Make sure softmax doesn't change the actual max

* perf tuning

* fix adding inputs arrays in matmul / srot

* format

* fix

.circleci/config.yml

@@ -16,6 +16,9 @@ parameters:
   linux_release:
     type: boolean
     default: false
+  cuda_release:
+    type: boolean
+    default: false
 
 jobs:
   build_documentation:
@@ -104,7 +107,7 @@ jobs:
           command: |
             echo "stubs"
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
@@ -162,7 +165,7 @@ jobs:
           command: |
             source env/bin/activate
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Run Python tests
           command: |
@@ -223,7 +226,6 @@ jobs:
           command: |
             sudo apt-get update
             sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
             python -m venv env
             source env/bin/activate
             CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
@@ -283,7 +285,7 @@ jobs:
           command: |
             source env/bin/activate
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
       - run:
           name: Build Python package
           command: |
@@ -342,7 +344,7 @@ jobs:
               CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               pip install . -v
             pip install typing_extensions
-            python setup.py generate_stubs 
+            python setup.py generate_stubs
             << parameters.extra_env >> \
               CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
               python -m build --wheel
@@ -356,6 +358,48 @@ jobs:
       - store_artifacts:
           path: wheelhouse/
 
+  build_cuda_release:
+    parameters:
+      python_version:
+        type: string
+        default: "3.9"
+      extra_env:
+        type: string
+        default: "DEV_RELEASE=1"
+    machine:
+      image: linux-cuda-12:default
+      resource_class: gpu.nvidia.small.gen2
+    steps:
+      - checkout
+      - run:
+          name: Build wheel
+          command: |
+            sudo apt-get update
+            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
+            python -m venv env
+            source env/bin/activate
+            pip install auditwheel
+            pip install patchelf
+            pip install build
+            pip install twine
+            << parameters.extra_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              pip install ".[dev]" -v
+            python setup.py generate_stubs
+            << parameters.extra_env >> \
+              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
+              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
+              python -m build --wheel
+            bash python/scripts/repair_cuda.sh
+      - run:
+          name: Upload package
+          command: |
+            source env/bin/activate
+            twine upload wheelhouse/*.whl
+      - store_artifacts:
+          path: wheelhouse/
+
 workflows:
   build_and_test:
     when:
@@ -625,3 +669,14 @@ workflows:
             parameters:
               python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
               extra_env: ["PYPI_RELEASE=1"]
+  cuda_test_release:
+    when:
+      and:
+        - equal: [ main, << pipeline.git.branch >> ]
+        - << pipeline.parameters.cuda_release >>
+    jobs:
+      - build_cuda_release:
+          matrix:
+            parameters:
+              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+              extra_env: ["PYPI_RELEASE=1"]

benchmarks/python/comparative/bench_torch.py

@@ -5,6 +5,7 @@ import os
 import time
 
 import torch
+import torch.cuda
 import torch.mps
 
 
@@ -44,8 +45,10 @@ def bench(f, *args):
 
 
 def sync_if_needed(x):
-    if x.device != torch.device("cpu"):
+    if x.device == torch.device("mps"):
         torch.mps.synchronize()
+    elif x.device == torch.device("cuda"):
+        torch.cuda.synchronize()
 
 
 @torch.no_grad()
@@ -99,6 +102,14 @@ def reduction(op, axis, x):
     sync_if_needed(x)
 
 
+@torch.no_grad()
+def sum_and_add(axis, x, y):
+    z = x.sum(axis=axis, keepdims=True)
+    for i in range(50):
+        z = (z + y).sum(axis=axis, keepdims=True)
+    sync_if_needed(x)
+
+
 @torch.no_grad()
 def softmax(axis, x):
     ys = []
@@ -340,7 +351,11 @@ if __name__ == "__main__":
         args.axis.pop(0)
 
     torch.set_num_threads(1)
-    device = "cpu" if args.cpu else "mps"
+    device = "mps"
+    if torch.cuda.is_available():
+        device = "cuda"
+    if args.cpu:
+        device = "cpu"
 
     types = args.dtype
     if not types:
@@ -460,5 +475,8 @@ if __name__ == "__main__":
     elif args.benchmark == "selu":
         print(bench(selu, x))
 
+    elif args.benchmark == "sum_and_add":
+        print(bench(sum_and_add, axis, *xs))
+
     else:
         raise ValueError(f"Unknown benchmark `{args.benchmark}`.")

docs/src/install.rst

@@ -30,6 +30,16 @@ MLX is also available on conda-forge. To install MLX with conda do:
 
    conda install conda-forge::mlx
 
+CUDA
+^^^^
+
+MLX has a CUDA backend which you can use on any Linux platform with CUDA 12
+and SM 7.0 (Volta) and up. To install MLX with CUDA support, run:
+
+.. code-block:: shell
+
+    pip install mlx-cuda
+
 
 Troubleshooting
 ^^^^^^^^^^^^^^^
@@ -65,6 +75,8 @@ Build Requirements
 Python API
 ^^^^^^^^^^
 
+.. _python install:
+
 To build and install the MLX python library from source, first, clone MLX from
 `its GitHub repo <https://github.com/ml-explore/mlx>`_:
 
@@ -107,6 +119,8 @@ IDE:
 C++ API
 ^^^^^^^
 
+.. _cpp install:
+
 Currently, MLX must be built and installed from source.
 
 Similarly to the python library, to build and install the MLX C++ library start
@@ -185,6 +199,7 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
+
 Binary Size Minimization
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -213,6 +228,50 @@ be anwywhere from a few hundred millisecond to a few seconds depending on the
 application. Once a kernel is compiled, it will be cached by the system. The
 Metal kernel cache persists across reboots.
 
+Linux
+^^^^^
+
+To build from source on Linux (CPU only), install the BLAS and LAPACK headers.
+For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   apt-get update -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+
+From here follow the instructions to install either the :ref:`Python <python
+install>` or :ref:`C++ <cpp install>` APIs.
+
+CUDA
+^^^^
+
+To build from source on Linux with CUDA, install the BLAS and LAPACK headers
+and the CUDA toolkit. For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+   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
+
+
+When building either the Python or C++ APIs make sure to pass the cmake flag
+``MLX_BUILD_CUDA=ON``. For example, to build the Python API run:
+
+.. code-block:: shell
+
+  CMAKE_BUILD_PARALLEL_LEVEL=8 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON" pip install -e ".[dev]"
+
+To build the C++ package run:
+
+.. code-block:: shell
+
+   mkdir -p build && cd build
+   cmake .. -DMLX_BUILD_CUDA=ON && make -j
+
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 

mlx/backend/common/reduce.cpp

@@ -5,11 +5,9 @@
 namespace mlx::core {
 
 std::pair<Shape, Strides> shapes_without_reduction_axes(
-    const array& x,
+    Shape shape,
+    Strides strides,
     const std::vector<int>& axes) {
-  auto shape = x.shape();
-  auto strides = x.strides();
-
   for (int i = axes.size() - 1; i >= 0; i--) {
     int a = axes[i];
     shape.erase(shape.begin() + a);
@@ -19,6 +17,15 @@ std::pair<Shape, Strides> shapes_without_reduction_axes(
   return std::make_pair(shape, strides);
 }
 
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes) {
+  auto shape = x.shape();
+  auto strides = x.strides();
+  return shapes_without_reduction_axes(
+      std::move(shape), std::move(strides), axes);
+}
+
 ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
   // The data is all there and we are reducing over everything
   if (x.size() == x.data_size() && axes.size() == x.ndim() &&

mlx/backend/common/reduce.h

@@ -51,5 +51,9 @@ ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes);
 std::pair<Shape, Strides> shapes_without_reduction_axes(
     const array& x,
     const std::vector<int>& axes);
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    Shape shape,
+    Strides strides,
+    const std::vector<int>& axes);
 
 } // namespace mlx::core

mlx/backend/cuda/CMakeLists.txt

@@ -29,9 +29,10 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/random.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/all_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/col_reduce.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/init_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
-          ${CMAKE_CURRENT_SOURCE_DIR}/reduce/segmented_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp

mlx/backend/cuda/allocator.cpp

@@ -3,6 +3,7 @@
 #include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/utils.h"
 #include "mlx/backend/cuda/worker.h"
+#include "mlx/utils.h"
 
 #include <cuda_runtime.h>
 #include <fmt/format.h>
@@ -14,9 +15,11 @@ namespace mlx::core {
 
 namespace cu {
 
+constexpr int page_size = 16384;
+
 CudaAllocator::CudaAllocator()
     : buffer_cache_(
-          getpagesize(),
+          page_size,
           [](CudaBuffer* buf) { return buf->size; },
           [this](CudaBuffer* buf) {
             cuda_free(buf->data);
@@ -31,7 +34,14 @@ CudaAllocator::CudaAllocator()
 
 Buffer CudaAllocator::malloc(size_t size) {
   // Find available buffer from cache.
+  auto orig_size = size;
   std::unique_lock lock(mutex_);
+  if (size < page_size) {
+    size = next_power_of_2(size);
+  } else {
+    size = page_size * ((size + page_size - 1) / page_size);
+  }
+
   CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
   if (!buf) {
     // If we have a lot of memory pressure or are over the maximum cache size,

mlx/backend/cuda/binary_two.cu

@@ -157,7 +157,7 @@ void binary_op_gpu_inplace(
               if (ndim <= 3) {
                 MLX_SWITCH_1_2_3(ndim, NDIM, {
                   auto kernel =
-                      &cu::binary_g_nd<Op, InType, OutType, IdxT, NDIM>;
+                      cu::binary_g_nd<Op, InType, OutType, IdxT, NDIM>;
                   auto [num_blocks, block_dims] =
                       get_launch_args(kernel, out_a, large);
                   kernel<<<num_blocks, block_dims, 0, stream>>>(

mlx/backend/cuda/copy.cu

@@ -24,7 +24,6 @@ void copy_gpu_inplace(
   auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
-
   if (ctype == CopyType::Scalar || ctype == CopyType::Vector) {
     copy_contiguous(encoder, ctype, in, out, offset_in, offset_out);
     return;

mlx/backend/cuda/device.cpp

@@ -114,7 +114,7 @@ void CommandEncoder::synchronize() {
   std::future<void> f = p->get_future();
   add_completed_handler([p = std::move(p)]() { p->set_value(); });
   worker_.end_batch();
-  worker_.commit();
+  commit();
   f.wait();
 }
 

mlx/backend/cuda/device/utils.cuh

@@ -155,8 +155,8 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_nd(
 #pragma unroll
   for (int i = NDIM - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
-    b_loc += dim_idx * b_strides[i];
+    b_loc += dim_idx * IdxT(b_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc);
@@ -175,9 +175,9 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_nd(
 #pragma unroll
   for (int i = NDIM - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
-    b_loc += dim_idx * b_strides[i];
+    b_loc += dim_idx * IdxT(b_strides[i]);
-    c_loc += dim_idx * c_strides[i];
+    c_loc += dim_idx * IdxT(c_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc, c_loc);
@@ -206,8 +206,8 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT> elem_to_loc_4d(
   IdxT b_loc = 0;
   for (int i = ndim - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
-    b_loc += dim_idx * b_strides[i];
+    b_loc += dim_idx * IdxT(b_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc);
@@ -226,9 +226,9 @@ inline __host__ __device__ cuda::std::tuple<IdxT, IdxT, IdxT> elem_to_loc_4d(
   IdxT c_loc = 0;
   for (int i = ndim - 1; i >= 0; --i) {
     int dim_idx = elem % shape[i];
-    a_loc += dim_idx * a_strides[i];
+    a_loc += dim_idx * IdxT(a_strides[i]);
-    b_loc += dim_idx * b_strides[i];
+    b_loc += dim_idx * IdxT(b_strides[i]);
-    c_loc += dim_idx * c_strides[i];
+    c_loc += dim_idx * IdxT(c_strides[i]);
     elem /= shape[i];
   }
   return cuda::std::make_tuple(a_loc, b_loc, c_loc);

mlx/backend/cuda/matmul.cpp

@@ -162,11 +162,15 @@ class MatMul {
       }
     }
 
-    array workspace(
+    void* workspace_ptr = nullptr;
-        allocator::malloc(heuristic_.workspaceSize),
+    if (heuristic_.workspaceSize > 0) {
-        {static_cast<int>(heuristic_.workspaceSize)},
+      array workspace(
-        int8);
+          allocator::malloc(heuristic_.workspaceSize),
-    encoder.add_temporary(workspace);
+          {static_cast<int>(heuristic_.workspaceSize)},
+          int8);
+      encoder.add_temporary(workspace);
+      workspace_ptr = workspace.data<void>();
+    }
 
     encoder.launch_kernel([&](cudaStream_t stream) {
       CHECK_CUBLAS_ERROR(cublasLtMatmul(
@@ -183,8 +187,8 @@ class MatMul {
           out,
           out_desc_,
           &heuristic_.algo,
-          workspace.data<void>(),
+          workspace_ptr,
-          workspace.nbytes(),
+          heuristic_.workspaceSize,
           stream));
     });
   }
@@ -358,9 +362,18 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
       a_batch_strides.back(),
       b_batch_strides.back());
 
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  auto nbatch = batch_count / batch_shape.back();
+  if (nbatch == 1) {
+    matmul.run(encoder, out.data<int8_t>(), a.data<int8_t>(), b.data<int8_t>());
+    return;
+  }
+
   ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
   ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
-  for (size_t i = 0; i < batch_count / batch_shape.back(); ++i) {
+  for (size_t i = 0; i < nbatch; ++i) {
     matmul.run(
         encoder,
         out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,
@@ -444,10 +457,28 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       b_batch_strides.back(),
       c_batch_strides.back());
 
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(c);
+  encoder.set_output_array(out);
+
+  auto nbatch = batch_count / batch_shape.back();
+  if (nbatch == 1) {
+    matmul.run(
+        encoder,
+        out.data<int8_t>(),
+        a.data<int8_t>(),
+        b.data<int8_t>(),
+        c.data<int8_t>(),
+        alpha_,
+        beta_);
+    return;
+  }
+
   ContiguousIterator a_it(batch_shape, a_batch_strides, batch_shape.size() - 1);
   ContiguousIterator b_it(batch_shape, b_batch_strides, batch_shape.size() - 1);
   ContiguousIterator c_it(batch_shape, c_batch_strides, batch_shape.size() - 1);
-  for (size_t i = 0; i < batch_count / batch_shape.back(); ++i) {
+  for (size_t i = 0; i < nbatch; ++i) {
     matmul.run(
         encoder,
         out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M * N,

mlx/backend/cuda/reduce.cu

@@ -21,28 +21,11 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(!axes_.empty());
   assert(out.size() != in.size());
 
-  out.set_data(allocator::malloc(out.nbytes()));
-
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  encoder.set_input_array(in);
-  encoder.set_output_array(out);
 
-  // Fill out with init value.
   if (in.size() == 0) {
-    encoder.launch_kernel([&](cudaStream_t stream) {
+    init_reduce(encoder, in, out, reduce_type_);
-      MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
-        MLX_SWITCH_REDUCE_OPS(reduce_type_, OP, {
-          using InType = cuda_type_t<CTYPE>;
-          using OutType = cu::ReduceResult<OP, InType>::type;
-          thrust::fill_n(
-              cu::thrust_policy(stream),
-              thrust::device_pointer_cast(out.data<OutType>()),
-              out.data_size(),
-              cu::ReduceInit<OP, InType>::value());
-        });
-      });
-    });
     return;
   }
 
@@ -59,9 +42,8 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
     plan = get_reduction_plan(in, axes_);
   }
 
-  if ((plan.type == ContiguousAllReduce) ||
+  if (plan.type == ContiguousAllReduce) {
-      (plan.type == ContiguousReduce && plan.shape.size() == 1)) {
+    all_reduce(encoder, in, out, reduce_type_);
-    segmented_reduce(encoder, in, out, reduce_type_, axes_, plan);
     return;
   }
 

mlx/backend/cuda/reduce/all_reduce.cu

@@ -0,0 +1,140 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/reduce/reduce.cuh"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <cub/block/block_load.cuh>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T, typename U, typename ReduceOp, int N = 4>
+__global__ void all_reduce(T* in, U* out, size_t block_step, size_t size) {
+  // TODO: Process multiple "rows" in each thread
+  constexpr int M = 1;
+
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  const U init = cu::ReduceInit<ReduceOp, T>::value();
+  ReduceOp op;
+
+  T vals[N];
+  U accs[M];
+  accs[0] = init;
+
+  size_t start = grid.block_rank() * block_step;
+  size_t end = start + block_step;
+  size_t check = min(end, size);
+
+  for (size_t i = start; i + block.size() * N <= check; i += block.size() * N) {
+    cub::LoadDirectBlockedVectorized<T, N>(block.thread_rank(), in + i, vals);
+    for (int j = 0; j < N; j++) {
+      accs[0] = op(accs[0], __cast<U, T>(vals[j]));
+    }
+  }
+
+  if (end > size) {
+    size_t offset = end - block.size() * N;
+    int block_end = size - offset;
+    cub::LoadDirectBlocked(
+        block.thread_rank(), in + offset, vals, block_end, __cast<T, U>(init));
+    for (int i = 0; i < N; i++) {
+      accs[0] = op(accs[0], __cast<U, T>(vals[i]));
+    }
+  }
+
+  __shared__ U shared_accumulators[32];
+  block_reduce(block, warp, accs, shared_accumulators, op, init);
+
+  if (block.thread_rank() == 0) {
+    out[grid.block_rank()] = accs[0];
+  }
+}
+
+} // namespace cu
+
+void all_reduce(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type) {
+  constexpr int N_READS = 8;
+
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto get_args = [](size_t size, int N) {
+    size_t reductions = size / N;
+    int threads = 512;
+    size_t full_blocks = (reductions + threads - 1) / threads;
+    int blocks;
+    if (full_blocks < 32) {
+      blocks = 1;
+    } else if (full_blocks < 128) {
+      blocks = 32;
+    } else if (full_blocks < 512) {
+      blocks = 128;
+    } else if (full_blocks < 1024) {
+      blocks = 512;
+    } else {
+      blocks = 1024;
+    }
+    size_t reductions_per_block = std::max(
+        static_cast<size_t>(threads), (reductions + blocks - 1) / blocks);
+    size_t block_step = reductions_per_block * N;
+
+    return std::make_tuple(blocks, threads, block_step);
+  };
+
+  int blocks, threads;
+  size_t block_step;
+  array x = in;
+
+  // Large array so allocate an intermediate and accumulate there
+  std::tie(blocks, threads, block_step) = get_args(x.size(), N_READS);
+  if (blocks > 1) {
+    std::tie(blocks, threads, block_step) = get_args(x.size(), N_READS);
+    array intermediate({blocks}, out.dtype(), nullptr, {});
+    intermediate.set_data(allocator::malloc(intermediate.nbytes()));
+    encoder.add_temporary(intermediate);
+    encoder.set_input_array(x);
+    encoder.set_output_array(intermediate);
+    encoder.launch_kernel([&](cudaStream_t stream) {
+      MLX_SWITCH_ALL_TYPES(x.dtype(), CTYPE, {
+        MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
+          using T = cuda_type_t<CTYPE>;
+          using U = cu::ReduceResult<OP, T>::type;
+          auto kernel = cu::all_reduce<T, U, OP, N_READS>;
+          kernel<<<blocks, threads, 0, stream>>>(
+              x.data<T>(), intermediate.data<U>(), block_step, x.size());
+        });
+      });
+    });
+
+    // Set the input for the next step and recalculate the blocks
+    x = intermediate;
+    std::tie(blocks, threads, block_step) = get_args(x.size(), N_READS);
+  }
+
+  encoder.set_input_array(x);
+  encoder.set_output_array(out);
+  encoder.launch_kernel([&](cudaStream_t stream) {
+    MLX_SWITCH_ALL_TYPES(x.dtype(), CTYPE, {
+      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
+        using T = cuda_type_t<CTYPE>;
+        using U = cu::ReduceResult<OP, T>::type;
+        auto kernel = cu::all_reduce<T, U, OP, N_READS>;
+        kernel<<<blocks, threads, 0, stream>>>(
+            x.data<T>(), out.data<U>(), block_step, x.size());
+      });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/reduce/col_reduce.cu

@@ -64,86 +64,6 @@ struct ColReduceArgs {
   }
 };
 
-template <typename T, typename U, typename Op, int NDIM, int N_READS = 4>
-__global__ void col_reduce_small(
-    const T* in,
-    U* out,
-    const __grid_constant__ ColReduceArgs args) {
-  auto grid = cg::this_grid();
-  auto block = cg::this_thread_block();
-
-  int column =
-      grid.block_index().x * block.dim_threads().x + block.thread_index().x;
-  if (column * N_READS >= args.reduction_stride) {
-    return;
-  }
-
-  int out_idx = grid.block_rank() / grid.dim_blocks().x;
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
-
-  Op op;
-  U totals[N_READS];
-  for (int i = 0; i < N_READS; i++) {
-    totals[i] = ReduceInit<Op, T>::value();
-  }
-
-  // Read input to local.
-  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-  loop.next(
-      block.thread_index().y,
-      args.reduce_shape.data(),
-      args.reduce_strides.data());
-  for (size_t r = block.thread_index().y;
-       r < args.non_col_reductions * args.reduction_size;
-       r += block.dim_threads().y) {
-    U vals[N_READS];
-    cub::LoadDirectBlocked(
-        column,
-        make_cast_iterator<U>(in + loop.location()),
-        vals,
-        args.reduction_stride,
-        ReduceInit<Op, T>::value());
-    for (int i = 0; i < N_READS; i++) {
-      totals[i] = op(vals[i], totals[i]);
-    }
-    loop.next(
-        block.dim_threads().y,
-        args.reduce_shape.data(),
-        args.reduce_strides.data());
-  }
-
-  // Do block reduce when each column has more than 1 element to reduce.
-  if (block.dim_threads().y > 1) {
-    __shared__ U shared_vals[32 * 8 * N_READS];
-    size_t col =
-        block.thread_index().y * block.dim_threads().x + block.thread_index().x;
-    for (int i = 0; i < N_READS; i++) {
-      shared_vals[col * N_READS + i] = totals[i];
-    }
-    block.sync();
-    if (block.thread_index().y == 0) {
-      for (int i = 0; i < N_READS; i++) {
-        totals[i] = shared_vals[block.thread_index().x * N_READS + i];
-      }
-      for (int j = 1; j < block.dim_threads().y; j++) {
-        col = j * block.dim_threads().x + block.thread_index().x;
-        for (int i = 0; i < N_READS; i++) {
-          totals[i] = op(shared_vals[col * N_READS + i], totals[i]);
-        }
-      }
-    }
-  }
-
-  // Write result.
-  if (block.thread_index().y == 0) {
-    cub::StoreDirectBlocked(
-        column,
-        out + out_idx * args.reduction_stride,
-        totals,
-        args.reduction_stride);
-  }
-}
-
 template <
     typename T,
     typename U,
@@ -152,67 +72,83 @@ template <
     int BM,
     int BN,
     int N_READS = 4>
-__global__ void col_reduce_looped(
+__global__ void
-    const T* in,
+col_reduce_looped(T* in, U* out, const __grid_constant__ ColReduceArgs args) {
-    U* out,
-    const __grid_constant__ ColReduceArgs args) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<WARP_SIZE>(block);
 
-  constexpr int n_warps = BN / N_READS;
+  constexpr int threads_per_row = BN / N_READS;
 
-  int out_idx = grid.block_rank() / grid.dim_blocks().x;
+  // Compute the indices for the tile
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
+  size_t tile_idx = grid.block_rank();
+  size_t tile_x = tile_idx % ((args.reduction_stride + BN - 1) / BN);
+  size_t tile_y = tile_idx / ((args.reduction_stride + BN - 1) / BN);
 
+  // Compute the indices for the thread within the tile
+  short thread_x = block.thread_rank() % threads_per_row;
+  short thread_y = block.thread_rank() / threads_per_row;
+
+  // Move the input pointer
+  in += elem_to_loc(tile_y, args.shape.data(), args.strides.data(), args.ndim) +
+      tile_x * BN;
+
+  // Initialize the running totals
   Op op;
   U totals[N_READS];
   for (int i = 0; i < N_READS; i++) {
     totals[i] = ReduceInit<Op, T>::value();
   }
 
-  // Read input to local.
-  int r = block.thread_rank() / n_warps;
-  int column = block.thread_rank() % n_warps;
-  int in_offset = grid.block_index().x * BN;
   LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-  loop.next(r, args.reduce_shape.data(), args.reduce_strides.data());
+  loop.next(thread_y, args.reduce_shape.data(), args.reduce_strides.data());
-  for (; r < args.non_col_reductions * args.reduction_size; r += BM) {
+  size_t total = args.non_col_reductions * args.reduction_size;
-    U vals[N_READS];
+  if (tile_x * BN + BN <= args.reduction_stride) {
-    cub::LoadDirectBlocked(
+    for (size_t r = thread_y; r < total; r += BM) {
-        column,
+      T vals[N_READS];
-        make_cast_iterator<U>(in + loop.location() + in_offset),
+      cub::LoadDirectBlockedVectorized(thread_x, in + loop.location(), vals);
-        vals,
+      for (int i = 0; i < N_READS; i++) {
-        args.reduction_stride - in_offset,
+        totals[i] = op(totals[i], __cast<U, T>(vals[i]));
-        ReduceInit<Op, T>::value());
+      }
-    for (int i = 0; i < N_READS; i++) {
+      loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
-      totals[i] = op(vals[i], totals[i]);
+    }
+  } else {
+    for (size_t r = thread_y; r < total; r += BM) {
+      T vals[N_READS];
+      cub::LoadDirectBlocked(
+          thread_x,
+          in + loop.location(),
+          vals,
+          args.reduction_stride - tile_x * BN,
+          __cast<T, U>(ReduceInit<Op, T>::value()));
+      for (int i = 0; i < N_READS; i++) {
+        totals[i] = op(totals[i], __cast<U, T>(vals[i]));
+      }
+      loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
     }
-    loop.next(BM, args.reduce_shape.data(), args.reduce_strides.data());
   }
 
   // Do warp reduce for each output.
-  constexpr int n_outputs = BN / n_warps;
+  constexpr int n_outputs = BN / threads_per_row;
   static_assert(BM == 32 && n_outputs == N_READS);
   __shared__ U shared_vals[BM * BN];
-  size_t col = block.thread_index().y * BN + block.thread_index().x * N_READS;
+  short s_idx = thread_y * BN + thread_x * N_READS;
   for (int i = 0; i < N_READS; i++) {
-    shared_vals[col + i] = totals[i];
+    shared_vals[s_idx + i] = totals[i];
   }
   block.sync();
-  col = warp.thread_rank() * BN + warp.meta_group_rank() * n_outputs;
+  s_idx = warp.thread_rank() * BN + warp.meta_group_rank() * n_outputs;
   for (int i = 0; i < n_outputs; i++) {
-    totals[i] = cg::reduce(warp, shared_vals[col + i], op);
+    totals[i] = cg::reduce(warp, shared_vals[s_idx + i], op);
   }
 
   // Write result.
   if (warp.thread_rank() == 0) {
-    size_t out_offset = grid.block_index().x * BN;
     cub::StoreDirectBlocked(
         warp.meta_group_rank(),
-        out + out_idx * args.reduction_stride + out_offset,
+        out + tile_y * args.reduction_stride + tile_x * BN,
         totals,
-        args.reduction_stride - out_offset);
+        args.reduction_stride - tile_x * BN);
   }
 }
 
@@ -230,6 +166,53 @@ inline auto output_grid_for_col_reduce(
   return get_2d_grid_dims(out_shape, out_strides);
 }
 
+void col_reduce_looped(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type,
+    const std::vector<int>& axes,
+    const ReductionPlan& plan,
+    cu::ColReduceArgs args) {
+  // Allocate data for the output using in's layout to access them as
+  // contiguously as possible.
+  allocate_same_layout(out, in, axes);
+
+  // Just a way to get out of the constness because cub doesn't like it ...
+  // (sigh)
+  array x = in;
+
+  encoder.set_input_array(x);
+  encoder.set_output_array(out);
+  encoder.launch_kernel([&](cudaStream_t stream) {
+    MLX_SWITCH_ALL_TYPES(x.dtype(), CTYPE, {
+      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
+        MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
+          using T = cuda_type_t<CTYPE>;
+          using U = cu::ReduceResult<OP, T>::type;
+
+          constexpr int N_READS = 4;
+          constexpr int BM = 32;
+          constexpr int BN = 32;
+          dim3 grid = output_grid_for_col_reduce(out, args);
+          size_t extra_blocks = cuda::ceil_div(args.reduction_stride, BN);
+          if (grid.x * extra_blocks < INT32_MAX) {
+            grid.x *= extra_blocks;
+          } else if (grid.y * extra_blocks < 65536) {
+            grid.y *= extra_blocks;
+          } else {
+            throw std::runtime_error(
+                "[col_reduce_looped] Need to factorize reduction_stride");
+          }
+          int blocks = BM * BN / N_READS;
+          auto kernel = cu::col_reduce_looped<T, U, OP, NDIM, BM, BN, N_READS>;
+          kernel<<<grid, blocks, 0, stream>>>(x.data<T>(), out.data<U>(), args);
+        });
+      });
+    });
+  });
+}
+
 void col_reduce(
     cu::CommandEncoder& encoder,
     const array& in,
@@ -237,42 +220,24 @@ void col_reduce(
     Reduce::ReduceType reduce_type,
     const std::vector<int>& axes,
     const ReductionPlan& plan) {
+  // Current col reduce options
+  //
+  // - col_reduce_looped
+  //
+  //   It is a general strided reduce. Each threadblock computes the output for
+  //   a subrow of the fast moving axis. For instance 32 elements.
+  //
+  // Notes: As in row reduce we opt to read as much in order as possible and
+  // leave
+  //        transpositions as they are (contrary to our Metal backend).
+  //
+  //        Moreover we need different kernels for short rows and tuning
+
+  // Make the args struct to help route to the best kernel
   cu::ColReduceArgs args(in, plan, axes);
 
-  encoder.launch_kernel([&](cudaStream_t stream) {
+  // Fallback col reduce
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
+  col_reduce_looped(encoder, in, out, reduce_type, axes, plan, args);
-      using InType = cuda_type_t<CTYPE>;
-      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
-        using OutType = cu::ReduceResult<OP, InType>::type;
-        MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
-          constexpr int N_READS = 4;
-          dim3 block_dims;
-          dim3 num_blocks = output_grid_for_col_reduce(out, args);
-          num_blocks.z = num_blocks.y;
-          num_blocks.y = num_blocks.x;
-          auto kernel =
-              cu::col_reduce_small<InType, OutType, OP, NDIM, N_READS>;
-          size_t total = args.non_col_reductions * args.reduction_size;
-          if (total < 32) {
-            size_t stride_blocks =
-                cuda::ceil_div(args.reduction_stride, N_READS);
-            block_dims.x = std::min(stride_blocks, 32ul);
-            block_dims.y = std::min(total, 8ul);
-            num_blocks.x = cuda::ceil_div(stride_blocks, block_dims.x);
-          } else {
-            constexpr int BM = 32;
-            constexpr int BN = 32;
-            block_dims.x = BM * BN / N_READS;
-            num_blocks.x = cuda::ceil_div(args.reduction_stride, BN);
-            kernel = cu::
-                col_reduce_looped<InType, OutType, OP, NDIM, BM, BN, N_READS>;
-          }
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in.data<InType>(), out.data<OutType>(), args);
-        });
-      });
-    });
-  });
 }
 
 } // namespace mlx::core

mlx/backend/cuda/reduce/init_reduce.cu

@@ -0,0 +1,51 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/reduce/reduce.cuh"
+
+#include <cooperative_groups.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename T, typename U, typename Op>
+__global__ void init_reduce(U* out, size_t size) {
+  auto index = cg::this_grid().thread_rank();
+  if (index < size) {
+    out[index] = ReduceInit<Op, T>::value();
+  }
+}
+
+} // namespace cu
+
+void init_reduce(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type) {
+  // Allocate if needed
+  if (out.data_shared_ptr() == nullptr) {
+    out.set_data(allocator::malloc(out.nbytes()));
+  }
+
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  encoder.launch_kernel([&](cudaStream_t stream) {
+    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
+      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
+        using T = cuda_type_t<CTYPE>;
+        using U = cu::ReduceResult<OP, T>::type;
+        auto kernel = cu::init_reduce<T, U, OP>;
+        dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
+        dim3 block(grid.x < 1024 ? grid.x : 1024, 1, 1);
+        grid.x = (grid.x + 1023) / 1024;
+        kernel<<<grid, block, 0, stream>>>(out.data<U>(), out.size());
+      });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/reduce/reduce.cuh

@@ -47,13 +47,11 @@ namespace mlx::core {
     throw std::invalid_argument("Unknown reduce type."); \
   }
 
-void segmented_reduce(
+void all_reduce(
     cu::CommandEncoder& encoder,
     const array& in,
     array& out,
-    Reduce::ReduceType reduce_type,
+    Reduce::ReduceType reduce_type);
-    const std::vector<int>& axes,
-    const ReductionPlan& plan);
 
 void row_reduce(
     cu::CommandEncoder& encoder,
@@ -71,4 +69,10 @@ void col_reduce(
     const std::vector<int>& axes,
     const ReductionPlan& plan);
 
+void init_reduce(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type);
+
 } // namespace mlx::core

mlx/backend/cuda/reduce/reduce_ops.cuh

@@ -3,48 +3,89 @@
 #pragma once
 
 #include "mlx/backend/cuda/device/utils.cuh"
+#include "mlx/backend/cuda/reduce/reduce_utils.cuh"
 
 namespace mlx::core::cu {
 
 // Reduce ops.
 struct And {
-  __device__ bool operator()(bool a, bool b) {
+  __device__ __forceinline__ bool operator()(bool a, bool b) {
     return a && b;
   }
+
+  __device__ void atomic_update(bool* x, bool y) {
+    atomic_reduce<bool, And>(x, y);
+  }
 };
 
 struct Or {
-  __device__ bool operator()(bool a, bool b) {
+  __device__ __forceinline__ bool operator()(bool a, bool b) {
     return a || b;
   }
+
+  __device__ void atomic_update(bool* x, bool y) {
+    atomic_reduce<bool, Or>(x, y);
+  }
 };
 
 struct Sum {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __device__ __forceinline__ T operator()(T a, T b) {
     return a + b;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Sum>(x, y);
+  }
+
+  __device__ void atomic_update(__nv_bfloat16* x, __nv_bfloat16 y) {
+    atomicAdd(x, y);
+  }
+
+  __device__ void atomic_update(int* x, int y) {
+    atomicAdd(x, y);
+  }
+
+  __device__ void atomic_update(float* x, float y) {
+    atomicAdd(x, y);
+  }
 };
 
 struct Prod {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __device__ __forceinline__ T operator()(T a, T b) {
     return a * b;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Prod>(x, y);
+  }
 };
 
 struct Min {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __device__ __forceinline__ T operator()(T a, T b) {
     return a < b ? a : b;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Min>(x, y);
+  }
 };
 
 struct Max {
   template <typename T>
-  __device__ T operator()(T a, T b) {
+  __device__ __forceinline__ T operator()(T a, T b) {
     return a > b ? a : b;
   }
+
+  template <typename T>
+  __device__ void atomic_update(T* x, T y) {
+    atomic_reduce<T, Max>(x, y);
+  }
 };
 
 // Traits to get the result type of reduce op.
@@ -120,7 +161,7 @@ template <typename T>
 struct ReduceInit<Prod, T> {
   static constexpr __host__ __device__ auto value() {
     if constexpr (cuda::std::is_same_v<T, cuComplex>) {
-      return T{1, 1};
+      return T{1, 0};
     } else {
       return typename ReduceResult<Prod, T>::type{1};
     }

mlx/backend/cuda/reduce/reduce_utils.cuh

@@ -0,0 +1,134 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device/utils.cuh"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <size_t N>
+struct uint_by_size;
+template <>
+struct uint_by_size<2> {
+  using type = uint16_t;
+};
+template <>
+struct uint_by_size<4> {
+  using type = uint32_t;
+};
+template <>
+struct uint_by_size<8> {
+  using type = unsigned long long int;
+};
+
+template <typename T, typename Op>
+__device__ void atomic_reduce(T* x, T y) {
+  if constexpr (sizeof(T) == 1) {
+    using U = uint16_t;
+    U* x_int = (U*)((char*)x - ((size_t)x % 2));
+    int shift = ((char*)x - (char*)x_int) * 8;
+    int mask = 0xff << shift;
+    U old_val, new_val;
+    do {
+      old_val = *x_int;
+      T result = Op{}(static_cast<T>((old_val >> shift) & 0xff), y);
+      new_val = (old_val & ~mask) | (result << shift);
+    } while (atomicCAS(x_int, old_val, new_val) != old_val);
+  } else {
+    using U = typename uint_by_size<sizeof(T)>::type;
+    U* x_int = (U*)(x);
+    U old_val, new_val;
+    do {
+      old_val = *x_int;
+      T result = Op{}(*((T*)&old_val), y);
+      new_val = *((U*)&result);
+    } while (atomicCAS(x_int, old_val, new_val) != old_val);
+  }
+}
+
+// TODO: Should make a custom complex type
+template <typename U, typename T>
+inline __device__ U __cast(T x) {
+  return static_cast<U>(x);
+}
+
+template <>
+inline __device__ bool __cast<bool, cuComplex>(cuComplex x) {
+  return x.x != 0 && x.y != 0;
+}
+
+template <>
+inline __device__ cuComplex __cast<cuComplex, bool>(bool x) {
+  return x ? make_cuFloatComplex(1, 1) : make_cuFloatComplex(0, 0);
+}
+
+template <typename T, int N, typename Block, typename Warp, typename Op>
+inline __device__ void
+block_reduce(Block block, Warp warp, T (&vals)[N], T* smem, Op op, T init) {
+  // First reduce in the current warp
+  for (int i = 0; i < N; i++) {
+    vals[i] = cg::reduce(warp, vals[i], op);
+  }
+
+  // Reduce across warps
+  if (warp.meta_group_size() > 1) {
+    if (warp.thread_rank() == 0) {
+      for (int i = 0; i < N; i++) {
+        smem[warp.meta_group_rank() * N + i] = vals[i];
+      }
+    }
+    block.sync();
+    if (warp.thread_rank() < warp.meta_group_size()) {
+      for (int i = 0; i < N; i++) {
+        vals[i] = smem[warp.thread_rank() * N + i];
+      }
+    } else {
+      for (int i = 0; i < N; i++) {
+        vals[i] = init;
+      }
+    }
+    for (int i = 0; i < N; i++) {
+      vals[i] = cg::reduce(warp, vals[i], op);
+    }
+  }
+}
+
+} // namespace cu
+
+inline void allocate_same_layout(
+    array& out,
+    const array& in,
+    const std::vector<int>& axes) {
+  // Initialize out such that it matches in's layout. Basically we keep any
+  // transpositions as it were and that allows us either to skip finding the
+  // location of the output that matches the input or simply contiguous read or
+  // writes.
+  auto out_strides = in.strides();
+  for (auto ax : axes) {
+    for (auto& s : out_strides) {
+      if (s > in.strides(ax)) {
+        s /= in.shape(ax);
+      }
+    }
+  }
+  auto [data_size, rc, cc] = check_contiguity(out.shape(), out_strides);
+  auto fl = in.flags();
+  fl.row_contiguous = rc;
+  fl.col_contiguous = cc;
+  fl.contiguous = data_size == out.size();
+  out.set_data(
+      allocator::malloc(out.nbytes()),
+      data_size,
+      out_strides,
+      fl,
+      allocator::free);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/reduce/row_reduce.cu

@@ -55,84 +55,88 @@ struct RowReduceArgs {
       non_row_reductions *= reduce_shape[i];
     }
   }
+
+  // Convert shape and strides as if in was contiguous
+  void convert_shapes_to_contiguous(
+      const array& in,
+      const std::vector<int>& axes) {
+    auto shape_vec = in.shape();
+    auto strides_vec = in.strides();
+    size_t s = 1;
+    for (int i = in.ndim() - 1; i >= 0; i--) {
+      strides_vec[i] = s;
+      s *= shape_vec[i];
+    }
+    std::tie(shape_vec, strides_vec) =
+        shapes_without_reduction_axes(shape_vec, strides_vec, axes);
+    std::tie(shape_vec, strides_vec) =
+        collapse_contiguous_dims(shape_vec, strides_vec);
+    shape = const_param(shape_vec);
+    strides = const_param(strides_vec);
+    ndim = shape_vec.size();
+  }
 };
 
-template <typename T, typename U, typename Op, int NDIM, int N_READS = 4>
+template <typename T, typename U, typename ReduceOp, int N = 4, int M = 1>
-__global__ void row_reduce_small(
+__global__ void row_reduce_simple(T* in, U* out, size_t n_rows, int size) {
-    const T* in,
-    U* out,
-    size_t out_size,
-    const __grid_constant__ RowReduceArgs args) {
-  size_t out_idx = cg::this_grid().thread_rank();
-  if (out_idx >= out_size) {
-    return;
-  }
-
-  Op op;
-
-  U total_val = ReduceInit<Op, T>::value();
-  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
-
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
-
-  for (size_t n = 0; n < args.non_row_reductions; n++) {
-    for (int r = 0; r < cuda::ceil_div(args.row_size, N_READS); r++) {
-      U vals[N_READS];
-      cub::LoadDirectBlocked(
-          r,
-          make_cast_iterator<U>(in + loop.location()),
-          vals,
-          args.row_size,
-          ReduceInit<Op, T>::value());
-      total_val = op(total_val, cub::ThreadReduce(vals, op));
-    }
-    loop.next(args.reduce_shape.data(), args.reduce_strides.data());
-  }
-
-  out[out_idx] = total_val;
-}
-
-template <typename T, typename U, typename Op, int NDIM, int N_READS = 4>
-__global__ void row_reduce_small_warp(
-    const T* in,
-    U* out,
-    size_t out_size,
-    const __grid_constant__ RowReduceArgs args) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
   auto warp = cg::tiled_partition<WARP_SIZE>(block);
 
-  size_t out_idx = grid.thread_rank() / WARP_SIZE;
+  const U init = cu::ReduceInit<ReduceOp, T>::value();
-  if (out_idx >= out_size) {
+  ReduceOp op;
-    return;
+
+  T vals[M][N];
+  U accs[M];
+  for (int i = 0; i < M; i++) {
+    accs[i] = init;
   }
 
-  Op op;
+  const size_t start_row =
+      min(n_rows - M, static_cast<size_t>(grid.block_rank() * M));
+  const size_t full_blocks = size / (block.size() * N);
+  const size_t final_offset = full_blocks * (block.size() * N);
+  in += start_row * size;
+  out += start_row;
 
-  U total_val = ReduceInit<Op, T>::value();
+  for (size_t r = 0; r < full_blocks; r++) {
-  LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
+    for (int k = 0; k < M; k++) {
-
+      cub::LoadDirectBlockedVectorized<T, N>(
-  in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
+          block.thread_rank(), in + k * size + r * (block.size() * N), vals[k]);
-
+      for (int j = 0; j < N; j++) {
-  for (size_t n = warp.thread_rank(); n < args.non_row_reductions;
+        accs[k] = op(accs[k], __cast<U, T>(vals[k][j]));
-       n += WARP_SIZE) {
+      }
-    for (int r = 0; r < cuda::ceil_div(args.row_size, N_READS); r++) {
-      U vals[N_READS];
-      cub::LoadDirectBlocked(
-          r,
-          make_cast_iterator<U>(in + loop.location()),
-          vals,
-          args.row_size,
-          ReduceInit<Op, T>::value());
-      total_val = op(total_val, cub::ThreadReduce(vals, op));
     }
-    loop.next(WARP_SIZE, args.reduce_shape.data(), args.reduce_strides.data());
   }
 
-  total_val = cg::reduce(warp, total_val, op);
+  if (final_offset < size) {
+    for (int k = 0; k < M; k++) {
+      cub::LoadDirectBlocked(
+          block.thread_rank(),
+          in + k * size + final_offset,
+          vals[k],
+          size,
+          __cast<T, U>(init));
+      for (int j = 0; j < N; j++) {
+        accs[k] = op(accs[k], __cast<U, T>(vals[k][j]));
+      }
+    }
+  }
 
-  if (warp.thread_rank() == 0) {
+  __shared__ U shared_accumulators[32 * M];
-    out[out_idx] = total_val;
+  block_reduce(block, warp, accs, shared_accumulators, op, init);
+
+  if (block.thread_rank() == 0) {
+    if (grid.block_rank() * M + M <= n_rows) {
+      for (int i = 0; i < M; i++) {
+        out[i] = accs[i];
+      }
+    } else {
+      short offset = grid.block_rank() * M + M - n_rows;
+      for (int i = offset; i < M; i++) {
+        out[i] = accs[i];
+      }
+    }
   }
 }
 
@@ -141,55 +145,167 @@ template <
     typename U,
     typename Op,
     int NDIM,
-    int BLOCK_DIM_X,
+    int BLOCK_DIM,
     int N_READS = 4>
 __global__ void row_reduce_looped(
-    const T* in,
+    T* in,
     U* out,
     size_t out_size,
     const __grid_constant__ RowReduceArgs args) {
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
 
-  size_t out_idx = grid.thread_rank() / BLOCK_DIM_X;
+  size_t out_idx = grid.block_rank();
-  if (out_idx >= out_size) {
-    return;
-  }
 
   Op op;
 
-  U total_val = ReduceInit<Op, T>::value();
+  U total[1];
+  U init = ReduceInit<Op, T>::value();
+  total[0] = init;
   LoopedElemToLoc<NDIM, (NDIM > 2)> loop(args.reduce_ndim);
+  size_t full_blocks = args.row_size / (BLOCK_DIM * N_READS);
+  size_t final_offset = full_blocks * BLOCK_DIM * N_READS;
 
   in += elem_to_loc(out_idx, args.shape.data(), args.strides.data(), args.ndim);
 
   for (size_t n = 0; n < args.non_row_reductions; n++) {
-    for (size_t r = 0; r < cuda::ceil_div(args.row_size, BLOCK_DIM_X * N_READS);
+    for (size_t r = 0; r < full_blocks; r++) {
-         r++) {
+      T vals[N_READS];
-      U vals[N_READS];
+      cub::LoadDirectBlockedVectorized<T, N_READS>(
-      cub::LoadDirectBlocked(
+          block.thread_rank(),
-          r * BLOCK_DIM_X + block.thread_index().x,
+          in + loop.location() + r * BLOCK_DIM * N_READS,
-          make_cast_iterator<U>(in + loop.location()),
+          vals);
-          vals,
+      for (int i = 0; i < N_READS; i++) {
-          args.row_size,
+        total[0] = op(total[0], __cast<U, T>(vals[i]));
-          ReduceInit<Op, T>::value());
+      }
-      total_val = op(total_val, cub::ThreadReduce(vals, op));
     }
+    if (final_offset < args.row_size) {
+      T vals[N_READS];
+      cub::LoadDirectBlocked(
+          block.thread_rank(),
+          in + loop.location() + final_offset,
+          vals,
+          args.row_size - final_offset,
+          __cast<T, U>(init));
+      for (int i = 0; i < N_READS; i++) {
+        total[0] = op(total[0], __cast<U, T>(vals[i]));
+      }
+    }
+    // TODO: Maybe block.sync() here?
     loop.next(args.reduce_shape.data(), args.reduce_strides.data());
   }
 
-  typedef cub::BlockReduce<U, BLOCK_DIM_X> BlockReduceT;
+  __shared__ U shared_accumulators[32];
-  __shared__ typename BlockReduceT::TempStorage temp;
+  block_reduce(block, warp, total, shared_accumulators, op, init);
-
-  total_val = BlockReduceT(temp).Reduce(total_val, op);
 
   if (block.thread_rank() == 0) {
-    out[out_idx] = total_val;
+    out[out_idx] = total[0];
   }
 }
 
 } // namespace cu
 
+void row_reduce_simple(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type,
+    const std::vector<int>& axes,
+    const ReductionPlan& plan) {
+  constexpr int N_READS = 8;
+
+  // Allocate data for the output using in's layout to avoid elem_to_loc in the
+  // kernel.
+  allocate_same_layout(out, in, axes);
+
+  // Just a way to get out of the constness because cub doesn't like it ...
+  // (sigh)
+  array x = in;
+
+  // TODO: If out.size() < 1024 which will be a common case then write this in
+  //       2 passes. Something like 32 * out.size() and then do a warp reduce.
+  encoder.set_input_array(x);
+  encoder.set_output_array(out);
+  encoder.launch_kernel([&](cudaStream_t stream) {
+    MLX_SWITCH_ALL_TYPES(x.dtype(), CTYPE, {
+      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
+        using T = cuda_type_t<CTYPE>;
+        using U = cu::ReduceResult<OP, T>::type;
+
+        // Calculate the grid and block dims
+        size_t reductions = (plan.shape.back() + N_READS - 1) / N_READS;
+        dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
+        int threads = std::min(1024UL, reductions);
+        threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+        dim3 block(threads, 1, 1);
+
+        // Pick the kernel
+        auto kernel = cu::row_reduce_simple<T, U, OP, N_READS>;
+        if (grid.x >= 1024) {
+          grid.x = (grid.x + 1) / 2;
+          kernel = cu::row_reduce_simple<T, U, OP, N_READS, 2>;
+        }
+
+        // Launch
+        kernel<<<grid, block, 0, stream>>>(
+            x.data<T>(), out.data<U>(), out.size(), plan.shape.back());
+      });
+    });
+  });
+}
+
+void row_reduce_looped(
+    cu::CommandEncoder& encoder,
+    const array& in,
+    array& out,
+    Reduce::ReduceType reduce_type,
+    const std::vector<int>& axes,
+    const ReductionPlan& plan,
+    cu::RowReduceArgs args) {
+  constexpr int N_READS = 8;
+
+  // Allocate data for the output using in's layout to access them as
+  // contiguously as possible.
+  allocate_same_layout(out, in, axes);
+
+  // Just a way to get out of the constness because cub doesn't like it ...
+  // (sigh)
+  array x = in;
+
+  encoder.set_input_array(x);
+  encoder.set_output_array(out);
+  encoder.launch_kernel([&](cudaStream_t stream) {
+    MLX_SWITCH_ALL_TYPES(x.dtype(), CTYPE, {
+      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
+        using T = cuda_type_t<CTYPE>;
+        using U = cu::ReduceResult<OP, T>::type;
+
+        // Calculate the grid and block dims
+        args.convert_shapes_to_contiguous(x, axes);
+        dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
+        size_t reductions = (args.row_size + N_READS - 1) / N_READS;
+        int threads = std::min(1024UL, reductions);
+        threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+        dim3 block(threads, 1, 1);
+
+        // Pick the kernel
+        auto kernel = cu::row_reduce_looped<T, U, OP, 1, 32, N_READS>;
+        MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
+          MLX_SWITCH_BLOCK_DIM(threads, THREADS, {
+            kernel = cu::row_reduce_looped<T, U, OP, NDIM, THREADS, N_READS>;
+            block.x = THREADS;
+          });
+        });
+
+        // Launch
+        kernel<<<grid, block, 0, stream>>>(
+            x.data<T>(), out.data<U>(), out.size(), args);
+      });
+    });
+  });
+}
+
 void row_reduce(
     cu::CommandEncoder& encoder,
     const array& in,
@@ -197,54 +313,35 @@ void row_reduce(
     Reduce::ReduceType reduce_type,
     const std::vector<int>& axes,
     const ReductionPlan& plan) {
+  // Current row reduction options
+  //
+  // - row_reduce_simple
+  //
+  //   That means that we are simply reducing across the fastest moving axis.
+  //   We are reducing 1 or 2 rows per threadblock depending on the size of
+  //   output.
+  //
+  // - row_reduce_looped
+  //
+  //   It is a general row reduction. We are computing 1 output per
+  //   threadblock. We read the fastest moving axis vectorized and loop over
+  //   the rest of the axes.
+  //
+  // Notes: We opt to read as much in order as possible and leave
+  //        transpositions as they are (contrary to our Metal backend).
+
+  // Simple row reduce means that we have 1 axis that we are reducing over and
+  // it has stride 1.
+  if (plan.shape.size() == 1) {
+    row_reduce_simple(encoder, in, out, reduce_type, axes, plan);
+    return;
+  }
+
+  // Make the args struct to help route to the best kernel
   cu::RowReduceArgs args(in, plan, axes);
 
-  encoder.launch_kernel([&](cudaStream_t stream) {
+  // Fallback row reduce
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
+  row_reduce_looped(encoder, in, out, reduce_type, axes, plan, std::move(args));
-      using InType = cuda_type_t<CTYPE>;
-      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
-        using OutType = cu::ReduceResult<OP, InType>::type;
-        MLX_SWITCH_REDUCE_NDIM(args.reduce_ndim, NDIM, {
-          constexpr size_t N_READS = 4;
-          dim3 out_dims = get_2d_grid_dims(out.shape(), out.strides());
-          dim3 block_dims, num_blocks;
-          auto kernel =
-              cu::row_reduce_small<InType, OutType, OP, NDIM, N_READS>;
-          if (args.row_size <= 64) {
-            if ((args.non_row_reductions < 32 && args.row_size <= 8) ||
-                (args.non_row_reductions <= 8)) {
-              block_dims.x = std::min(out_dims.x, 1024u);
-              num_blocks.x = cuda::ceil_div(out_dims.x, block_dims.x);
-              num_blocks.y = out_dims.y;
-            } else {
-              block_dims.x = WARP_SIZE;
-              num_blocks.y = out_dims.x;
-              num_blocks.z = out_dims.y;
-              kernel =
-                  cu::row_reduce_small_warp<InType, OutType, OP, NDIM, N_READS>;
-            }
-          } else {
-            size_t num_threads = cuda::ceil_div(args.row_size, N_READS);
-            num_threads = cuda::ceil_div(num_threads, WARP_SIZE) * WARP_SIZE;
-            MLX_SWITCH_BLOCK_DIM(num_threads, BLOCK_DIM_X, {
-              num_blocks.y = out_dims.x;
-              num_blocks.z = out_dims.y;
-              block_dims.x = BLOCK_DIM_X;
-              kernel = cu::row_reduce_looped<
-                  InType,
-                  OutType,
-                  OP,
-                  NDIM,
-                  BLOCK_DIM_X,
-                  N_READS>;
-            });
-          }
-          kernel<<<num_blocks, block_dims, 0, stream>>>(
-              in.data<InType>(), out.data<OutType>(), out.size(), args);
-        });
-      });
-    });
-  });
 }
 
 } // namespace mlx::core

mlx/backend/cuda/reduce/segmented_reduce.cu

@@ -1,84 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/device/cast_op.cuh"
-#include "mlx/backend/cuda/reduce/reduce.cuh"
-
-#include <thrust/device_ptr.h>
-#include <cub/device/device_reduce.cuh>
-#include <cub/device/device_segmented_reduce.cuh>
-
-namespace mlx::core {
-
-template <typename... Args>
-void cub_all_reduce(cu::CommandEncoder& encoder, Args&&... args) {
-  // Allocate temporary storage.
-  size_t size;
-  CHECK_CUDA_ERROR(cub::DeviceReduce::Reduce(nullptr, size, args...));
-  array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
-  encoder.add_temporary(temp);
-  // Run op.
-  CHECK_CUDA_ERROR(cub::DeviceReduce::Reduce(temp.data<void>(), size, args...));
-}
-
-template <typename... Args>
-void cub_segmented_reduce(cu::CommandEncoder& encoder, Args&&... args) {
-  // Allocate temporary storage.
-  size_t size;
-  CHECK_CUDA_ERROR(cub::DeviceSegmentedReduce::Reduce(nullptr, size, args...));
-  array temp(allocator::malloc(size), {static_cast<int>(size)}, uint8);
-  encoder.add_temporary(temp);
-  // Run op.
-  CHECK_CUDA_ERROR(
-      cub::DeviceSegmentedReduce::Reduce(temp.data<void>(), size, args...));
-}
-
-struct MultiplyOp {
-  int factor;
-  __device__ int operator()(int i) {
-    return i * factor;
-  }
-};
-
-void segmented_reduce(
-    cu::CommandEncoder& encoder,
-    const array& in,
-    array& out,
-    Reduce::ReduceType reduce_type,
-    const std::vector<int>& axes,
-    const ReductionPlan& plan) {
-  encoder.launch_kernel([&](cudaStream_t stream) {
-    MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
-      MLX_SWITCH_REDUCE_OPS(reduce_type, OP, {
-        using InType = cuda_type_t<CTYPE>;
-        using OutType = cu::ReduceResult<OP, InType>::type;
-        auto in_iter = cu::make_cast_iterator<OutType>(
-            thrust::device_pointer_cast(in.data<InType>()));
-        auto out_ptr = thrust::device_pointer_cast(out.data<OutType>());
-        auto init = cu::ReduceInit<OP, InType>::value();
-
-        if (plan.type == ContiguousAllReduce) {
-          cub_all_reduce(
-              encoder, in_iter, out_ptr, in.data_size(), OP(), init, stream);
-        } else if (plan.type == ContiguousReduce) {
-          auto offsets = thrust::make_transform_iterator(
-              thrust::make_counting_iterator(0), MultiplyOp{plan.shape.back()});
-          cub_segmented_reduce(
-              encoder,
-              in_iter,
-              out_ptr,
-              out.size(),
-              offsets,
-              offsets + 1,
-              OP(),
-              init,
-              stream);
-        } else {
-          throw std::runtime_error("Unsupported plan in segmented_reduce.");
-        }
-      });
-    });
-  });
-}
-
-} // namespace mlx::core

mlx/backend/cuda/softmax.cu

@@ -51,7 +51,7 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
         make_cast_iterator<AccT>(in),
         vals,
         axis_size,
-        Limits<AccT>::finite_min());
+        Limits<AccT>::min());
     prevmax = maxval;
     maxval = max_op(maxval, cub::ThreadReduce(vals, max_op));
     // Online normalizer calculation for softmax:
@@ -79,7 +79,7 @@ __global__ void softmax(const T* in, T* out, int axis_size) {
   block.sync();
   maxval = warp.thread_rank() < warp.meta_group_size()
       ? local_max[warp.thread_rank()]
-      : Limits<AccT>::finite_min();
+      : Limits<AccT>::min();
   maxval = cg::reduce(warp, maxval, max_op);
   normalizer = normalizer * softmax_exp(prevmax - maxval);
   if (warp.thread_rank() == 0) {

mlx/backend/cuda/sort.cu

@@ -79,9 +79,6 @@ void segmented_sort(cu::CommandEncoder& encoder, Args&&... args) {
 void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
   array out = out_;
   auto& encoder = cu::get_command_encoder(s);
-  encoder.set_input_array(in);
-  encoder.set_output_array(out);
-
   if (axis < 0) {
     axis += in.ndim();
   }
@@ -106,6 +103,8 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
         in.flags());
   }
 
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
   encoder.launch_kernel([&](cudaStream_t stream) {
     MLX_SWITCH_ALL_TYPES(in.dtype(), CTYPE, {
       if constexpr (!std::is_same_v<CTYPE, complex64_t>) {

python/mlx/nn/layers/base.py

@@ -413,7 +413,7 @@ class Module(dict):
                             f'Module does not have sub-module named "{k}".'
                         )
             elif isinstance(modules, list):
-                for i in range(len(dst)):
+                for i in range(len(modules)):
                     current_value = dst[i]
                     new_value = modules[i]
                     if self.is_module(current_value) and self.is_module(new_value):

python/scripts/repair_cuda.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+
+auditwheel repair dist/* \
+  --plat manylinux_2_35_x86_64 \
+  --exclude libcublas* \
+  --exclude libnvrtc*
+
+cd wheelhouse
+repaired_wheel=$(find . -name "*.whl" -print -quit)
+unzip -q "${repaired_wheel}"
+core_so=$(find mlx -name "core*.so" -print -quit)
+rpath=$(patchelf --print-rpath "${core_so}")
+rpath=$rpath:\$ORIGIN/../nvidia/cublas/lib:\$ORIGIN/../nvidia/cuda_nvrtc/lib
+patchelf --force-rpath --set-rpath "$rpath" "$core_so"
+
+# Re-zip the repaired wheel
+zip -r -q "${repaired_wheel}" .

python/tests/cuda_skip.py

@@ -13,7 +13,6 @@ cuda_skip = {
     "TestLayers.test_upsample",
     "TestOps.test_complex_ops",
     "TestOps.test_dynamic_slicing",
-    "TestOps.test_softmax",
     "TestReduce.test_axis_permutation_sums",
     "TestReduce.test_dtypes",
     "TestReduce.test_expand_sums",

python/tests/test_nn.py

@@ -259,6 +259,11 @@ class TestBase(mlx_tests.MLXTestCase):
         with self.assertRaises(ValueError):
             m = m.update_modules({"list": ["hi"]})
 
+        # Allow updating a strict subset
+        m = nn.Sequential(nn.Linear(3, 3), nn.Linear(3, 3))
+        m.update_modules({"layers": [{}, nn.Linear(3, 4)]})
+        self.assertEqual(m.layers[1].weight.shape, (4, 3))
+
 
 class TestLayers(mlx_tests.MLXTestCase):
     def test_identity(self):

setup.py

@@ -174,20 +174,26 @@ if __name__ == "__main__":
     )
     package_dir = {"": "python"}
     package_data = {"mlx": ["lib/*", "include/*", "share/*"], "mlx.core": ["*.pyi"]}
+    install_requires = []
+    build_cuda = "MLX_BUILD_CUDA=ON" in os.environ.get("CMAKE_ARGS", "")
+    if build_cuda:
+        install_requires = ["nvidia-cublas-cu12", "nvidia-cuda-nvrtc-cu12"]
 
     setup(
-        name="mlx",
+        name="mlx-cuda" if build_cuda else "mlx",
         version=get_version(),
         author="MLX Contributors",
         author_email="mlx@group.apple.com",
         description="A framework for machine learning on Apple silicon.",
         long_description=long_description,
         long_description_content_type="text/markdown",
+        license="MIT",
         url="https://github.com/ml-explore/mlx",
         packages=packages,
         package_dir=package_dir,
         package_data=package_data,
         include_package_data=True,
+        install_requires=install_requires,
         extras_require={
             "dev": [
                 "nanobind==2.4.0",