Support unaligned M

* speedup scalars in cuda

* comment

.circleci/config.yml

@@ -272,6 +272,7 @@ jobs:
           name: Build Python package
           command: |
             source env/bin/activate
+            python setup.py clean --all
             << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
       - when:
           condition:
@@ -333,6 +334,7 @@ jobs:
             << parameters.build_env >> pip install ".[dev]" -v
             pip install typing_extensions
             python setup.py generate_stubs
+            python setup.py clean --all
             MLX_BUILD_STAGE=1 << parameters.build_env >> python -m build -w
             bash python/scripts/repair_linux.sh
       - when:
@@ -364,7 +366,7 @@ jobs:
         type: string
         default: ""
     machine:
-      image: linux-cuda-12:default
+      image: linux-cuda-12:2024.11.1
       resource_class: gpu.nvidia.small.gen2
     steps:
       - checkout

ACKNOWLEDGMENTS.md

@@ -19,6 +19,7 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />

CMakeLists.txt

@@ -22,7 +22,7 @@ project(
 
 # ----------------------------- Setup -----------------------------
 set(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)

docs/src/python/optimizers/common_optimizers.rst

@@ -19,3 +19,4 @@ Common Optimizers
    Adamax
    Lion
    MultiOptimizer
+   Muon

mlx/backend/cuda/CMakeLists.txt

@@ -42,7 +42,9 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/unary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/quantized.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/qmm.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
 target_compile_definitions(mlx PRIVATE MLX_USE_CUDA)
@@ -90,7 +92,7 @@ target_compile_options(
 # Compute capability 7 is required for synchronization between CPU/GPU with
 # managed memory. TODO: Add more architectures for potential performance gain.
 set(MLX_CUDA_ARCHITECTURES
-    "70;80"
+    "80"
     CACHE STRING "CUDA architectures")
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
@@ -130,3 +132,12 @@ target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
 # Install CCCL headers for JIT.
 install(DIRECTORY ${cccl_SOURCE_DIR}/include/cuda
         DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/cccl)
+
+# Make Thunderkittens available
+FetchContent_Declare(
+  kittens
+  GIT_REPOSITORY https://github.com/HazyResearch/ThunderKittens.git
+  GIT_TAG aaab847f430ed313ed466e64b25b9177babd1db8
+  GIT_SHALLOW TRUE)
+FetchContent_MakeAvailable(kittens)
+target_include_directories(mlx BEFORE PRIVATE "${kittens_SOURCE_DIR}/include")

mlx/backend/cuda/allocator.cpp

@@ -17,6 +17,52 @@ namespace cu {
 
 constexpr int page_size = 16384;
 
+// Any allocations smaller than this will try to use the small pool
+constexpr int small_block_size = 8;
+
+// The small pool size in bytes. This should be a multiple of the host page
+// size and small_block_size.
+constexpr int small_pool_size = 4 * page_size;
+
+SmallSizePool::SmallSizePool() {
+  CHECK_CUDA_ERROR(cudaMallocManaged(&buffer_, small_pool_size));
+  end_ = reinterpret_cast<void*>(
+      reinterpret_cast<char*>(buffer_) + small_pool_size);
+  next_free_ = reinterpret_cast<Block*>(buffer_);
+
+  auto num_blocks = small_pool_size / small_block_size;
+  auto curr = next_free_;
+  for (size_t i = 0; i < num_blocks - 1; ++i) {
+    curr->next = reinterpret_cast<Block*>(
+        reinterpret_cast<char*>(buffer_) + (i + 1) * small_block_size);
+    curr = curr->next;
+  }
+  curr->next = nullptr;
+}
+
+SmallSizePool::~SmallSizePool() {
+  CHECK_CUDA_ERROR(cudaFree(buffer_));
+}
+
+void* SmallSizePool::malloc() {
+  if (next_free_ == nullptr) {
+    return nullptr;
+  }
+  Block* b = next_free_;
+  next_free_ = next_free_->next;
+  return static_cast<void*>(b);
+}
+
+void SmallSizePool::free(void* p) {
+  auto b = static_cast<Block*>(p);
+  b->next = next_free_;
+  next_free_ = b;
+}
+
+bool SmallSizePool::in_pool(void* p) {
+  return (p >= buffer_) && (p < end_);
+}
+
 CudaAllocator::CudaAllocator()
     : buffer_cache_(
           page_size,
@@ -36,7 +82,9 @@ Buffer CudaAllocator::malloc(size_t size) {
   // Find available buffer from cache.
   auto orig_size = size;
   std::unique_lock lock(mutex_);
-  if (size < page_size) {
+  if (size <= small_block_size) {
+    size = 8;
+  } else if (size < page_size) {
     size = next_power_of_2(size);
   } else {
     size = page_size * ((size + page_size - 1) / page_size);
@@ -53,11 +101,19 @@ Buffer CudaAllocator::malloc(size_t size) {
 
     lock.unlock();
     buf = new CudaBuffer{nullptr, size};
-    cudaError_t err = cudaMallocManaged(&buf->data, size);
+
-    if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
+    // Try the scalar pool first
-      throw std::runtime_error(fmt::format(
+    if (size <= small_block_size) {
-          "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+      buf->data = scalar_pool_.malloc();
     }
+    if (!buf->data) {
+      cudaError_t err = cudaMallocManaged(&buf->data, size);
+      if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
+        throw std::runtime_error(fmt::format(
+            "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+      }
+    }
+
     lock.lock();
   }
   active_memory_ += size;
@@ -116,7 +172,11 @@ void CudaAllocator::cuda_free(void* buf) {
       return;
     }
   }
-  cudaFree(buf);
+  if (scalar_pool_.in_pool(buf)) {
+    scalar_pool_.free(buf);
+  } else {
+    cudaFree(buf);
+  }
 }
 
 size_t CudaAllocator::get_active_memory() const {

mlx/backend/cuda/allocator.h

@@ -22,6 +22,28 @@ struct CudaBuffer {
   size_t size;
 };
 
+class SmallSizePool {
+ private:
+  struct Block {
+    Block* next;
+  };
+
+  void* buffer_{nullptr};
+  Block* next_free_{nullptr};
+  void* end_{nullptr};
+
+ public:
+  SmallSizePool();
+  ~SmallSizePool();
+
+  SmallSizePool(const SmallSizePool&) = delete;
+  SmallSizePool& operator=(const SmallSizePool&) = delete;
+
+  void* malloc();
+  void free(void* p);
+  bool in_pool(void* p);
+};
+
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
@@ -60,6 +82,7 @@ class CudaAllocator : public allocator::Allocator {
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
+  SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();

mlx/backend/cuda/arg_reduce.cu

@@ -166,6 +166,7 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
           kernel,
           num_blocks,
           block_dim(),
+          0,
           in.data<T>(),
           out.data<uint32_t>(),
           out.size(),

mlx/backend/cuda/binary.cu

@@ -219,6 +219,7 @@ void binary_op_gpu_inplace(
                         kernel,
                         num_blocks,
                         block_dims,
+                        0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out.data<OutType>(),
@@ -235,6 +236,7 @@ void binary_op_gpu_inplace(
                       kernel,
                       num_blocks,
                       block_dims,
+                      0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out.data<OutType>(),
@@ -269,6 +271,7 @@ void binary_op_gpu_inplace(
                 kernel,
                 num_blocks,
                 block_dims,
+                0,
                 a.data<InType>(),
                 b.data<InType>(),
                 out.data<OutType>(),

mlx/backend/cuda/binary_two.cu

@@ -239,6 +239,7 @@ void binary_two_op_gpu_inplace(
                         kernel,
                         num_blocks,
                         block_dims,
+                        0,
                         a.data<InType>(),
                         b.data<InType>(),
                         out_a.data<OutType>(),
@@ -256,6 +257,7 @@ void binary_two_op_gpu_inplace(
                       kernel,
                       num_blocks,
                       block_dims,
+                      0,
                       a.data<InType>(),
                       b.data<InType>(),
                       out_a.data<OutType>(),
@@ -291,6 +293,7 @@ void binary_two_op_gpu_inplace(
                 kernel,
                 num_blocks,
                 block_dims,
+                0,
                 a.data<InType>(),
                 b.data<InType>(),
                 out_a.data<OutType>(),

mlx/backend/cuda/compiled.cpp

@@ -53,9 +53,10 @@ struct FusedKernelBuilder {
 
     // Build function signature.
     if (contiguous) {
-      os += "template <typename IdxT = uint32_t>\n";
+      os += "template <typename IdxT = uint32_t, int work_per_thread = 1>\n";
     } else {
-      os += "template <int NDIM, typename IdxT = uint32_t>\n";
+      os +=
+          "template <int NDIM, typename IdxT = uint32_t, int work_per_thread = 1>\n";
     }
     os += fmt::format("__global__ void {}(\n", kernel_name + name);
     for (size_t i = 0; i < params.size(); ++i) {
@@ -67,12 +68,46 @@ struct FusedKernelBuilder {
     }
     os += ") {\n";
 
-    // Index.
+    // Index. For non contiguous kernels we create a separate index
+    // variable per variable otherwise everyone uses `index`.
     os +=
-        "  IdxT index = cg::this_grid().thread_rank();\n"
+        "  IdxT index = cg::this_grid().thread_rank() * work_per_thread;\n"
         "  if (index >= size) {\n"
         "    return;\n"
         "  }\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "  IdxT " + xname + "_idx = 0;\n";
+      }
+      os += "  {\n";
+      os += "    IdxT loc = index;\n";
+      os +=
+          "    #pragma unroll\n"
+          "    for (int i = NDIM - 1; i >= 0; i--) {\n";
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "      " + xname + "_idx += (loc \% shape[i]) * IdxT(" + xname +
+            "_strides[i]);\n";
+      }
+      os +=
+          "      loc /= shape[i];\n"
+          "    }\n"
+          "  }\n";
+    }
+
+    // Work loop
+    os +=
+        "\n"
+        "  for (int i = 0; i < work_per_thread && index < size; i++) {\n";
 
     // Read inputs.
     for (size_t i = 0; i < inputs.size(); ++i) {
@@ -89,12 +124,9 @@ struct FusedKernelBuilder {
       } else if (contiguous) {
         value = fmt::format("{}[index]", xname);
       } else {
-        std::string index = fmt::format(
+        value = fmt::format("{}[{}_idx]", xname, xname);
-            "elem_to_loc_nd<NDIM>(index, shape.data(), {}_strides.data())",
-            xname);
-        value = fmt::format("{}[{}]", xname, index);
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write tape.
@@ -113,14 +145,30 @@ struct FusedKernelBuilder {
         }
         value += fmt::format("tmp_{})", namer.get_name(x.inputs().back()));
       }
-      os += fmt::format("  {} tmp_{} = {};\n", type, xname, value);
+      os += fmt::format("    {} tmp_{} = {};\n", type, xname, value);
     }
 
     // Write output.
     for (const auto& x : outputs) {
-      os += fmt::format("  {0}[index] = tmp_{0};\n", namer.get_name(x));
+      os += fmt::format("    {0}[index] = tmp_{0};\n", namer.get_name(x));
     }
 
+    // End of work loop
+    os +=
+        "\n"
+        "    index++;\n";
+    if (!contiguous) {
+      for (size_t i = 0; i < inputs.size(); ++i) {
+        const auto& x = inputs[i];
+        const std::string& xname = namer.get_name(x);
+        if (is_scalar(x) || is_constant(i)) {
+          continue;
+        }
+        os += "    " + xname + "_idx += " + xname + "_strides[NDIM - 1];\n";
+      }
+    }
+    os += "  }\n";
+
     os += "}\n";
   }
 };
@@ -156,15 +204,28 @@ void Compiled::eval_gpu(
     builder.build("_strided", false);
     builder.os += "\n} // namespace mlx::core::cu\n";
     // Build kernel names.
-    std::vector<std::string> kernel_names = {
+    std::vector<std::string> kernel_names;
-        fmt::format("mlx::core::cu::{}_contiguous<uint32_t>", lib_name()),
+    for (auto work_per_thread : std::array<int, 2>{1, 4}) {
-        fmt::format("mlx::core::cu::{}_contiguous<int64_t>", lib_name()),
-    };
-    for (int i = 1; i <= MAX_NDIM; ++i) {
       kernel_names.push_back(fmt::format(
-          "mlx::core::cu::{}_strided<{}, uint32_t>", lib_name(), i));
+          "mlx::core::cu::{}_contiguous<uint32_t, {}>",
-      kernel_names.push_back(
+          lib_name(),
-          fmt::format("mlx::core::cu::{}_strided<{}, int64_t>", lib_name(), i));
+          work_per_thread));
+      kernel_names.push_back(fmt::format(
+          "mlx::core::cu::{}_contiguous<int64_t, {}>",
+          lib_name(),
+          work_per_thread));
+      for (int i = 1; i <= MAX_NDIM; ++i) {
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, uint32_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+        kernel_names.push_back(fmt::format(
+            "mlx::core::cu::{}_strided<{}, int64_t, {}>",
+            lib_name(),
+            i,
+            work_per_thread));
+      }
     }
     return std::make_pair(std::move(builder.os), std::move(kernel_names));
   });
@@ -207,13 +268,21 @@ void Compiled::eval_gpu(
     args.append<uint32_t>(outputs[0].data_size());
   }
 
+  // Choose work per thread
+  int work_per_thread = 4;
+  if (!contiguous && shape.back() % work_per_thread != 0) {
+    work_per_thread = 1;
+  }
+
   // Launch kernel.
   const char* index_type = large ? "int64_t" : "uint32_t";
   std::string kernel_name = fmt::format("mlx::core::cu::{}", lib_name());
   if (contiguous) {
-    kernel_name += fmt::format("_contiguous<{}>", index_type);
+    kernel_name +=
+        fmt::format("_contiguous<{}, {}>", index_type, work_per_thread);
   } else {
-    kernel_name += fmt::format("_strided<{}, {}>", shape.size(), index_type);
+    kernel_name += fmt::format(
+        "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
   }
   auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
@@ -224,8 +293,9 @@ void Compiled::eval_gpu(
   }
 
   auto kernel = mod.get_kernel(kernel_name);
-  auto [num_blocks, block_dims] = get_launch_args(kernel, outputs[0], large);
+  auto [num_blocks, block_dims] =
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+      get_launch_args(kernel, outputs[0], large, work_per_thread);
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 } // namespace mlx::core

mlx/backend/cuda/copy/copy_contiguous.cu

@@ -82,6 +82,7 @@ void copy_contiguous(
             kernel,
             num_blocks,
             block_dims,
+            0,
             in.data<InType>() + in_offset,
             out.data<OutType>() + out_offset,
             out.data_size());

mlx/backend/cuda/copy/copy_general.cu

@@ -79,6 +79,7 @@ void copy_general(
                     kernel,
                     num_blocks,
                     block_dims,
+                    0,
                     in_ptr,
                     out_ptr,
                     data_size,
@@ -94,6 +95,7 @@ void copy_general(
                   kernel,
                   num_blocks,
                   block_dims,
+                  0,
                   in_ptr,
                   out_ptr,
                   data_size,

mlx/backend/cuda/copy/copy_general_dynamic.cu

@@ -82,6 +82,7 @@ void copy_general_dynamic(
                     kernel,
                     num_blocks,
                     block_dims,
+                    0,
                     in_ptr,
                     out_ptr,
                     out.size(),
@@ -99,6 +100,7 @@ void copy_general_dynamic(
                   kernel,
                   num_blocks,
                   block_dims,
+                  0,
                   in_ptr,
                   out_ptr,
                   out.size(),

mlx/backend/cuda/copy/copy_general_input.cu

@@ -71,6 +71,7 @@ void copy_general_input(
                     kernel,
                     num_blocks,
                     block_dims,
+                    0,
                     in_ptr,
                     out_ptr,
                     out.size(),
@@ -85,6 +86,7 @@ void copy_general_input(
                   kernel,
                   num_blocks,
                   block_dims,
+                  0,
                   in_ptr,
                   out_ptr,
                   out.size(),

mlx/backend/cuda/device.cpp

@@ -66,7 +66,6 @@ CommandEncoder& Device::get_command_encoder(Stream s) {
 }
 
 CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
-  CHECK_CUDA_ERROR(cudaGraphCreate(&graph, 0));
   CHECK_CUDA_ERROR(
       cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
 }
@@ -216,12 +215,14 @@ void CommandEncoder::add_kernel_node(
     void* func,
     dim3 grid_dim,
     dim3 block_dim,
+    uint32_t smem_bytes,
     void** params) {
   cudaKernelNodeParams kernel_params = {0};
   kernel_params.func = func;
   kernel_params.gridDim = grid_dim;
   kernel_params.blockDim = block_dim;
   kernel_params.kernelParams = params;
+  kernel_params.sharedMemBytes = smem_bytes;
   cudaGraphNode_t node;
   CHECK_CUDA_ERROR(
       cudaGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));
@@ -232,6 +233,7 @@ void CommandEncoder::add_kernel_node(
     CUfunction func,
     dim3 grid_dim,
     dim3 block_dim,
+    uint32_t smem_bytes,
     void** params) {
   CUDA_KERNEL_NODE_PARAMS kernel_params = {0};
   kernel_params.func = func;
@@ -242,6 +244,7 @@ void CommandEncoder::add_kernel_node(
   kernel_params.blockDimY = block_dim.y;
   kernel_params.blockDimZ = block_dim.z;
   kernel_params.kernelParams = params;
+  kernel_params.sharedMemBytes = smem_bytes;
   CUgraphNode node;
   CHECK_CUDA_ERROR(
       cuGraphAddKernelNode(&node, graph_, NULL, 0, &kernel_params));

mlx/backend/cuda/device.h

@@ -45,25 +45,34 @@ class CommandEncoder {
   void set_output_array(const array& arr);
 
   template <typename F, typename... Params>
-  void
+  void add_kernel_node(
-  add_kernel_node(F* func, dim3 grid_dim, dim3 block_dim, Params&&... params) {
+      F* func,
+      dim3 grid_dim,
+      dim3 block_dim,
+      uint32_t smem_bytes,
+      Params&&... params) {
     constexpr size_t num = sizeof...(Params);
     void* ptrs[num];
     size_t i = 0;
     ([&](auto&& p) { ptrs[i++] = static_cast<void*>(&p); }(
          std::forward<Params>(params)),
      ...);
-    add_kernel_node((void*)func, grid_dim, block_dim, ptrs);
+    add_kernel_node((void*)func, grid_dim, block_dim, smem_bytes, ptrs);
   }
 
   void add_kernel_node(
       CUfunction func,
       dim3 grid_dim,
       dim3 block_dim,
+      uint32_t smem_bytes,
       void** params);
 
-  void
+  void add_kernel_node(
-  add_kernel_node(void* func, dim3 grid_dim, dim3 block_dim, void** params);
+      void* func,
+      dim3 grid_dim,
+      dim3 block_dim,
+      uint32_t smem_bytes,
+      void** params);
 
   void add_temporary(const array& arr) {
     temporaries_.push_back(arr.data_shared_ptr());

mlx/backend/cuda/indexing.cpp

@@ -129,7 +129,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto kernel = mod.get_kernel(kernel_name);
   auto [num_blocks, block_dims] = get_launch_args(kernel, out, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -230,7 +230,7 @@ void Scatter::eval_gpu(const std::vector<array>& inputs, array& out) {
   encoder.set_output_array(out);
   auto kernel = mod.get_kernel(kernel_name);
   auto [num_blocks, block_dims] = get_launch_args(kernel, upd, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -318,7 +318,7 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   encoder.set_output_array(out);
   auto kernel = mod.get_kernel(kernel_name);
   auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
@@ -422,7 +422,7 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   encoder.set_output_array(out);
   auto kernel = mod.get_kernel(kernel_name);
   auto [num_blocks, block_dims] = get_launch_args(kernel, idx, large);
-  encoder.add_kernel_node(kernel, num_blocks, block_dims, args.args());
+  encoder.add_kernel_node(kernel, num_blocks, block_dims, 0, args.args());
 }
 
 } // namespace mlx::core

mlx/backend/cuda/jit_module.cpp

@@ -52,13 +52,29 @@ const std::string& cuda_home() {
 }
 
 // Return the location of CCCL headers shipped with the distribution.
-bool get_cccl_include(std::string* out) {
+const std::string& cccl_dir() {
-  auto cccl_headers = current_binary_dir().parent_path() / "include" / "cccl";
+  static std::string dir = []() {
-  if (!std::filesystem::exists(cccl_headers)) {
+    std::filesystem::path path;
-    return false;
+#if defined(MLX_CCCL_DIR)
-  }
+    // First search the install dir if defined.
-  *out = fmt::format("--include-path={}", cccl_headers.string());
+    path = MLX_CCCL_DIR;
-  return true;
+    if (std::filesystem::exists(path)) {
+      return path.string();
+    }
+#endif
+    // Then search dynamically from the dir of libmlx.so file.
+    path = current_binary_dir().parent_path() / "include" / "cccl";
+    if (std::filesystem::exists(path)) {
+      return path.string();
+    }
+    // Finally check the environment variable.
+    path = std::getenv("MLX_CCCL_DIR");
+    if (!path.empty() && std::filesystem::exists(path)) {
+      return path.string();
+    }
+    return std::string();
+  }();
+  return dir;
 }
 
 // Get the cache directory for storing compiled results.
@@ -121,7 +137,8 @@ void write_cached_ptx(
     const std::filesystem::path& cache_dir,
     const std::string& module_name,
     const std::vector<char>& ptx,
-    const std::vector<std::pair<std::string, std::string>>& ptx_kernels) {
+    const std::vector<std::pair<std::string, std::string>>& ptx_kernels,
+    const std::string& source_code) {
   if (cache_dir.empty()) {
     return;
   }
@@ -134,6 +151,9 @@ void write_cached_ptx(
   for (const auto& [name, mangled] : ptx_kernels) {
     txt_file << name << "\t" << mangled << std::endl;
   }
+
+  std::ofstream source_file(cache_dir / (module_name + ".cu"));
+  source_file << source_code;
 }
 
 // Return if |device|'s version is not newer than |major|.|minor| version.
@@ -234,8 +254,9 @@ JitModule::JitModule(
         device.compute_capability_major(),
         device.compute_capability_minor());
     args.push_back(compute.c_str());
-    std::string cccl_include;
+    std::string cccl_include = cccl_dir();
-    if (get_cccl_include(&cccl_include)) {
+    if (!cccl_include.empty()) {
+      cccl_include = fmt::format("--include-path={}", cccl_include);
       args.push_back(cccl_include.c_str());
     }
     std::string cuda_include =
@@ -272,7 +293,8 @@ JitModule::JitModule(
     } else {
       CHECK_NVRTC_ERROR(nvrtcGetPTX(prog, ptx.data()));
     }
-    write_cached_ptx(ptx_cache_dir(), module_name, ptx, ptx_kernels);
+    write_cached_ptx(
+        ptx_cache_dir(), module_name, ptx, ptx_kernels, source_code);
   }
 
   // Load module.

mlx/backend/cuda/layer_norm.cu

@@ -237,8 +237,7 @@ void LayerNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -267,6 +266,7 @@ void LayerNorm::eval_gpu(
           kernel,
           n_rows,
           block_dim(),
+          0,
           x.data<DataType>(),
           w.data<DataType>(),
           b.data<DataType>(),
@@ -295,9 +295,7 @@ void LayerNormVJP::eval_gpu(
       return x;
     }
     copied = true;
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    return contiguous_copy_gpu(x, s);
-    copy_gpu(x, x_copy, CopyType::General, s);
-    return x_copy;
   };
   bool donate_x = inputs[0].is_donatable();
   bool donate_g = inputs[3].is_donatable();
@@ -381,6 +379,7 @@ void LayerNormVJP::eval_gpu(
                 kernel,
                 n_rows,
                 block_dim(),
+                0,
                 x.data<DataType>(),
                 w.data<DataType>(),
                 g.data<DataType>(),

mlx/backend/cuda/logsumexp.cu

@@ -108,8 +108,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
     if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       encoder.add_temporary(x_copy);
       return x_copy;
     }
@@ -152,6 +151,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
           kernel,
           n_rows,
           block_dim(),
+          0,
           in.data<DataType>(),
           out.data<DataType>(),
           axis_size);

mlx/backend/cuda/matmul.cpp

@@ -27,6 +27,35 @@ void check_cublas_error(const char* name, cublasStatus_t err) {
   }
 }
 
+struct CublasPreference {
+  CublasPreference(Device& device) {
+    // The recommended cublas workspace size is 4 MiB for pre-Hopper and 32 MiB
+    // for Hopper+:
+    // https://docs.nvidia.com/cuda/cublas/#cublassetworkspace
+    uint64_t MiB = 1024 * 1024;
+    uint64_t workspace_size =
+        device.compute_capability_major() >= 9 ? 32 * MiB : 4 * MiB;
+
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceCreate(&pref_));
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceSetAttribute(
+        pref_,
+        CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
+        &workspace_size,
+        sizeof(uint64_t)));
+  }
+
+  ~CublasPreference() {
+    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceDestroy(pref_));
+  }
+
+  cublasLtMatmulPreference_t pref_{nullptr};
+};
+
+cublasLtMatmulPreference_t cublas_preference(Device& device) {
+  static CublasPreference pref(device);
+  return pref.pref_;
+}
+
 class MatMul {
  public:
   MatMul(
@@ -43,7 +72,7 @@ class MatMul {
       int32_t batch_count,
       int64_t a_batch_stride,
       int64_t b_batch_stride)
-      : handle_(device.lt_handle()) {
+      : handle_(device.lt_handle()), pref_(cublas_preference(device)) {
     heuristic_.state = CUBLAS_STATUS_NOT_INITIALIZED;
 
     auto scale_type = dtype_to_cuda_type(dtype);
@@ -77,20 +106,6 @@ class MatMul {
         type, b_rows, b_cols, b_transposed, ldb, batch_count, b_batch_stride);
     out_desc_ = create_matrix_layout(
         type, a_rows, b_cols, false, b_cols, batch_count, a_rows * b_cols);
-
-    // The recommended cublas workspace size is 4 MiB for pre-Hopper and 32 MiB
-    // for Hopper+:
-    // https://docs.nvidia.com/cuda/cublas/#cublassetworkspace
-    uint64_t MiB = 1024 * 1024;
-    uint64_t workspace_size =
-        device.compute_capability_major() >= 9 ? 32 * MiB : 4 * MiB;
-
-    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceCreate(&pref_));
-    CHECK_CUBLAS_ERROR(cublasLtMatmulPreferenceSetAttribute(
-        pref_,
-        CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
-        &workspace_size,
-        sizeof(uint64_t)));
   }
 
   MatMul(
@@ -104,7 +119,6 @@ class MatMul {
       uint64_t b_rows,
       uint64_t b_cols,
       int64_t ldb,
-      bool c_transposed,
       int64_t ldc,
       int32_t batch_count,
       int64_t a_batch_stride,
@@ -126,15 +140,15 @@ class MatMul {
             b_batch_stride) {
     auto type = dtype_to_cuda_type(dtype);
     c_desc_ = create_matrix_layout(
-        type, a_rows, b_cols, c_transposed, ldc, batch_count, c_batch_stride);
+        type, a_rows, b_cols, false, ldc, batch_count, c_batch_stride);
   }
 
   ~MatMul() {
-    cublasLtMatrixLayoutDestroy(a_desc_);
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(a_desc_));
-    cublasLtMatrixLayoutDestroy(b_desc_);
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(b_desc_));
-    cublasLtMatrixLayoutDestroy(c_desc_);
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(c_desc_));
-    cublasLtMatrixLayoutDestroy(out_desc_);
+    CHECK_CUBLAS_ERROR(cublasLtMatrixLayoutDestroy(out_desc_));
-    cublasLtMatmulDescDestroy(matmul_desc_);
+    CHECK_CUBLAS_ERROR(cublasLtMatmulDescDestroy(matmul_desc_));
   }
 
   void run(
@@ -259,9 +273,9 @@ class MatMul {
     return desc;
   }
 
+  cublasLtMatmulPreference_t pref_{nullptr};
   cublasLtHandle_t handle_{nullptr};
   cublasLtMatmulDesc_t matmul_desc_{nullptr};
-  cublasLtMatmulPreference_t pref_{nullptr};
   cublasLtMatrixLayout_t a_desc_{nullptr};
   cublasLtMatrixLayout_t b_desc_{nullptr};
   cublasLtMatrixLayout_t c_desc_{nullptr};
@@ -282,8 +296,7 @@ check_transpose(cu::CommandEncoder& enc, const Stream& s, const array& arr) {
   } else if (stx == 1 && sty == arr.shape(-2)) {
     return std::make_tuple(true, sty, arr);
   } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+    array arr_copy = contiguous_copy_gpu(arr, s);
-    copy_gpu(arr, arr_copy, CopyType::General, s);
     enc.add_temporary(arr_copy);
     return std::make_tuple(false, arr.shape(-1), arr_copy);
   }
@@ -389,9 +402,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 3);
   auto& a_pre = inputs[0];
   auto& b_pre = inputs[1];
-  auto& c_pre = inputs[2];
+  auto c = inputs[2];
-
-  out.set_data(allocator::malloc(out.nbytes()));
 
   /////////////////////////////////////////////////////////////////////////////
   // Init checks and prep
@@ -404,7 +415,24 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
   // the arrays
   auto [a_transposed, lda, a] = check_transpose(encoder, s, a_pre);
   auto [b_transposed, ldb, b] = check_transpose(encoder, s, b_pre);
-  auto [c_transposed, ldc, c] = check_transpose(encoder, s, c_pre);
+
+  int64_t ldc;
+  {
+    auto stx = c.strides()[c.ndim() - 2];
+    auto sty = c.strides()[c.ndim() - 1];
+    if (sty == 1 && stx == c.shape(-1)) {
+      ldc = stx;
+      out.set_data(allocator::malloc(out.nbytes()));
+    } else if (sty == 1 && stx == 0) {
+      ldc = 0;
+      out.set_data(allocator::malloc(out.nbytes()));
+    } else {
+      // Copy C into out and set C to out
+      ldc = c.shape(-1);
+      copy_gpu(c, out, CopyType::General, s);
+      c = out;
+    }
+  }
 
   /////////////////////////////////////////////////////////////////////////////
   // Check and collapse batch dimensions
@@ -442,7 +470,6 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
       K,
       N,
       ldb,
-      c_transposed,
       ldc,
       batch_shape.back(),
       a_batch_strides.back(),

mlx/backend/cuda/matmul/mma.cuh

@@ -0,0 +1,108 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/matmul/tiles.cuh"
+
+namespace mlx::core::cu {
+
+template <typename U, typename T>
+__device__ inline void
+mma_t(Tile16x16<U>& C, Tile16x16<T>& A, Tile16x16<T>& B) {}
+
+/**
+ * Multiply the 16x16 bfloat16 tiles and accumulate the result in one 16x16
+ * float tile.
+ *
+ * We actually perform C += A @ B.T
+ */
+__device__ inline void mma_t(
+    Tile16x16<float>& C,
+    Tile16x16<__nv_bfloat16>& A,
+    Tile16x16<__nv_bfloat16>& B) {
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0, %1, %2, %3}, "
+      "{%4, %5, %6, %7}, "
+      "{%8, %9}, "
+      "{%10, %11, %12, %13};"
+
+      // D matrix
+      : "+f"(C.values[0].x),
+        "+f"(C.values[0].y),
+        "+f"(C.values[1].x),
+        "+f"(C.values[1].y)
+
+      // A matrix
+      : "r"(*(uint32_t*)(&A.values[0])),
+        "r"(*(uint32_t*)(&A.values[1])),
+        "r"(*(uint32_t*)(&A.values[2])),
+        "r"(*(uint32_t*)(&A.values[3])),
+
+        // B matrix
+        "r"(*(uint32_t*)(&B.values[0])),
+        "r"(*(uint32_t*)(&B.values[2])),
+
+        // C matrix
+        "f"(C.values[0].x),
+        "f"(C.values[0].y),
+        "f"(C.values[1].x),
+        "f"(C.values[1].y));
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0, %1, %2, %3}, "
+      "{%4, %5, %6, %7}, "
+      "{%8, %9}, "
+      "{%10, %11, %12, %13};"
+
+      // D matrix
+      : "+f"(C.values[2].x),
+        "+f"(C.values[2].y),
+        "+f"(C.values[3].x),
+        "+f"(C.values[3].y)
+
+      // A matrix
+      : "r"(*(uint32_t*)(&A.values[0])),
+        "r"(*(uint32_t*)(&A.values[1])),
+        "r"(*(uint32_t*)(&A.values[2])),
+        "r"(*(uint32_t*)(&A.values[3])),
+
+        // B matrix
+        "r"(*(uint32_t*)(&B.values[1])),
+        "r"(*(uint32_t*)(&B.values[3])),
+
+        // C matrix
+        "f"(C.values[2].x),
+        "f"(C.values[2].y),
+        "f"(C.values[3].x),
+        "f"(C.values[3].y));
+}
+
+/**
+ * Multiply larger register tiles by delegating to mma_t.
+ */
+template <typename U, typename T, int M, int N, int K>
+__device__ inline void mma_t(
+    RegisterTile<U, M, N>& C,
+    RegisterTile<T, M, K>& A,
+    RegisterTile<T, N, K>& B) {
+  constexpr int TILES_M = RegisterTile<T, M, K>::TILES_Y;
+  constexpr int TILES_K = RegisterTile<T, M, K>::TILES_X;
+  constexpr int TILES_N = RegisterTile<T, N, K>::TILES_Y;
+
+  MLX_UNROLL
+  for (int k = 0; k < TILES_K; k++) {
+    MLX_UNROLL
+    for (int m = 0; m < TILES_M; m++) {
+      MLX_UNROLL
+      for (int n = 0; n < TILES_N; n++) {
+        mma_t(
+            C.data[m * TILES_N + n],
+            A.data[m * TILES_K + k],
+            B.data[n * TILES_K + k]);
+      }
+    }
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/matmul/tiles.cuh

@@ -0,0 +1,419 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#define MLX_UNROLL _Pragma("unroll")
+
+namespace mlx::core::cu {
+
+// Map types to their vector of 2 type float -> float2, double -> double2 etc
+template <typename T>
+struct Vector2;
+template <>
+struct Vector2<double> {
+  using type = double2;
+};
+template <>
+struct Vector2<float> {
+  using type = float2;
+};
+template <>
+struct Vector2<__half> {
+  using type = __half2;
+};
+template <>
+struct Vector2<__nv_bfloat16> {
+  using type = __nv_bfloat162;
+};
+template <typename T>
+using Vector2_t = typename Vector2<T>::type;
+
+/**
+ * The basic building block for Ampere mmas. A 16x16 tile distributed across
+ * the warp.
+ *
+ * Each thread holds 8 values. They are distributed according to
+ * https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-fragment-mma-16816-float
+ *
+ * For use instructions see the individual methods eg load().
+ */
+template <typename T>
+struct Tile16x16 {
+  using T2 = Vector2_t<T>;
+
+  T2 values[4];
+
+  __device__ inline void fill(T v) {
+    T2 v2 = {v, v};
+    for (int i = 0; i < 4; i++) {
+      values[i] = v2;
+    }
+  }
+
+  /**
+   * Load a 16x16 tile from shared memory.
+   *
+   * The instruction is a bit weird in the sense that the address provided by
+   * each thread and the elements loaded are not the same.
+   *
+   * We load 4 8x8 tiles. The tile rows are stored contiguously in memory. As a
+   * result the warp provides 4*8 = 32 addresses one per row.
+   *
+   * Threads 0-7 provide the addresses for the first tile, 8-15 for the second
+   * and so on. For instance to load a non swizzled tile we would do
+   *
+   *    base_addr + (laneid % 16) * BK + (laneid / 2) * 8
+   *
+   * See
+   * https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-instructions-ldmatrix
+   */
+  __device__ inline void load(uint32_t row_address) {
+    if constexpr (
+        std::is_same_v<T2, __nv_bfloat162> || std::is_same_v<T2, __half2>) {
+      asm volatile(
+          "ldmatrix.sync.aligned.m8n8.x4.shared::cta.b16 {%0, %1, %2, %3}, [%4];\n"
+          : "=r"(*(uint32_t*)&(values[0])),
+            "=r"(*(uint32_t*)&(values[1])),
+            "=r"(*(uint32_t*)&(values[2])),
+            "=r"(*(uint32_t*)&(values[3]))
+          : "r"(row_address));
+    }
+  }
+
+  /**
+   * Store the tile to the address pointed to by `x`.
+   *
+   * The provided pointer is a generic pointer but this is meant to be used to
+   * store to global memory. For storing to shared memory we should use
+   * `stmatrix`.
+   *
+   * This also showcases the format of the tile quite nicely. Each register is
+   * holding to adjacent values. The indices are
+   *
+   *    row + 0, col + 0
+   *    row + 8, col + 0
+   *    row + 0, col + 8
+   *    row + 8, col + 8
+   *
+   * Given that we are dealing with Vector2_t<U> the column offsets are 4
+   * instead of 8.
+   */
+  template <typename U>
+  __device__ inline void store_global(U* x, int N) {
+    using U2 = Vector2_t<U>;
+    U2* x2 = reinterpret_cast<U2*>(x);
+    const int laneid = threadIdx.x % 32;
+    const int row = laneid / 4;
+    const int col = laneid % 4;
+    if constexpr (std::is_same_v<U2, T2>) {
+      x2[(row + 0) * (N / 2) + col + 0] = values[0];
+      x2[(row + 0) * (N / 2) + col + 4] = values[2];
+      x2[(row + 8) * (N / 2) + col + 0] = values[1];
+      x2[(row + 8) * (N / 2) + col + 4] = values[3];
+    } else if constexpr (
+        std::is_same_v<T2, float2> && std::is_same_v<U, __nv_bfloat16>) {
+      x2[(row + 0) * (N / 2) + col + 0] =
+          __floats2bfloat162_rn(values[0].x, values[0].y);
+      x2[(row + 0) * (N / 2) + col + 4] =
+          __floats2bfloat162_rn(values[2].x, values[2].y);
+      x2[(row + 8) * (N / 2) + col + 0] =
+          __floats2bfloat162_rn(values[1].x, values[1].y);
+      x2[(row + 8) * (N / 2) + col + 4] =
+          __floats2bfloat162_rn(values[3].x, values[3].y);
+    }
+  }
+
+  template <typename U>
+  __device__ inline void store_global_safe(U* x, int N, int max_rows) {
+    const int laneid = threadIdx.x % 32;
+    const int row = laneid / 4;
+    const int col = laneid % 4;
+    if (row < max_rows) {
+      x[(row + 0) * N + 2 * col + 0] = static_cast<U>(values[0].x);
+      x[(row + 0) * N + 2 * col + 1] = static_cast<U>(values[0].y);
+      x[(row + 0) * N + 2 * col + 8] = static_cast<U>(values[2].x);
+      x[(row + 0) * N + 2 * col + 9] = static_cast<U>(values[2].y);
+    }
+    if (row + 8 < max_rows) {
+      x[(row + 8) * N + 2 * col + 0] = static_cast<U>(values[1].x);
+      x[(row + 8) * N + 2 * col + 1] = static_cast<U>(values[1].y);
+      x[(row + 8) * N + 2 * col + 8] = static_cast<U>(values[3].x);
+      x[(row + 8) * N + 2 * col + 9] = static_cast<U>(values[3].y);
+    }
+  }
+};
+
+/**
+ * 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 U>
+  __device__ inline void
+  store_global_safe(U* x, int N, int row, int col, int max_rows) {
+    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_safe(
+            x + (row + i * 16) * N + col + j * 16, N, max_rows - row - i * 16);
+      }
+    }
+  }
+};
+
+template <typename T, int ROWS_, int COLS_>
+struct SharedTile {
+  static constexpr int ROWS = ROWS_;
+  static constexpr int COLS = COLS_;
+  static constexpr int TILES_X = COLS / 16;
+  static constexpr int TILES_Y = ROWS / 16;
+  static constexpr int NUMEL = ROWS * COLS;
+
+  // Swizzle taken from ThunderKittens.
+  //
+  // See inludes/types/shared/st.cuh
+  //
+  // I do feel that it is too math heavy and can be improved. Also the math is
+  // done every time although the addresses don't change from load to load. I
+  // guess we are expecting the compiler to figure that out.
+  static constexpr int swizzle_bytes =
+      (sizeof(T) == 2 ? (TILES_X % 4 == 0 ? 128 : (TILES_X % 2 == 0 ? 64 : 32))
+                      : (sizeof(T) == 4 ? (TILES_X % 2 == 0 ? 128 : 64) : 0));
+
+  T data[ROWS * COLS];
+
+  // Return a pointer to the element at (row, col) using the swizzle.
+  __device__ static inline T* ptr(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 uint64_t addr =
+          (uint64_t)(&ptr
+                         [outer_idx * ROWS * subtile_cols + row * subtile_cols +
+                          col % subtile_cols]);
+      const int swizzle = ((addr % swizzle_repeat) >> 7) << 4;
+      return (T*)(addr ^ swizzle);
+    } else {
+      return ptr + 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) {
+    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 +
+          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);
+    }
+  }
+
+  // Convenience functions to edit elements going through the swizzle.
+  __device__ inline T& operator()(int row, int col) {
+    return *ptr(data, row, col);
+  }
+  __device__ inline void store(float4& v, int row, int col) {
+    *(reinterpret_cast<float4*>(ptr(data, row, col))) = v;
+  }
+  __device__ inline void store(float2& v, int row, int col) {
+    *(reinterpret_cast<float2*>(ptr(data, row, col))) = v;
+  }
+  __device__ inline void store(float& v, int row, int col) {
+    *(reinterpret_cast<float*>(ptr(data, row, col))) = v;
+  }
+  template <int N>
+  __device__ inline void store(T (&v)[N], int row, int col) {
+    if constexpr (sizeof(T) * N == 4) {
+      store(*(reinterpret_cast<float*>(&v[0])), row, col);
+    } else if constexpr (sizeof(T) * N == 8) {
+      store(*(reinterpret_cast<float2*>(&v[0])), row, col);
+    } else if constexpr (sizeof(T) * N == 16) {
+      store(*(reinterpret_cast<float4*>(&v[0])), row, col);
+    } else {
+      MLX_UNROLL
+      for (int i = 0; i < N; i++) {
+        *ptr(data, row, col + i) = v[i];
+      }
+    }
+  }
+};
+
+/**
+ * Load the tile from global memory by loading 16 bytes at a time and storing
+ * them immediately.
+ */
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void load(Tile& tile, const T* x, int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  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;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    float4 tmp;
+    tmp = *(reinterpret_cast<const float4*>(&x[i * STEP_ROWS * N]));
+    tile.store(tmp, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+  }
+}
+
+/**
+ * Copy 16 bytes from the globale memory address pointed to by x to the smem
+ * address pointed to by row_address.
+ *
+ * A simple wrapper over the PTX.
+ */
+template <typename T>
+__device__ inline void cp_async_16(uint32_t row_address, const T* x) {
+  asm volatile(
+      "cp.async.ca.shared::cta.global [%0], [%1], 16;\n" ::"r"(row_address),
+      "l"(reinterpret_cast<const int4*>(x)));
+}
+
+/**
+ * Submit all the previous async copies to be executed.
+ */
+__device__ inline void cp_async_commit() {
+  asm volatile("cp.async.commit_group;\n" ::);
+}
+
+/**
+ * Wait for all the async copies to finish.
+ */
+__device__ inline void cp_async_wait_all() {
+  asm volatile("cp.async.wait_all;\n" ::);
+}
+
+/**
+ * The asynchronous equivalent of load.
+ *
+ * Loads the tile from global memory by submitting a bunch of async copy
+ * instructions. The copy won't start until commit is called and we don't have
+ * a guarantee it will finish until wait is called.
+ *
+ * It should be used as follows
+ *
+ *    load(...)
+ *    load(...)
+ *    cp_async_commit()
+ *    do_other_stuff()
+ *    cp_async_wait_all()
+ *    do_stuff_with_shmem()
+ */
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void
+load_async(Tile& tile, uint32_t base_address, const T* x, int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  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;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    cp_async_16(
+        tile.loc(base_address, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD),
+        x + i * STEP_ROWS * N);
+  }
+}
+
+template <int NUM_WARPS, typename T, typename Tile>
+__device__ inline void load_async_safe(
+    Tile& tile,
+    uint32_t base_address,
+    const T* x,
+    int N,
+    int max_rows) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(float4) / sizeof(T);
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+
+  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;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    if (row + i * STEP_ROWS < max_rows) {
+      cp_async_16(
+          tile.loc(base_address, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD),
+          x + i * STEP_ROWS * N);
+    } else {
+      float4 tmp = {0, 0, 0, 0};
+      tile.store(tmp, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+    }
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/primitives.cu

@@ -81,7 +81,6 @@ NO_GPU(Hadamard)
 NO_GPU(Load)
 NO_GPU_MULTI(LUF)
 NO_GPU_MULTI(QRF)
-NO_GPU(QuantizedMatmul)
 NO_GPU(SegmentedMM)
 NO_GPU_MULTI(SVD)
 NO_GPU(Inverse)

mlx/backend/cuda/quantized/affine_quantize.cu

@@ -2,30 +2,17 @@
 
 #include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/kernel_utils.cuh"
-#include "mlx/backend/gpu/copy.h"
+#include "mlx/backend/cuda/quantized/quantized_utils.cuh"
 #include "mlx/dtype_utils.h"
-#include "mlx/fast_primitives.h"
 
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
-#include <nvtx3/nvtx3.hpp>
 
 namespace mlx::core {
 namespace cu {
 
 namespace cg = cooperative_groups;
 
-template <int bits, int wsize = 8>
-inline constexpr __device__ short get_pack_factor() {
-  return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
-}
-
-template <int bits, int wsize = 8>
-inline constexpr __device__ short get_bytes_per_pack() {
-  constexpr int power_of_2_bits = (bits & (bits - 1)) == 0;
-  return power_of_2_bits ? (wsize / 8) : (bits == 5 ? 5 : 3);
-}
-
 template <typename T, int group_size, int bits>
 __global__ void
 affine_quantize(const T* w, uint8_t* out, T* scales, T* biases, size_t size) {
@@ -240,145 +227,102 @@ __global__ void affine_dequantize(
 }
 
 } // namespace cu
-namespace {
 
-inline array ensure_row_contiguous(
+void affine_quantize(
-    const array& x,
+    const array& w,
+    array& wq,
+    array& scales,
+    array& biases,
+    int group_size_,
+    int bits_,
     cu::CommandEncoder& enc,
     const Stream& s) {
-  if (!x.flags().row_contiguous) {
+  // Calculate the number of elements per thread
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+  int per_thread = group_size_ / WARP_SIZE;
-    copy_gpu(x, x_copy, CopyType::General, s);
+  size_t size = w.size() / per_thread;
-    enc.add_temporary(x_copy);
-    return x_copy;
-  } else {
-    return x;
-  }
-}
-
-} // namespace
-
-template <typename F>
-void dispatch_groups(int group_size, F&& f) {
-  switch (group_size) {
-    case 32:
-      f(std::integral_constant<int, 32>{});
-      break;
-    case 64:
-      f(std::integral_constant<int, 64>{});
-      break;
-    case 128:
-      f(std::integral_constant<int, 128>{});
-      break;
-  }
-}
-
-template <typename F>
-void dispatch_bits(int bits, F&& f) {
-  switch (bits) {
-    case 2:
-      f(std::integral_constant<int, 2>{});
-      break;
-    case 3:
-      f(std::integral_constant<int, 3>{});
-      break;
-    case 4:
-      f(std::integral_constant<int, 4>{});
-      break;
-    case 5:
-      f(std::integral_constant<int, 5>{});
-      break;
-    case 6:
-      f(std::integral_constant<int, 6>{});
-      break;
-    case 8:
-      f(std::integral_constant<int, 8>{});
-      break;
-  }
-}
-
-void fast::AffineQuantize::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  auto& w_pre = inputs[0];
-  auto& out = outputs[0];
-  out.set_data(allocator::malloc(out.nbytes()));
-
-  auto& s = stream();
-  auto& d = cu::device(s.device);
-  auto& enc = d.get_command_encoder(s);
-
-  auto w = ensure_row_contiguous(w_pre, enc, s);
-  enc.set_input_array(w);
-  if (dequantize_) {
-    auto scales = ensure_row_contiguous(inputs[1], enc, s);
-    auto biases = ensure_row_contiguous(inputs[2], enc, s);
-    enc.set_input_array(scales);
-    enc.set_input_array(biases);
-    enc.set_output_array(out);
-  } else {
-    auto& scales = outputs[1];
-    auto& biases = outputs[2];
-    scales.set_data(allocator::malloc(scales.nbytes()));
-    biases.set_data(allocator::malloc(biases.nbytes()));
-    enc.set_output_array(out);
-    enc.set_output_array(scales);
-    enc.set_output_array(biases);
-  }
-
-  auto dtype = dequantize_ ? outputs[0].dtype() : inputs[0].dtype();
-
-  // Treat uint32 as uint8 in kernel
-  int uint8_per_uint32 = 4;
-  int packs_per_int = (bits_ == 3 || bits_ == 5) ? 8
-      : bits_ == 6                               ? 4
-                                                 : 8 / bits_;
-  int per_thread = dequantize_ ? packs_per_int : group_size_ / WARP_SIZE;
-  size_t size =
-      dequantize_ ? out.size() / packs_per_int : w.size() / per_thread;
 
+  // Calculate the thread grid that we need to launch
   bool large = size > UINT_MAX;
   auto grid_shape = w.shape();
+  grid_shape.back() /= per_thread;
 
-  if (dequantize_) {
+  enc.set_input_array(w);
-    grid_shape.back() *= uint8_per_uint32;
+  enc.set_output_array(wq);
-  } else {
+  enc.set_output_array(scales);
-    grid_shape.back() /= per_thread;
+  enc.set_output_array(biases);
-  }
+  dispatch_float_types(w.dtype(), "affine_quantize", [&](auto type_tag) {
-
-  dispatch_float_types(dtype, "affine_quantize", [&](auto type_tag) {
     dispatch_groups(group_size_, [&](auto group_size) {
       dispatch_bits(bits_, [&](auto bits) {
-        using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
-        if (dequantize_) {
+        auto kernel = cu::affine_quantize<T, group_size.value, bits.value>;
-          auto kernel =
+        auto [num_blocks, block_dims] =
-              cu::affine_dequantize<DataType, group_size.value, bits.value>;
+            get_launch_args(kernel, size, grid_shape, w.strides(), large);
-          auto [num_blocks, block_dims] =
+        enc.add_kernel_node(
-              get_launch_args(kernel, size, grid_shape, w.strides(), large);
+            kernel,
-          enc.add_kernel_node(
+            num_blocks,
-              kernel,
+            block_dims,
-              num_blocks,
+            0,
-              block_dims,
+            w.data<T>(),
-              w.data<uint8_t>(),
+            wq.data<uint8_t>(),
-              inputs[1].data<DataType>(),
+            scales.data<T>(),
-              inputs[2].data<DataType>(),
+            biases.data<T>(),
-              out.data<DataType>(),
+            w.size());
-              out.size());
+      });
-        } else {
+    });
-          auto kernel =
+  });
-              cu::affine_quantize<DataType, group_size.value, bits.value>;
+}
-          auto [num_blocks, block_dims] =
+
-              get_launch_args(kernel, size, grid_shape, w.strides(), large);
+void affine_dequantize(
-          enc.add_kernel_node(
+    const array& wq,
-              kernel,
+    const array& scales,
-              num_blocks,
+    const array& biases,
-              block_dims,
+    array& w,
-              w.data<DataType>(),
+    int group_size_,
-              out.data<uint8_t>(),
+    int bits_,
-              outputs[1].data<DataType>(),
+    cu::CommandEncoder& enc,
-              outputs[2].data<DataType>(),
+    const Stream& s) {
-              w.size());
+  // Calculate how many numbers we pack together. For 2, 4, 8 bits we pack in
-        }
+  // one uint8, for 3, 6 in 3 uint8 and for 5 in 5 uint8.
+  constexpr int uint8_per_uint32 = 4;
+  int packs_per_int;
+  switch (bits_) {
+    case 3:
+    case 5:
+      packs_per_int = 8;
+      break;
+    case 6:
+      packs_per_int = 4;
+      break;
+    default:
+      packs_per_int = 8 / bits_;
+  }
+
+  size_t size = w.size() / packs_per_int;
+  bool large = size > UINT_MAX;
+  auto grid_shape = w.shape();
+  grid_shape.back() *= uint8_per_uint32;
+
+  enc.set_input_array(wq);
+  enc.set_input_array(scales);
+  enc.set_input_array(biases);
+  enc.set_output_array(w);
+  dispatch_float_types(w.dtype(), "affine_quantize", [&](auto type_tag) {
+    dispatch_groups(group_size_, [&](auto group_size) {
+      dispatch_bits(bits_, [&](auto bits) {
+        using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+        auto kernel = cu::affine_dequantize<T, group_size.value, bits.value>;
+        auto [num_blocks, block_dims] =
+            get_launch_args(kernel, size, grid_shape, w.strides(), large);
+        enc.add_kernel_node(
+            kernel,
+            num_blocks,
+            block_dims,
+            0,
+            wq.data<uint8_t>(),
+            scales.data<T>(),
+            biases.data<T>(),
+            w.data<T>(),
+            w.size());
       });
     });
   });

mlx/backend/cuda/quantized/qmm.cu

@@ -0,0 +1,228 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/matmul/mma.cuh"
+#include "mlx/backend/cuda/matmul/tiles.cuh"
+#include "mlx/backend/cuda/quantized/quantized_utils.cuh"
+#include "mlx/dtype_utils.h"
+
+namespace mlx::core {
+
+namespace cu {
+
+template <int NUM_WARPS, int group_size, int bits, typename T, typename Tile>
+__device__ inline void load_quantized(
+    Tile& tile,
+    const uint8_t* x,
+    const T* scales,
+    const T* biases,
+    int N) {
+  constexpr int NUM_THREADS = NUM_WARPS * 32;
+  constexpr int ELEMENTS_PER_LOAD = sizeof(uint32_t) * get_pack_factor<bits>();
+  constexpr int NUM_LOADS = Tile::NUMEL / ELEMENTS_PER_LOAD;
+  constexpr int NUM_LOADS_PER_THREAD = NUM_LOADS / NUM_THREADS;
+  constexpr int NUM_LOADS_PER_ROW = Tile::COLS / ELEMENTS_PER_LOAD;
+  constexpr int STEP_ROWS = NUM_THREADS / NUM_LOADS_PER_ROW;
+  constexpr int MASK = (1 << bits) - 1;
+
+  const int row = threadIdx.x / NUM_LOADS_PER_ROW;
+  const int col = threadIdx.x % NUM_LOADS_PER_ROW;
+
+  const int Nx = N / get_pack_factor<bits>();
+  const int Ng = N / group_size;
+
+  x += row * Nx + col * (ELEMENTS_PER_LOAD / get_pack_factor<bits>());
+  scales += row * Ng + col * ELEMENTS_PER_LOAD / group_size;
+  biases += row * Ng + col * ELEMENTS_PER_LOAD / group_size;
+
+  MLX_UNROLL
+  for (int i = 0; i < NUM_LOADS_PER_THREAD; i++) {
+    T vs[ELEMENTS_PER_LOAD];
+    uint32_t w = *reinterpret_cast<const uint32_t*>(x + i * STEP_ROWS * Nx);
+    T s = scales[i * STEP_ROWS * Ng];
+    T b = biases[i * STEP_ROWS * Ng];
+    MLX_UNROLL
+    for (int j = 0; j < ELEMENTS_PER_LOAD; j++) {
+      vs[j] = static_cast<T>((w >> (j * bits)) & MASK) * s + b;
+    }
+    tile.store(vs, row + i * STEP_ROWS, col * ELEMENTS_PER_LOAD);
+  }
+}
+
+template <
+    typename T,
+    int BM,
+    int BN,
+    int BK,
+    int group_size,
+    int bits,
+    bool aligned_M>
+__global__ void qmm_t(
+    const T* x,
+    const uint8_t* w,
+    const T* scales,
+    const T* biases,
+    T* y,
+    int M,
+    int N,
+    int K) {
+  constexpr int WARPS_M = 2;
+  constexpr int WARPS_N = 4;
+  constexpr int NUM_WARPS = WARPS_M * WARPS_N;
+  constexpr int WARP_STEP_M = BM / WARPS_M;
+  constexpr int WARP_STEP_N = BN / WARPS_N;
+
+  const int warpid = threadIdx.x / 32;
+  const int laneid = threadIdx.x % 32;
+  const int wm = warpid / WARPS_N;
+  const int wn = warpid % WARPS_N;
+  const int offset_m = wm * WARP_STEP_M;
+  const int offset_n = wn * WARP_STEP_N;
+
+  extern __shared__ char shmem[];
+  SharedTile<T, BM, BK>(&xs)[1] = *(SharedTile<T, BM, BK>(*)[1])(&shmem[0]);
+  SharedTile<T, BN, BK>(&ws)[1] =
+      *(SharedTile<T, BN, BK>(*)[1])(&shmem[1 * sizeof(T) * BM * BK]);
+
+  RegisterTile<float, BM / WARPS_M, BN / WARPS_N> C;
+  RegisterTile<T, BM / WARPS_M, 16> A;
+  RegisterTile<T, BN / WARPS_N, 16> B;
+
+  const int max_rows = M - blockIdx.y * BM;
+
+  x += blockIdx.y * BM * K;
+  w += blockIdx.x * BN * K / get_pack_factor<bits>();
+  scales += blockIdx.x * BN * K / group_size;
+  biases += blockIdx.x * BN * K / group_size;
+  y += blockIdx.y * BM * N + blockIdx.x * BN;
+
+  C.fill(0);
+
+  int tic = 0;
+  uint32_t base_addr_xs[1], base_addr_ws[1];
+  base_addr_xs[0] = __cvta_generic_to_shared(&xs[0].data[0]);
+  base_addr_ws[0] = __cvta_generic_to_shared(&ws[0].data[0]);
+
+  if (aligned_M || max_rows >= BM) {
+    for (int k_block = 0; k_block < K; k_block += BK) {
+      load_async<NUM_WARPS>(xs[tic], base_addr_xs[tic], x + k_block, K);
+      cp_async_commit();
+      load_quantized<NUM_WARPS, group_size, bits>(
+          ws[tic],
+          w + k_block / get_pack_factor<bits>(),
+          scales + k_block / group_size,
+          biases + k_block / group_size,
+          K);
+      cp_async_wait_all();
+      __syncthreads();
+
+      MLX_UNROLL
+      for (int k = 0; k < BK / 16; k++) {
+        A.load(
+            xs[tic],
+            base_addr_xs[tic],
+            offset_m + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        B.load(
+            ws[tic],
+            base_addr_ws[tic],
+            offset_n + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        mma_t(C, A, B);
+      }
+    }
+    C.store_global(y, N, offset_m, offset_n);
+  } else {
+    for (int k_block = 0; k_block < K; k_block += BK) {
+      load_async_safe<NUM_WARPS>(
+          xs[tic], base_addr_xs[tic], x + k_block, K, max_rows);
+      cp_async_commit();
+      load_quantized<NUM_WARPS, group_size, bits>(
+          ws[tic],
+          w + k_block / get_pack_factor<bits>(),
+          scales + k_block / group_size,
+          biases + k_block / group_size,
+          K);
+      cp_async_wait_all();
+      __syncthreads();
+
+      MLX_UNROLL
+      for (int k = 0; k < BK / 16; k++) {
+        A.load(
+            xs[tic],
+            base_addr_xs[tic],
+            offset_m + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        B.load(
+            ws[tic],
+            base_addr_ws[tic],
+            offset_n + laneid % 16,
+            k * 16 + laneid / 16 * 8);
+        mma_t(C, A, B);
+      }
+    }
+    C.store_global_safe(y, N, offset_m, offset_n, max_rows);
+  }
+}
+
+} // namespace cu
+
+void qmm(
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& biases,
+    array& out,
+    bool transpose_,
+    int group_size_,
+    int bits_,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  if (x.dtype() != bfloat16) {
+    throw std::invalid_argument("[qmm] Only bfloat16 is supported for now");
+  }
+  if (!transpose_) {
+    throw std::invalid_argument(
+        "[qmm] Only transposed matmul is supported for now");
+  }
+
+  dispatch_float_types(x.dtype(), "qmm", [&](auto type_tag) {
+    dispatch_groups(group_size_, [&](auto group_size) {
+      dispatch_bits(bits_, [&](auto bits) {
+        using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
+
+        constexpr int BM = 128;
+        constexpr int BN = 128;
+        constexpr int BK = 32;
+        auto kernel =
+            cu::qmm_t<DataType, BM, BN, BK, group_size.value, bits.value, true>;
+        if (M % BM != 0) {
+          kernel = cu::
+              qmm_t<DataType, BM, BN, BK, group_size.value, bits.value, false>;
+        }
+
+        dim3 grid((N + BN - 1) / BN, (M + BM - 1) / BM);
+
+        enc.add_kernel_node(
+            kernel,
+            grid,
+            2 * 4 * 32,
+            1 * sizeof(DataType) * (BM * BK + BN * BK),
+            x.data<DataType>(),
+            w.data<uint8_t>(),
+            scales.data<DataType>(),
+            biases.data<DataType>(),
+            out.data<DataType>(),
+            M,
+            N,
+            K);
+      });
+    });
+  });
+}
+
+} // namespace mlx::core

mlx/backend/cuda/quantized/quantized.cu

@@ -0,0 +1,113 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/backend/cuda/quantized/quantized.cuh"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/fast_primitives.h"
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace {
+
+inline array ensure_row_contiguous(
+    const array& x,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  if (!x.flags().row_contiguous) {
+    array x_copy = contiguous_copy_gpu(x, s);
+    enc.add_temporary(x_copy);
+    return x_copy;
+  } else {
+    return x;
+  }
+}
+
+inline array ensure_row_contiguous_matrix(
+    const array& x,
+    cu::CommandEncoder& enc,
+    const Stream& s) {
+  auto stride_0 = x.strides()[x.ndim() - 2];
+  auto stride_1 = x.strides()[x.ndim() - 1];
+  if (stride_0 == x.shape(-1) && stride_1 == 1) {
+    return x;
+  } else {
+    array x_copy = contiguous_copy_gpu(x, s);
+    enc.add_temporary(x_copy);
+    return x_copy;
+  }
+}
+
+} // namespace
+
+void QuantizedMatmul::eval_gpu(const std::vector<array>& inputs, array& out) {
+  auto& s = stream();
+  auto& d = cu::device(s.device);
+  auto& enc = d.get_command_encoder(s);
+
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  // Make sure the last two dims of x and w, s, b are contiguous. This should
+  // be relaxed for x.
+  array x = ensure_row_contiguous_matrix(inputs[0], enc, s);
+  array w = ensure_row_contiguous_matrix(inputs[1], enc, s);
+  array scales = ensure_row_contiguous_matrix(inputs[2], enc, s);
+  array biases = ensure_row_contiguous_matrix(inputs[3], enc, s);
+
+  // Extract the matmul shapes
+  bool non_batched = w.ndim() == 2 && x.flags().row_contiguous;
+  int K = x.shape(-1);
+  int M = non_batched ? x.size() / K : x.shape(-2);
+  int N = out.shape(-1);
+
+  qmm(x,
+      w,
+      scales,
+      biases,
+      out,
+      transpose_,
+      group_size_,
+      bits_,
+      M,
+      N,
+      K,
+      enc,
+      s);
+}
+
+void fast::AffineQuantize::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& s = stream();
+  auto& d = cu::device(s.device);
+  auto& enc = d.get_command_encoder(s);
+
+  if (dequantize_) {
+    auto wq = ensure_row_contiguous(inputs[0], enc, s);
+    auto scales = ensure_row_contiguous(inputs[1], enc, s);
+    auto biases = ensure_row_contiguous(inputs[2], enc, s);
+    auto& w = outputs[0];
+
+    w.set_data(allocator::malloc(w.nbytes()));
+
+    affine_dequantize(wq, scales, biases, w, group_size_, bits_, enc, s);
+  } else {
+    auto w = ensure_row_contiguous(inputs[0], enc, s);
+    auto& wq = outputs[0];
+    auto& scales = outputs[1];
+    auto& biases = outputs[2];
+
+    wq.set_data(allocator::malloc(wq.nbytes()));
+    scales.set_data(allocator::malloc(scales.nbytes()));
+    biases.set_data(allocator::malloc(biases.nbytes()));
+
+    affine_quantize(w, wq, scales, biases, group_size_, bits_, enc, s);
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/quantized/quantized.cuh

@@ -0,0 +1,42 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core {
+
+void affine_quantize(
+    const array& w,
+    array& wq,
+    array& scales,
+    array& biases,
+    int group_size_,
+    int bits_,
+    cu::CommandEncoder& enc,
+    const Stream& s);
+
+void affine_dequantize(
+    const array& wq,
+    const array& scales,
+    const array& biases,
+    array& w,
+    int group_size_,
+    int bits_,
+    cu::CommandEncoder& enc,
+    const Stream& s);
+
+void qmm(
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& biases,
+    array& out,
+    bool transpose_,
+    int group_size_,
+    int bits_,
+    int M,
+    int N,
+    int K,
+    cu::CommandEncoder& enc,
+    const Stream& s);
+
+} // namespace mlx::core

mlx/backend/cuda/quantized/quantized_utils.cuh

@@ -0,0 +1,59 @@
+// Copyright © 2025 Apple Inc.
+
+namespace mlx::core {
+
+namespace cu {
+
+template <int bits, int wsize = 8>
+inline constexpr __device__ short get_pack_factor() {
+  return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
+}
+
+template <int bits, int wsize = 8>
+inline constexpr __device__ short get_bytes_per_pack() {
+  constexpr int power_of_2_bits = (bits & (bits - 1)) == 0;
+  return power_of_2_bits ? (wsize / 8) : (bits == 5 ? 5 : 3);
+}
+
+} // namespace cu
+
+template <typename F>
+void dispatch_groups(int group_size, F&& f) {
+  switch (group_size) {
+    case 32:
+      f(std::integral_constant<int, 32>{});
+      break;
+    case 64:
+      f(std::integral_constant<int, 64>{});
+      break;
+    case 128:
+      f(std::integral_constant<int, 128>{});
+      break;
+  }
+}
+
+template <typename F>
+void dispatch_bits(int bits, F&& f) {
+  switch (bits) {
+    case 2:
+      f(std::integral_constant<int, 2>{});
+      break;
+    case 3:
+      f(std::integral_constant<int, 3>{});
+      break;
+    case 4:
+      f(std::integral_constant<int, 4>{});
+      break;
+    case 5:
+      f(std::integral_constant<int, 5>{});
+      break;
+    case 6:
+      f(std::integral_constant<int, 6>{});
+      break;
+    case 8:
+      f(std::integral_constant<int, 8>{});
+      break;
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/random.cu

@@ -170,6 +170,7 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
         cu::rbitsc,
         grid,
         block,
+        0,
         keys.data<uint32_t>(),
         out.data<uint8_t>(),
         grid_dims,
@@ -180,6 +181,7 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
         cu::rbits,
         grid,
         block,
+        0,
         keys.data<uint32_t>(),
         out.data<uint8_t>(),
         grid_dims,

mlx/backend/cuda/reduce.cu

@@ -47,8 +47,7 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
     }
   }
   if (plan.type == GeneralReduce || broadcasted || !in.flags().contiguous) {
-    array in_copy(in.shape(), in.dtype(), nullptr, {});
+    array in_copy = contiguous_copy_gpu(in, s);
-    copy_gpu(in, in_copy, CopyType::General, s);
     encoder.add_temporary(in_copy);
     in = in_copy;
     plan = get_reduction_plan(in, axes_);

mlx/backend/cuda/reduce/all_reduce.cu

@@ -120,6 +120,7 @@ void all_reduce(
             kernel,
             blocks,
             threads,
+            0,
             static_cast<T*>(indata),
             intermediate.data<U>(),
             block_step,
@@ -146,6 +147,7 @@ void all_reduce(
           kernel,
           blocks,
           threads,
+          0,
           static_cast<T*>(indata),
           out.data<U>(),
           block_step,

mlx/backend/cuda/reduce/col_reduce.cu

@@ -230,7 +230,7 @@ void col_reduce_looped(
         auto kernel =
             cu::col_reduce_looped<T, U, OP, reduce_ndim(), BM, BN, N_READS>;
         encoder.add_kernel_node(
-            kernel, grid, blocks, indata, out.data<U>(), args);
+            kernel, grid, blocks, 0, indata, out.data<U>(), args);
       });
     });
   });

mlx/backend/cuda/reduce/init_reduce.cu

@@ -41,7 +41,8 @@ void init_reduce(
       dim3 grid = get_2d_grid_dims(out.shape(), out.strides());
       dim3 block(grid.x < 1024 ? grid.x : 1024, 1, 1);
       grid.x = (grid.x + 1023) / 1024;
-      encoder.add_kernel_node(kernel, grid, block, out.data<U>(), out.size());
+      encoder.add_kernel_node(
+          kernel, grid, block, 0, out.data<U>(), out.size());
     });
   });
 }

mlx/backend/cuda/reduce/row_reduce.cu

@@ -269,7 +269,7 @@ void row_reduce_simple(
 
       int size = plan.shape.back();
       encoder.add_kernel_node(
-          kernel, grid, block, indata, out.data<U>(), out.size(), size);
+          kernel, grid, block, 0, indata, out.data<U>(), out.size(), size);
     });
   });
 }
@@ -322,7 +322,7 @@ void row_reduce_looped(
       });
 
       encoder.add_kernel_node(
-          kernel, grid, block, indata, out.data<U>(), out.size(), args);
+          kernel, grid, block, 0, indata, out.data<U>(), out.size(), args);
     });
   });
 }

mlx/backend/cuda/rms_norm.cu

@@ -206,8 +206,7 @@ void RMSNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -233,6 +232,7 @@ void RMSNorm::eval_gpu(
           kernel,
           n_rows,
           block_dim(),
+          0,
           x.data<DataType>(),
           w.data<DataType>(),
           out.data<DataType>(),
@@ -259,9 +259,7 @@ void RMSNormVJP::eval_gpu(
       return x;
     }
     copied = true;
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    return contiguous_copy_gpu(x, s);
-    copy_gpu(x, x_copy, CopyType::General, s);
-    return x_copy;
   };
   bool donate_x = inputs[0].is_donatable();
   bool donate_g = inputs[2].is_donatable();
@@ -330,6 +328,7 @@ void RMSNormVJP::eval_gpu(
                 kernel,
                 n_rows,
                 block_dim(),
+                0,
                 x.data<DataType>(),
                 w.data<DataType>(),
                 g.data<DataType>(),

mlx/backend/cuda/rope.cu

@@ -325,6 +325,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),
@@ -341,6 +342,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),
@@ -360,6 +362,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),
@@ -381,6 +384,7 @@ void RoPE::eval_gpu(
               kernel,
               grid,
               block,
+              0,
               (donated ? out : in).data<DataType>(),
               out.data<DataType>(),
               offset.data<int32_t>(),

mlx/backend/cuda/scan.cu

@@ -379,9 +379,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
           in.flags());
     }
   } else {
-    array arr_copy(in.shape(), in.dtype(), nullptr, {});
+    in = contiguous_copy_gpu(in, s);
-    copy_gpu(in, arr_copy, CopyType::General, s);
-    in = std::move(arr_copy);
     out.copy_shared_buffer(in);
   }
 
@@ -416,6 +414,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
                   kernel,
                   in.data_size() / axis_size,
                   block_dim,
+                  0,
                   in.data<T>(),
                   out.data<U>(),
                   axis_size);
@@ -445,6 +444,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
                   kernel,
                   num_blocks,
                   block_dim,
+                  0,
                   in.data<T>(),
                   out.data<U>(),
                   axis_size,

mlx/backend/cuda/softmax.cu

@@ -125,8 +125,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -153,6 +152,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
           kernel,
           n_rows,
           block_dim(),
+          0,
           in.data<DataType>(),
           out.data<DataType>(),
           axis_size);

mlx/backend/cuda/sort.cu

@@ -72,8 +72,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
   bool is_segmented_sort = in.flags().contiguous && in.strides()[axis] == 1;
   if (!is_segmented_sort) {
     array trans = swapaxes_in_eval(in, axis, last_dim);
-    in = array(trans.shape(), trans.dtype(), nullptr, {});
+    in = contiguous_copy_gpu(trans, s);
-    copy_gpu(trans, in, CopyType::General, s);
     encoder.add_temporary(in);
     out = array(allocator::malloc(out.nbytes()), in.shape(), out.dtype());
     encoder.add_temporary(out);

mlx/backend/cuda/ternary.cu

@@ -133,6 +133,7 @@ void ternary_op_gpu_inplace(
                     kernel,
                     num_blocks,
                     block_dims,
+                    0,
                     a.data<bool>(),
                     b.data<DType>(),
                     c.data<DType>(),
@@ -151,6 +152,7 @@ void ternary_op_gpu_inplace(
                   kernel,
                   num_blocks,
                   block_dims,
+                  0,
                   a.data<bool>(),
                   b.data<DType>(),
                   c.data<DType>(),
@@ -180,6 +182,7 @@ void ternary_op_gpu_inplace(
             kernel,
             num_blocks,
             block_dims,
+            0,
             a.data<bool>(),
             b.data<DType>(),
             c.data<DType>(),

mlx/backend/cuda/unary.cu

@@ -142,6 +142,7 @@ void unary_op_gpu_inplace(
                 kernel,
                 num_blocks,
                 block_dims,
+                0,
                 in.data<InType>(),
                 out.data<OutType>(),
                 out.data_size());
@@ -154,6 +155,7 @@ void unary_op_gpu_inplace(
                 kernel,
                 num_blocks,
                 block_dims,
+                0,
                 in.data<InType>(),
                 out.data<OutType>(),
                 out.data_size(),

mlx/backend/gpu/copy.cpp

@@ -46,4 +46,10 @@ void copy_gpu_inplace(
       in, out, in.shape(), i_strides, out.strides(), i_offset, 0, ctype, s);
 }
 
+array contiguous_copy_gpu(const array& arr, const Stream& s) {
+  array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+  copy_gpu(arr, arr_copy, CopyType::General, s);
+  return arr_copy;
+}
+
 } // namespace mlx::core

mlx/backend/gpu/copy.h

@@ -43,4 +43,7 @@ void copy_gpu_inplace(
 // Fill the output with the scalar val
 void fill_gpu(const array& val, array& out, const Stream& s);
 
+// Return a contiguous array with same shape that copies the data of |arr|.
+array contiguous_copy_gpu(const array& arr, const Stream& s);
+
 } // namespace mlx::core

mlx/backend/metal/conv.cpp

@@ -149,8 +149,7 @@ void explicit_gemm_conv_group_ND_gpu(
       wt, {wt.strides(0), 1, C_per_group}, wt.flags(), wt.size());
 
   // Materialize
-  auto wt_transpose = array(wt_view.shape(), wt_view.dtype(), nullptr, {});
+  array wt_transpose = contiguous_copy_gpu(wt_view, s);
-  copy_gpu(wt_view, wt_transpose, CopyType::General, s);
 
   // Perform gemm
   std::vector<array> copies = {in_unfolded, wt_transpose};
@@ -961,16 +960,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto in = inputs[0];
   auto wt = inputs[1];
   if (!in.flags().row_contiguous) {
-    array arr_copy(in.shape(), in.dtype(), nullptr, {});
+    in = contiguous_copy_gpu(in, s);
-    copy_gpu(in, arr_copy, CopyType::General, s);
+    copies.push_back(in);
-    copies.push_back(arr_copy);
-    in = arr_copy;
   }
   if (!wt.flags().row_contiguous) {
-    array arr_copy(wt.shape(), wt.dtype(), nullptr, {});
+    wt = contiguous_copy_gpu(wt, s);
-    copy_gpu(wt, arr_copy, CopyType::General, s);
+    copies.push_back(wt);
-    copies.push_back(arr_copy);
-    wt = arr_copy;
   }
 
   // 3D conv

mlx/backend/metal/logsumexp.cpp

@@ -25,8 +25,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
     if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       d.add_temporary(x_copy, s.index);
       return x_copy;
     }

mlx/backend/metal/matmul.cpp

@@ -33,8 +33,7 @@ std::tuple<bool, int64_t, array> check_transpose(
   } else if (stx == 1 && (!is_vector || sty == arr.shape(-2))) {
     return std::make_tuple(true, sty, arr);
   } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+    array arr_copy = contiguous_copy_gpu(arr, s);
-    copy_gpu(arr, arr_copy, CopyType::General, s);
     copies.push_back(arr_copy);
     return std::make_tuple(false, arr.shape(-1), arr_copy);
   }
@@ -43,8 +42,7 @@ std::tuple<bool, int64_t, array> check_transpose(
 inline array
 ensure_row_contiguous(const array& x, metal::Device& d, const Stream& s) {
   if (!x.flags().row_contiguous) {
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    array x_copy = contiguous_copy_gpu(x, s);
-    copy_gpu(x, x_copy, CopyType::General, s);
     d.add_temporary(x_copy, s.index);
     return x_copy;
   } else {
@@ -75,8 +73,7 @@ ensure_batch_contiguous(const array& x, metal::Device& d, const Stream& s) {
     }
   }
 
-  array x_copy(x.shape(), x.dtype(), nullptr, {});
+  array x_copy = contiguous_copy_gpu(x, s);
-  copy_gpu(x, x_copy, CopyType::General, s);
   d.add_temporary(x_copy, s.index);
   return std::make_tuple(false, x_copy.strides()[x_copy.ndim() - 2], x_copy);
 }
@@ -1894,8 +1891,7 @@ void segmented_mm(
       return std::make_tuple(false, x);
     }
 
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    array x_copy = contiguous_copy_gpu(x, s);
-    copy_gpu(x, x_copy, CopyType::General, s);
     d.add_temporary(x_copy, s.index);
     return std::make_tuple(true, x_copy);
   };

mlx/backend/metal/normalization.cpp

@@ -40,8 +40,7 @@ void RMSNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -107,9 +106,7 @@ void RMSNormVJP::eval_gpu(
     if (x.flags().row_contiguous) {
       return {x, false};
     }
-
+    array x_copy = contiguous_copy_gpu(x, s);
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
-    copy_gpu(x, x_copy, CopyType::General, s);
     return {x_copy, true};
   };
   bool donate_x = inputs[0].is_donatable();
@@ -241,8 +238,7 @@ void LayerNorm::eval_gpu(
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }
@@ -319,8 +315,7 @@ void LayerNormVJP::eval_gpu(
     if (x.flags().row_contiguous) {
       return {x, false};
     }
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    array x_copy = contiguous_copy_gpu(x, s);
-    copy_gpu(x, x_copy, CopyType::General, s);
     return {x_copy, true};
   };
   bool donate_x = inputs[0].is_donatable();

mlx/backend/metal/quantized.cpp

@@ -20,8 +20,7 @@ namespace {
 inline array
 ensure_row_contiguous(const array& x, metal::Device& d, const Stream& s) {
   if (!x.flags().row_contiguous) {
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    array x_copy = contiguous_copy_gpu(x, s);
-    copy_gpu(x, x_copy, CopyType::General, s);
     d.add_temporary(x_copy, s.index);
     return x_copy;
   } else {
@@ -38,8 +37,7 @@ inline array ensure_row_contiguous_matrix(
   if (stride_0 == x.shape(-1) && stride_1 == 1) {
     return x;
   } else {
-    array x_copy(x.shape(), x.dtype(), nullptr, {});
+    array x_copy = contiguous_copy_gpu(x, s);
-    copy_gpu(x, x_copy, CopyType::General, s);
     d.add_temporary(x_copy, s.index);
     return x_copy;
   }

mlx/backend/metal/reduce.cpp

@@ -989,8 +989,7 @@ void Reduce::eval_gpu(const std::vector<array>& inputs, array& out) {
     //       input for the axes with stride smaller than the minimum reduction
     //       stride.
     if (plan.type == GeneralReduce) {
-      array in_copy(in.shape(), in.dtype(), nullptr, {});
+      array in_copy = contiguous_copy_gpu(in, s);
-      copy_gpu(in, in_copy, CopyType::General, s);
       d.add_temporary(in_copy, s.index);
       in = in_copy;
       plan = get_reduction_plan(in, axes_);

mlx/backend/metal/scaled_dot_product_attention.cpp

@@ -398,8 +398,7 @@ void ScaledDotProductAttention::eval_gpu(
   auto copy_unless = [&copies, &s](
                          auto predicate, const array& arr) -> const array& {
     if (!predicate(arr)) {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+      array arr_copy = contiguous_copy_gpu(arr, s);
-      copy_gpu(arr, arr_copy, CopyType::General, s);
       copies.push_back(std::move(arr_copy));
       return copies.back();
     } else {

mlx/backend/metal/scan.cpp

@@ -30,9 +30,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
           in.flags());
     }
   } else {
-    array arr_copy(in.shape(), in.dtype(), nullptr, {});
+    in = contiguous_copy_gpu(in, s);
-    copy_gpu(in, arr_copy, CopyType::General, s);
-    in = std::move(arr_copy);
     out.copy_shared_buffer(in);
   }
 

mlx/backend/metal/softmax.cpp

@@ -35,8 +35,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
       }
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
+      array x_copy = contiguous_copy_gpu(x, s);
-      copy_gpu(x, x_copy, CopyType::General, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }

mlx/version.h

@@ -4,7 +4,7 @@
 
 #define MLX_VERSION_MAJOR 0
 #define MLX_VERSION_MINOR 26
-#define MLX_VERSION_PATCH 3
+#define MLX_VERSION_PATCH 5
 #define MLX_VERSION_NUMERIC \
   (100000 * MLX_VERSION_MAJOR + 1000 * MLX_VERSION_MINOR + MLX_VERSION_PATCH)
 

python/mlx/optimizers/optimizers.py

@@ -848,6 +848,106 @@ class Adafactor(Optimizer):
         return parameter - update
 
 
+class Muon(Optimizer):
+    r"""The Muon optimizer.
+
+    Our Muon (MomentUm Orthogonalized by Newton-schulz) optimizer follows the
+    original implementation: `Muon: An optimizer for hidden layers in neural
+    networks <https://kellerjordan.github.io/posts/muon/>`_
+
+    Note:
+        - Muon may be sub-optimal for the embedding layer, the final fully
+          connected layer, or any 0D/1D parameters. Those should be optimized
+          by a different method (e.g., :class:`AdamW`).
+        - For 4D convolutional filters, it works by flattening their last
+          dimensions.
+
+    Args:
+        learning_rate (float or callable): The learning rate.
+        momentum (float, optional): The momentum strength. Default: ``0.95``
+        weight_decay (float, optional): The weight decay (L2 penalty).
+            Default: ``0.01``
+        nesterov (bool, optional): Enables Nesterov momentum. Recommended for
+            better performance.  Default: ``True``
+        ns_steps (int, optional): Number of Newton-Schulz iteration steps for
+            orthogonalization.  Default: ``5``
+    """
+
+    def __init__(
+        self,
+        learning_rate: Union[float, Callable[[mx.array], mx.array]],
+        momentum: float = 0.95,
+        weight_decay: float = 0.01,
+        nesterov: bool = True,
+        ns_steps: int = 5,
+    ):
+        super().__init__()
+
+        self._maybe_schedule("learning_rate", learning_rate)
+        self.momentum = momentum
+        self.weight_decay = weight_decay
+        self.nesterov = nesterov
+        self.ns_steps = ns_steps
+
+    def init_single(self, parameter: mx.array, state: dict):
+        """Initialize optimizer state"""
+        state["v"] = mx.zeros_like(parameter)
+
+    def _zeropower_via_newtonschulz5(self, X, steps: int):
+        assert (
+            X.ndim == 2
+        ), f"Expected a 2D array for Newton-Schulz iteration, got shape {X.shape} instead."
+        a, b, c = (3.4445, -4.7750, 2.0315)
+        transpose_needed = X.shape[-2] > X.shape[-1]
+
+        if transpose_needed:
+            X = X.T
+
+        X = X / (mx.linalg.norm(X, keepdims=True) + 1e-7)
+
+        for _ in range(steps):
+            A = X @ X.T
+            B = mx.addmm(b * A, A, A, beta=1.0, alpha=c)
+            X = mx.addmm(a * X, B, X, beta=1.0, alpha=1.0)
+
+        if transpose_needed:
+            X = X.T
+        return X
+
+    def apply_single(self, gradient: mx.array, parameter: mx.array, state: dict):
+        """Performs the Muon parameter update"""
+
+        if self.weight_decay != 0:
+            gradient = gradient + self.weight_decay * parameter
+
+        v = self.momentum * state["v"]
+        v = v + (1 - self.momentum) * gradient
+        state["v"] = v
+
+        if self.nesterov:
+            update = gradient * (1 - self.momentum) + v * self.momentum
+        else:
+            update = v
+
+        lr = self.learning_rate.astype(gradient.dtype)
+
+        if update.ndim >= 2:
+            original_shape = update.shape
+            reshape_needed = update.ndim > 2
+
+            if reshape_needed:
+                update = mx.reshape(update, (update.shape[0], -1))
+
+            update = self._zeropower_via_newtonschulz5(update, steps=self.ns_steps)
+
+            if reshape_needed:
+                update = mx.reshape(update, original_shape)
+
+            lr *= max(1, update.shape[-2] / update.shape[-1]) ** 0.5
+
+        return parameter - lr * update
+
+
 def clip_grad_norm(grads, max_norm):
     """Clips the global norm of the gradients.
 

python/scripts/repair_cuda.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
 auditwheel repair dist/* \
-  --plat manylinux_2_39_x86_64 \
+  --plat manylinux_2_35_x86_64 \
   --exclude libcublas* \
   --exclude libnvrtc* \
   -w wheel_tmp

python/tests/test_blas.py

@@ -691,6 +691,21 @@ class TestBlas(mlx_tests.MLXTestCase):
             self.assertListEqual(list(d_npy.shape), list(d_mlx.shape))
             self.assertTrue(np.allclose(d_mlx, d_npy, atol=1e-5))
 
+        # Transposed c
+        a = mx.ones((10, 5)).T
+        b = mx.ones((5, 5))
+        out = mx.addmm(a, b, a, beta=1.5, alpha=0.5)
+        expected = 1.5 * a + 0.5 * (b @ a)
+        self.assertTrue(mx.allclose(expected, out))
+
+        # Broadcast c
+        a = mx.ones((5, 5))
+        b = mx.ones((5, 5))
+        c = mx.ones((1, 5))
+        out = mx.addmm(c, a, b, beta=1.5, alpha=0.5)
+        expected = 1.5 * c + 0.5 * (a @ b)
+        self.assertTrue(mx.allclose(expected, out))
+
     def test_addmm_grad(self):
         def make_ref_addmm(alpha, beta):
             return lambda c, a, b: alpha * (a @ b) + beta * c

python/tests/test_optimizers.py

@@ -286,6 +286,53 @@ class TestOptimizers(mlx_tests.MLXTestCase):
             self.assertEqual(xp["x"].shape, x.shape)
         self.assertEqual(optimizer.state["step"], 2)
 
+    def test_muon(self):
+        params = {
+            "first": [mx.zeros((10, 5)), mx.zeros((1,))],
+            "second": mx.zeros((3, 3)),
+            "conv": mx.zeros((16, 8, 3, 3)),
+        }
+        grads = tree_map(lambda x: mx.ones_like(x), params)
+
+        # Explicit init
+        optim = opt.Muon(learning_rate=1e-2, momentum=0.95, nesterov=True)
+        optim.init(params)
+        self.assertTrue(
+            tree_equal(
+                lambda p, s: mx.array_equal(s["v"], mx.zeros_like(p)),
+                params,
+                optim.state,
+            )
+        )
+
+        # Test update
+        updated_params = optim.apply_gradients(grads, params)
+
+        # Check that shapes are preserved
+        self.assertTrue(
+            tree_equal(
+                lambda p, u: p.shape == u.shape,
+                params,
+                updated_params,
+            )
+        )
+
+        # Check that parameters actually changed
+        self.assertFalse(
+            tree_equal(
+                lambda p, u: mx.array_equal(p, u),
+                params,
+                updated_params,
+            )
+        )
+
+        # Test with different configurations
+        optim_no_nesterov = opt.Muon(learning_rate=1e-2, momentum=0.95, nesterov=False)
+        optim_no_nesterov.apply_gradients(grads, params)
+
+        optim_no_momentum = opt.Muon(learning_rate=1e-2, momentum=0.0)
+        optim_no_momentum.apply_gradients(grads, params)
+
     def test_compiled_optimizer(self):
         model = nn.Linear(10, 10)
         x = mx.random.uniform(shape=(2, 10))

tests/CMakeLists.txt

@@ -39,6 +39,14 @@ target_sources(
           linalg_tests.cpp
           ${METAL_TEST_SOURCES})
 
+if(MLX_BUILD_CUDA)
+  # Find the CCCL headers in install dir.
+  target_compile_definitions(
+    mlx
+    PRIVATE
+      MLX_CCCL_DIR="${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_INCLUDEDIR}/cccl")
+endif()
+
 target_link_libraries(tests PRIVATE mlx doctest)
 doctest_discover_tests(tests)
 add_test(NAME tests COMMAND tests)