////////////////////////////////////////////////////////

mlx/backend/metal/kernels/fp4_quantized.metal

@@ -0,0 +1,127 @@
+// Copyright © 2025 Apple Inc.
+
+// clang-format off
+#include "mlx/backend/metal/kernels/utils.h"
+#include "mlx/backend/metal/kernels/steel/gemm/gemm.h"
+#include "mlx/backend/metal/kernels/quantized_utils.h"
+#include "mlx/backend/metal/kernels/fp4_quantized.h"
+
+#define instantiate_quantized(name, type) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4", \
+      name, \
+      type, \
+      32,   \
+      uint8_t)
+
+#define instantiate_quantized_batched(name, type, batched) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_batch_" #batched, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      batched)
+
+#define instantiate_quantized_aligned(name, type, aligned) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_alN_" #aligned, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      aligned)
+
+#define instantiate_quantized_aligned_batched(name, type, aligned, batched) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_alN_" #aligned "_batch_" #batched, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      aligned, \
+      batched)
+
+#define instantiate_quantized_quad(name, type, D, batched) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_d_" #D "_batch_" #batched, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      D,       \
+      batched)
+
+#define instantiate_quantized_split_k(name, type, split_k) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_spk_" #split_k, \
+      name,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      split_k)
+
+#define instantiate_gather_qmm_rhs(func, name, type, bm, bn, bk, wm, wn, transpose) \
+  instantiate_kernel( \
+      #name "_" #type "_gs_32_b_4_bm_" #bm "_bn_" #bn "_bk_" #bk "_wm_" #wm "_wn_" #wn, \
+      func,    \
+      type,    \
+      32,      \
+      uint8_t, \
+      bm,      \
+      bn,      \
+      bk,      \
+      wm,      \
+      wn,      \
+      transpose)
+
+#define instantiate_quantized_batched_wrap(name, type) \
+  instantiate_quantized_batched(name, type, 1)         \
+  instantiate_quantized_batched(name, type, 0)
+
+#define instantiate_quantized_all_batched(type) \
+  instantiate_quantized_batched_wrap(mxfp4_qmv_fast, type) \
+  instantiate_quantized_batched_wrap(mxfp4_qmv, type)      \
+  instantiate_quantized_batched_wrap(mxfp4_qvm, type)      \
+  instantiate_quantized_batched_wrap(mxfp4_qmm_n, type)
+
+#define instantiate_quantized_all_single(type) \
+  instantiate_quantized(mxfp4_gather_qmv_fast, type) \
+  instantiate_quantized(mxfp4_gather_qmv, type)      \
+  instantiate_quantized(mxfp4_gather_qvm, type)      \
+  instantiate_quantized(mxfp4_gather_qmm_n, type)
+
+#define instantiate_quantized_all_aligned(type) \
+  instantiate_quantized_aligned(mxfp4_gather_qmm_t, type, true)      \
+  instantiate_quantized_aligned(mxfp4_gather_qmm_t, type, false)     \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, true, 1)  \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, true, 0)  \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, false, 1) \
+  instantiate_quantized_aligned_batched(mxfp4_qmm_t, type, false, 0)
+
+#define instantiate_quantized_all_quad(type) \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 64, 1)  \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 64, 0)  \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 128, 1) \
+  instantiate_quantized_quad(mxfp4_qmv_quad, type, 128, 0)
+
+#define instantiate_quantized_all_splitk(type) \
+  instantiate_quantized_split_k(mxfp4_qvm_split_k, type, 8) \
+  instantiate_quantized_split_k(mxfp4_qvm_split_k, type, 32)
+
+#define instantiate_quantized_all_rhs(type) \
+  instantiate_gather_qmm_rhs(mxfp4_gather_qmm_rhs, mxfp4_gather_qmm_rhs_nt, type, 16, 32, 32, 1, 2, true) \
+  instantiate_gather_qmm_rhs(mxfp4_gather_qmm_rhs, mxfp4_gather_qmm_rhs_nn, type, 16, 32, 32, 1, 2, false)
+
+#define instantiate_quantized_types(type) \
+  instantiate_quantized_all_batched(type) \
+  instantiate_quantized_all_quad(type)    \
+  instantiate_quantized_all_splitk(type)  \
+  instantiate_quantized_all_single(type)  \
+  instantiate_quantized_all_aligned(type) \
+  instantiate_quantized_all_rhs(type)
+
+instantiate_quantized_types(float)
+instantiate_quantized_types(bfloat16_t)
+instantiate_quantized_types(float16_t)
+    // clang-format on

mlx/backend/metal/kernels/gemv_masked.h

@@ -262,36 +262,37 @@ struct GEMVKernel {
       vec_mask_offset += vec_mask_step;
     }
 
-    if (leftover > 0 &&
-        (!has_operand_mask ||
-         (bool(mat_mask[mat_mask_offset]) &&
-          bool(vec_mask[vec_mask_offset])))) {
-      T block_scale{1};
-      if (has_mul_operand_mask) {
-        block_scale =
-            T(mat_mask[mat_mask_offset]) * T(vec_mask[vec_mask_offset]);
-      }
-
-      load_safe<AccT>(in_vec, v_coeff, bn, in_size);
-
-      // Apply scale
-      if (has_mul_operand_mask) {
-        MLX_MTL_PRAGMA_UNROLL
-        for (int tn = 0; tn < TN; tn++) {
-          v_coeff[tn] *= block_scale;
+    if (leftover > 0) {
+      if (!has_operand_mask ||
+          (bool(mat_mask[mat_mask_offset]) &&
+           bool(vec_mask[vec_mask_offset]))) {
+        T block_scale{1};
+        if (has_mul_operand_mask) {
+          block_scale =
+              T(mat_mask[mat_mask_offset]) * T(vec_mask[vec_mask_offset]);
         }
-      }
 
-      // Per thread work loop
-      MLX_MTL_PRAGMA_UNROLL
-      for (int tm = 0; tm < TM; tm++) {
-        // Load for the row
-        load_safe(&mat[tm * matrix_ld], inter, bn, in_size);
+        load_safe<AccT>(in_vec, v_coeff, bn, in_size);
 
-        // Accumulate results
+        // Apply scale
+        if (has_mul_operand_mask) {
+          MLX_MTL_PRAGMA_UNROLL
+          for (int tn = 0; tn < TN; tn++) {
+            v_coeff[tn] *= block_scale;
+          }
+        }
+
+        // Per thread work loop
         MLX_MTL_PRAGMA_UNROLL
-        for (int tn = 0; tn < TN; tn++) {
-          result[tm] += inter[tn] * v_coeff[tn];
+        for (int tm = 0; tm < TM; tm++) {
+          // Load for the row
+          load_safe(&mat[tm * matrix_ld], inter, bn, in_size);
+
+          // Accumulate results
+          MLX_MTL_PRAGMA_UNROLL
+          for (int tn = 0; tn < TN; tn++) {
+            result[tm] += inter[tn] * v_coeff[tn];
+          }
         }
       }
     }
@@ -544,31 +545,32 @@ struct GEMVTKernel {
         vec_mask_offset += vec_mask_step;
       }
 
-      if (leftover > 0 &&
-          (!has_operand_mask ||
-           (bool(mat_mask[mat_mask_offset]) &&
-            bool(vec_mask[vec_mask_offset])))) {
-        T block_scale{1};
-        if (has_mul_operand_mask) {
-          block_scale =
-              T(mat_mask[mat_mask_offset]) * T(vec_mask[vec_mask_offset]);
-        }
-
-        for (int tm = 0; tm < TM && bm + tm < in_vec_size; tm++) {
-          v_coeff[tm] = static_cast<AccT>(in_vec[bm + tm]);
-
+      if (leftover > 0) {
+        if (!has_operand_mask ||
+            (bool(mat_mask[mat_mask_offset]) &&
+             bool(vec_mask[vec_mask_offset]))) {
+          T block_scale{1};
           if (has_mul_operand_mask) {
-            v_coeff[tm] *= block_scale;
+            block_scale =
+                T(mat_mask[mat_mask_offset]) * T(vec_mask[vec_mask_offset]);
           }
 
-          MLX_MTL_PRAGMA_UNROLL
-          for (int tn = 0; tn < TN; tn++) {
-            inter[tn] = mat[(bm + tm) * marix_ld + out_col + tn];
-          }
+          for (int tm = 0; tm < TM && bm + tm < in_vec_size; tm++) {
+            v_coeff[tm] = static_cast<AccT>(in_vec[bm + tm]);
 
-          MLX_MTL_PRAGMA_UNROLL
-          for (int tn = 0; tn < TN; tn++) {
-            result[tn] += v_coeff[tm] * inter[tn];
+            if (has_mul_operand_mask) {
+              v_coeff[tm] *= block_scale;
+            }
+
+            MLX_MTL_PRAGMA_UNROLL
+            for (int tn = 0; tn < TN; tn++) {
+              inter[tn] = mat[(bm + tm) * marix_ld + out_col + tn];
+            }
+
+            MLX_MTL_PRAGMA_UNROLL
+            for (int tn = 0; tn < TN; tn++) {
+              result[tn] += v_coeff[tm] * inter[tn];
+            }
           }
         }
       }

mlx/backend/metal/kernels/indexing/gather.h

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/backend/metal/kernels/indexing.h"
+#include "mlx/backend/metal/kernels/indexing/indexing.h"
 
 template <typename T, typename IdxT, int NIDX, int IDX_NDIM, typename LocT>
 METAL_FUNC void gather_impl(

mlx/backend/metal/kernels/indexing/gather_front.h

@@ -0,0 +1,24 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/metal/kernels/indexing/indexing.h"
+
+template <typename T, typename IdxT, typename LocT, int N>
+[[kernel]] void gather_front(
+    const device T* src,
+    const device IdxT* indices,
+    device T* out,
+    const constant int64_t& stride,
+    const constant int& size,
+    uint2 index [[thread_position_in_grid]],
+    uint2 grid_dim [[threads_per_grid]]) {
+  auto idx = offset_neg_idx(indices[index.y], size);
+  LocT src_idx = static_cast<LocT>(stride) * idx;
+  LocT out_idx = static_cast<LocT>(stride) * index.y;
+
+  int s_idx = N * index.x;
+  for (int i = 0; i < N && s_idx < stride; ++i, ++s_idx) {
+    out[out_idx + s_idx] = src[src_idx + s_idx];
+  }
+}

mlx/backend/metal/kernels/indexing/scatter.h

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/backend/metal/kernels/indexing.h"
+#include "mlx/backend/metal/kernels/indexing/indexing.h"
 
 template <
     typename T,

mlx/backend/metal/kernels/quantized.h

@@ -1434,7 +1434,7 @@ METAL_FUNC void adjust_matrix_offsets(
 }
 
 template <typename T, int group_size, int bits, int D, bool batched>
-[[kernel]] void qmv_quad(
+[[kernel]] void affine_qmv_quad(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1486,7 +1486,7 @@ template <typename T, int group_size, int bits, int D, bool batched>
 }
 
 template <typename T, int group_size, int bits, bool batched>
-[[kernel]] void qmv_fast(
+[[kernel]] void affine_qmv_fast(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1538,7 +1538,7 @@ template <typename T, int group_size, int bits, bool batched>
 }
 
 template <typename T, const int group_size, const int bits, bool batched>
-[[kernel]] void qmv(
+[[kernel]] void affine_qmv(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1590,7 +1590,7 @@ template <typename T, const int group_size, const int bits, bool batched>
 }
 
 template <typename T, const int group_size, const int bits, bool batched>
-[[kernel]] void qvm(
+[[kernel]] void affine_qvm(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1642,7 +1642,7 @@ template <typename T, const int group_size, const int bits, bool batched>
 }
 
 template <typename T, const int group_size, const int bits, int split_k = 32>
-[[kernel]] void qvm_split_k(
+[[kernel]] void affine_qvm_split_k(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1706,7 +1706,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void qmm_t(
+[[kernel]] void affine_qmm_t(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1764,7 +1764,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void qmm_n(
+[[kernel]] void affine_qmm_n(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1817,7 +1817,7 @@ template <
 }
 
 template <typename T, int group_size, int bits>
-[[kernel]] void gather_qmv_fast(
+[[kernel]] void affine_gather_qmv_fast(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1879,7 +1879,7 @@ template <typename T, int group_size, int bits>
 }
 
 template <typename T, int group_size, int bits>
-[[kernel]] void gather_qmv(
+[[kernel]] void affine_gather_qmv(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -1941,7 +1941,7 @@ template <typename T, int group_size, int bits>
 }
 
 template <typename T, int group_size, int bits>
-[[kernel]] void gather_qvm(
+[[kernel]] void affine_gather_qvm(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -2010,7 +2010,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void gather_qmm_t(
+[[kernel]] void affine_gather_qmm_t(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -2077,7 +2077,7 @@ template <
     const int BM = 32,
     const int BK = 32,
     const int BN = 32>
-[[kernel]] void gather_qmm_n(
+[[kernel]] void affine_gather_qmm_n(
     const device uint32_t* w [[buffer(0)]],
     const device T* scales [[buffer(1)]],
     const device T* biases [[buffer(2)]],
@@ -2138,92 +2138,6 @@ template <
       w, scales, biases, x, y, Xs, Ws, K, N, M, tid, lid, simd_gid, simd_lid);
 }
 
-template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
-METAL_FUNC void gemm_loop_aligned(
-    threadgroup T* As,
-    threadgroup T* Bs,
-    thread mma_t& mma_op,
-    thread loader_a_t& loader_a,
-    thread loader_b_t& loader_b,
-    const int k_iterations) {
-  for (int k = 0; k < k_iterations; k++) {
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Load elements into threadgroup memory
-    loader_a.load_unsafe();
-    loader_b.load_unsafe();
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Multiply and accumulate threadgroup elements
-    mma_op.mma(As, Bs);
-
-    // Prepare for next iteration
-    loader_a.next();
-    loader_b.next();
-  }
-}
-
-template <
-    bool rows_aligned,
-    bool cols_aligned,
-    bool transpose,
-    typename T,
-    typename mma_t,
-    typename loader_a_t,
-    typename loader_b_t>
-METAL_FUNC void gemm_loop_unaligned(
-    threadgroup T* As,
-    threadgroup T* Bs,
-    thread mma_t& mma_op,
-    thread loader_a_t& loader_a,
-    thread loader_b_t& loader_b,
-    const int k_iterations,
-    const short tgp_bm,
-    const short tgp_bn,
-    const short tgp_bk) {
-  for (int k = 0; k < k_iterations; k++) {
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Load elements into threadgroup memory
-    if (rows_aligned) {
-      loader_a.load_unsafe();
-    } else {
-      loader_a.load_safe(short2(tgp_bk, tgp_bm));
-    }
-    if (cols_aligned) {
-      loader_b.load_unsafe();
-    } else {
-      loader_b.load_safe(
-          transpose ? short2(tgp_bk, tgp_bn) : short2(tgp_bn, tgp_bk));
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    // Multiply and accumulate threadgroup elements
-    mma_op.mma(As, Bs);
-
-    // Prepare for next iteration
-    loader_a.next();
-    loader_b.next();
-  }
-}
-
-template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
-METAL_FUNC void gemm_loop_finalize(
-    threadgroup T* As,
-    threadgroup T* Bs,
-    thread mma_t& mma_op,
-    thread loader_a_t& loader_a,
-    thread loader_b_t& loader_b,
-    const short2 tile_a,
-    const short2 tile_b) {
-  loader_a.load_safe(tile_a);
-  loader_b.load_safe(tile_b);
-  threadgroup_barrier(mem_flags::mem_threadgroup);
-  mma_op.mma(As, Bs);
-}
-
 template <
     typename T,
     int group_size,
@@ -2234,7 +2148,7 @@ template <
     int WM,
     int WN,
     bool transpose>
-[[kernel]] void gather_qmm_rhs(
+[[kernel]] void affine_gather_qmm_rhs(
     const device T* x [[buffer(0)]],
     const device uint32_t* w [[buffer(1)]],
     const device T* scales [[buffer(2)]],

mlx/backend/metal/kernels/quantized.metal

@@ -3,6 +3,7 @@
 // clang-format off
 #include "mlx/backend/metal/kernels/utils.h"
 #include "mlx/backend/metal/kernels/steel/gemm/gemm.h"
+#include "mlx/backend/metal/kernels/quantized_utils.h"
 #include "mlx/backend/metal/kernels/quantized.h"
 
 #define instantiate_quantized(name, type, group_size, bits)     \
@@ -79,40 +80,40 @@
   instantiate_quantized_batched(name, type, group_size, bits, 0)
 
 #define instantiate_quantized_all_batched(type, group_size, bits) \
-  instantiate_quantized_batched_wrap(qmv_fast, type, group_size, bits)     \
-  instantiate_quantized_batched_wrap(qmv, type, group_size, bits)     \
-  instantiate_quantized_batched_wrap(qvm, type, group_size, bits)     \
-  instantiate_quantized_batched_wrap(qmm_n, type, group_size, bits)
+  instantiate_quantized_batched_wrap(affine_qmv_fast, type, group_size, bits)     \
+  instantiate_quantized_batched_wrap(affine_qmv, type, group_size, bits)     \
+  instantiate_quantized_batched_wrap(affine_qvm, type, group_size, bits)     \
+  instantiate_quantized_batched_wrap(affine_qmm_n, type, group_size, bits)
 
 #define instantiate_quantized_all_single(type, group_size, bits) \
   instantiate_quantized(affine_quantize, type, group_size, bits) \
   instantiate_quantized(affine_dequantize, type, group_size, bits)     \
-  instantiate_quantized(gather_qmv_fast, type, group_size, bits)     \
-  instantiate_quantized(gather_qmv, type, group_size, bits)     \
-  instantiate_quantized(gather_qvm, type, group_size, bits)     \
-  instantiate_quantized(gather_qmm_n, type, group_size, bits)
+  instantiate_quantized(affine_gather_qmv_fast, type, group_size, bits)     \
+  instantiate_quantized(affine_gather_qmv, type, group_size, bits)     \
+  instantiate_quantized(affine_gather_qvm, type, group_size, bits)     \
+  instantiate_quantized(affine_gather_qmm_n, type, group_size, bits)
 
 #define instantiate_quantized_all_aligned(type, group_size, bits)   \
-  instantiate_quantized_aligned(gather_qmm_t, type, group_size, bits, true) \
-  instantiate_quantized_aligned(gather_qmm_t, type, group_size, bits, false) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, true, 1) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, true, 0) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, false, 1) \
-  instantiate_quantized_aligned_batched(qmm_t, type, group_size, bits, false, 0)
+  instantiate_quantized_aligned(affine_gather_qmm_t, type, group_size, bits, true) \
+  instantiate_quantized_aligned(affine_gather_qmm_t, type, group_size, bits, false) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, true, 1) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, true, 0) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, false, 1) \
+  instantiate_quantized_aligned_batched(affine_qmm_t, type, group_size, bits, false, 0)
 
 #define instantiate_quantized_all_quad(type, group_size, bits)   \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 64, 1)   \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 64, 0)  \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 128, 1)  \
-  instantiate_quantized_quad(qmv_quad, type, group_size, bits, 128, 0)
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 64, 1)   \
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 64, 0)   \
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 128, 1)  \
+  instantiate_quantized_quad(affine_qmv_quad, type, group_size, bits, 128, 0)
 
 #define instantiate_quantized_all_splitk(type, group_size, bits)   \
-  instantiate_quantized_split_k(qvm_split_k, type, group_size, bits, 8)   \
-  instantiate_quantized_split_k(qvm_split_k, type, group_size, bits, 32)
+  instantiate_quantized_split_k(affine_qvm_split_k, type, group_size, bits, 8)   \
+  instantiate_quantized_split_k(affine_qvm_split_k, type, group_size, bits, 32)
 
 #define instantiate_quantized_all_rhs(type, group_size, bits) \
-  instantiate_gather_qmm_rhs(gather_qmm_rhs, gather_qmm_rhs_nt, type, group_size, bits, 16, 32, 32, 1, 2, true) \
-  instantiate_gather_qmm_rhs(gather_qmm_rhs, gather_qmm_rhs_nn, type, group_size, bits, 16, 32, 32, 1, 2, false)
+  instantiate_gather_qmm_rhs(affine_gather_qmm_rhs, affine_gather_qmm_rhs_nt, type, group_size, bits, 16, 32, 32, 1, 2, true) \
+  instantiate_gather_qmm_rhs(affine_gather_qmm_rhs, affine_gather_qmm_rhs_nn, type, group_size, bits, 16, 32, 32, 1, 2, false)
 
 #define instantiate_quantized_funcs(type, group_size, bits) \
   instantiate_quantized_all_single(type, group_size, bits)  \

mlx/backend/metal/kernels/quantized_utils.h

@@ -0,0 +1,90 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include <metal_simdgroup>
+#include <metal_stdlib>
+
+template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
+METAL_FUNC void gemm_loop_aligned(
+    threadgroup T* As,
+    threadgroup T* Bs,
+    thread mma_t& mma_op,
+    thread loader_a_t& loader_a,
+    thread loader_b_t& loader_b,
+    const int k_iterations) {
+  for (int k = 0; k < k_iterations; k++) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Load elements into threadgroup memory
+    loader_a.load_unsafe();
+    loader_b.load_unsafe();
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Multiply and accumulate threadgroup elements
+    mma_op.mma(As, Bs);
+
+    // Prepare for next iteration
+    loader_a.next();
+    loader_b.next();
+  }
+}
+
+template <
+    bool rows_aligned,
+    bool cols_aligned,
+    bool transpose,
+    typename T,
+    typename mma_t,
+    typename loader_a_t,
+    typename loader_b_t>
+METAL_FUNC void gemm_loop_unaligned(
+    threadgroup T* As,
+    threadgroup T* Bs,
+    thread mma_t& mma_op,
+    thread loader_a_t& loader_a,
+    thread loader_b_t& loader_b,
+    const int k_iterations,
+    const short tgp_bm,
+    const short tgp_bn,
+    const short tgp_bk) {
+  for (int k = 0; k < k_iterations; k++) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Load elements into threadgroup memory
+    if (rows_aligned) {
+      loader_a.load_unsafe();
+    } else {
+      loader_a.load_safe(short2(tgp_bk, tgp_bm));
+    }
+    if (cols_aligned) {
+      loader_b.load_unsafe();
+    } else {
+      loader_b.load_safe(
+          transpose ? short2(tgp_bk, tgp_bn) : short2(tgp_bn, tgp_bk));
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // Multiply and accumulate threadgroup elements
+    mma_op.mma(As, Bs);
+
+    // Prepare for next iteration
+    loader_a.next();
+    loader_b.next();
+  }
+}
+
+template <typename T, typename mma_t, typename loader_a_t, typename loader_b_t>
+METAL_FUNC void gemm_loop_finalize(
+    threadgroup T* As,
+    threadgroup T* Bs,
+    thread mma_t& mma_op,
+    thread loader_a_t& loader_a,
+    thread loader_b_t& loader_b,
+    const short2 tile_a,
+    const short2 tile_b) {
+  loader_a.load_safe(tile_a);
+  loader_b.load_safe(tile_b);
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  mma_op.mma(As, Bs);
+}

mlx/backend/metal/kernels/sort.h

@@ -45,7 +45,9 @@ struct ThreadSort {
       for (short j = i & 1; j < N_PER_THREAD - 1; j += 2) {
         if (op(vals[j + 1], vals[j])) {
           thread_swap(vals[j + 1], vals[j]);
-          thread_swap(idxs[j + 1], idxs[j]);
+          if (ARG_SORT) {
+            thread_swap(idxs[j + 1], idxs[j]);
+          }
         }
       }
     }
@@ -111,7 +113,9 @@ struct BlockMergeSort {
       bool pred = (b_idx < B_sz) && (a_idx >= A_sz || op(b, a));
 
       vals[i] = pred ? b : a;
-      idxs[i] = pred ? Bs_idx[b_idx] : As_idx[a_idx];
+      if (ARG_SORT) {
+        idxs[i] = pred ? Bs_idx[b_idx] : As_idx[a_idx];
+      }
 
       b_idx += short(pred);
       a_idx += short(!pred);

mlx/backend/metal/kernels/steel/conv/loaders/loader_channel_n.h

@@ -83,7 +83,7 @@ struct Conv2DInputBlockLoaderSmallChannels {
   const constant MLXConvParams<2>* params;
   const constant ImplicitGemmConv2DParams* gemm_params;
 
-  short weight_hw;
+  int weight_hw;
 
   const device T* src[n_rows];
 

mlx/backend/metal/no_metal.cpp

@@ -26,15 +26,15 @@ device_info() {
 
 namespace fast {
 
-MetalKernelFunction metal_kernel(
+CustomKernelFunction metal_kernel(
     const std::string&,
     const std::vector<std::string>&,
     const std::vector<std::string>&,
     const std::string&,
     const std::string&,
-    bool ensure_row_contiguous,
-    bool atomic_outputs) {
-  throw std::runtime_error("[metal_kernel] No GPU back-end.");
+    bool,
+    bool) {
+  throw std::runtime_error("[metal_kernel] No Metal back-end.");
 }
 
 } // namespace fast

mlx/backend/metal/nojit_kernels.cpp

@@ -283,6 +283,7 @@ MTL::ComputePipelineState* get_fft_kernel(
 MTL::ComputePipelineState* get_quantized_kernel(
     metal::Device& d,
     const std::string& kernel_name,
+    const std::string&,
     const std::string&) {
   return d.get_kernel(kernel_name);
 }
@@ -295,6 +296,7 @@ MTL::ComputePipelineState* get_gather_qmm_kernel(
     const array&,
     int,
     int,
+    const std::string&,
     int,
     int,
     int,

mlx/backend/metal/primitives.cpp

@@ -4,7 +4,6 @@
 #include <numeric>
 #include <sstream>
 
-#include "mlx/backend/common/compiled.h"
 #include "mlx/backend/common/slicing.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/gpu/copy.h"
@@ -25,60 +24,6 @@ void arange_set_scalars(T start, T next, metal::CommandEncoder& enc) {
   enc.set_bytes(step, 1);
 }
 
-static array compute_dynamic_offset(
-    const array& indices,
-    const Strides& strides,
-    const std::vector<int>& axes,
-    Stream s) {
-  auto& d = metal::device(s.device);
-
-  // Kernel to compute offset here.
-  array offset({1}, int64, nullptr, {});
-  bool donate = indices.is_donatable() &&
-      (indices.data_size() * indices.itemsize()) >= offset.itemsize();
-  if (donate) {
-    offset.copy_shared_buffer(indices);
-  } else {
-    offset.set_data(allocator::malloc(offset.itemsize()));
-  }
-  d.add_temporary(offset, s.index);
-
-  auto dtype = indices.dtype();
-  std::string lib_name = "compute_dynamic_offset_" + type_to_name(dtype);
-  auto lib = d.get_library(lib_name, [dtype]() {
-    return fmt::format(
-        R"(
-        [[kernel]] void compute_dynamic_offset_{0}(
-            constant const {1}* indices [[buffer(0)]],
-            device int64_t& offset [[buffer(1)]],
-            constant const int64_t* strides [[buffer(2)]],
-            constant const int* axes [[buffer(3)]],
-            constant const int& n_axes [[buffer(4)]],
-            uint index [[thread_position_in_grid]]) {{
-          int64_t acc = 0;
-          for (int i = 0; i < n_axes; ++i) {{
-            acc += indices[i] * strides[axes[i]];
-          }}
-          offset = acc;
-        }})",
-        type_to_name(dtype),
-        get_type_string(dtype));
-  });
-  auto kernel = d.get_kernel(lib_name, lib);
-
-  auto& compute_encoder = d.get_command_encoder(s.index);
-  compute_encoder.set_compute_pipeline_state(kernel);
-  compute_encoder.set_input_array(indices, 0);
-  compute_encoder.set_output_array(offset, 1);
-  compute_encoder.set_vector_bytes(strides, 2);
-  compute_encoder.set_vector_bytes(axes, 3);
-  int n_axes = axes.size();
-  compute_encoder.set_bytes(n_axes, 4);
-  MTL::Size dims = MTL::Size(1, 1, 1);
-  compute_encoder.dispatch_threads(dims, dims);
-  return offset;
-}
-
 void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 0);
   out.set_data(allocator::malloc(out.nbytes()));
@@ -256,72 +201,6 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   compute_encoder.dispatch_threads(grid_dims, group_dims);
 }
 
-void DynamicSlice::eval_gpu(const std::vector<array>& inputs, array& out) {
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-
-  auto& in = inputs[0];
-  auto& start = inputs[1];
-  out.set_data(allocator::malloc(out.nbytes()));
-  auto s = stream();
-  auto in_offset = compute_dynamic_offset(start, in.strides(), axes_, s);
-  copy_gpu_inplace(
-      /* const array& src = */ in,
-      /* array& dst = */ out,
-      /* const Shape& data_shape = */ out.shape(),
-      /* const Strides& i_strides = */ in.strides(),
-      /* const Strides& o_strides = */ out.strides(),
-      /* int64_t i_offset = */ 0,
-      /* int64_t o_offset = */ 0,
-      /* CopyType ctype = */ CopyType::GeneralGeneral,
-      /* const Stream& s = */ s,
-      /* const std::optional<array>& dynamic_i_offset = */ in_offset,
-      /* const std::optional<array>& dynamic_o_offset = */ std::nullopt);
-}
-
-void DynamicSliceUpdate::eval_gpu(
-    const std::vector<array>& inputs,
-    array& out) {
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-
-  auto& in = inputs[0];
-  auto& upd = inputs[1];
-  auto& start_indices = inputs[2];
-
-  if (upd.size() == 0) {
-    out.copy_shared_buffer(in);
-    return;
-  }
-
-  // Copy or donate input to output
-  auto s = stream();
-  auto& d = metal::device(s.device);
-  auto ctype = in.flags().contiguous && in.size() == in.data_size()
-      ? CopyType::Vector
-      : CopyType::General;
-  copy_gpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, s);
-
-  auto out_offset =
-      compute_dynamic_offset(start_indices, out.strides(), axes_, s);
-  copy_gpu_inplace(
-      /* const array& src = */ upd,
-      /* array& dst = */ out,
-      /* const Shape& data_shape = */ upd.shape(),
-      /* const Strides& i_strides = */ upd.strides(),
-      /* const Strides& o_strides = */ out.strides(),
-      /* int64_t i_offset = */ 0,
-      /* int64_t o_offset = */ 0,
-      /* CopyType ctype = */ CopyType::GeneralGeneral,
-      /* const Stream& s = */ s,
-      /* const std::optional<array>& dynamic_i_offset = */ std::nullopt,
-      /* const std::optional<array>& dynamic_o_offset = */ out_offset);
-}
-
 void QRF::eval_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {

mlx/backend/metal/quantized.cpp

@@ -1,7 +1,5 @@
 // Copyright © 2023-2024 Apple Inc.
 
-#include <cassert>
-
 #include "mlx/backend/common/broadcasting.h"
 #include "mlx/backend/common/compiled.h"
 #include "mlx/backend/gpu/copy.h"
@@ -17,6 +15,28 @@ namespace mlx::core {
 
 namespace {
 
+template <typename... Args>
+auto get_quantized_kernel_wrapped(
+    metal::Device& d,
+    const std::string& name,
+    const std::string& func,
+    const std::string& mode,
+    const std::string& type,
+    int group_size,
+    int bits,
+    Args... args) {
+  std::string template_def;
+  auto fname = mode + "_" + func;
+  if (mode == "affine") {
+    template_def = get_template_definition(
+        name, fname, type, group_size, bits, std::forward<Args>(args)...);
+  } else {
+    template_def = get_template_definition(
+        name, fname, type, group_size, "uint8_t", std::forward<Args>(args)...);
+  }
+  return get_quantized_kernel(d, name, template_def, mode);
+}
+
 inline array
 ensure_row_contiguous(const array& x, metal::Device& d, const Stream& s) {
   if (!x.flags().row_contiguous) {
@@ -99,7 +119,7 @@ inline int add_strides_and_shapes(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     int offset) {
   if (skip) {
     return 0;
@@ -109,16 +129,18 @@ inline int add_strides_and_shapes(
 
   int x_batch_ndims = x.ndim() - 2;
   int w_batch_ndims = w.ndim() - 2;
-  compute_encoder.set_bytes(x_batch_ndims, offset);
-  compute_encoder.set_vector_bytes(x.shape(), offset + 1);
-  compute_encoder.set_vector_bytes(x.strides(), offset + 2);
-  compute_encoder.set_bytes(w_batch_ndims, offset + 3);
-  compute_encoder.set_vector_bytes(w.shape(), offset + 4);
-  compute_encoder.set_vector_bytes(w.strides(), offset + 5);
-  compute_encoder.set_vector_bytes(scales.strides(), offset + 6);
-  compute_encoder.set_vector_bytes(biases.strides(), offset + 7);
+  compute_encoder.set_bytes(x_batch_ndims, offset++);
+  compute_encoder.set_vector_bytes(x.shape(), offset++);
+  compute_encoder.set_vector_bytes(x.strides(), offset++);
+  compute_encoder.set_bytes(w_batch_ndims, offset++);
+  compute_encoder.set_vector_bytes(w.shape(), offset++);
+  compute_encoder.set_vector_bytes(w.strides(), offset++);
+  compute_encoder.set_vector_bytes(scales.strides(), offset++);
+  if (biases) {
+    compute_encoder.set_vector_bytes(biases->strides(), offset++);
+  }
 
-  return 8;
+  return offset;
 }
 
 inline int add_gather_strides_and_shapes(
@@ -130,12 +152,12 @@ inline int add_gather_strides_and_shapes(
       lhs_indices.shape(), {lhs_indices.strides(), rhs_indices.strides()});
   int ndims = shape.size();
 
-  compute_encoder.set_bytes(ndims, offset);
-  compute_encoder.set_vector_bytes(shape, offset + 1);
-  compute_encoder.set_vector_bytes(strides[0], offset + 2);
-  compute_encoder.set_vector_bytes(strides[1], offset + 3);
+  compute_encoder.set_bytes(ndims, offset++);
+  compute_encoder.set_vector_bytes(shape, offset++);
+  compute_encoder.set_vector_bytes(strides[0], offset++);
+  compute_encoder.set_vector_bytes(strides[1], offset++);
 
-  return 4;
+  return offset;
 }
 
 } // namespace
@@ -144,7 +166,7 @@ void qmv_quad(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     array& out,
     int group_size,
     int bits,
@@ -152,7 +174,8 @@ void qmv_quad(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int B = out.size() / M / N;
 
   constexpr int quads_per_simd = 8;
@@ -165,9 +188,10 @@ void qmv_quad(
   std::string kname;
   kname.reserve(64);
   std::string type_string = get_type_string(x.dtype());
+
   concatenate(
       kname,
-      "qmv_quad_",
+      mode + "_qmv_quad_",
       type_string,
       "_gs_",
       group_size,
@@ -176,21 +200,22 @@ void qmv_quad(
       "_d_",
       K,
       B > 1 ? "_batch_1" : "_batch_0");
-  auto template_def = get_template_definition(
-      kname, "qmv_quad", type_string, group_size, bits, K, B > 1);
-
-  auto kernel = get_quantized_kernel(d, kname, template_def);
+  auto kernel = get_quantized_kernel_wrapped(
+      d, kname, "qmv_quad", mode, type_string, group_size, bits, K, B > 1);
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_output_array(out, 4);
-  compute_encoder.set_bytes(K, 5);
-  compute_encoder.set_bytes(N, 6);
-  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, 7);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(K, c++);
+  compute_encoder.set_bytes(N, c++);
+  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, c++);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -199,7 +224,7 @@ void qmv(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     array& out,
     int group_size,
     int bits,
@@ -207,7 +232,8 @@ void qmv(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int B = out.size() / M / N;
 
   int bn = 8;
@@ -219,30 +245,40 @@ void qmv(
   kname.reserve(64);
   std::string type_string = get_type_string(x.dtype());
   bool fast = N % bn == 0 && K % 512 == 0;
+
   concatenate(
       kname,
-      fast ? "qmv_fast_" : "qmv_",
+      mode + (fast ? "_qmv_fast_" : "_qmv_"),
       type_string,
       "_gs_",
       group_size,
       "_b_",
       bits,
       B > 1 ? "_batch_1" : "_batch_0");
-  auto template_def = get_template_definition(
-      kname, fast ? "qmv_fast" : "qmv", type_string, group_size, bits, B > 1);
+  auto kernel = get_quantized_kernel_wrapped(
+      d,
+      kname,
+      (fast ? "qmv_fast" : "qmv"),
+      mode,
+      type_string,
+      group_size,
+      bits,
+      B > 1);
 
-  auto kernel = get_quantized_kernel(d, kname, template_def);
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_output_array(out, 4);
-  compute_encoder.set_bytes(K, 5);
-  compute_encoder.set_bytes(N, 6);
-  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, 7);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(K, c++);
+  compute_encoder.set_bytes(N, c++);
+  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, c);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -251,7 +287,7 @@ void qvm_split_k(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     array& out,
     int group_size,
     int bits,
@@ -259,7 +295,8 @@ void qvm_split_k(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int split_k = K > 8192 ? 32 : 8;
   int split_D = (K + split_k - 1) / split_k;
   int B = out.size() / M / N;
@@ -283,7 +320,6 @@ void qvm_split_k(
   auto w_shape = w.shape();
   auto w_strides = w.strides();
   auto s_strides = scales.strides();
-  auto b_strides = biases.strides();
 
   // Add split_k dim with reshapes
   x_shape.insert(x_shape.end() - 2, split_k);
@@ -297,7 +333,6 @@ void qvm_split_k(
   w_strides.insert(w_strides.end() - 2, split_D * w.shape(-1));
   w_batch_ndims += 1;
   s_strides.insert(s_strides.end() - 2, split_D * scales.shape(-1));
-  b_strides.insert(b_strides.end() - 2, split_D * biases.shape(-1));
 
   int final_block_size = K - (split_k - 1) * split_D;
 
@@ -315,7 +350,7 @@ void qvm_split_k(
   kname.reserve(64);
   concatenate(
       kname,
-      "qvm_split_k_",
+      mode + "_qvm_split_k_",
       type_string,
       "_gs_",
       group_size,
@@ -323,31 +358,38 @@ void qvm_split_k(
       bits,
       "_spk_",
       split_k);
-  auto template_def = get_template_definition(
-      kname, "qvm_split_k", type_string, group_size, bits, split_k);
 
   // Encode and dispatch kernel
-  auto kernel = get_quantized_kernel(d, kname, template_def);
+  auto kernel = get_quantized_kernel_wrapped(
+      d, kname, "qvm_split_k", mode, type_string, group_size, bits, split_k);
+
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_output_array(intermediate, 4);
-  compute_encoder.set_bytes(split_D, 5);
-  compute_encoder.set_bytes(N, 6);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_output_array(intermediate, c++);
+  compute_encoder.set_bytes(split_D, c++);
+  compute_encoder.set_bytes(N, c++);
 
-  compute_encoder.set_bytes(x_batch_ndims, 7);
-  compute_encoder.set_vector_bytes(x_shape, 8);
-  compute_encoder.set_vector_bytes(x_strides, 9);
-  compute_encoder.set_bytes(w_batch_ndims, 10);
-  compute_encoder.set_vector_bytes(w_shape, 11);
-  compute_encoder.set_vector_bytes(w_strides, 12);
-  compute_encoder.set_vector_bytes(s_strides, 13);
-  compute_encoder.set_vector_bytes(b_strides, 14);
-  compute_encoder.set_bytes(final_block_size, 15);
+  compute_encoder.set_bytes(x_batch_ndims, c++);
+  compute_encoder.set_vector_bytes(x_shape, c++);
+  compute_encoder.set_vector_bytes(x_strides, c++);
+  compute_encoder.set_bytes(w_batch_ndims, c++);
+  compute_encoder.set_vector_bytes(w_shape, c++);
+  compute_encoder.set_vector_bytes(w_strides, c++);
+  compute_encoder.set_vector_bytes(s_strides, c++);
+  if (biases) {
+    auto b_strides = biases->strides();
+    b_strides.insert(b_strides.end() - 2, split_D * biases->shape(-1));
+    compute_encoder.set_vector_bytes(b_strides, c++);
+  }
+  compute_encoder.set_bytes(final_block_size, c++);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 
@@ -364,7 +406,7 @@ void qvm(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     array& out,
     int group_size,
     int bits,
@@ -372,7 +414,8 @@ void qvm(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int B = out.size() / M / N;
 
   int bn = 64;
@@ -385,28 +428,29 @@ void qvm(
   std::string type_string = get_type_string(x.dtype());
   concatenate(
       kname,
-      "qvm_",
+      mode + "_qvm_",
       type_string,
       "_gs_",
       group_size,
       "_b_",
       bits,
       B > 1 ? "_batch_1" : "_batch_0");
-  auto template_def = get_template_definition(
-      kname, "qvm", type_string, group_size, bits, B > 1);
-
-  auto kernel = get_quantized_kernel(d, kname, template_def);
+  auto kernel = get_quantized_kernel_wrapped(
+      d, kname, "qvm", mode, type_string, group_size, bits, B > 1);
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_output_array(out, 4);
-  compute_encoder.set_bytes(K, 5);
-  compute_encoder.set_bytes(N, 6);
-  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, 7);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(K, c++);
+  compute_encoder.set_bytes(N, c++);
+  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, c++);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -415,7 +459,7 @@ void qmm(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     array& out,
     bool transpose,
     int group_size,
@@ -424,7 +468,8 @@ void qmm(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int B = out.size() / M / N;
 
   int wm = 2;
@@ -441,7 +486,7 @@ void qmm(
   std::string type_string = get_type_string(x.dtype());
   concatenate(
       kname,
-      transpose ? "qmm_t_" : "qmm_n_",
+      mode + (transpose ? "_qmm_t_" : "_qmm_n_"),
       type_string,
       "_gs_",
       group_size,
@@ -450,27 +495,37 @@ void qmm(
       transpose ? (aligned ? "_alN_true" : "_alN_false") : "",
       batched ? "_batch_1" : "_batch_0");
   std::string template_def;
+  MTL::ComputePipelineState* kernel;
   if (transpose) {
-    template_def = get_template_definition(
-        kname, "qmm_t", type_string, group_size, bits, aligned, batched);
+    kernel = get_quantized_kernel_wrapped(
+        d,
+        kname,
+        "qmm_t",
+        mode,
+        type_string,
+        group_size,
+        bits,
+        aligned,
+        batched);
   } else {
-    template_def = get_template_definition(
-        kname, "qmm_n", type_string, group_size, bits, batched);
+    kernel = get_quantized_kernel_wrapped(
+        d, kname, "qmm_n", mode, type_string, group_size, bits, batched);
   }
-
-  auto kernel = get_quantized_kernel(d, kname, template_def);
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_output_array(out, 4);
-  compute_encoder.set_bytes(K, 5);
-  compute_encoder.set_bytes(N, 6);
-  compute_encoder.set_bytes(M, 7);
-  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, 8);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(K, c++);
+  compute_encoder.set_bytes(N, c++);
+  compute_encoder.set_bytes(M, c++);
+  add_strides_and_shapes(compute_encoder, B <= 1, x, w, scales, biases, c);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -479,7 +534,7 @@ void gather_qmm(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     const array& lhs_indices,
     const array& rhs_indices,
     array& out,
@@ -490,7 +545,8 @@ void gather_qmm(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int B = out.size() / M / N;
 
   int wm = 2;
@@ -503,44 +559,43 @@ void gather_qmm(
   std::string kname;
   kname.reserve(64);
   bool aligned = N % 32 == 0;
-  bool batched = B > 1;
   std::string type_string = get_type_string(x.dtype());
   concatenate(
       kname,
-      transpose ? "gather_qmm_t_" : "gather_qmm_n_",
+      mode + (transpose ? "_gather_qmm_t_" : "_gather_qmm_n_"),
       type_string,
       "_gs_",
       group_size,
       "_b_",
       bits,
       transpose ? (aligned ? "_alN_true" : "_alN_false") : "");
-  std::string template_def;
+  MTL::ComputePipelineState* kernel;
   if (transpose) {
-    template_def = get_template_definition(
-        kname, "gather_qmm_t", type_string, group_size, bits, aligned);
+    kernel = get_quantized_kernel_wrapped(
+        d, kname, "gather_qmm_t", mode, type_string, group_size, bits, aligned);
   } else {
-    template_def = get_template_definition(
-        kname, "gather_qmm_n", type_string, group_size, bits);
+    kernel = get_quantized_kernel_wrapped(
+        d, kname, "gather_qmm_n", mode, type_string, group_size, bits);
   }
 
-  auto kernel = get_quantized_kernel(d, kname, template_def);
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_input_array(lhs_indices, 4);
-  compute_encoder.set_input_array(rhs_indices, 5);
-  compute_encoder.set_output_array(out, 6);
-  compute_encoder.set_bytes(K, 7);
-  compute_encoder.set_bytes(N, 8);
-  compute_encoder.set_bytes(M, 9);
-  int n =
-      add_strides_and_shapes(compute_encoder, false, x, w, scales, biases, 10);
-  add_gather_strides_and_shapes(
-      compute_encoder, lhs_indices, rhs_indices, 10 + n);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_input_array(lhs_indices, c++);
+  compute_encoder.set_input_array(rhs_indices, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(K, c++);
+  compute_encoder.set_bytes(N, c++);
+  compute_encoder.set_bytes(M, c++);
+  c = add_strides_and_shapes(compute_encoder, false, x, w, scales, biases, c);
+  add_gather_strides_and_shapes(compute_encoder, lhs_indices, rhs_indices, c);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -549,7 +604,7 @@ void gather_qmv(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     const array& lhs_indices,
     const array& rhs_indices,
     array& out,
@@ -559,7 +614,8 @@ void gather_qmv(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int B = out.size() / M / N;
 
   int bn = 8;
@@ -573,36 +629,39 @@ void gather_qmv(
   bool fast = N % bn == 0 && K % 512 == 0;
   concatenate(
       kname,
-      fast ? "gather_qmv_fast_" : "gather_qmv_",
+      mode + (fast ? "_gather_qmv_fast_" : "_gather_qmv_"),
       type_string,
       "_gs_",
       group_size,
       "_b_",
       bits);
-  auto template_def = get_template_definition(
+
+  auto kernel = get_quantized_kernel_wrapped(
+      d,
       kname,
-      fast ? "gather_qmv_fast" : "gather_qmv",
+      (fast ? "gather_qmv_fast" : "gather_qmv"),
+      mode,
       type_string,
       group_size,
       bits);
 
-  auto kernel = get_quantized_kernel(d, kname, template_def);
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_input_array(lhs_indices, 4);
-  compute_encoder.set_input_array(rhs_indices, 5);
-  compute_encoder.set_output_array(out, 6);
-  compute_encoder.set_bytes(K, 7);
-  compute_encoder.set_bytes(N, 8);
-  int n =
-      add_strides_and_shapes(compute_encoder, false, x, w, scales, biases, 9);
-  add_gather_strides_and_shapes(
-      compute_encoder, lhs_indices, rhs_indices, 9 + n);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_input_array(lhs_indices, c++);
+  compute_encoder.set_input_array(rhs_indices, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(K, c++);
+  compute_encoder.set_bytes(N, c++);
+  c = add_strides_and_shapes(compute_encoder, false, x, w, scales, biases, c);
+  add_gather_strides_and_shapes(compute_encoder, lhs_indices, rhs_indices, c);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -611,7 +670,7 @@ void gather_qvm(
     const array& x,
     const array& w,
     const array& scales,
-    const array& biases,
+    const std::optional<array>& biases,
     const array& lhs_indices,
     const array& rhs_indices,
     array& out,
@@ -621,7 +680,8 @@ void gather_qvm(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string& mode) {
   int B = out.size() / M / N;
 
   int bn = 64;
@@ -633,27 +693,32 @@ void gather_qvm(
   kname.reserve(64);
   std::string type_string = get_type_string(x.dtype());
   concatenate(
-      kname, "gather_qvm_", type_string, "_gs_", group_size, "_b_", bits);
-  auto template_def = get_template_definition(
-      kname, "gather_qvm", type_string, group_size, bits);
-
-  auto kernel = get_quantized_kernel(d, kname, template_def);
+      kname,
+      mode + "_gather_qvm_",
+      type_string,
+      "_gs_",
+      group_size,
+      "_b_",
+      bits);
+  auto kernel = get_quantized_kernel_wrapped(
+      d, kname, "gather_qvm", mode, type_string, group_size, bits);
   auto& compute_encoder = d.get_command_encoder(s.index);
   compute_encoder.set_compute_pipeline_state(kernel);
 
-  compute_encoder.set_input_array(w, 0);
-  compute_encoder.set_input_array(scales, 1);
-  compute_encoder.set_input_array(biases, 2);
-  compute_encoder.set_input_array(x, 3);
-  compute_encoder.set_input_array(lhs_indices, 4);
-  compute_encoder.set_input_array(rhs_indices, 5);
-  compute_encoder.set_output_array(out, 6);
-  compute_encoder.set_bytes(K, 7);
-  compute_encoder.set_bytes(N, 8);
-  int n =
-      add_strides_and_shapes(compute_encoder, false, x, w, scales, biases, 9);
-  add_gather_strides_and_shapes(
-      compute_encoder, lhs_indices, rhs_indices, 9 + n);
+  int c = 0;
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases) {
+    compute_encoder.set_input_array(*biases, c++);
+  }
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_input_array(lhs_indices, c++);
+  compute_encoder.set_input_array(rhs_indices, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(K, c++);
+  compute_encoder.set_bytes(N, c++);
+  c = add_strides_and_shapes(compute_encoder, false, x, w, scales, biases, c++);
+  add_gather_strides_and_shapes(compute_encoder, lhs_indices, rhs_indices, c);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -662,7 +727,7 @@ void gather_qmm_rhs(
     const array& x_,
     const array& w_,
     const array& scales_,
-    const array& biases_,
+    const std::optional<array>& biases_,
     const array& indices_,
     array& out,
     bool transpose,
@@ -672,7 +737,8 @@ void gather_qmm_rhs(
     int N,
     int K,
     metal::Device& d,
-    const Stream& s) {
+    const Stream& s,
+    const std::string mode) {
   // Start by normalizing the indices
   array indices = ensure_row_contiguous(indices_, d, s);
 
@@ -697,7 +763,6 @@ void gather_qmm_rhs(
   array x = broadcast_with_indices(x_);
   array w = ensure_row_contiguous(w_, d, s);
   array scales = ensure_row_contiguous(scales_, d, s);
-  array biases = ensure_row_contiguous(biases_, d, s);
 
   // TODO: Tune the block sizes
   int bm = 16, bn = 32, bk = 32;
@@ -713,7 +778,7 @@ void gather_qmm_rhs(
   std::string type_string = get_type_string(x.dtype());
   concatenate(
       kname,
-      transpose ? "gather_qmm_rhs_nt_" : "gather_qmm_rhs_nn_",
+      mode + (transpose ? "_gather_qmm_rhs_nt_" : "_gather_qmm_rhs_nn_"),
       type_string,
       "_gs_",
       group_size,
@@ -759,6 +824,7 @@ void gather_qmm_rhs(
       x,
       group_size,
       bits,
+      mode,
       bm,
       bn,
       bk,
@@ -770,15 +836,19 @@ void gather_qmm_rhs(
   MTL::Size group_dims(32, wn, wm);
   MTL::Size grid_dims((N + bn - 1) / bn, (M + bm - 1) / bm, 1);
 
-  compute_encoder.set_input_array(x, 0);
-  compute_encoder.set_input_array(w, 1);
-  compute_encoder.set_input_array(scales, 2);
-  compute_encoder.set_input_array(biases, 3);
-  compute_encoder.set_input_array(indices, 4);
-  compute_encoder.set_output_array(out, 5);
-  compute_encoder.set_bytes(M, 6);
-  compute_encoder.set_bytes(N, 7);
-  compute_encoder.set_bytes(K, 8);
+  int c = 0;
+  compute_encoder.set_input_array(x, c++);
+  compute_encoder.set_input_array(w, c++);
+  compute_encoder.set_input_array(scales, c++);
+  if (biases_) {
+    array biases = ensure_row_contiguous(*biases_, d, s);
+    compute_encoder.set_input_array(biases, c++);
+  }
+  compute_encoder.set_input_array(indices, c++);
+  compute_encoder.set_output_array(out, c++);
+  compute_encoder.set_bytes(M, c++);
+  compute_encoder.set_bytes(N, c++);
+  compute_encoder.set_bytes(K, c++);
 
   compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 }
@@ -794,7 +864,10 @@ void QuantizedMatmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   array x = ensure_row_contiguous_matrix(inputs[0], d, s);
   array w = ensure_row_contiguous_matrix(inputs[1], d, s);
   array scales = ensure_row_contiguous_matrix(inputs[2], d, s);
-  array biases = ensure_row_contiguous_matrix(inputs[3], d, s);
+  std::optional<array> biases = std::nullopt;
+  if (inputs.size() == 4) {
+    biases = ensure_row_contiguous_matrix(inputs[3], d, s);
+  }
 
   // Extract the matmul shapes
   bool non_batched = w.ndim() == 2 && x.flags().row_contiguous;
@@ -803,7 +876,7 @@ void QuantizedMatmul::eval_gpu(const std::vector<array>& inputs, array& out) {
   int N = out.shape(-1);
 
   int vector_limit = transpose_ ? get_qmv_batch_limit(K, N, d) : 4;
-
+  auto mode = quantization_mode_to_string(mode_);
   // It is a matrix matrix product.
   if (M >= vector_limit) {
     qmm(x,
@@ -818,30 +891,33 @@ void QuantizedMatmul::eval_gpu(const std::vector<array>& inputs, array& out) {
         N,
         K,
         d,
-        s);
+        s,
+        mode);
     return;
   }
 
   // It is a qmv with a small inner dimension so route to qmv_quad kernel
   if (transpose_ && (K == 128 || K == 64) && is_power_of_2(bits_)) {
-    qmv_quad(x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s);
+    qmv_quad(
+        x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s, mode);
     return;
   }
 
   // Run of the mill qmv
   if (transpose_) {
-    qmv(x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s);
+    qmv(x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s, mode);
     return;
   }
 
   // Run of the mill qvm
   if (K < 1024) {
-    qvm(x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s);
+    qvm(x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s, mode);
     return;
   }
 
   // Qvm with large dimension so route to a split K kernel for more parallelism
-  qvm_split_k(x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s);
+  qvm_split_k(
+      x, w, scales, biases, out, group_size_, bits_, M, N, K, d, s, mode);
   return;
 }
 
@@ -854,9 +930,12 @@ void GatherQMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   array x = ensure_row_contiguous_matrix(inputs[0], d, s);
   array w = ensure_row_contiguous_matrix(inputs[1], d, s);
   array scales = ensure_row_contiguous_matrix(inputs[2], d, s);
-  array biases = ensure_row_contiguous_matrix(inputs[3], d, s);
-  const array& lhs_indices = inputs[4];
-  const array& rhs_indices = inputs[5];
+  std::optional<array> biases = std::nullopt;
+  if (inputs.size() == 6) {
+    biases = ensure_row_contiguous_matrix(inputs[3], d, s);
+  }
+  const array& lhs_indices = inputs[inputs.size() - 2];
+  const array& rhs_indices = inputs[inputs.size() - 1];
 
   int K = x.shape(-1);
   int M = x.shape(-2);
@@ -864,6 +943,7 @@ void GatherQMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   int B = out.size() / M / N;
   int E = w.size() / w.shape(-1) / w.shape(-2);
   int vector_limit = transpose_ ? get_qmv_batch_limit(K, N, d) : 4;
+  auto mode = quantization_mode_to_string(mode_);
 
   // We are walking x in order and w is also in order so we can batch up the
   // matmuls and reuse reading x and w.
@@ -884,7 +964,8 @@ void GatherQMM::eval_gpu(const std::vector<array>& inputs, array& out) {
         N,
         K,
         d,
-        s);
+        s,
+        mode);
     return;
   }
 
@@ -905,7 +986,8 @@ void GatherQMM::eval_gpu(const std::vector<array>& inputs, array& out) {
         N,
         K,
         d,
-        s);
+        s,
+        mode);
     return;
   }
 
@@ -924,7 +1006,8 @@ void GatherQMM::eval_gpu(const std::vector<array>& inputs, array& out) {
         N,
         K,
         d,
-        s);
+        s,
+        mode);
     return;
   }
 
@@ -942,10 +1025,11 @@ void GatherQMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       N,
       K,
       d,
-      s);
+      s,
+      mode);
 }
 
-void fast::AffineQuantize::eval_gpu(
+void fast::Quantize::eval_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
   auto& w_pre = inputs[0];
@@ -974,15 +1058,27 @@ void fast::AffineQuantize::eval_gpu(
     compute_encoder.set_output_array(biases, 3);
   }
 
-  std::ostringstream kname;
   auto type_string = dequantize_ ? get_type_string(out.dtype())
                                  : get_type_string(w_pre.dtype());
-  auto kernel_func = dequantize_ ? "affine_dequantize" : "affine_quantize";
-  kname << kernel_func << "_" << type_string << "_gs_" << group_size_ << "_b_"
-        << bits_;
-  auto template_def = get_template_definition(
-      kname.str(), kernel_func, type_string, group_size_, bits_);
-  auto kernel = get_quantized_kernel(d, kname.str(), template_def);
+  std::string kname;
+  concatenate(
+      kname,
+      dequantize_ ? "affine_dequantize" : "affine_quantize",
+      "_",
+      type_string,
+      "_gs_",
+      group_size_,
+      "_b_",
+      bits_);
+  auto kernel = get_quantized_kernel_wrapped(
+      d,
+      kname,
+      dequantize_ ? "dequantize" : "quantize",
+      "affine",
+      type_string,
+      group_size_,
+      bits_);
+
   compute_encoder.set_compute_pipeline_state(kernel);
 
   // Treat uint32 as uint8 in kernel

mlx/backend/metal/slicing.cpp

@@ -2,9 +2,12 @@
 
 #include <numeric>
 
+#include "mlx/backend/common/compiled.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/gpu/slicing.h"
 #include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/kernels.h"
+#include "mlx/backend/metal/utils.h"
 
 namespace mlx::core {
 
@@ -39,4 +42,58 @@ void concatenate_gpu(
   }
 }
 
+array compute_dynamic_offset(
+    const array& indices,
+    const Strides& strides,
+    const std::vector<int>& axes,
+    const Stream& s) {
+  auto& d = metal::device(s.device);
+
+  // Kernel to compute offset here.
+  array offset({1}, int64, nullptr, {});
+  bool donate = indices.is_donatable() &&
+      (indices.data_size() * indices.itemsize()) >= offset.itemsize();
+  if (donate) {
+    offset.copy_shared_buffer(indices);
+  } else {
+    offset.set_data(allocator::malloc(offset.itemsize()));
+  }
+  d.add_temporary(offset, s.index);
+
+  auto dtype = indices.dtype();
+  std::string lib_name = "compute_dynamic_offset_" + type_to_name(dtype);
+  auto lib = d.get_library(lib_name, [dtype]() {
+    return fmt::format(
+        R"(
+        [[kernel]] void compute_dynamic_offset_{0}(
+            constant const {1}* indices [[buffer(0)]],
+            device int64_t& offset [[buffer(1)]],
+            constant const int64_t* strides [[buffer(2)]],
+            constant const int* axes [[buffer(3)]],
+            constant const int& n_axes [[buffer(4)]],
+            uint index [[thread_position_in_grid]]) {{
+          int64_t acc = 0;
+          for (int i = 0; i < n_axes; ++i) {{
+            acc += indices[i] * strides[axes[i]];
+          }}
+          offset = acc;
+        }})",
+        type_to_name(dtype),
+        get_type_string(dtype));
+  });
+  auto kernel = d.get_kernel(lib_name, lib);
+
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder.set_compute_pipeline_state(kernel);
+  compute_encoder.set_input_array(indices, 0);
+  compute_encoder.set_output_array(offset, 1);
+  compute_encoder.set_vector_bytes(strides, 2);
+  compute_encoder.set_vector_bytes(axes, 3);
+  int n_axes = axes.size();
+  compute_encoder.set_bytes(n_axes, 4);
+  MTL::Size dims = MTL::Size(1, 1, 1);
+  compute_encoder.dispatch_threads(dims, dims);
+  return offset;
+}
+
 } // namespace mlx::core

mlx/backend/no_cpu/primitives.cpp

@@ -129,7 +129,7 @@ NO_CPU(Inverse)
 NO_CPU(View)
 
 namespace fast {
-NO_CPU_MULTI(AffineQuantize)
+NO_CPU_MULTI(Quantize)
 } // namespace fast
 
 namespace distributed {

mlx/backend/no_gpu/primitives.cpp

@@ -154,7 +154,7 @@ NO_GPU_USE_FALLBACK(RMSNorm)
 NO_GPU_MULTI(RMSNormVJP)
 NO_GPU_USE_FALLBACK(RoPE)
 NO_GPU(ScaledDotProductAttention)
-NO_GPU_MULTI(AffineQuantize)
+NO_GPU_MULTI(Quantize)
 NO_GPU_MULTI(CustomKernel)
 } // namespace fast
 

mlx/distributed/CMakeLists.txt

@@ -6,3 +6,4 @@ target_sources(
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/mpi)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ring)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/nccl)

mlx/distributed/distributed.cpp

@@ -2,15 +2,21 @@
 
 #include <unordered_map>
 
+#include "mlx/backend/cuda/cuda.h"
 #include "mlx/distributed/distributed.h"
 #include "mlx/distributed/distributed_impl.h"
 #include "mlx/distributed/mpi/mpi.h"
+#include "mlx/distributed/nccl/nccl.h"
 #include "mlx/distributed/ring/ring.h"
 
 namespace mlx::core::distributed {
 
 namespace detail {
 
+Stream communication_stream(Group group, StreamOrDevice s /* = {} */) {
+  return group.raw_group()->communication_stream(s);
+}
+
 void all_sum(Group group, const array& input, array& output, Stream stream) {
   group.raw_group()->all_sum(input, output, stream);
 }
@@ -37,6 +43,10 @@ void recv(Group group, array& out, int src, Stream stream) {
 
 class EmptyGroup : public GroupImpl {
  public:
+  Stream communication_stream(StreamOrDevice s) override {
+    return to_stream(s);
+  }
+
   int rank() override {
     return 0;
   }
@@ -80,7 +90,7 @@ class EmptyGroup : public GroupImpl {
 } // namespace detail
 
 bool is_available() {
-  return mpi::is_available() || ring::is_available();
+  return mpi::is_available() || ring::is_available() || nccl::is_available();
 }
 
 int Group::rank() const {
@@ -105,15 +115,23 @@ Group init(bool strict /* = false */, const std::string& bk /* = "any" */) {
   }
 
   // Create the requested communication group
-  std::shared_ptr<detail::GroupImpl> group;
+  std::shared_ptr<detail::GroupImpl> group{nullptr};
   std::string bk_ = bk;
   if (bk == "mpi") {
     group = mpi::init(strict);
   } else if (bk == "ring") {
     group = ring::init(strict);
+  } else if (bk == "nccl") {
+    group = nccl::init(strict);
   } else if (bk == "any") {
-    group = ring::init(false);
-    bk_ = "ring";
+    if (mlx::core::cu::is_available()) {
+      group = nccl::init(false);
+      bk_ = "nccl";
+    }
+    if (group == nullptr) {
+      group = ring::init(false);
+      bk_ = "ring";
+    }
     if (group == nullptr) {
       group = mpi::init(false);
       bk_ = "mpi";

mlx/distributed/distributed.h

@@ -5,6 +5,7 @@
 #include <memory>
 
 #include "mlx/array.h"
+#include "mlx/utils.h"
 
 namespace mlx::core::distributed {
 

mlx/distributed/distributed_impl.h

@@ -13,10 +13,15 @@ class GroupImpl {
  public:
   virtual ~GroupImpl() {}
 
+  // Choose the stream this communication group can operate on
+  virtual Stream communication_stream(StreamOrDevice s = {}) = 0;
+
+  // Group operations
   virtual int rank() = 0;
   virtual int size() = 0;
   virtual std::shared_ptr<GroupImpl> split(int color, int key = -1) = 0;
 
+  // Actual communication operations
   virtual void all_sum(const array& input, array& output, Stream stream) = 0;
   virtual void all_gather(const array& input, array& output, Stream stream) = 0;
   virtual void send(const array& input, int dst, Stream stream) = 0;
@@ -25,6 +30,9 @@ class GroupImpl {
   virtual void all_min(const array& input, array& output, Stream stream) = 0;
 };
 
+/* Define the MLX stream that the communication should happen in. */
+Stream communication_stream(Group group, StreamOrDevice s = {});
+
 /* Perform an all reduce sum operation */
 void all_sum(Group group, const array& input, array& output, Stream stream);
 

mlx/distributed/mpi/mpi.cpp

@@ -349,6 +349,10 @@ class MPIGroup : public GroupImpl {
     }
   }
 
+  Stream communication_stream(StreamOrDevice s) override {
+    return to_stream(s, Device::cpu);
+  }
+
   int rank() override {
     if (rank_ < 0) {
       mpi().rank(comm_, &rank_);

mlx/distributed/nccl/CMakeLists.txt

@@ -0,0 +1,26 @@
+if(MLX_BUILD_CUDA)
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/nccl.cpp)
+  find_package(NCCL)
+  if(NCCL_FOUND)
+    target_link_libraries(mlx PRIVATE ${NCCL_LIBRARIES})
+    target_include_directories(mlx PRIVATE ${NCCL_INCLUDE_DIRS})
+  else()
+    message(
+      STATUS
+        "NCCL not found, using stubs. To run distributed with NCCL backend, install NCCL."
+    )
+
+    include(ExternalProject)
+    ExternalProject_Add(
+      nccl_stub
+      SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/nccl_stub"
+      BUILD_COMMAND ${CMAKE_COMMAND} --build .
+      INSTALL_COMMAND "")
+    set(NCCL_PATH
+        "${CMAKE_CURRENT_BINARY_DIR}/nccl_stub-prefix/src/nccl_stub-build/")
+    target_link_libraries(mlx PRIVATE ${NCCL_PATH}/libnccl.so)
+    target_include_directories(mlx PRIVATE ${NCCL_PATH})
+  endif()
+else()
+  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/no_nccl.cpp)
+endif()

mlx/distributed/nccl/nccl.cpp

@@ -0,0 +1,372 @@
+#include <arpa/inet.h>
+#include <cuda_runtime.h>
+#include <nccl.h>
+#include <netdb.h>
+#include <sys/socket.h>
+#include <unistd.h>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+#include <mutex>
+#include <stdexcept>
+#include <string>
+#include <type_traits>
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/distributed/distributed.h"
+#include "mlx/distributed/distributed_impl.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/utils.h"
+
+namespace mlx::core::distributed::nccl {
+
+#define CHECK_CUDA(cmd)              \
+  do {                               \
+    cudaError_t e = cmd;             \
+    if (e != cudaSuccess) {          \
+      fprintf(                       \
+          stderr,                    \
+          "CUDA error %s:%d '%s'\n", \
+          __FILE__,                  \
+          __LINE__,                  \
+          cudaGetErrorString(e));    \
+      exit(1);                       \
+    }                                \
+  } while (0)
+
+#define CHECK_NCCL(cmd)              \
+  do {                               \
+    ncclResult_t r = cmd;            \
+    if (r != ncclSuccess) {          \
+      fprintf(                       \
+          stderr,                    \
+          "NCCL error %s:%d '%s'\n", \
+          __FILE__,                  \
+          __LINE__,                  \
+          ncclGetErrorString(r));    \
+      exit(1);                       \
+    }                                \
+  } while (0)
+
+#define MLX_NCCL_TYPE_LIST(X) \
+  X(int8_t, ncclChar)         \
+  X(uint8_t, ncclUint8)       \
+  X(int32_t, ncclInt)         \
+  X(uint32_t, ncclUint32)     \
+  X(int64_t, ncclInt64)       \
+  X(uint64_t, ncclUint64)     \
+  X(float16_t, ncclHalf)      \
+  X(bfloat16_t, ncclBfloat16) \
+  X(float, ncclFloat)         \
+  X(double, ncclDouble)
+
+template <class>
+struct nccl_map {
+  static constexpr bool ok = false; // default: unsupported
+};
+
+#define MLX_DEF_NCCL_MAP(T, E)                 \
+  template <>                                  \
+  struct nccl_map<T> {                         \
+    static constexpr bool ok = true;           \
+    static constexpr ncclDataType_t value = E; \
+  };
+
+MLX_NCCL_TYPE_LIST(MLX_DEF_NCCL_MAP)
+#undef MLX_DEF_NCCL_MAP
+
+namespace detail {
+
+template <typename F>
+void dispatch_dtype(const array& arr, F&& f) {
+  dispatch_all_types(arr.dtype(), [&](auto type_tag) {
+    using T = MLX_GET_TYPE(type_tag);
+    if constexpr (nccl_map<T>::ok) {
+      f(type_tag, nccl_map<T>::value);
+    } else {
+      throw std::invalid_argument("[nccl] Unknown or unsupported dtype");
+    }
+  });
+}
+
+inline void sendAll(int sock, const void* buf, size_t len) {
+  const char* ptr = reinterpret_cast<const char*>(buf);
+  while (len > 0) {
+    ssize_t sent = send(sock, ptr, len, 0);
+    if (sent <= 0) {
+      perror("send");
+      exit(1);
+    }
+    ptr += sent;
+    len -= sent;
+  }
+}
+
+inline void recvAll(int sock, void* buf, size_t len) {
+  char* ptr = reinterpret_cast<char*>(buf);
+  while (len > 0) {
+    ssize_t rec = recv(sock, ptr, len, 0);
+    if (rec <= 0) {
+      perror("recv");
+      exit(1);
+    }
+    ptr += rec;
+    len -= rec;
+  }
+}
+
+inline void bootstrap_unique_id(
+    ncclUniqueId& id,
+    int rank,
+    int size,
+    const std::string& initMethod) {
+  // Parse the init method to extract the host and port
+  if (initMethod.rfind("tcp://", 0) != 0)
+    throw;
+  auto hostport = initMethod.substr(6);
+  auto colon = hostport.find(':');
+  std::string host = hostport.substr(0, colon);
+  int port = std::stoi(hostport.substr(colon + 1));
+
+  if (rank == 0) {
+    // create a unique id on the rank 0
+    CHECK_NCCL(ncclGetUniqueId(&id));
+
+    // create a socket to send the unique id to all other ranks
+    int sock = socket(AF_INET, SOCK_STREAM, 0);
+
+    if (sock < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] Couldn't create socket (error: " << errno << ")";
+      throw std::runtime_error(msg.str());
+    }
+
+    sockaddr_in serv = {};
+    serv.sin_family = AF_INET;
+    serv.sin_addr.s_addr = htonl(INADDR_ANY);
+    serv.sin_port = htons(port);
+
+    int reuse = 1;
+    // Without this, if rank-0 crashes or restarts process quickly,
+    // the OS might refuse to let binding to the same port, so reuse
+
+    if (setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, &reuse, sizeof(reuse)) < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] setsockopt() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+
+    if (bind(sock, reinterpret_cast<sockaddr*>(&serv), sizeof(serv)) < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] bind() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+    if (listen(sock, size - 1) < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] listen() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+
+    for (int peer = 1; peer < size; ++peer) {
+      int conn = accept(sock, nullptr, nullptr);
+      if (conn < 0) {
+        std::ostringstream msg;
+        msg << "[nccl] accept() failed: " << strerror(errno);
+        throw std::runtime_error(msg.str());
+      }
+      sendAll(conn, &id, sizeof(id));
+      close(conn);
+    }
+    close(sock);
+
+  } else {
+    // Here just wanted to make show that rank 0 has enough time to bind
+    // so we will retry to connect until max attempts
+
+    int sock = socket(AF_INET, SOCK_STREAM, 0);
+    if (sock < 0) {
+      std::ostringstream msg;
+      msg << "[nccl] socket() failed: " << strerror(errno);
+      throw std::runtime_error(msg.str());
+    }
+
+    hostent* he = gethostbyname(host.c_str());
+    if (!he) {
+      throw std::runtime_error("[nccl] lookup failed for host: " + host);
+    }
+    sockaddr_in serv = {};
+    serv.sin_family = AF_INET;
+    memcpy(&serv.sin_addr, he->h_addr_list[0], he->h_length);
+    serv.sin_port = htons(port);
+
+    const int max_retries = 30;
+    int attempt = 0;
+    bool connected = false;
+
+    bool do_log = std::getenv("NCCL_DEBUG") == "INFO";
+    for (attempt = 0; attempt < max_retries; ++attempt) {
+      if (connect(sock, reinterpret_cast<sockaddr*>(&serv), sizeof(serv)) ==
+          0) {
+        connected = true;
+        if (do_log) {
+          std::cout << "[Rank " << rank
+                    << "] Connected successfully on attempt " << attempt + 1
+                    << std::endl;
+          break;
+        }
+      }
+      if (errno != ECONNREFUSED) {
+        break;
+      }
+      std::this_thread::sleep_for(std::chrono::milliseconds(500));
+    }
+
+    if (!connected) {
+      std::ostringstream msg;
+      msg << "[Rank " << rank << "] connect() failed after " << attempt
+          << " retries: " << strerror(errno);
+      close(sock);
+      throw std::runtime_error(msg.str());
+    }
+    recvAll(sock, &id, sizeof(id));
+    close(sock);
+  }
+}
+
+} // namespace detail
+
+using GroupImpl = mlx::core::distributed::detail::GroupImpl;
+class NCCLGroup : public GroupImpl {
+ public:
+  NCCLGroup(int worldRank, int worldSize, const std::string initMethod)
+      : rank_(worldRank),
+        size_(worldSize),
+        comm_(nullptr),
+        initMethod_(initMethod) {
+    if (initialized_)
+      return;
+    int ndev;
+    CHECK_CUDA(cudaGetDeviceCount(&ndev));
+    CHECK_CUDA(cudaSetDevice(rank_ % ndev));
+    detail::bootstrap_unique_id(uniqueId_, rank_, size_, initMethod_);
+    CHECK_NCCL(ncclCommInitRank(&comm_, size_, uniqueId_, rank_));
+    initialized_ = true;
+  }
+
+  ~NCCLGroup() {
+    ncclCommDestroy(comm_);
+    ncclGroupEnd();
+    initialized_ = false;
+  }
+
+  Stream communication_stream(StreamOrDevice s) override {
+    return to_stream(s, Device::gpu);
+  }
+
+  int rank() override {
+    return rank_;
+  }
+
+  int size() override {
+    return size_;
+  }
+
+  void all_sum(const array& input, array& output, Stream stream) override {
+    detail::dispatch_dtype(input, [&](auto type_tag, ncclDataType_t dt) {
+      using T = typename decltype(type_tag)::type;
+      all_reduce_impl<T>(input, output, stream, dt, ncclSum);
+    });
+  }
+
+  virtual std::shared_ptr<GroupImpl> split(int color, int key = -1) override {
+    throw std::runtime_error("[nccl] Group split not supported.");
+  }
+
+  void all_gather(const array& input, array& output, Stream stream) override {
+    throw std::runtime_error(
+        "[nccl] All gather not supported in NCCL backend.");
+  }
+
+  void send(const array& input, int dst, Stream stream) override {
+    throw std::runtime_error("[nccl] Send not supported in NCCL backend.");
+  }
+
+  void recv(array& output, int src, Stream stream) override {
+    throw std::runtime_error("[nccl] Recv not supported in NCCL backend.");
+  }
+
+  void all_max(const array& input, array& output, Stream stream) override {
+    throw std::runtime_error("[nccl] All max not supported in NCCL backend.");
+  }
+
+  void all_min(const array& input, array& output, Stream stream) override {
+    throw std::runtime_error("[nccl] All min not supported in NCCL backend.");
+  }
+
+  template <typename T>
+  void all_reduce_impl(
+      const array& input,
+      array& output,
+      Stream stream,
+      ncclDataType_t dt,
+      ncclRedOp_t op) {
+    auto& encoder = cu::get_command_encoder(stream);
+
+    CHECK_NCCL(ncclAllReduce(
+        input.data<T>(),
+        output.data<T>(),
+        input.size(),
+        dt,
+        op,
+        comm_,
+        encoder.stream()));
+  }
+
+  int rank_, size_;
+  std::string initMethod_;
+  ncclUniqueId uniqueId_;
+  ncclComm_t comm_;
+  bool initialized_ = false;
+};
+
+bool is_available() {
+  return true;
+}
+
+namespace detail {
+std::string get_env_var_or_throw(const char* env_var_name, bool strict) {
+  const char* value = std::getenv(env_var_name);
+  if (value == nullptr && strict) {
+    std::ostringstream msg;
+    msg << "[nccl] Required environment variable '" << env_var_name
+        << "' is not set. "
+        << "Please set it before initializing the distributed backend.";
+    throw std::runtime_error(msg.str());
+  }
+  if (value == nullptr) {
+    return "";
+  }
+  return std::string(value);
+}
+} // namespace detail
+
+std::shared_ptr<GroupImpl> init(bool strict /* = false */) {
+  std::string host = detail::get_env_var_or_throw("NCCL_HOST_IP", strict);
+  std::string port = detail::get_env_var_or_throw("NCCL_PORT", strict);
+  std::string rank_str = detail::get_env_var_or_throw("MLX_RANK", strict);
+  std::string n_nodes_str =
+      detail::get_env_var_or_throw("MLX_WORLD_SIZE", strict);
+  if (!strict &&
+      (host.empty() || port.empty() || rank_str.empty() ||
+       n_nodes_str.empty())) {
+    return nullptr;
+  }
+
+  int rank = std::stoi(rank_str);
+  int n_nodes = std::stoi(n_nodes_str);
+  std::string init_method = "tcp://" + host + ":" + port;
+
+  return std::make_shared<NCCLGroup>(rank, n_nodes, init_method);
+}
+} // namespace mlx::core::distributed::nccl

mlx/distributed/nccl/nccl.h

@@ -0,0 +1,12 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/distributed/distributed.h"
+
+namespace mlx::core::distributed::nccl {
+
+using GroupImpl = mlx::core::distributed::detail::GroupImpl;
+
+bool is_available();
+std::shared_ptr<GroupImpl> init(bool strict = false);
+
+} // namespace mlx::core::distributed::nccl

mlx/distributed/nccl/nccl_stub/CMakeLists.txt

@@ -0,0 +1,13 @@
+cmake_minimum_required(VERSION 3.25)
+
+project(nccl LANGUAGES C CXX)
+
+file(
+  DOWNLOAD
+  "https://raw.githubusercontent.com/NVIDIA/nccl/refs/tags/v2.27.5-1/src/nccl.h.in"
+  "${CMAKE_CURRENT_BINARY_DIR}/nccl.h")
+
+add_library(nccl SHARED nccl_stubs.cpp)
+find_package(CUDAToolkit REQUIRED)
+target_include_directories(nccl PRIVATE ${CUDAToolkit_INCLUDE_DIRS})
+target_include_directories(nccl PRIVATE ${CMAKE_CURRENT_BINARY_DIR})

mlx/distributed/nccl/nccl_stub/nccl_stubs.cpp

@@ -0,0 +1,7 @@
+#include <arpa/inet.h>
+#include <cuda_runtime.h>
+#include <nccl.h>
+
+ncclResult_t ncclGetUniqueId(ncclUniqueId*) {
+  return ncclSuccess;
+}

mlx/distributed/nccl/no_nccl.cpp

@@ -0,0 +1,20 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/distributed/nccl/nccl.h"
+
+namespace mlx::core::distributed::nccl {
+
+using GroupImpl = mlx::core::distributed::detail::GroupImpl;
+
+bool is_available() {
+  return false;
+}
+
+std::shared_ptr<GroupImpl> init(bool strict /* = false */) {
+  if (strict) {
+    throw std::runtime_error("Cannot initialize nccl distributed backend.");
+  }
+  return nullptr;
+}
+
+} // namespace mlx::core::distributed::nccl

mlx/distributed/ops.cpp

@@ -2,6 +2,9 @@
 
 #include <sstream>
 
+#include "mlx/backend/cuda/cuda.h"
+#include "mlx/backend/metal/metal.h"
+#include "mlx/distributed/distributed_impl.h"
 #include "mlx/distributed/ops.h"
 #include "mlx/distributed/primitives.h"
 
@@ -28,11 +31,12 @@ array all_sum(
   if (group.size() == 1) {
     return x;
   }
+  auto stream = detail::communication_stream(group, s);
+
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(
-          to_stream(s, Device::cpu), group, AllReduce::Sum),
+      std::make_shared<AllReduce>(stream, group, AllReduce::Sum),
       {x});
 }
 
@@ -45,11 +49,12 @@ array all_max(
   if (group.size() == 1) {
     return x;
   }
+  auto stream = detail::communication_stream(group, s);
+
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(
-          to_stream(s, Device::cpu), group, AllReduce::Max),
+      std::make_shared<AllReduce>(stream, group, AllReduce::Max),
       {x});
 }
 
@@ -62,11 +67,12 @@ array all_min(
   if (group.size() == 1) {
     return x;
   }
+  auto stream = detail::communication_stream(group, s);
+
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(
-          to_stream(s, Device::cpu), group, AllReduce::Min),
+      std::make_shared<AllReduce>(stream, group, AllReduce::Min),
       {x});
 }
 
@@ -79,6 +85,7 @@ array all_gather(
   if (group.size() == 1) {
     return x;
   }
+  auto stream = detail::communication_stream(group, s);
 
   auto result_shape = x.shape();
   if (result_shape.size() == 0) {
@@ -89,7 +96,7 @@ array all_gather(
   return array(
       std::move(result_shape),
       x.dtype(),
-      std::make_shared<AllGather>(to_stream(s, Device::cpu), group),
+      std::make_shared<AllGather>(stream, group),
       {x});
 }
 
@@ -103,6 +110,7 @@ array send(
   if (group.size() == 1) {
     throw std::invalid_argument("Cannot send to a singleton group");
   }
+  auto stream = detail::communication_stream(group, s);
 
   if (dst < 0 || dst >= group.size()) {
     std::ostringstream msg;
@@ -112,10 +120,7 @@ array send(
   }
 
   return array(
-      x.shape(),
-      x.dtype(),
-      std::make_shared<Send>(to_stream(s, Device::cpu), group, dst),
-      {x});
+      x.shape(), x.dtype(), std::make_shared<Send>(stream, group, dst), {x});
 }
 
 array recv(
@@ -129,6 +134,7 @@ array recv(
   if (group.size() == 1) {
     throw std::invalid_argument("Cannot recv from a singleton group");
   }
+  auto stream = detail::communication_stream(group, s);
 
   if (src < 0 || src >= group.size()) {
     std::ostringstream msg;
@@ -139,7 +145,7 @@ array recv(
   return array(
       std::move(shape),
       std::move(dtype),
-      std::make_shared<Recv>(to_stream(s, Device::cpu), group, src),
+      std::make_shared<Recv>(stream, group, src),
       std::vector<array>{});
 }
 

mlx/distributed/ring/ring.cpp

@@ -619,6 +619,10 @@ class RingGroup : public GroupImpl {
     }
   }
 
+  Stream communication_stream(StreamOrDevice s) override {
+    return to_stream(s, Device::cpu);
+  }
+
   int rank() override {
     return rank_;
   }

mlx/export.cpp

@@ -335,7 +335,7 @@ struct PrimitiveFactory {
       SERIALIZE_PRIMITIVE(Cholesky),
       SERIALIZE_PRIMITIVE(Eig),
       SERIALIZE_PRIMITIVE(Eigh),
-      SERIALIZE_PRIMITIVE(AffineQuantize),
+      SERIALIZE_PRIMITIVE(Quantize),
       SERIALIZE_PRIMITIVE(RMSNorm),
       SERIALIZE_PRIMITIVE(RMSNormVJP),
       SERIALIZE_PRIMITIVE(LayerNorm),

mlx/fast.cpp

@@ -762,255 +762,14 @@ bool ScaledDotProductAttention::is_equivalent(const Primitive& other) const {
   return scale_ == a_other.scale_ && do_causal_ == a_other.do_causal_;
 }
 
-array pack_and_quantize(
-    array& packed_w,
-    const array& scales,
-    const array& biases,
-    int bits,
-    const Stream& s) {
-  int el_per_int = 32 / bits;
-  array zero(0, packed_w.dtype());
-  array n_bins((1 << bits) - 1, packed_w.dtype()); // 2**bits - 1
-  packed_w = astype(
-      clip(
-          round(divide(subtract(packed_w, biases, s), scales, s), s),
-          zero,
-          n_bins,
-          s),
-      uint32,
-      s);
-  if (is_power_of_2(bits)) {
-    array shifts = power(array(2, uint32), arange(0, 32, bits, uint32, s), s);
-    packed_w = reshape(packed_w, {packed_w.shape(0), -1, el_per_int}, s);
-    packed_w = sum(
-        multiply(packed_w, shifts, s), /* axis= */ 2, /* keepdims= */ false, s);
-  } else {
-    // This is slow but we have fast GPU/CPU versions of this function so we
-    // shouldn't be here often.
-    packed_w = expand_dims(packed_w, /* axis= */ -1, s);
-    packed_w = bitwise_and(
-        right_shift(packed_w, arange(bits, uint32, s), s),
-        array({1}, uint32),
-        s);
-    auto new_shape = packed_w.shape();
-    new_shape[new_shape.size() - 2] = -1;
-    new_shape.back() = 32;
-    packed_w = reshape(packed_w, new_shape, s);
-    array shifts = arange(32, uint32, s);
-    packed_w =
-        sum(left_shift(packed_w, shifts, s),
-            /* axis= */ -1,
-            /* keepdims= */ false,
-            s);
-  }
-  return packed_w;
-}
-
-std::tuple<array, array, array>
-affine_quantize(const array& w, int group_size, int bits, StreamOrDevice s_) {
-  auto s = to_stream(s_);
-
-  if (group_size != 32 && group_size != 64 && group_size != 128) {
-    std::ostringstream msg;
-    msg << "[quantize] The requested group size " << group_size
-        << " is not supported. The supported group sizes are 32, 64, and 128.";
-    throw std::invalid_argument(msg.str());
-  }
-
-  if (bits < 2 || bits > 8 || bits == 7) {
-    std::ostringstream msg;
-    msg << "[quantize] The requested number of bits " << bits
-        << " is not supported. The supported bits are 2, 3, 4, 5, 6 and 8.";
-    throw std::invalid_argument(msg.str());
-  }
-
-  if (w.ndim() < 2) {
-    std::ostringstream msg;
-    msg << "[quantize] The matrix to be quantized must have at least 2 dimension "
-        << "but it has only " << w.ndim() << ".";
-    throw std::invalid_argument(msg.str());
-  }
-
-  if ((w.shape(-1) % group_size) != 0) {
-    std::ostringstream msg;
-    msg << "[quantize] The last dimension of the matrix needs to be divisible by "
-        << "the quantization group size " << group_size
-        << ". However the provided " << " matrix has shape " << w.shape();
-    throw std::invalid_argument(msg.str());
-  }
-
-  auto fallback = [group_size, bits, s](
-                      const std::vector<array>& inputs) -> std::vector<array> {
-    auto& w = inputs[0];
-    auto wshape = w.shape();
-    wshape.back() = -1;
-
-    array zero(0, float32);
-    array n_bins((1 << bits) - 1, float32); // 2**bits - 1
-    array eps(1e-7, float32);
-
-    array packed_w = reshape(w, {-1, w.shape(-1) / group_size, group_size}, s);
-
-    array w_max = max(packed_w, /* axis= */ -1, /* keepdims= */ true, s);
-    array w_min = min(packed_w, /* axis= */ -1, /* keepdims= */ true, s);
-    w_max = astype(w_max, float32, s);
-    w_min = astype(w_min, float32, s);
-
-    array mask = greater(abs(w_min, s), abs(w_max, s), s);
-    array scales =
-        maximum(divide(subtract(w_max, w_min, s), n_bins, s), eps, s);
-    scales = where(mask, scales, negative(scales, s), s);
-    array edge = where(mask, w_min, w_max, s);
-    array q0 = round(divide(edge, scales, s), s);
-    scales = where(not_equal(q0, zero, s), divide(edge, q0, s), scales);
-    array biases = where(equal(q0, zero, s), zero, edge, s);
-
-    packed_w = pack_and_quantize(packed_w, scales, biases, bits, s);
-
-    scales = astype(scales, w.dtype(), s);
-    biases = astype(biases, w.dtype(), s);
-    return {
-        reshape(packed_w, wshape, s),
-        reshape(scales, wshape, s),
-        reshape(biases, wshape, s),
-    };
-  };
-
-  auto wq_shape = w.shape();
-  wq_shape.back() = w.shape(-1) * bits / 32;
-  auto sshape = w.shape();
-  sshape.back() = w.shape(-1) / group_size;
-  auto outputs = array::make_arrays(
-      {std::move(wq_shape), sshape, sshape},
-      {uint32, w.dtype(), w.dtype()},
-      std::make_shared<AffineQuantize>(s, fallback, group_size, bits, false),
-      {w});
-  return {outputs[0], outputs[1], outputs[2]};
-}
-
-array affine_dequantize(
-    const array& w,
-    const array& scales,
-    const array& biases,
-    int group_size,
-    int bits,
-    StreamOrDevice s_) {
-  if (bits <= 0) {
-    std::ostringstream msg;
-    msg << "[dequantize] Invalid value for bits: " << bits;
-    throw std::invalid_argument(msg.str());
-  }
-  if (group_size <= 0) {
-    std::ostringstream msg;
-    msg << "[dequantize] Invalid value for group_size: " << group_size;
-    throw std::invalid_argument(msg.str());
-  }
-  if (w.ndim() < 2 || scales.ndim() < 2 || biases.ndim() < 2) {
-    std::ostringstream msg;
-    msg << "[quantize] The matrix to be quantized must have at least 2 dimension "
-        << "but it has only " << w.ndim() << ".";
-    throw std::invalid_argument(msg.str());
-  }
-
-  auto wshape = w.shape();
-  auto sshape = scales.shape();
-  auto bshape = biases.shape();
-  wshape.back() = -1;
-  sshape.back() = -1;
-  bshape.back() = -1;
-
-  if (wshape != sshape || wshape != bshape) {
-    throw std::invalid_argument(
-        "[dequantize] Shape of scales and biases does not match the matrix");
-  }
-
-  if (w.dtype() != uint32) {
-    throw std::invalid_argument(
-        "[dequantize] The matrix should be given as a uint32");
-  }
-
-  // Packing into uint32
-  int out_size = w.shape(-1) * 32 / bits;
-
-  if (out_size != scales.shape(-1) * group_size) {
-    std::ostringstream msg;
-    msg << "[dequantize] Shape of scales and biases does not match the matrix "
-        << "given the quantization parameters. Provided matrix of shape "
-        << w.shape() << " and scales/biases of shape " << scales.shape()
-        << " with group_size=" << group_size << " and bits=" << bits << ".";
-    throw std::invalid_argument(msg.str());
-  }
-
-  auto s = to_stream(s_);
-
-  auto fallback =
-      [wshape = std::move(wshape),
-       sshape = std::move(sshape),
-       group_size,
-       bits,
-       s](const std::vector<array>& inputs) mutable -> std::vector<array> {
-    auto w = inputs[0];
-    auto& scales = inputs[1];
-    auto& biases = inputs[2];
-    if (is_power_of_2(bits)) {
-      std::vector<array> parts;
-      for (int start = 0; start < 32; start += bits) {
-        int shift_left = 32 - (start + bits);
-        int shift_right = shift_left + start;
-
-        parts.push_back(expand_dims(
-            right_shift(
-                left_shift(w, array(32 - (start + bits), uint32), s),
-                array(32 - bits, uint32),
-                s),
-            -1,
-            s));
-      }
-      w = concatenate(parts, -1, s);
-    } else {
-      w = expand_dims(w, /* axis= */ -1, s);
-      w = bitwise_and(
-          right_shift(w, arange(32, uint32, s), s), array({1}, uint32), s);
-      auto new_shape = w.shape();
-      new_shape[new_shape.size() - 2] = -1;
-      new_shape.back() = bits;
-      w = reshape(w, new_shape, s);
-      array shifts = arange(bits, uint32, s);
-      w = sum(
-          left_shift(w, shifts, s), /* axis= */ -1, /* keepdims= */ false, s);
-    }
-
-    // Dequantize
-    wshape.push_back(group_size);
-    w = reshape(w, wshape, s);
-    w = multiply(w, expand_dims(scales, -1, s), s);
-    w = add(w, expand_dims(biases, -1, s), s);
-    w = reshape(w, sshape, s);
-
-    return {w};
-  };
-
-  if (s.device == Device::gpu) {
-    auto out_shape = w.shape();
-    out_shape.back() = out_size;
-    return array(
-        std::move(out_shape),
-        scales.dtype(),
-        std::make_shared<AffineQuantize>(s, fallback, group_size, bits, true),
-        {w, scales, biases});
-  }
-  return fallback({w, scales, biases})[0];
-}
-
-bool AffineQuantize::is_equivalent(const Primitive& other) const {
-  const AffineQuantize& p_other = static_cast<const AffineQuantize&>(other);
+bool Quantize::is_equivalent(const Primitive& other) const {
+  const Quantize& p_other = static_cast<const Quantize&>(other);
   return (
       p_other.group_size_ == group_size_ && p_other.bits_ == bits_ &&
-      p_other.dequantize_ == dequantize_);
+      p_other.mode_ == mode_ && p_other.dequantize_ == dequantize_);
 }
 
-std::vector<Shape> AffineQuantize::output_shapes(
-    const std::vector<array>& inputs) {
+std::vector<Shape> Quantize::output_shapes(const std::vector<array>& inputs) {
   auto& w = inputs[0];
   if (dequantize_) {
     auto out_size = w.shape(-1) * 32 / bits_;
@@ -1022,8 +781,12 @@ std::vector<Shape> AffineQuantize::output_shapes(
     wq_shape.back() = w.shape(-1) * bits_ / 32;
     auto sshape = w.shape();
     sshape.back() = w.shape(-1) / group_size_;
-    auto bshape = sshape;
-    return {std::move(wq_shape), std::move(sshape), std::move(bshape)};
+    if (inputs.size() == 2) {
+      return {std::move(wq_shape), std::move(sshape)};
+    } else {
+      auto bshape = sshape;
+      return {std::move(wq_shape), std::move(sshape), std::move(bshape)};
+    }
   }
 }
 

mlx/fast.h

@@ -52,23 +52,10 @@ array scaled_dot_product_attention(
     const std::vector<array>& mask_arrs = {},
     StreamOrDevice s = {});
 
-std::tuple<array, array, array> affine_quantize(
-    const array& w,
-    int group_size = 64,
-    int bits = 4,
-    StreamOrDevice s = {});
+using TemplateArg = std::variant<int, bool, Dtype>;
+using ScalarArg = std::variant<bool, int, float>;
 
-array affine_dequantize(
-    const array& w,
-    const array& scales,
-    const array& biases,
-    int group_size = 64,
-    int bits = 4,
-    StreamOrDevice s = {});
-
-typedef std::variant<int, bool, Dtype> TemplateArg;
-
-typedef std::function<std::vector<array>(
+using CustomKernelFunction = std::function<std::vector<array>(
     const std::vector<array>&,
     const std::vector<Shape>&,
     const std::vector<Dtype>&,
@@ -77,10 +64,9 @@ typedef std::function<std::vector<array>(
     std::vector<std::pair<std::string, TemplateArg>>,
     std::optional<float>,
     bool,
-    StreamOrDevice)>
-    MetalKernelFunction;
+    StreamOrDevice)>;
 
-MetalKernelFunction metal_kernel(
+CustomKernelFunction metal_kernel(
     const std::string& name,
     const std::vector<std::string>& input_names,
     const std::vector<std::string>& output_names,
@@ -89,4 +75,27 @@ MetalKernelFunction metal_kernel(
     bool ensure_row_contiguous = true,
     bool atomic_outputs = false);
 
+CustomKernelFunction cuda_kernel(
+    const std::string& name,
+    const std::vector<std::string>& input_names,
+    const std::vector<std::string>& output_names,
+    const std::string& source,
+    const std::string& header = "",
+    bool ensure_row_contiguous = true,
+    int shared_memory = 0);
+
+std::vector<array> precompiled_cuda_kernel(
+    const std::string& name,
+    const std::string& compiled_source,
+    const std::vector<array>& inputs,
+    const std::vector<Shape>& output_shapes,
+    const std::vector<Dtype>& output_dtypes,
+    const std::vector<ScalarArg>& scalars,
+    std::tuple<int, int, int> grid,
+    std::tuple<int, int, int> threadgroup,
+    int shared_memory = 0,
+    std::optional<float> init_value = std::nullopt,
+    bool ensure_row_contiguous = false,
+    StreamOrDevice s = {});
+
 } // namespace mlx::core::fast

mlx/fast_primitives.h

@@ -1,6 +1,7 @@
 // Copyright © 2024 Apple Inc.
 
 #include <optional>
+#include <variant>
 
 #include "mlx/primitives.h"
 
@@ -244,17 +245,19 @@ class ScaledDotProductAttention : public Custom {
   bool do_causal_;
 };
 
-class AffineQuantize : public Custom {
+class Quantize : public Custom {
  public:
-  explicit AffineQuantize(
+  explicit Quantize(
       Stream stream,
       std::function<std::vector<array>(std::vector<array>)> fallback,
       int group_size,
       int bits,
+      QuantizationMode mode,
       bool dequantize)
       : Custom(stream, fallback),
         group_size_(group_size),
         bits_(bits),
+        mode_(mode),
         dequantize_(dequantize) {}
 
   void eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs)
@@ -263,17 +266,18 @@ class AffineQuantize : public Custom {
   void eval_gpu(const std::vector<array>& inputs, std::vector<array>& outputs)
       override;
 
-  DEFINE_NAME(AffineQuantize);
+  DEFINE_NAME(Quantize);
 
   bool is_equivalent(const Primitive& other) const override;
   std::vector<Shape> output_shapes(const std::vector<array>& inputs) override;
   auto state() const {
-    return std::make_tuple(nullptr, group_size_, bits_, dequantize_);
+    return std::make_tuple(nullptr, group_size_, bits_, mode_, dequantize_);
   }
 
  private:
   int group_size_;
   int bits_;
+  QuantizationMode mode_;
   bool dequantize_;
 };
 
@@ -283,6 +287,8 @@ struct CustomKernelShapeInfo {
   bool ndim = false;
 };
 
+using ScalarArg = std::variant<bool, int, float>;
+
 class CustomKernel : public Primitive {
  public:
   CustomKernel(
@@ -293,7 +299,10 @@ class CustomKernel : public Primitive {
       std::tuple<int, int, int> threadgroup,
       std::vector<CustomKernelShapeInfo> shape_infos,
       bool ensure_row_contiguous,
-      std::optional<float> init_value)
+      std::optional<float> init_value,
+      std::vector<ScalarArg> scalar_arguments,
+      bool is_precompiled,
+      int shared_memory)
       : Primitive(stream),
         source_(std::move(source)),
         name_(std::move(name)),
@@ -301,11 +310,14 @@ class CustomKernel : public Primitive {
         threadgroup_(threadgroup),
         shape_infos_(std::move(shape_infos)),
         ensure_row_contiguous_(ensure_row_contiguous),
-        init_value_(init_value) {}
+        init_value_(init_value),
+        scalar_arguments_(std::move(scalar_arguments)),
+        is_precompiled_(is_precompiled),
+        shared_memory_(shared_memory) {}
 
   void eval_cpu(const std::vector<array>& inputs, std::vector<array>& outputs)
       override {
-    throw std::runtime_error("Custom Metal kernels only run on GPU.");
+    throw std::runtime_error("Custom kernels only run on GPU.");
   }
 
   void eval_gpu(const std::vector<array>& inputs, std::vector<array>& outputs)
@@ -321,6 +333,9 @@ class CustomKernel : public Primitive {
   std::vector<CustomKernelShapeInfo> shape_infos_;
   bool ensure_row_contiguous_;
   std::optional<float> init_value_;
+  std::vector<ScalarArg> scalar_arguments_;
+  bool is_precompiled_;
+  int shared_memory_;
 };
 
 } // namespace mlx::core::fast

mlx/io/CMakeLists.txt

@@ -11,7 +11,7 @@ if(MLX_BUILD_GGUF)
   FetchContent_Declare(
     gguflib
     GIT_REPOSITORY https://github.com/antirez/gguf-tools/
-    GIT_TAG af7d88d808a7608a33723fba067036202910acb3)
+    GIT_TAG 8fa6eb65236618e28fd7710a0fba565f7faa1848)
   FetchContent_MakeAvailable(gguflib)
   target_include_directories(mlx
                              PRIVATE $<BUILD_INTERFACE:${gguflib_SOURCE_DIR}>)