Improve the cutlass gemm

cmake/FindNCCL.cmake

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

docs/src/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 libcudnn9-dev-cuda-12 -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
 
 
 When building either the Python or C++ APIs make sure to pass the cmake flag

docs/src/python/cuda.rst

@@ -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

mlx/backend/cuda/CMakeLists.txt

@@ -20,12 +20,14 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/conv/gemm_grouped_conv.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/cudnn_utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/custom_kernel.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/gemv.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cutlass_gemm.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/gemms/simple_gemm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/jit_module.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
@@ -88,6 +90,9 @@ target_include_directories(mlx PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/gen")
 target_compile_options(mlx
                        PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>")
 
+# Keep ptx around for inspection
+target_compile_options(mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--keep>")
+
 # Enable calling host constexpr functions from device. This is needed because
 # the constexpr version of isnan is host only.
 target_compile_options(
@@ -173,3 +178,12 @@ target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
 # Install CCCL headers for JIT.
 install(DIRECTORY ${cccl_SOURCE_DIR}/include/cuda
         DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/cccl)
+
+# Fetch and make available cutlass
+FetchContent_Declare(
+  cutlass
+  GIT_REPOSITORY https://github.com/NVIDIA/cutlass.git
+  GIT_TAG v4.1.0)
+FetchContent_Populate(cutlass)
+target_include_directories(
+  mlx PRIVATE $<BUILD_INTERFACE:${cutlass_SOURCE_DIR}/include>)

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/cudnn_utils.cpp

@@ -23,6 +23,24 @@ inline cudnn_frontend::Tensor build_cudnn_tensor(
       .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);
@@ -33,8 +51,9 @@ auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
   return std::make_tuple(std::move(shape), std::move(strides));
 }
 
-auto nhwc_to_nchw(const array& x) {
+inline auto nhwc_to_nchw(const array& x) {
-  return nhwc_to_nchw(convert_vector<int64_t>(x.shape()), x.strides());
+  return nhwc_to_nchw(
+      convert_vector<int64_t>(x.shape()), normalized_strides(x));
 }
 
 // Return available engines for a |op_graph|.
@@ -140,7 +159,7 @@ bool prepare_cudnn_plan(
 
 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, x.strides());
+  return build_cudnn_tensor(id, x, shape, normalized_strides(x));
 }
 
 cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x) {
@@ -160,7 +179,8 @@ cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x) {
     return build_cudnn_tensor(id, x, shape, strides);
   }
   if (x.ndim() == 2) {
-    int64_t s = x.strides(0);
+    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);

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/distributed.cu

@@ -1,51 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cuda/device.h"
-#include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/distributed/primitives.h"
-#include "mlx/primitives.h"
-
-#include <cassert>
-
-namespace mlx::core {
-namespace distributed {
-void AllReduce::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  assert(inputs.size() == 1);
-  assert(outputs.size() == 1);
-
-  auto& input = inputs[0];
-  auto& output = outputs[0];
-
-  auto& encoder = cu::get_command_encoder(stream());
-
-  if (input.is_donatable()) {
-    output.copy_shared_buffer(input);
-  } else {
-    output.set_data(allocator::malloc(output.nbytes()));
-  }
-
-  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 distributed
-} // namespace mlx::core

mlx/backend/cuda/gemms/cutlass_gemm.cu

@@ -0,0 +1,396 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+#include <cute/tensor.hpp>
+#include <cutlass/arch/arch.h>
+#include <cutlass/cutlass.h>
+#include <cutlass/gemm/device/gemm.h>
+#include <cutlass/layout/matrix.h>
+#include <cutlass/numeric_types.h>
+
+#include <iostream>
+
+namespace mlx::core::cu {
+
+namespace {
+
+using namespace cute;
+using bf16 = cute::bfloat16_t;
+
+template <typename Kernel>
+void configure_matmul(Kernel kernel, int smem_size) {
+  static bool initialized = false;
+  if (!initialized) {
+    initialized = true;
+    cudaFuncSetAttribute(
+        kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
+  }
+}
+
+template <bool transpose, typename Tiler>
+constexpr int get_feature_size(Tiler smem) {
+  int feature_size = (transpose) ? size<0>(smem) : size<1>(smem);
+  return (feature_size >= 64) ? 64 : feature_size;
+}
+
+constexpr int constexpr_log2(int x) {
+  return (x > 0) ? 1 + constexpr_log2(x >> 1) : -1;
+}
+
+template <int feature_size, int itemsize, int copy_bits>
+constexpr int get_swizzle_bits() {
+  constexpr int swizzle_bits =
+      constexpr_log2(feature_size * itemsize / copy_bits);
+  return (swizzle_bits > 3) ? 3 : swizzle_bits;
+}
+
+template <int itemsize, bool transpose, int copy_bits, typename Tiler>
+constexpr auto make_smem_layout(Tiler smem) {
+  constexpr int feature_size = get_feature_size<transpose>(smem);
+  constexpr int swizzle_bits =
+      get_swizzle_bits<feature_size, itemsize, copy_bits>();
+
+  using F = Int<feature_size>;
+  using BaseLayout = std::conditional_t<
+      transpose,
+      Layout<cute::Shape<F, _8>, cute::Stride<_1, F>>,
+      Layout<cute::Shape<_8, F>, cute::Stride<F, _1>>>;
+
+  auto swizzled =
+      make_composed_layout(Swizzle<swizzle_bits, 3, 3>{}, 0, BaseLayout{});
+
+  return tile_to_shape(swizzled, smem);
+}
+
+template <int itemsize, bool transpose, int copy_bits, typename Tiler>
+constexpr auto make_result_smem_layout(Tiler smem) {
+  constexpr int feature_size = get_feature_size<transpose>(smem);
+  constexpr int swizzle_bits =
+      get_swizzle_bits<feature_size, itemsize, copy_bits>();
+
+  using F = Int<feature_size>;
+  using BaseLayout = std::conditional_t<
+      transpose,
+      Layout<cute::Shape<F, _8>, cute::Stride<_1, F>>,
+      Layout<cute::Shape<_8, F>, cute::Stride<F, _1>>>;
+
+  auto swizzled = make_composed_layout(
+      Swizzle<transpose ? 0 : swizzle_bits, 3, 4>{}, 0, BaseLayout{});
+
+  return tile_to_shape(swizzled, smem);
+}
+
+template <
+    int num_threads,
+    int itemsize,
+    bool transpose,
+    int copy_bits,
+    typename Copier,
+    typename Tiler>
+constexpr auto make_tiled_copy(Copier copy_op, Tiler smem) {
+  constexpr int num_elements = copy_bits / itemsize;
+  constexpr int feature_size = transpose ? size<0>(smem) : size<1>(smem);
+  constexpr int copies_per_feature = feature_size / num_elements;
+
+  using E = Int<num_elements>;
+  using C = Int<copies_per_feature>;
+  using R = Int<num_threads / copies_per_feature>;
+
+  using ThreadLayout = std::conditional_t<
+      transpose,
+      Layout<cute::Shape<C, R>, cute::Stride<_1, C>>,
+      Layout<cute::Shape<R, C>, cute::Stride<C, _1>>>;
+  using ValueLayout = std::conditional_t<
+      transpose,
+      Layout<cute::Shape<E, _1>>,
+      Layout<cute::Shape<_1, E>>>;
+
+  return make_tiled_copy(copy_op, ThreadLayout{}, ValueLayout{});
+}
+
+template <int rasterization_factor>
+__device__ inline int2 raster_tile(int x, int y) {
+  return {
+      x / rasterization_factor,
+      (x % rasterization_factor) + y * rasterization_factor};
+}
+
+template <
+    typename T,
+    typename SLayoutA,
+    typename SLayoutB,
+    typename SLayoutC,
+    typename CopyA,
+    typename CopyB,
+    typename CopyC,
+    typename MMA,
+    int rasterization_factor>
+__global__ static __launch_bounds__(decltype(size(MMA{}))::value) void matmul_kernel(
+    const T* __restrict__ A,
+    const T* __restrict__ B,
+    T* __restrict__ C,
+    SLayoutA SA,
+    SLayoutB SB,
+    SLayoutC SC,
+    CopyA copy_a,
+    CopyB copy_b,
+    CopyC copy_c,
+    MMA mma,
+    int M,
+    int N,
+    int K) {
+  constexpr auto BM = size<0>(SA);
+  constexpr auto BN = size<0>(SB);
+  constexpr auto BK = size<1>(SA);
+  constexpr auto PIPE = size<2>(SA);
+
+  const int2 tile = raster_tile<rasterization_factor>(blockIdx.x, blockIdx.y);
+  const int blocks_m = ceil_div(M, BM);
+  const int blocks_n = ceil_div(N, BN);
+
+  // Exit early if the tile is OOB
+  if (tile.x >= blocks_m || tile.y >= blocks_n) {
+    return;
+  }
+
+  // Make the full tensors
+  Tensor full_A =
+      make_tensor(make_gmem_ptr(A), make_shape(M, K), make_stride(K, _1{}));
+  Tensor full_B =
+      make_tensor(make_gmem_ptr(B), make_shape(N, K), make_stride(K, _1{}));
+  Tensor full_C =
+      make_tensor(make_gmem_ptr(C), make_shape(M, N), make_stride(N, _1{}));
+
+  // Partition the tensors into tiles and select the ones for this threadblock
+  Tensor local_A =
+      local_tile(full_A, make_shape(BM, BK), make_coord(tile.x, _));
+  Tensor local_B =
+      local_tile(full_B, make_shape(BN, BK), make_coord(tile.y, _));
+  Tensor local_C =
+      local_tile(full_C, make_shape(BM, BN), make_coord(tile.x, tile.y));
+
+  // Make shared memory tensors
+  extern __shared__ char shared_memory[];
+  T* shared_A_ptr = reinterpret_cast<T*>(shared_memory);
+  T* shared_B_ptr =
+      reinterpret_cast<T*>(shared_memory + cosize(SA) * sizeof(T));
+  T* shared_C_ptr = reinterpret_cast<T*>(shared_memory);
+  Tensor shared_A = make_tensor(make_smem_ptr(shared_A_ptr), SA);
+  Tensor shared_B = make_tensor(make_smem_ptr(shared_B_ptr), SB);
+  Tensor shared_C = make_tensor(make_smem_ptr(shared_C_ptr), SC);
+
+  // Get the copies that correspond to this thread
+  auto thread_copy_a = copy_a.get_slice(threadIdx.x);
+  Tensor local_A_src = thread_copy_a.partition_S(local_A);
+  Tensor local_A_dst = thread_copy_a.partition_D(shared_A);
+  auto thread_copy_b = copy_b.get_slice(threadIdx.x);
+  Tensor local_B_src = thread_copy_a.partition_S(local_B);
+  Tensor local_B_dst = thread_copy_a.partition_D(shared_B);
+  auto thread_copy_c = copy_c.get_slice(threadIdx.x);
+  Tensor local_C_src = thread_copy_c.partition_S(shared_C);
+  Tensor local_C_dst = thread_copy_c.partition_D(local_C);
+
+  // Start fetches
+  int k_tile_count = size<2>(local_A);
+  int k_tile_next = 0;
+  CUTE_UNROLL
+  for (int k = 0; k < PIPE - 1; k++) {
+    copy(copy_a, local_A_src(_, _, _, k_tile_next), local_A_dst(_, _, _, k));
+    copy(copy_b, local_B_src(_, _, _, k_tile_next), local_B_dst(_, _, _, k));
+    cp_async_fence();
+    k_tile_count--;
+    k_tile_next += (k_tile_count > 0);
+  }
+
+  // Get the MMA that corresponds to this thread and allocate registers
+  auto thread_mma = mma.get_slice(threadIdx.x);
+  Tensor mma_shared_A = thread_mma.partition_A(shared_A);
+  Tensor mma_shared_B = thread_mma.partition_B(shared_B);
+  Tensor mma_shared_C = thread_mma.partition_C(shared_C);
+  Tensor mma_global_C = thread_mma.partition_C(local_C);
+  Tensor mma_frag_A = mma.make_fragment_A(mma_shared_A(_, _, _, 0));
+  Tensor mma_frag_B = mma.make_fragment_B(mma_shared_B(_, _, _, 0));
+  Tensor mma_frag_C = mma.make_fragment_C(mma_global_C);
+  clear(mma_frag_C);
+
+  // Make shared to register copies
+  Copy_Atom<SM75_U32x4_LDSM_N, bf16> s2r_atom_a;
+  Copy_Atom<SM75_U32x4_LDSM_N, bf16> s2r_atom_b;
+  auto s2r_copy_a = make_tiled_copy_A(s2r_atom_a, mma);
+  auto s2r_thread_copy_a = s2r_copy_a.get_slice(threadIdx.x);
+  auto s2r_copy_b = make_tiled_copy_B(s2r_atom_b, mma);
+  auto s2r_thread_copy_b = s2r_copy_b.get_slice(threadIdx.x);
+  Tensor mma_A_src = s2r_thread_copy_a.partition_S(shared_A);
+  Tensor mma_A_dst = s2r_thread_copy_a.retile_D(mma_frag_A);
+  Tensor mma_B_src = s2r_thread_copy_b.partition_S(shared_B);
+  Tensor mma_B_dst = s2r_thread_copy_b.retile_D(mma_frag_B);
+
+  constexpr auto RPIPE = size<2>(mma_shared_A);
+  int smem_read = 0;
+  int smem_write = PIPE - 1;
+  Tensor mma_A_src_p = mma_A_src(_, _, _, smem_read);
+  Tensor mma_B_src_p = mma_B_src(_, _, _, smem_read);
+
+  // Start the register pipeline
+  if constexpr (RPIPE > 1) {
+    cp_async_wait<PIPE - 2>();
+    __syncthreads();
+    copy(s2r_copy_a, mma_A_src_p(_, _, Int<0>{}), mma_A_dst(_, _, Int<0>{}));
+    copy(s2r_copy_b, mma_B_src_p(_, _, Int<0>{}), mma_B_dst(_, _, Int<0>{}));
+  }
+
+  CUTE_NO_UNROLL
+  while (k_tile_count > -(PIPE - 1)) {
+    CUTE_UNROLL
+    for (int k_block = 0; k_block < RPIPE; k_block++) {
+      if (k_block == RPIPE - 1) {
+        mma_A_src_p = mma_A_src(_, _, _, smem_read);
+        mma_B_src_p = mma_B_src(_, _, _, smem_read);
+        cp_async_wait<PIPE - 2>();
+        __syncthreads();
+      }
+
+      // Load the next register tile
+      auto k_block_next = (k_block + 1) % RPIPE;
+      copy(
+          s2r_copy_a,
+          mma_A_src_p(_, _, k_block_next),
+          mma_A_dst(_, _, k_block_next));
+      copy(
+          s2r_copy_b,
+          mma_B_src_p(_, _, k_block_next),
+          mma_B_dst(_, _, k_block_next));
+
+      if (k_block == 0) {
+        copy(
+            copy_a,
+            local_A_src(_, _, _, k_tile_next),
+            local_A_dst(_, _, _, smem_write));
+        copy(
+            copy_b,
+            local_B_src(_, _, _, k_tile_next),
+            local_B_dst(_, _, _, smem_write));
+        cp_async_fence();
+        k_tile_count--;
+        k_tile_next += (k_tile_count > 0);
+        smem_write = smem_read;
+        smem_read = (smem_read == PIPE - 1) ? 0 : (smem_read + 1);
+      }
+
+      gemm(
+          mma,
+          mma_frag_A(_, _, k_block),
+          mma_frag_B(_, _, k_block),
+          mma_frag_C);
+    }
+  }
+
+  copy(mma_frag_C, mma_shared_C);
+  __syncthreads();
+  copy(copy_c, local_C_src, local_C_dst);
+
+  // if (threadIdx.x == 0) {
+  //   print("fC: "); print(mma_frag_C); print("\n");
+  //   print("sC: "); print(mma_shared_C); print("\n");
+  //   print("dC: "); print(local_C_dst); print("\n");
+  //
+  //   print(s2r_atom_a); print("\n");
+  // }
+}
+
+} // namespace
+
+void cutlass_gemm(
+    const array& a,
+    const array& b,
+    array& out,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc) {
+  enc.set_input_array(a);
+  enc.set_input_array(b);
+  enc.set_output_array(out);
+  dispatch_float_types(a.dtype(), "simple_gemm", [&](auto type_tag) {
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    if constexpr (std::is_same_v<DataType, __nv_bfloat16>) {
+      using namespace cute;
+
+      // Tile definitions
+      auto BM = Int<128>{};
+      auto BN = Int<128>{};
+      auto BK = Int<64>{};
+      auto BP = Int<3>{};
+      auto GM = Int<8>{};
+
+      // Thread definitions
+      using TM = Int<2>;
+      using TN = Int<2>;
+      using TK = Int<1>;
+      constexpr int num_threads = TM::value * TN::value * 32;
+
+      auto SA = make_smem_layout<16, false, 128>(make_shape(BM, BK, BP));
+      auto SB = make_smem_layout<16, false, 128>(make_shape(BN, BK, BP));
+      auto SC = make_result_smem_layout<16, false, 128>(make_shape(BM, BN));
+
+      constexpr auto smem_size = (cosize(SA) + cosize(SB)) * sizeof(bf16);
+
+      auto async_copy_op =
+          Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<uint128_t>, bf16>{};
+      auto tiled_copy_a = make_tiled_copy<num_threads, 16, false, 128>(
+          async_copy_op, make_shape(BM, BK));
+      auto tiled_copy_b = make_tiled_copy<num_threads, 16, false, 128>(
+          async_copy_op, make_shape(BN, BK));
+
+      auto sync_copy_op = Copy_Atom<UniversalCopy<uint128_t>, bf16>{};
+      auto tiled_copy_c = make_tiled_copy<num_threads, 16, false, 128>(
+          sync_copy_op, make_shape(BM, BN));
+
+      auto mma_op = SM80_16x8x16_F32BF16BF16F32_TN{};
+      auto tiled_mma = make_tiled_mma(
+          mma_op, Layout<cute::Shape<TM, TN, TK>>{}, Tile<_32, _32, _16>{});
+
+      auto kernel = matmul_kernel<
+          bf16,
+          decltype(SA),
+          decltype(SB),
+          decltype(SC),
+          decltype(tiled_copy_a),
+          decltype(tiled_copy_b),
+          decltype(tiled_copy_c),
+          decltype(tiled_mma),
+          GM.value>;
+      configure_matmul(kernel, smem_size);
+
+      dim3 block(size(tiled_mma));
+      dim3 grid(
+          size(ceil_div(M, BM) * GM), size(ceil_div(ceil_div(N, BN), GM)));
+
+      enc.add_kernel_node(
+          kernel,
+          grid,
+          block,
+          smem_size,
+          a.data<bf16>(),
+          b.data<bf16>(),
+          out.data<bf16>(),
+          SA,
+          SB,
+          SC,
+          tiled_copy_a,
+          tiled_copy_b,
+          tiled_copy_c,
+          tiled_mma,
+          M,
+          N,
+          K);
+    } else {
+      throw std::runtime_error("Only bfloat16 supported");
+    }
+  });
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/cutlass_gemm.h

@@ -0,0 +1,18 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core::cu {
+
+void cutlass_gemm(
+    const array& a,
+    const array& b,
+    array& out,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc);
+
+}

mlx/backend/cuda/gemms/simple_gemm.cu

@@ -0,0 +1,69 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/steel/gemm.cuh"
+#include "mlx/dtype_utils.h"
+
+#include <iostream>
+
+namespace mlx::core::cu {
+
+namespace {
+
+template <typename Kernel>
+static void configure_smem(Kernel kernel, int SM) {
+  static bool done = false;
+  if (done) {
+    return;
+  }
+  std::cout << "configuring" << std::endl;
+  cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, SM);
+  cudaFuncSetAttribute(
+      kernel,
+      cudaFuncAttributePreferredSharedMemoryCarveout,
+      cudaSharedmemCarveoutMaxShared);
+  done = true;
+}
+
+} // namespace
+
+void simple_gemm(
+    const array& a,
+    const array& b,
+    array& out,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc) {
+  enc.set_input_array(a);
+  enc.set_input_array(b);
+  enc.set_output_array(out);
+  dispatch_float_types(a.dtype(), "simple_gemm", [&](auto type_tag) {
+    using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+    constexpr int BM = 128;
+    constexpr int BN = 128;
+    constexpr int BK = 32;
+    constexpr int PIPE = 3;
+    constexpr int SM = PIPE * sizeof(DataType) * (BM * BK + BN * BK);
+    constexpr int WM = 2;
+    constexpr int WN = 4;
+
+    auto kernel = ab_t_aligned<DataType, BM, BN, BK, WM, WN, PIPE>;
+    configure_smem(kernel, SM);
+
+    dim3 grid(N / BN, M / BM);
+    enc.add_kernel_node(
+        kernel,
+        grid,
+        WM * WN * WARP_SIZE,
+        SM,
+        a.data<DataType>(),
+        b.data<DataType>(),
+        out.data<DataType>(),
+        N,
+        K);
+  });
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/gemms/simple_gemm.h

@@ -0,0 +1,18 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core::cu {
+
+void simple_gemm(
+    const array& a,
+    const array& b,
+    array& out,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc);
+
+}

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;
@@ -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;
@@ -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

@@ -101,8 +101,8 @@ const std::filesystem::path& ptx_cache_dir() {
 bool read_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
-    std::vector<char>* ptx,
+    std::string& ptx,
-    std::vector<std::pair<std::string, std::string>>* ptx_kernels) {
+    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
   if (cache_dir.empty()) {
     return false;
   }
@@ -117,15 +117,15 @@ bool read_cached_ptx(
   if (!ptx_file.good()) {
     return false;
   }
-  ptx->resize(ptx_size);
+  ptx.resize(ptx_size);
-  ptx_file.read(ptx->data(), ptx_size);
+  ptx_file.read(ptx.data(), ptx_size);
 
   std::ifstream txt_file(cache_dir / (module_name + ".txt"), std::ios::binary);
   std::string line;
   while (std::getline(txt_file, line)) {
     auto tab = line.find('\t');
     if (tab != std::string::npos) {
-      ptx_kernels->emplace_back(line.substr(0, tab), line.substr(tab + 1));
+      ptx_kernels.emplace_back(line.substr(0, tab), line.substr(tab + 1));
     }
   }
   return true;
@@ -135,7 +135,7 @@ 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()) {
@@ -217,22 +217,18 @@ constexpr const char* g_headers[] = {
     jit_source_utils,
 };
 
-} // namespace
+void compile(
-
-JitModule::JitModule(
     Device& device,
     const std::string& module_name,
-    const KernelBuilder& builder) {
+    const std::string& source,
-  // Check cache.
+    const std::vector<std::string>& kernel_names,
-  std::vector<char> ptx;
+    std::string& ptx,
-  std::vector<std::pair<std::string, std::string>> ptx_kernels;
+    std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
-  if (!read_cached_ptx(ptx_cache_dir(), module_name, &ptx, &ptx_kernels)) {
+  // Create the program
-    // Create program.
-    auto [source_code, kernel_names] = builder();
   nvrtcProgram prog;
   CHECK_NVRTC_ERROR(nvrtcCreateProgram(
       &prog,
-        source_code.c_str(),
+      source.c_str(),
       (module_name + ".cu").c_str(),
       std::size(g_headers),
       g_headers,
@@ -286,16 +282,20 @@ JitModule::JitModule(
   } else {
     CHECK_NVRTC_ERROR(nvrtcGetPTXSize(prog, &ptx_size));
   }
-    ptx.resize(ptx_size, 0);
+  ptx.resize(ptx_size);
   if (use_sass) {
     CHECK_NVRTC_ERROR(nvrtcGetCUBIN(prog, ptx.data()));
   } else {
     CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
   }
-    write_cached_ptx(
-        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
 }
 
+void load_module(
+    const std::string& module_name,
+    const std::string& ptx,
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
+    CUmodule& module_,
+    std::unordered_map<std::string, std::pair<CUfunction, bool>>& kernels) {
   // Load module.
   char jit_log[4089] = {};
   CUjit_option options[] = {
@@ -312,21 +312,69 @@ 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 +385,12 @@ 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()>;
@@ -63,14 +64,16 @@ struct KernelArgs {
  private:
   std::vector<void*> args_;
 
-  // The cuLaunchKernel API requires passing pointers to arguments so store
+  // The cuGraphAddKernelNode API requires passing pointers to arguments so
-  // temporary values untill kernel is launched.
+  // 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 +85,19 @@ 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 +105,7 @@ 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/matmul.cpp

@@ -3,7 +3,9 @@
 #include "mlx/backend/common/matmul.h"
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/gemms/cublas_gemm.h"
+#include "mlx/backend/cuda/gemms/cutlass_gemm.h"
 #include "mlx/backend/cuda/gemms/gemv.h"
+#include "mlx/backend/cuda/gemms/simple_gemm.h"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/primitives.h"
 
@@ -11,8 +13,14 @@
 #include <numeric>
 
 namespace mlx::core {
+
 namespace {
 
+int get_test_gemm() {
+  static int t = env::get_var("MLX_ENABLE_TEST_GEMM", 0);
+  return t;
+}
+
 std::tuple<bool, int64_t, array>
 check_transpose(cu::CommandEncoder& enc, const Stream& s, const array& arr) {
   auto stx = arr.strides()[arr.ndim() - 2];
@@ -95,6 +103,18 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
+  if (M % 512 == 0 && N % 512 == 0 && K % 512 == 0 && !a_transposed &&
+      b_transposed && batch_count == 1 && get_test_gemm() == 1) {
+    cu::simple_gemm(a, b, out, M, N, K, encoder);
+    return;
+  }
+
+  if (M % 512 == 0 && N % 512 == 0 && K % 512 == 0 && !a_transposed &&
+      b_transposed && batch_count == 1 && get_test_gemm() == 2) {
+    cu::cutlass_gemm(a, b, out, M, N, K, encoder);
+    return;
+  }
+
   /////////////////////////////////////////////////////////////////////////////
   // Invoke cublasLt
   CublasGemm gemm(

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

@@ -41,11 +41,8 @@ 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/steel/gemm.cuh

@@ -4,95 +4,189 @@
 
 namespace mlx::core::cu {
 
+template <typename T, int BM, int BN, int BK, int WM, int WN>
+__device__ inline void gemm_ab_t(
+    RegisterTile<float, BM / WM, BN / WN>& C,
+    SharedTile<T, BM, BK>& As,
+    SharedTile<T, BN, BK>& Bs,
+    RegisterTileLoader<SharedTile<T, BM, BK>>& rloader_a,
+    RegisterTileLoader<SharedTile<T, BN, BK>>& rloader_b) {
+  RegisterTile<T, BM / WM, 16> A[2];
+  RegisterTile<T, BN / WN, 16> B[2];
+
+  rloader_a.load(A[0], As.base_addr(), 0);
+  rloader_b.load(B[0], Bs.base_addr(), 0);
+
+  MLX_UNROLL
+  for (int k = 1; k < BK / 16; k++) {
+    rloader_a.load(A[k & 1], As.base_addr(), k);
+    rloader_b.load(B[k & 1], Bs.base_addr(), k);
+
+    mma_t(C, A[(k - 1) & 1], B[(k - 1) & 1]);
+  }
+  mma_t(C, A[(BK / 16 - 1) & 1], B[(BK / 16 - 1) & 1]);
+}
+
 /**
  * An example gemm written with the utils.
  *
  * Computes A @ B.T when A and B are all aligned with the block sizes.
  */
-template <typename T, int BM, int BN, int BK>
+// template <typename T, int BM, int BN, int BK, int WM, int WN, int PIPE>
-__global__ void ab_t_aligned(const T* a, const T* b, T* y, int N, int K) {
+//__global__ __launch_bounds__(WM * WN * WARP_SIZE, 1)
-  constexpr int WARPS_M = 2;
+// void ab_t_aligned(const T* a, const T* b, T* y, int N, int K) {
-  constexpr int WARPS_N = 2;
+//   constexpr int NUM_WARPS = WM * WN;
-  constexpr int NUM_WARPS = WARPS_M * WARPS_N;
+//   constexpr int WARP_STEP_M = BM / WM;
-  constexpr int WARP_STEP_M = BM / WARPS_M;
+//   constexpr int WARP_STEP_N = BN / WN;
-  constexpr int WARP_STEP_N = BN / WARPS_N;
+//
+//   // Precompute some offsets for each thread
+//   const int warpid = threadIdx.x / 32;
+//   const int laneid = threadIdx.x % 32;
+//   const int wm = warpid / WN;
+//   const int wn = warpid % WN;
+//   const int offset_m = wm * WARP_STEP_M;
+//   const int offset_n = wn * WARP_STEP_N;
+//
+//   // Allocate shared memory
+//   extern __shared__ char shmem[];
+//   SharedTile<T, BM, BK>(&as)[PIPE] =
+//       *(SharedTile<T, BM, BK>(*)[PIPE])(&shmem[0]);
+//   SharedTile<T, BN, BK>(&bs)[PIPE] =
+//       *(SharedTile<T, BN, BK>(*)[PIPE])(&shmem[sizeof(T) * PIPE * BM * BK]);
+//
+//   // Move the global pointers to the tile
+//   a += blockIdx.y * BM * K;
+//   b += blockIdx.x * BN * K;
+//   y += blockIdx.y * BM * N + blockIdx.x * BN;
+//
+//   // Make the loaders to/from SMEM
+//   SharedTileLoader<NUM_WARPS, SharedTile<T, BM, BK>> sloader_a(a, K);
+//   SharedTileLoader<NUM_WARPS, SharedTile<T, BN, BK>> sloader_b(b, K);
+//   RegisterTileLoader<SharedTile<T, BM, BK>> rloader_a(offset_m, laneid);
+//   RegisterTileLoader<SharedTile<T, BN, BK>> rloader_b(offset_n, laneid);
+//
+//   // Start the SM pipeline
+//   MLX_UNROLL
+//   for (int i = 0; i < PIPE - 1; i++) {
+//     sloader_a.load_async(as[i].base_addr());
+//     sloader_b.load_async(bs[i].base_addr());
+//     cp_async_commit();
+//     sloader_a.next();
+//     sloader_b.next();
+//   }
+//
+//   // Allocate and zero the MMA accumulator
+//   RegisterTile<float, BM / WM, BN / WN> C;
+//   C.fill(0);
+//
+//   // Matmul loop
+//   int num_blocks = K / BK;
+//   int sread = 0;
+//   int swrite = PIPE - 1;
+//   for (int i = 0; i < num_blocks; i++) {
+//     cp_async_wait<PIPE - 1>();
+//
+//     gemm_ab_t<T, BM, BN, BK, WM, WN>(
+//         C, as[sread], bs[sread], rloader_a, rloader_b);
+//
+//     sloader_a.load_async(as[swrite].base_addr());
+//     sloader_b.load_async(bs[swrite].base_addr());
+//     cp_async_commit();
+//     sloader_a.next(i + PIPE < num_blocks);
+//     sloader_b.next(i + PIPE < num_blocks);
+//
+//     swrite = sread;
+//     sread = (sread + 1) % PIPE;
+//   }
+//
+//   C.store_global(y, N, offset_m, offset_n);
+// }
+
+template <typename T, int BM, int BN, int BK, int WM, int WN, int PIPE>
+__global__ __launch_bounds__(
+    WM* WN* WARP_SIZE,
+    1) void ab_t_aligned(const T* a, const T* b, T* y, int N, int K) {
+  constexpr int NUM_WARPS = WM * WN;
+  constexpr int WARP_STEP_M = BM / WM;
+  constexpr int WARP_STEP_N = BN / WN;
 
   // Precompute some offsets for each thread
   const int warpid = threadIdx.x / 32;
   const int laneid = threadIdx.x % 32;
-  const int wm = warpid / WARPS_N;
+  const int wm = warpid / WN;
-  const int wn = warpid % WARPS_N;
+  const int wn = warpid % WN;
   const int offset_m = wm * WARP_STEP_M;
   const int offset_n = wn * WARP_STEP_N;
 
   // Allocate shared memory
   extern __shared__ char shmem[];
-  SharedTile<T, BM, BK>(&as)[2] = *(SharedTile<T, BM, BK>(*)[2])(&shmem[0]);
+  SharedTile<T, BM, BK>(&as)[PIPE] =
-  SharedTile<T, BN, BK>(&bs)[2] =
+      *(SharedTile<T, BM, BK>(*)[PIPE])(&shmem[0]);
-      *(SharedTile<T, BN, BK>(*)[2])(&shmem[sizeof(T) * 2 * BM * BK]);
+  SharedTile<T, BN, BK>(&bs)[PIPE] =
-
+      *(SharedTile<T, BN, BK>(*)[PIPE])(&shmem[sizeof(T) * PIPE * BM * BK]);
-  // Allocate registers for the MMA
-  RegisterTile<float, BM / WARPS_M, BN / WARPS_N> C;
-  RegisterTile<T, BM / WARPS_M, 16> A;
-  RegisterTile<T, BN / WARPS_N, 16> B;
 
   // Move the global pointers to the tile
   a += blockIdx.y * BM * K;
   b += blockIdx.x * BN * K;
   y += blockIdx.y * BM * N + blockIdx.x * BN;
 
-  // Zero the accumulators
+  // Make the loaders to/from SMEM
-  C.fill(0);
+  using sloader = SharedTileLoader<NUM_WARPS, SharedTile<T, BM, BK>>;
+  constexpr int SSTEP = sloader::STEP_ROWS * sizeof(T) * BK;
+  const int srow = threadIdx.x / sloader::NUM_LOADS_PER_ROW;
+  const int scol =
+      (threadIdx.x % sloader::NUM_LOADS_PER_ROW) * sloader::ELEMENTS_PER_LOAD;
+  a += srow * K + scol;
+  b += srow * K + scol;
+  uint32_t sm_offsets[PIPE][2];
+  MLX_UNROLL
+  for (int s = 0; s < PIPE; s++) {
+    sm_offsets[s][0] = as[s].loc(as[s].base_addr(), srow, scol);
+    sm_offsets[s][1] = bs[s].loc(bs[s].base_addr(), srow, scol);
+  }
+  RegisterTileLoader<SharedTile<T, BM, BK>> rloader_a(offset_m, laneid);
+  RegisterTileLoader<SharedTile<T, BN, BK>> rloader_b(offset_n, laneid);
 
   // Start the SM pipeline
-  load_async<NUM_WARPS>(as[0], as[0].base_addr(), a, K);
-  load_async<NUM_WARPS>(bs[0], bs[0].base_addr(), b, K);
-  cp_async_commit();
-
-  int tic = 0;
-  for (int k_block = BK; k_block < K; k_block += BK) {
-    load_async<NUM_WARPS>(as[tic ^ 1], as[tic ^ 1].base_addr(), a + k_block, K);
-    load_async<NUM_WARPS>(bs[tic ^ 1], bs[tic ^ 1].base_addr(), b + k_block, K);
-    cp_async_commit();
-    cp_async_wait<1>();
-    __syncthreads();
-
   MLX_UNROLL
-    for (int k = 0; k < BK / 16; k++) {
+  for (int s = 0; s < PIPE - 1; s++) {
-      A.load(
+    MLX_UNROLL
-          as[tic],
+    for (int l = 0; l < sloader::NUM_LOADS_PER_THREAD; l++) {
-          as[tic].base_addr(),
+      cp_async<16>(sm_offsets[s][0] + l * SSTEP, a);
-          offset_m + laneid % 16,
+      cp_async<16>(sm_offsets[s][1] + l * SSTEP, b);
-          k * 16 + laneid / 16 * 8);
+      a += sloader::STEP_ROWS * K;
-      B.load(
+      b += sloader::STEP_ROWS * K;
-          bs[tic],
+    }
-          bs[tic].base_addr(),
+    cp_async_commit();
-          offset_n + laneid % 16,
-          k * 16 + laneid / 16 * 8);
-
-      mma_t(C, A, B);
   }
 
-    tic ^= 1;
+  // Allocate and zero the MMA accumulator
-  }
+  RegisterTile<float, BM / WM, BN / WN> C;
+  C.fill(0);
 
-  // Empty the pipeline
+  // Matmul loop
-  cp_async_wait_all();
+  int num_blocks = K / BK;
-  __syncthreads();
+  int sread = 0;
+  int swrite = PIPE - 1;
+  for (int i = 0; i < num_blocks; i++) {
+    cp_async_wait<PIPE - 1>();
+
+    gemm_ab_t<T, BM, BN, BK, WM, WN>(
+        C, as[sread], bs[sread], rloader_a, rloader_b);
+
+    if (false) {
       MLX_UNROLL
-  for (int k = 0; k < BK / 16; k++) {
+      for (int l = 0; l < sloader::NUM_LOADS_PER_THREAD; l++) {
-    A.load(
+        cp_async<16>(sm_offsets[swrite][0] + l * SSTEP, a);
-        as[tic],
+        cp_async<16>(sm_offsets[swrite][1] + l * SSTEP, b);
-        as[tic].base_addr(),
+        a += sloader::STEP_ROWS * K;
-        offset_m + laneid % 16,
+        b += sloader::STEP_ROWS * K;
-        k * 16 + laneid / 16 * 8);
+      }
-    B.load(
+    }
-        bs[tic],
+    cp_async_commit();
-        bs[tic].base_addr(),
-        offset_n + laneid % 16,
-        k * 16 + laneid / 16 * 8);
 
-    mma_t(C, A, B);
+    swrite = sread;
+    sread = (sread + 1) % PIPE;
   }
 
   C.store_global(y, N, offset_m, offset_n);

mlx/backend/cuda/steel/tiles.cuh

@@ -223,59 +223,10 @@ struct RegisterTile {
   }
 };
 
-/**
- * A simple container of multiple Tile16x16.
- *
- * Provides utility functions for loading and manipulating collections of basic
- * tiles.
- */
-template <typename T, int ROWS_, int COLS_>
-struct RegisterTile {
-  static constexpr int ROWS = ROWS_;
-  static constexpr int COLS = COLS_;
-  static constexpr int TILES_X = COLS / 16;
-  static constexpr int TILES_Y = ROWS / 16;
-
-  Tile16x16<T> data[TILES_X * TILES_Y];
-
-  __device__ inline void fill(T v) {
-    MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
-        data[i * TILES_X + j].fill(v);
-      }
-    }
-  }
-
-  template <typename Tile>
-  __device__ inline void
-  load(Tile& tile, uint32_t base_address, int row, int col) {
-    MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
-        data[i * TILES_X + j].load(
-            tile.loc(base_address, row + i * 16, col + j * 16));
-      }
-    }
-  }
-
-  template <typename U>
-  __device__ inline void store_global(U* x, int N, int row, int col) {
-    MLX_UNROLL
-    for (int i = 0; i < TILES_Y; i++) {
-      MLX_UNROLL
-      for (int j = 0; j < TILES_X; j++) {
-        data[i * TILES_X + j].store_global(
-            x + (row + i * 16) * N + col + j * 16, N);
-      }
-    }
-  }
-};
-
 template <typename T, int ROWS_, int COLS_>
 struct SharedTile {
+  using value_type = T;
+
   static constexpr int ROWS = ROWS_;
   static constexpr int COLS = COLS_;
   static constexpr int TILES_X = COLS / 16;
@@ -317,23 +268,26 @@ struct SharedTile {
     }
   }
 
-  // Return the location of the element at (row, col) using the swizzle.
+  __device__ static inline uint32_t offset(int row, int col) {
-  __device__ static inline uint32_t loc(uint32_t ptr, int row, int col) {
     if constexpr (swizzle_bytes > 0) {
       static constexpr int swizzle_repeat = swizzle_bytes * 8;
       static constexpr int subtile_cols = swizzle_bytes / sizeof(T);
       const int outer_idx = col / subtile_cols;
-      const uint32_t addr = ptr +
+      const uint32_t addr = sizeof(T) *
-          sizeof(T) *
           (outer_idx * ROWS * subtile_cols + row * subtile_cols +
            col % subtile_cols);
       const int swizzle = ((addr % swizzle_repeat) >> 7) << 4;
       return (addr ^ swizzle);
     } else {
-      return ptr + sizeof(T) * (row * COLS + col);
+      return sizeof(T) * (row * COLS + col);
     }
   }
 
+  // Return the location of the element at (row, col) using the swizzle.
+  __device__ static inline uint32_t loc(uint32_t ptr, int row, int col) {
+    return ptr + offset(row, col);
+  }
+
   // Convenience functions to edit elements going through the swizzle.
   __device__ inline T& operator()(int row, int col) {
     return *ptr(data, row, col);
@@ -364,6 +318,76 @@ struct SharedTile {
   }
 };
 
+template <int NUM_WARPS, typename Tile>
+struct SharedTileLoader {
+  using T = typename Tile::value_type;
+
+  static constexpr int NUM_THREADS = NUM_WARPS * 32;
+  static constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  static constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  static constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  static constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  static constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  const T* x_;
+  int N_;
+  uint32_t offset_;
+
+  __device__ SharedTileLoader(const T* x, int N) : x_(x), N_(N) {
+    const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+    const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+    x_ += row * N + col * ELEMENTS_PER_LOAD;
+    offset_ = Tile::offset(row, col * ELEMENTS_PER_LOAD);
+  }
+
+  __device__ inline void load_async(uint32_t base_address) {
+    MLX_UNROLL
+    for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+      cp_async<16>(
+          base_address + offset_ + i * STEP_ROWS * sizeof(T) * Tile::COLS,
+          x_ + i * STEP_ROWS * N_);
+    }
+  }
+
+  __device__ inline void next() {
+    x_ += Tile::COLS;
+  }
+};
+
+template <typename Tile>
+struct RegisterTileLoader {
+  using T = typename Tile::value_type;
+
+  uint32_t offset_[Tile::COLS / 16];
+
+  __device__ RegisterTileLoader(int offset_row, int laneid) {
+    const int row = offset_row + laneid & 15;
+    const int col = (laneid >> 4) << 3;
+
+    MLX_UNROLL
+    for (int i = 0; i < Tile::COLS / 16; i++) {
+      offset_[i] = Tile::offset(row, col + i * 16);
+    }
+  }
+
+  template <typename T, int ROWS, int COLS>
+  __device__ inline void
+  load(RegisterTile<T, ROWS, COLS>& x, uint32_t base_address, int col) {
+    constexpr int TILES_Y = RegisterTile<T, ROWS, COLS>::TILES_Y;
+    constexpr int TILES_X = RegisterTile<T, ROWS, COLS>::TILES_X;
+
+    MLX_UNROLL
+    for (int i = 0; i < TILES_Y; i++) {
+      MLX_UNROLL
+      for (int j = 0; j < TILES_X; j++) {
+        x.data[i * TILES_X + j].load(
+            base_address + offset_[j + col] + i * 16 * Tile::COLS * sizeof(T));
+      }
+    }
+  }
+};
+
 /**
  * Load the tile from global memory by loading 16 bytes at a time and storing
  * them immediately.

mlx/backend/cuda/steel/utils.cuh

@@ -21,15 +21,15 @@ __device__ inline void cp_async(uint32_t row_address, const T* x) {
 #if defined(MLX_CUDA_SM_80_ENABLED)
   if constexpr (N == 16) {
     asm volatile(
-        "cp.async.ca.shared::cta.global [%0], [%1], 16;\n" ::"r"(row_address),
+        "cp.async.cg.shared::cta.global [%0], [%1], 16;\n" ::"r"(row_address),
         "l"(reinterpret_cast<const int4*>(x)));
   } else if constexpr (N == 8) {
     asm volatile(
-        "cp.async.ca.shared::cta.global [%0], [%1], 8;\n" ::"r"(row_address),
+        "cp.async.cg.shared::cta.global [%0], [%1], 8;\n" ::"r"(row_address),
         "l"(reinterpret_cast<const int2*>(x)));
   } else if constexpr (N == 4) {
     asm volatile(
-        "cp.async.ca.shared::cta.global [%0], [%1], 4;\n" ::"r"(row_address),
+        "cp.async.cg.shared::cta.global [%0], [%1], 4;\n" ::"r"(row_address),
         "l"(reinterpret_cast<const int*>(x)));
   }
 #endif

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/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,
-    bool atomic_outputs) {
+    bool) {
-  throw std::runtime_error("[metal_kernel] No GPU back-end.");
+  throw std::runtime_error("[metal_kernel] No Metal back-end.");
 }
 
 } // namespace fast

mlx/distributed/CMakeLists.txt

@@ -6,4 +6,3 @@ 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

@@ -2,11 +2,9 @@
 
 #include <unordered_map>
 
-#include <iostream>
 #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 {
@@ -82,7 +80,7 @@ class EmptyGroup : public GroupImpl {
 } // namespace detail
 
 bool is_available() {
-  return mpi::is_available() || ring::is_available() || nccl::is_available();
+  return mpi::is_available() || ring::is_available();
 }
 
 int Group::rank() const {
@@ -113,8 +111,6 @@ Group init(bool strict /* = false */, const std::string& bk /* = "any" */) {
     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";

mlx/distributed/distributed.h

@@ -3,6 +3,7 @@
 #pragma once
 
 #include <memory>
+
 #include "mlx/array.h"
 
 namespace mlx::core::distributed {

mlx/distributed/nccl/CMakeLists.txt

@@ -1,8 +0,0 @@
-if(MLX_BUILD_CUDA)
-  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/nccl.cpp)
-  find_package(NCCL REQUIRED)
-  target_link_libraries(mlx PRIVATE ${NCCL_LIBRARIES})
-  target_include_directories(mlx PRIVATE ${NCCL_INCLUDE_DIRS})
-else()
-  target_sources(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/no_nccl.cpp)
-endif()

mlx/distributed/nccl/nccl.cpp

@@ -1,354 +0,0 @@
-#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"
-
-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;
-
-    for (attempt = 0; attempt < max_retries; ++attempt) {
-      if (connect(sock, reinterpret_cast<sockaddr*>(&serv), sizeof(serv)) ==
-          0) {
-        connected = true;
-        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;
-  }
-
-  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 {
-static std::string get_env_var_or_throw(const char* env_var_name) {
-  const char* value = std::getenv(env_var_name);
-  if (value == nullptr) {
-    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());
-  }
-  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");
-  std::string port = detail::get_env_var_or_throw("NCCL_PORT");
-  std::string rank_str = detail::get_env_var_or_throw("MLX_RANK");
-  std::string n_nodes_str = detail::get_env_var_or_throw("MLX_WORLD_SIZE");
-
-  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

@@ -1,12 +0,0 @@
-// 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/no_nccl.cpp

@@ -1,20 +0,0 @@
-// 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

@@ -2,20 +2,9 @@
 
 #include <sstream>
 
-#include "mlx/backend/cuda/cuda.h"
-#include "mlx/backend/metal/metal.h"
 #include "mlx/distributed/ops.h"
 #include "mlx/distributed/primitives.h"
 
-inline mlx::core::Device get_device() {
-  if (mlx::core::metal::is_available()) {
-    return mlx::core::Device::cpu;
-  } else if (mlx::core::cu::is_available()) {
-    return mlx::core::Device::gpu;
-  }
-  throw std::runtime_error("No available device for distributed operations.");
-}
-
 namespace mlx::core::distributed {
 
 namespace {
@@ -35,7 +24,6 @@ array all_sum(
     std::optional<Group> group_ /* = std::nullopt */,
     StreamOrDevice s /* = {} */) {
   auto group = to_group(group_);
-  auto dev = get_device();
 
   if (group.size() == 1) {
     return x;
@@ -43,7 +31,8 @@ array all_sum(
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(to_stream(s, dev), group, AllReduce::Sum),
+      std::make_shared<AllReduce>(
+          to_stream(s, Device::cpu), group, AllReduce::Sum),
       {x});
 }
 
@@ -52,7 +41,6 @@ array all_max(
     std::optional<Group> group_ /* = std::nullopt */,
     StreamOrDevice s /* = {} */) {
   auto group = to_group(group_);
-  auto dev = get_device();
 
   if (group.size() == 1) {
     return x;
@@ -60,7 +48,8 @@ array all_max(
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(to_stream(s, dev), group, AllReduce::Max),
+      std::make_shared<AllReduce>(
+          to_stream(s, Device::cpu), group, AllReduce::Max),
       {x});
 }
 
@@ -69,7 +58,6 @@ array all_min(
     std::optional<Group> group_ /* = std::nullopt */,
     StreamOrDevice s /* = {} */) {
   auto group = to_group(group_);
-  auto dev = get_device();
 
   if (group.size() == 1) {
     return x;
@@ -77,7 +65,8 @@ array all_min(
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<AllReduce>(to_stream(s, dev), group, AllReduce::Min),
+      std::make_shared<AllReduce>(
+          to_stream(s, Device::cpu), group, AllReduce::Min),
       {x});
 }
 
@@ -86,7 +75,6 @@ array all_gather(
     std::optional<Group> group_ /* = std::nullopt */,
     StreamOrDevice s /* = {} */) {
   auto group = to_group(group_);
-  auto dev = get_device();
 
   if (group.size() == 1) {
     return x;
@@ -101,7 +89,7 @@ array all_gather(
   return array(
       std::move(result_shape),
       x.dtype(),
-      std::make_shared<AllGather>(to_stream(s, dev), group),
+      std::make_shared<AllGather>(to_stream(s, Device::cpu), group),
       {x});
 }
 
@@ -111,7 +99,6 @@ array send(
     std::optional<Group> group_ /* = std::nullopt */,
     StreamOrDevice s /* = {} */) {
   auto group = to_group(group_);
-  auto dev = get_device();
 
   if (group.size() == 1) {
     throw std::invalid_argument("Cannot send to a singleton group");
@@ -127,7 +114,7 @@ array send(
   return array(
       x.shape(),
       x.dtype(),
-      std::make_shared<Send>(to_stream(s, dev), group, dst),
+      std::make_shared<Send>(to_stream(s, Device::cpu), group, dst),
       {x});
 }
 
@@ -138,7 +125,6 @@ array recv(
     std::optional<Group> group_ /* = std::nullopt */,
     StreamOrDevice s /* = {} */) {
   auto group = to_group(group_);
-  auto dev = get_device();
 
   if (group.size() == 1) {
     throw std::invalid_argument("Cannot recv from a singleton group");
@@ -153,7 +139,7 @@ array recv(
   return array(
       std::move(shape),
       std::move(dtype),
-      std::make_shared<Recv>(to_stream(s, dev), group, src),
+      std::make_shared<Recv>(to_stream(s, Device::cpu), group, src),
       std::vector<array>{});
 }
 

mlx/fast.h

@@ -66,9 +66,10 @@ array affine_dequantize(
     int bits = 4,
     StreamOrDevice s = {});
 
-typedef std::variant<int, bool, Dtype> TemplateArg;
+using TemplateArg = std::variant<int, bool, Dtype>;
+using ScalarArg = std::variant<bool, int, float>;
 
-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 +78,9 @@ typedef std::function<std::vector<array>(
     std::vector<std::pair<std::string, TemplateArg>>,
     std::optional<float>,
     bool,
-    StreamOrDevice)>
+    StreamOrDevice)>;
-    MetalKernelFunction;
 
-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 +89,27 @@ 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"
 
@@ -283,6 +284,8 @@ struct CustomKernelShapeInfo {
   bool ndim = false;
 };
 
+using ScalarArg = std::variant<bool, int, float>;
+
 class CustomKernel : public Primitive {
  public:
   CustomKernel(
@@ -293,7 +296,10 @@ 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 +307,14 @@ 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 +330,9 @@ 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

python/mlx/distributed_run.py

@@ -415,48 +415,6 @@ def launch_mpi(parser, hosts, args, command):
             pass
 
 
-def launch_nccl(parser, hosts, args, command):
-    master_host = hosts[0].ips[0]
-
-    if master_host != "127.0.0.1":
-        raise ValueError("The NCCL backend only supports localhost for now. ")
-    master_port = args.nccl_port
-    world_size = len(hosts)
-
-    base_env = os.environ.copy()
-    base_env.update(
-        {
-            "NCCL_DEBUG": "INFO",
-            "NCCL_SOCKET_IFNAME": "lo",  # Use loopback for local communication
-            "NCCL_HOST_IP": master_host,
-            "NCCL_PORT": str(master_port),
-            "MLX_WORLD_SIZE": str(world_size),
-        }
-    )
-    procs = []
-    try:
-        for rank in range(world_size):
-            env = base_env.copy()
-            env["MLX_RANK"] = str(rank)
-            env["CUDA_VISIBLE_DEVICES"] = str(rank % args.nproc_per_node)
-            p = Popen(command, env=env)
-            procs.append(p)
-
-        for p in procs:
-            ret = p.wait()
-            if ret != 0:
-                raise RuntimeError(f"Rank process exited with {ret}")
-
-    except (RuntimeError, KeyboardInterrupt) as err:
-        for p in procs:
-            if p.poll() is None:
-                try:
-                    p.kill()
-                except Exception:
-                    pass
-        raise
-
-
 def check_ssh_connections(hosts):
     results = [False] * len(hosts)
 
@@ -707,7 +665,7 @@ def distributed_config():
     )
     parser.add_argument(
         "--backend",
-        choices=["ring", "mpi", "nccl"],
+        choices=["ring", "mpi"],
         default="ring",
         help="Which distributed backend to configure",
     )
@@ -779,7 +737,7 @@ def main():
     parser.add_argument("--hostfile", help="The file containing the hosts")
     parser.add_argument(
         "--backend",
-        choices=["ring", "mpi", "nccl"],
+        choices=["ring", "mpi"],
         default="ring",
         help="Which distributed backend to launch",
     )
@@ -811,13 +769,6 @@ def main():
     parser.add_argument(
         "--cwd", help="Set the working directory on each node to the provided one"
     )
-    parser.add_argument(
-        "--nccl-port",
-        type=int,
-        default=12345,
-        help="The port to use for the NCCL communication (only for nccl backend)",
-    )
-
     args, rest = parser.parse_known_args()
     if rest[0] == "--":
         rest.pop(0)
@@ -848,10 +799,8 @@ def main():
     # Launch
     if args.backend == "ring":
         launch_ring(parser, hosts, args, rest)
-    if args.backend == "mpi":
+    elif args.backend == "mpi":
         launch_mpi(parser, hosts, args, rest)
-    if args.backend == "nccl":
-        launch_nccl(parser, hosts, args, rest)
 
 
 if __name__ == "__main__":

python/mlx/nn/utils.py

@@ -76,7 +76,6 @@ def average_gradients(
     group: Optional[mx.distributed.Group] = None,
     all_reduce_size: int = 32 * 1024**2,
     communication_type: Optional[mx.Dtype] = None,
-    stream: mx.Stream = mx.cpu,
 ):
     """Average the gradients across the distributed processes in the passed group.
 
@@ -95,7 +94,6 @@ def average_gradients(
         communication_type (Optional[mlx.core.Dtype]): If provided cast to this
             type before performing the communication. Typically cast to a
             smaller float to reduce the communication size. Default: ``None``.
-        stream (mlx.core.Stream): The stream to use for the reduction. Default: ``mlx.cpu``.
     """
     group = group or mx.distributed.init()
     N = group.size()
@@ -106,7 +104,7 @@ def average_gradients(
     def _average(x):
         dt = x.dtype
         x = x.astype(communication_type) if communication_type is not None else x
-        return mx.distributed.all_sum(x, stream=stream).astype(dt) / N
+        return mx.distributed.all_sum(x, stream=mx.cpu).astype(dt) / N
 
     if all_reduce_size <= 0:
         return tree_map(_average, gradients)

python/src/CMakeLists.txt

@@ -17,6 +17,7 @@ nanobind_add_module(
   ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/metal.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/cuda.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/memory.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/mlx_func.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/ops.cpp

python/src/cuda.cpp

@@ -0,0 +1,19 @@
+// Copyright © 2023-2025 Apple Inc.
+
+#include <nanobind/nanobind.h>
+
+#include "mlx/backend/cuda/cuda.h"
+
+namespace mx = mlx::core;
+namespace nb = nanobind;
+
+void init_cuda(nb::module_& m) {
+  nb::module_ cuda = m.def_submodule("cuda", "mlx.cuda");
+
+  cuda.def(
+      "is_available",
+      &mx::cu::is_available,
+      R"pbdoc(
+      Check if the CUDA back-end is available.
+      )pbdoc");
+}

python/src/distributed.cpp

@@ -79,7 +79,7 @@ void init_distributed(nb::module_& parent_module) {
             in case ``mx.distributed.is_available()`` returns False otherwise
             it throws a runtime error. Default: ``False``
           backend (str, optional): Which distributed backend to initialize.
-            Possible values ``mpi``, ``ring``, ``nccl``, ``any``. If set to ``any`` all
+            Possible values ``mpi``, ``ring``, ``any``. If set to ``any`` all
             available backends are tried and the first one that succeeds
             becomes the global group which will be returned in subsequent
             calls. Default: ``any``

python/src/fast.cpp

@@ -17,6 +17,66 @@ namespace mx = mlx::core;
 namespace nb = nanobind;
 using namespace nb::literals;
 
+namespace {
+
+struct PyCustomKernelFunction {
+  PyCustomKernelFunction(mx::fast::CustomKernelFunction kernel, const char* tag)
+      : kernel_(std::move(kernel)), tag_(tag) {}
+
+  std::vector<mx::array> operator()(
+      const std::vector<ScalarOrArray>& inputs_,
+      const std::vector<mx::Shape>& output_shapes,
+      const std::vector<mx::Dtype>& output_dtypes,
+      std::tuple<int, int, int> grid,
+      std::tuple<int, int, int> threadgroup,
+      const std::optional<std::vector<std::pair<std::string, nb::object>>>&
+          template_args_ = std::nullopt,
+      std::optional<float> init_value = std::nullopt,
+      bool verbose = false,
+      mx::StreamOrDevice s = {}) const {
+    std::vector<mx::array> inputs;
+    for (const auto& value : inputs_) {
+      inputs.push_back(to_array(value, std::nullopt));
+    }
+    std::vector<std::pair<std::string, mx::fast::TemplateArg>> template_args;
+    if (template_args_) {
+      for (const auto& [name, value] : template_args_.value()) {
+        // Handle bool, int and dtype template args
+        if (nb::isinstance<bool>(value)) {
+          bool bool_val = nb::cast<bool>(value);
+          template_args.emplace_back(name, bool_val);
+        } else if (nb::isinstance<int>(value)) {
+          int int_val = nb::cast<int>(value);
+          template_args.emplace_back(name, int_val);
+        } else if (nb::isinstance<mx::Dtype>(value)) {
+          mx::Dtype dtype = nb::cast<mx::Dtype>(value);
+          template_args.emplace_back(name, dtype);
+        } else {
+          std::ostringstream msg;
+          msg << tag_
+              << " Invalid template argument. Must be `mlx.core.Dtype`, `int` or `bool`.";
+          throw std::invalid_argument(msg.str());
+        }
+      }
+    }
+    return kernel_(
+        inputs,
+        output_shapes,
+        output_dtypes,
+        grid,
+        threadgroup,
+        template_args,
+        init_value,
+        verbose,
+        s);
+  }
+
+  mx::fast::CustomKernelFunction kernel_;
+  const char* tag_;
+};
+
+} // namespace
+
 void init_fast(nb::module_& parent_module) {
   auto m =
       parent_module.def_submodule("fast", "mlx.core.fast: fast operations");
@@ -240,53 +300,7 @@ void init_fast(nb::module_& parent_module) {
             ensure_row_contiguous,
             atomic_outputs);
         return nb::cpp_function(
-            [kernel = std::move(kernel)](
+            PyCustomKernelFunction(std::move(kernel), "[metal_kernel]"),
-                const std::vector<ScalarOrArray>& inputs_,
-                const std::vector<mx::Shape>& output_shapes,
-                const std::vector<mx::Dtype>& output_dtypes,
-                std::tuple<int, int, int> grid,
-                std::tuple<int, int, int> threadgroup,
-                const std::optional<
-                    std::vector<std::pair<std::string, nb::object>>>&
-                    template_args_ = std::nullopt,
-                std::optional<float> init_value = std::nullopt,
-                bool verbose = false,
-                mx::StreamOrDevice s = {}) {
-              std::vector<mx::array> inputs;
-              for (const auto& value : inputs_) {
-                inputs.push_back(to_array(value, std::nullopt));
-              }
-              std::vector<std::pair<std::string, mx::fast::TemplateArg>>
-                  template_args;
-              if (template_args_) {
-                for (const auto& [name, value] : template_args_.value()) {
-                  // Handle bool, int and dtype template args
-                  if (nb::isinstance<bool>(value)) {
-                    bool bool_val = nb::cast<bool>(value);
-                    template_args.emplace_back(name, bool_val);
-                  } else if (nb::isinstance<int>(value)) {
-                    int int_val = nb::cast<int>(value);
-                    template_args.emplace_back(name, int_val);
-                  } else if (nb::isinstance<mx::Dtype>(value)) {
-                    mx::Dtype dtype = nb::cast<mx::Dtype>(value);
-                    template_args.emplace_back(name, dtype);
-                  } else {
-                    throw std::invalid_argument(
-                        "[metal_kernel] Invalid template argument. Must be `mlx.core.Dtype`, `int` or `bool`.");
-                  }
-                }
-              }
-              return kernel(
-                  inputs,
-                  output_shapes,
-                  output_dtypes,
-                  grid,
-                  threadgroup,
-                  template_args,
-                  init_value,
-                  verbose,
-                  s);
-            },
             nb::kw_only(),
             "inputs"_a,
             "output_shapes"_a,
@@ -384,4 +398,216 @@ void init_fast(nb::module_& parent_module) {
           b = exp_elementwise(a)
           assert mx.allclose(b, mx.exp(a))
      )pbdoc");
+
+  m.def(
+      "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_mem) {
+        auto kernel = mx::fast::cuda_kernel(
+            name,
+            input_names,
+            output_names,
+            source,
+            header,
+            ensure_row_contiguous,
+            shared_mem);
+        return nb::cpp_function(
+            PyCustomKernelFunction(std::move(kernel), "[cuda_kernel]"),
+            nb::kw_only(),
+            "inputs"_a,
+            "output_shapes"_a,
+            "output_dtypes"_a,
+            "grid"_a,
+            "threadgroup"_a,
+            "template"_a = nb::none(),
+            "init_value"_a = nb::none(),
+            "verbose"_a = false,
+            "stream"_a = nb::none(),
+            nb::sig(
+                "def __call__(self, *, inputs: List[Union[scalar, array]], output_shapes: List[Sequence[int]], output_dtypes: List[Dtype], grid: tuple[int, int, int], threadgroup: tuple[int, int, int], template: Optional[List[Tuple[str, Union[bool, int, Dtype]]]] = None, init_value: Optional[float] = None, verbose: bool = false, stream: Union[None, Stream, Device] = None)"),
+            R"pbdoc(
+            Run the kernel.
+
+            Args:
+              inputs (List[array]): The inputs passed to the CUDA kernel.
+              output_shapes (List[Sequence[int]]): The list of shapes for each output in ``output_names``.
+              output_dtypes (List[Dtype]): The list of data types for each output in ``output_names``.
+              grid (tuple[int, int, int]): 3-tuple specifying the grid to launch the kernel with.
+                For compatibility with :func:`metal_kernel` the grid is in threads and not in threadgroups.
+              threadgroup (tuple[int, int, int]): 3-tuple specifying the threadgroup size to use.
+              template (List[Tuple[str, Union[bool, int, Dtype]]], optional): Template arguments.
+                  These will be added as template arguments to the kernel definition. Default: ``None``.
+              init_value (float, optional): Optional value to use to initialize all of the output arrays.
+                  By default, output arrays are uninitialized. Default: ``None``.
+              verbose (bool, optional): Whether to print the full generated source code of the kernel
+                  when it is run. Default: ``False``.
+              stream (mx.stream, optional): Stream to run the kernel on. Default: ``None``.
+
+            Returns:
+              List[array]: The list of output arrays.)pbdoc");
+      },
+      "name"_a,
+      "input_names"_a,
+      "output_names"_a,
+      "source"_a,
+      "header"_a = "",
+      "ensure_row_contiguous"_a = true,
+      "shared_memory"_a = 0,
+      R"pbdoc(
+      A jit-compiled custom CUDA kernel defined from a source string.
+
+      This is the CUDA equivalent of :ref:`custom_metal_kernels`.
+
+      Args:
+        name (str): Name for the kernel.
+        input_names (List[str]): The parameter names of the inputs in the
+           function signature.
+        output_names (List[str]): The parameter names of the outputs in the
+           function signature.
+        source (str): Source code. This is the body of a function in CUDA,
+           the function signature will be automatically generated.
+        header (str): Header source code to include before the main function.
+           Useful for helper functions or includes that should live outside of
+           the main function body.
+        ensure_row_contiguous (bool): Whether to ensure the inputs are row contiguous
+           before the kernel runs. Default: ``True``.
+        shared_memory (int): The dynamic shared memory to request for the
+          kernel. A value of 0 means no dynamic shared memory. Default: ``0``.
+
+      Returns:
+        Callable ``cuda_kernel``.
+
+      Example:
+
+        .. code-block:: python
+
+          def exp_elementwise(a: mx.array):
+              source = '''
+                  auto elem = cooperative_groups::this_grid().thread_rank();
+                  T tmp = inp[elem];
+                  out[elem] = exp(tmp);
+              '''
+
+              kernel = mx.fast.cuda_kernel(
+                  name="myexp",
+                  input_names=["inp"],
+                  output_names=["out"],
+                  source=source
+              )
+              outputs = kernel(
+                  inputs=[a],
+                  template=[("T", mx.float32)],
+                  grid=(a.size, 1, 1),
+                  threadgroup=(256, 1, 1),
+                  output_shapes=[a.shape],
+                  output_dtypes=[a.dtype],
+                  verbose=True,
+              )
+              return outputs[0]
+
+          a = mx.random.normal(shape=(16, 16)).astype(mx.float16)
+          b = exp_elementwise(a)
+          assert mx.allclose(b, mx.exp(a))
+     )pbdoc");
+
+  m.def(
+      "precompiled_cuda_kernel",
+      [](const std::string& name,
+         const nb::bytes compiled_source,
+         const std::vector<ScalarOrArray>& inputs_,
+         const std::vector<mx::Shape>& output_shapes,
+         const std::vector<mx::Dtype>& output_dtypes,
+         const std::vector<nb::object>& scalars_,
+         std::tuple<int, int, int> grid,
+         std::tuple<int, int, int> threadgroup,
+         int shared_memory,
+         std::optional<float> init_value = std::nullopt,
+         bool ensure_row_contiguous = false,
+         mx::StreamOrDevice s = {}) {
+        // Collect the inputs and cast them to array
+        std::vector<mx::array> inputs;
+        for (const auto& value : inputs_) {
+          inputs.push_back(to_array(value, std::nullopt));
+        }
+
+        // Collect the scalar inputs
+        std::vector<mx::fast::ScalarArg> scalars;
+        scalars.reserve(scalars_.size());
+        for (const auto& v : scalars_) {
+          if (nb::isinstance<bool>(v)) {
+            scalars.push_back(nb::cast<bool>(v));
+          } else if (nb::isinstance<int>(v)) {
+            scalars.push_back(nb::cast<int>(v));
+          } else if (nb::isinstance<float>(v)) {
+            scalars.push_back(nb::cast<float>(v));
+          } else {
+            nb::object vtype = v.attr("__class__");
+            std::string vtype_name =
+                nb::cast<std::string>(vtype.attr("__name__"));
+            std::ostringstream msg;
+            msg << "[precompiled_cuda_kernel] Invalid scalar argument type. "
+                << "Received " << vtype_name
+                << " but must be one of bool, int or float";
+            throw std::invalid_argument(msg.str());
+          }
+        }
+
+        return mx::fast::precompiled_cuda_kernel(
+            name,
+            std::string(
+                static_cast<const char*>(compiled_source.data()),
+                compiled_source.size()),
+            inputs,
+            output_shapes,
+            output_dtypes,
+            scalars,
+            grid,
+            threadgroup,
+            shared_memory,
+            init_value,
+            ensure_row_contiguous,
+            s);
+      },
+      nb::kw_only(),
+      "name"_a,
+      "compiled_source"_a,
+      "inputs"_a,
+      "output_shapes"_a,
+      "output_dtypes"_a,
+      "scalars"_a,
+      "grid"_a,
+      "threadgroup"_a,
+      "shared_memory"_a = 0,
+      "init_value"_a = nb::none(),
+      "ensure_row_contiguous"_a = false,
+      "stream"_a = nb::none(),
+      R"pbdoc(
+      Run a precompiled CUDA kernel defined from PTX or cubin.
+
+      This op is still experimental and various parts of the API may change.
+
+      Args:
+        name (str): Name for the kernel
+        compiled_source (bytes): The precompiled kernel in raw bytes.
+        inputs (List[array]): The inputs passed to the CUDA kernel.
+        output_shapes (List[Sequence[int]]): The list of shapes for each output.
+        output_dtypes (List[Dtype]): The list of data types for each output.
+        scalars (List[Union[bool, int, float]]): A list of scalar arguments to
+          pass to the kernel.
+        grid (tuple[int, int, int]): 3-tuple specifying the grid to launch the kernel with.
+          For compatibility with :func:`metal_kernel` the grid is in threads and not in threadblocks.
+        threadgroup (tuple[int, int, int]): 3-tuple specifying the threadgroup size to use.
+        shared_memory (int): The dynamic shared memory to request for the
+          kernel. A value of 0 means no dynamic shared memory. Default: ``0``.
+        init_value (float, optional): Optional value to use to initialize all of the output arrays.
+            By default, output arrays are uninitialized. Default: ``None``.
+        ensure_row_contiguous (bool): Whether to ensure the inputs are row contiguous
+           before the kernel runs. Default: ``False``.
+        stream (mx.stream, optional): Stream to run the kernel on. Default: ``None``.
+      )pbdoc");
 }

python/src/mlx.cpp

@@ -12,6 +12,7 @@ void init_array(nb::module_&);
 void init_device(nb::module_&);
 void init_stream(nb::module_&);
 void init_metal(nb::module_&);
+void init_cuda(nb::module_&);
 void init_memory(nb::module_&);
 void init_ops(nb::module_&);
 void init_transforms(nb::module_&);
@@ -35,6 +36,7 @@ NB_MODULE(core, m) {
   init_stream(m);
   init_array(m);
   init_metal(m);
+  init_cuda(m);
   init_memory(m);
   init_ops(m);
   init_transforms(m);

python/tests/cuda_skip.py

@@ -13,8 +13,6 @@ cuda_skip = {
     # Hadamard NYI
     "TestOps.test_hadamard",
     "TestOps.test_hadamard_grad_vmap",
-    # Convolutions NYI
-    "TestConv.test_1d_conv_with_2d",
     # FFTs NYI
     "TestFFT.test_fft",
     "TestFFT.test_fft_big_powers_of_two",

python/tests/nccl_test_distributed.py

@@ -1,284 +0,0 @@
-# Copyright © 2024 Apple Inc.
-import mlx.core as mx
-import mlx.nn as nn
-import mlx_tests
-from mlx.nn.layers.distributed import shard_inplace, shard_linear
-from mlx.nn.utils import average_gradients
-
-
-class TestNCCLDistributed(mlx_tests.MLXTestCase):
-    @classmethod
-    def setUpClass(cls):
-        world = mx.distributed.init(strict=True, backend="nccl")
-        rank = world.rank()
-        mx.set_default_device(mx.Device(mx.gpu, rank % 8))
-
-    def test_all_reduce(self):
-        world = mx.distributed.init()
-        dtypes = [
-            (mx.int8, 0),
-            (mx.uint8, 0),
-            (mx.int32, 0),
-            (mx.uint32, 0),
-            (mx.float32, 1e-6),
-            (mx.float16, 5e-3),
-            (mx.bfloat16, 1e-1),
-        ]
-        sizes = [
-            (7,),
-            (10,),
-            (1024,),
-            (1024, 1024),
-        ]
-        key = mx.random.key(0)
-
-        for dt, rtol in dtypes:
-            for sh in sizes:
-                x = (
-                    mx.random.uniform(shape=(world.size(),) + sh, key=key) * 10
-                ).astype(dt)
-
-                # All sum
-                y = mx.distributed.all_sum(x[world.rank()])
-                z = x.sum(0)
-                maxrelerror = (y - z).abs()
-                if rtol > 0:
-                    maxrelerror /= z.abs()
-                maxrelerror = maxrelerror.max()
-                self.assertLessEqual(maxrelerror, rtol)
-
-    def test_average_gradients(self):
-        original_all_sum = mx.distributed.all_sum
-        n_calls = 0
-        xtype = None
-
-        def new_all_sum(x, **kwargs):
-            nonlocal n_calls
-            nonlocal xtype
-
-            n_calls += 1
-            if xtype is not None:
-                self.assertEqual(xtype, x.dtype)
-
-            return original_all_sum(x, **kwargs)
-
-        mx.distributed.all_sum = new_all_sum
-        try:
-            grads = [mx.ones(10) for i in range(10)]
-            new_grads = average_gradients(grads, stream=mx.gpu)
-            mx.eval(new_grads)
-            self.assertEqual(len(new_grads), 10)
-            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
-            self.assertEqual(n_calls, 1)
-
-            n_calls = 0
-            new_grads = average_gradients(grads, all_reduce_size=4 * 50, stream=mx.gpu)
-            mx.eval(new_grads)
-            self.assertEqual(len(new_grads), 10)
-            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
-            self.assertEqual(n_calls, 2)
-
-            n_calls = 0
-            new_grads = average_gradients(grads, all_reduce_size=0, stream=mx.gpu)
-            mx.eval(new_grads)
-            self.assertEqual(len(new_grads), 10)
-            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
-            self.assertEqual(n_calls, 10)
-
-            n_calls = 0
-            xtype = mx.float16
-            new_grads = average_gradients(
-                grads,
-                all_reduce_size=2 * 50,
-                communication_type=mx.float16,
-                stream=mx.gpu,
-            )
-            mx.eval(new_grads)
-            self.assertEqual(len(new_grads), 10)
-            self.assertTrue(all(g.dtype == mx.float32 for g in new_grads))
-            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
-            self.assertEqual(n_calls, 2)
-
-        finally:
-            mx.distributed.all_sum = original_all_sum
-
-    def test_donation(self):
-        x = mx.random.normal((1024,))
-        mx.eval(x)
-        mx.synchronize()
-
-        mx.reset_peak_memory()
-        scale = mx.array(2.0)
-        y = mx.distributed.all_sum(x)
-        mx.eval(y)
-        mx.synchronize()
-        all_sum_only = mx.get_peak_memory()
-        y = mx.distributed.all_sum(x) * scale
-        mx.eval(y)
-        mx.synchronize()
-        all_sum_with_binary = mx.get_peak_memory()
-
-        self.assertEqual(all_sum_only, all_sum_with_binary)
-
-    def test_shard_linear(self):
-        # Seed the prng to have the same inputs and weights generated everywhere
-        mx.random.seed(0xF0F0F0F0)
-
-        # Prepare inputs
-        world = mx.distributed.init()
-        part = (
-            slice(None),
-            slice(
-                world.rank() * 1024 // world.size(),
-                (world.rank() + 1) * 1024 // world.size(),
-            ),
-        )
-        x = mx.random.normal((4, 1024))
-
-        # Create and shard some linear layers
-        lin = nn.Linear(1024, 1024, bias=True)
-        slin1 = shard_linear(lin, "all-to-sharded")
-        slin2 = shard_linear(lin, "sharded-to-all")
-        y = lin(x)
-        y1 = slin1(x)
-        y2 = slin2(x[part])
-        self.assertTrue(mx.allclose(y, y2, atol=1e-4, rtol=1e-4))
-        self.assertTrue(mx.allclose(y[part], y1, atol=1e-4, rtol=1e-4))
-
-        # Check the backward works as expected
-        def dummy_loss(model, x, y):
-            return (model(x) * y).sum()
-
-        mod = nn.Sequential(
-            nn.Linear(128, 128),
-            nn.Linear(128, 128),
-            nn.Linear(128, 128),
-            nn.Linear(128, 128),
-        )
-        smod = nn.Sequential(
-            shard_linear(mod.layers[0], "all-to-sharded"),
-            shard_linear(mod.layers[1], "sharded-to-all"),
-            shard_linear(mod.layers[2], "all-to-sharded"),
-            shard_linear(mod.layers[3], "sharded-to-all"),
-        )
-
-        grad1 = nn.value_and_grad(mod, dummy_loss)
-        grad2 = nn.value_and_grad(smod, dummy_loss)
-
-        x = mx.random.normal((4, 128))
-        y = mx.random.normal((4, 128))
-
-        l1, g1 = grad1(mod, x, y)
-        l2, g2 = grad2(smod, x, y)
-        mx.eval(l1, g1, l2, g2)
-
-        part = slice(
-            world.rank() * 128 // world.size(), (world.rank() + 1) * 128 // world.size()
-        )
-
-        self.assertTrue(mx.allclose(l1, l2))
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][0]["weight"][part],
-                g2["layers"][0]["weight"],
-                atol=1e-6,
-                rtol=1e-4,
-            )
-        )
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][2]["weight"][part],
-                g2["layers"][2]["weight"],
-                atol=1e-6,
-                rtol=1e-4,
-            )
-        )
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][1]["weight"][:, part],
-                g2["layers"][1]["weight"],
-                atol=1e-6,
-                rtol=1e-4,
-            )
-        )
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][3]["weight"][:, part],
-                g2["layers"][3]["weight"],
-                atol=1e-6,
-                rtol=1e-4,
-            )
-        )
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][0]["bias"][part],
-                g2["layers"][0]["bias"],
-                atol=1e-6,
-                rtol=1e-4,
-            )
-        )
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][2]["bias"][part],
-                g2["layers"][2]["bias"],
-                atol=1e-6,
-                rtol=1e-4,
-            )
-        )
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][1]["bias"], g2["layers"][1]["bias"], atol=1e-6, rtol=1e-4
-            )
-        )
-        self.assertTrue(
-            mx.allclose(
-                g1["layers"][3]["bias"], g2["layers"][3]["bias"], atol=1e-6, rtol=1e-4
-            )
-        )
-
-    def test_shard_predicate(self):
-        mx.random.seed(0xF0F0F0F0)
-
-        class MyConv(nn.Module):
-            def __init__(self, *args, **kwargs):
-                super().__init__()
-                self.aggregate = kwargs.pop("aggregate", False)
-                self.conv = nn.Conv2d(*args, **kwargs)
-
-            def __call__(self, x):
-                x = self.conv(x)
-                if self.aggregate:
-                    x = mx.distributed.all_sum(x)
-                return x
-
-        def sharding(path, weight):
-            parts = path.split(".")
-            even = int(parts[1]) % 2 == 0
-            if even:
-                return 0
-            else:
-                return -1 if parts[-1] != "bias" else None
-
-        mod = nn.Sequential(
-            MyConv(3, 128, kernel_size=3),
-            MyConv(128, 128, kernel_size=3),
-            MyConv(128, 128, kernel_size=3),
-            MyConv(128, 3, kernel_size=3),
-        )
-        smod = nn.Sequential(
-            MyConv(3, 128, kernel_size=3),
-            MyConv(128, 128, kernel_size=3, aggregate=True),
-            MyConv(128, 128, kernel_size=3),
-            MyConv(128, 3, kernel_size=3, aggregate=True),
-        )
-        smod.update(mod.parameters())
-        shard_inplace(smod, sharding)
-
-        x = mx.random.normal((4, 16, 16, 3))
-        y1 = mod(x)
-        y2 = smod(x)
-        self.assertTrue(mx.allclose(y1, y2, atol=1e-6, rtol=1e-4))
-
-
-if __name__ == "__main__":
-    mlx_tests.MLXTestRunner()

python/tests/test_conv.py

@@ -1186,6 +1186,13 @@ class TestConv(mlx_tests.MLXTestCase):
         y_hat = mx.conv2d(x, w)
         self.assertTrue(mx.allclose(y, y_hat))
 
+    def test_conv2d_large_filter_small_channels(self):
+        x = mx.random.normal(shape=(1, 181, 181, 1))
+        w = mx.random.normal(shape=(1, 182, 182, 1))
+        y = mx.conv2d(x, w, (1, 1), (1, 1), stream=mx.cpu)
+        y_hat = mx.conv2d(x, w, (1, 1), (1, 1))
+        self.assertTrue(mx.allclose(y, y_hat, rtol=1e-3, atol=1e-3))
+
 
 if __name__ == "__main__":
     mlx_tests.MLXTestRunner()

python/tests/test_fast.py

@@ -581,18 +581,28 @@ class TestFast(mlx_tests.MLXTestCase):
         )(x)
         self.assertTrue(mx.allclose(vmap_out, vmap_fast_out))
 
-    @unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
+    @unittest.skipIf(not mx.is_available(mx.gpu), "No GPU available")
     def test_custom_kernel_basic(self):
-        mx.random.seed(7)
+        if mx.metal.is_available():
-        a = mx.random.normal(shape=(2, 2))
-        kernel = mx.fast.metal_kernel(
-            name="basic",
-            input_names=["a"],
-            output_names=["out1"],
             source = """
                 uint elem = thread_position_in_grid.x;
                 out1[elem] = a[elem];
-            """,
+            """
+            custom_kernel = mx.fast.metal_kernel
+        elif mx.cuda.is_available():
+            source = """
+                auto elem = cooperative_groups::this_grid().thread_rank();
+                out1[elem] = a[elem];
+            """
+            custom_kernel = mx.fast.cuda_kernel
+
+        mx.random.seed(7)
+        a = mx.random.normal(shape=(2, 2))
+        kernel = custom_kernel(
+            name="basic",
+            input_names=["a"],
+            output_names=["out1"],
+            source=source,
         )
         out = kernel(
             inputs=[a],
@@ -604,16 +614,9 @@ class TestFast(mlx_tests.MLXTestCase):
         )
         self.assertTrue(mx.allclose(out[0], a))
 
-    @unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
+    @unittest.skipIf(not mx.is_available(mx.gpu), "No GPU available")
     def test_custom_kernel_args(self):
-        mx.random.seed(7)
+        if mx.metal.is_available():
-        a = mx.random.normal(shape=(3, 6))
-        c = mx.random.normal(shape=(2, 2)).astype(mx.bfloat16)
-
-        kernel = mx.fast.metal_kernel(
-            name="arg_test",
-            input_names=["a", "b", "c", "d"],
-            output_names=["out1", "out2"],
             source = """
                 uint elem = thread_position_in_grid.x;
                 T tmp = a[0];
@@ -623,7 +626,30 @@ class TestFast(mlx_tests.MLXTestCase):
                     out1[elem] = 1;
                 }
                 out2[elem] = a[1] + b[2] + c[1] - d;
-            """,
+            """
+            custom_kernel = mx.fast.metal_kernel
+        elif mx.cuda.is_available():
+            source = """
+                auto elem = cooperative_groups::this_grid().thread_rank();
+                T tmp = a[0];
+                if (e) {
+                    out1[elem] = a[1] + b[2] + static_cast<float>(c[3]) + d[0] + f;
+                } else {
+                    out1[elem] = 1;
+                }
+                out2[elem] = a[1] + b[2] + static_cast<float>(c[1]) - d[0];
+            """
+            custom_kernel = mx.fast.cuda_kernel
+
+        mx.random.seed(7)
+        a = mx.random.normal(shape=(3, 6))
+        c = mx.random.normal(shape=(2, 2)).astype(mx.bfloat16)
+
+        kernel = custom_kernel(
+            name="arg_test",
+            input_names=["a", "b", "c", "d"],
+            output_names=["out1", "out2"],
+            source=source,
         )
         out = kernel(
             inputs=[
@@ -647,10 +673,9 @@ class TestFast(mlx_tests.MLXTestCase):
         self.assertTrue(mx.allclose(out[0], mx.full((3, 2), 14.0484)))
         self.assertTrue(mx.allclose(out[1], mx.full((3, 2), -2, dtype=mx.int32)))
 
-    @unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
+    @unittest.skipIf(not mx.is_available(mx.gpu), "No GPU available")
     def test_custom_kernel_strides(self):
-        mx.random.seed(7)
+        if mx.metal.is_available():
-        a = mx.random.normal(shape=(3, 6))
             source = """
                 uint elem = thread_position_in_grid.x;
                 uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
@@ -662,12 +687,29 @@ class TestFast(mlx_tests.MLXTestCase):
                 T tmp = inp[elem];
                 out[elem] = metal::precise::exp(tmp) * threads_per_simdgroup;
             """
+            custom_kernel = mx.fast.metal_kernel
+        elif mx.cuda.is_available():
+            source = """
+                auto elem = cooperative_groups::this_grid().thread_rank();
+                auto loc = elem_to_loc(elem, inp_shape.data(), inp_strides.data(), inp_ndim);
+                T tmp = inp[loc];
+                out[elem] = exp(tmp) * WARP_SIZE;
+            """
+            source_contig = """
+                auto elem = cooperative_groups::this_grid().thread_rank();
+                T tmp = inp[elem];
+                out[elem] = exp(tmp) * WARP_SIZE;
+            """
+            custom_kernel = mx.fast.cuda_kernel
+
+        mx.random.seed(7)
+        a = mx.random.normal(shape=(3, 6))
 
         # non contiguous
         a = mx.tile(a[::2], [4, 1])
 
         for contig in [True, False]:
-            kernel = mx.fast.metal_kernel(
+            kernel = custom_kernel(
                 name="myexp" + str(contig),
                 input_names=["inp"],
                 output_names=["out"],
@@ -685,24 +727,41 @@ class TestFast(mlx_tests.MLXTestCase):
             )
             self.assertTrue(mx.allclose(mx.exp(a) * 32, outputs[0]))
 
-    @unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
+    @unittest.skipIf(not mx.is_available(mx.gpu), "No GPU available")
     def test_custom_kernel_helper(self):
-        mx.random.seed(7)
+        if mx.metal.is_available():
-        a = mx.random.normal(shape=(2, 2))
-        kernel = mx.fast.metal_kernel(
-            name="helper",
-            input_names=["a"],
-            output_names=["out1"],
             header = """
             template <typename T>
             T do_exp(T x) {
                 return metal::precise::exp(x);
             }
-            """,
+            """
             source = """
                 uint elem = thread_position_in_grid.x;
                 out1[elem] = do_exp(a[elem]);
-            """,
+            """
+            custom_kernel = mx.fast.metal_kernel
+        elif mx.cuda.is_available():
+            header = """
+            template <typename T>
+            __device__ T do_exp(T x) {
+                return exp(x);
+            }
+            """
+            source = """
+                auto elem = cooperative_groups::this_grid().thread_rank();
+                out1[elem] = do_exp(a[elem]);
+            """
+            custom_kernel = mx.fast.cuda_kernel
+
+        mx.random.seed(7)
+        a = mx.random.normal(shape=(2, 2))
+        kernel = custom_kernel(
+            name="helper",
+            input_names=["a"],
+            output_names=["out1"],
+            header=header,
+            source=source,
         )
         out = kernel(
             inputs=[a],
@@ -714,16 +773,21 @@ class TestFast(mlx_tests.MLXTestCase):
         )
         self.assertTrue(mx.allclose(out[0], mx.exp(a)))
 
-    @unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
+    @unittest.skipIf(not mx.is_available(mx.gpu), "No GPU available")
     def test_custom_kernel_attributes(self):
+        if mx.metal.is_available():
+            source = "out[0] = threads_per_threadgroup.x;"
+            custom_kernel = mx.fast.metal_kernel
+        elif mx.cuda.is_available():
+            source = "out[0] = blockDim.x;"
+            custom_kernel = mx.fast.cuda_kernel
+
         a = mx.zeros(shape=(1, 1))
-        kernel = mx.fast.metal_kernel(
+        kernel = custom_kernel(
             name="test_fun",
             input_names=["a"],
             output_names=["out"],
-            source="""
+            source=source,
-                out[0] = threads_per_threadgroup.x;
-            """,
         )
         out = kernel(
             inputs=[a],