Fix four step fft

* fix bw for elementwise ops

* add compile

* fix

* fix

* fix

* fix

.gitignore

@@ -36,6 +36,7 @@ share/python-wheels/
 .installed.cfg
 *.egg
 MANIFEST
+uv.lock
 
 # vim
 *.swp

CMakeLists.txt

@@ -34,6 +34,7 @@ option(MLX_BUILD_BENCHMARKS "Build benchmarks for mlx" OFF)
 option(MLX_BUILD_PYTHON_BINDINGS "Build python bindings for mlx" OFF)
 option(MLX_BUILD_METAL "Build metal backend" ON)
 option(MLX_BUILD_CPU "Build cpu backend" ON)
+option(MLX_BUILD_CUDA "Build cuda backend" OFF)
 option(MLX_METAL_DEBUG "Enhance metal debug workflow" OFF)
 option(MLX_ENABLE_X64_MAC "Enable building for x64 macOS" OFF)
 option(MLX_BUILD_GGUF "Include support for GGUF format" ON)
@@ -83,6 +84,10 @@ if(MLX_BUILD_METAL)
   set(QUARTZ_LIB "-framework QuartzCore")
 endif()
 
+if(MLX_BUILD_CUDA)
+  enable_language(CUDA)
+endif()
+
 if(MLX_BUILD_METAL AND NOT METAL_LIB)
   message(STATUS "Metal not found. Unable to build GPU")
   set(MLX_BUILD_METAL OFF)

MANIFEST.in

@@ -1,4 +1,6 @@
 include CMakeLists.txt
+include mlx.pc.in
 recursive-include mlx/ *
+include cmake/*
 include python/src/*
 include python/mlx/py.typed # support type hinting as in PEP-561

docs/src/python/fft.rst

@@ -20,3 +20,5 @@ FFT
   irfft2
   rfftn
   irfftn
+  fftshift
+  ifftshift

mlx/CMakeLists.txt

@@ -49,5 +49,16 @@ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/io)
 if(MLX_BUILD_METAL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/metal)
 else()
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_metal)
+  target_sources(mlx
+                 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/metal/no_metal.cpp)
+endif()
+
+if(MLX_BUILD_CUDA)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda)
+endif()
+
+if(MLX_BUILD_METAL OR MLX_BUILD_CUDA)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/gpu)
+else()
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/no_gpu)
 endif()

mlx/array.h

@@ -356,7 +356,7 @@ class array {
   }
 
   enum Status {
-    // The ouptut of a computation which has not been scheduled.
+    // The output of a computation which has not been scheduled.
     // For example, the status of `x` in `auto x = a + b`.
     unscheduled,
 

mlx/backend/common/CMakeLists.txt

@@ -6,4 +6,5 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/transpose.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp)

mlx/backend/common/common.cpp

@@ -2,6 +2,7 @@
 #include <cassert>
 
 #include "mlx/backend/common/broadcasting.h"
+#include "mlx/backend/common/transpose.h"
 #include "mlx/backend/common/utils.h"
 #include "mlx/primitives.h"
 
@@ -19,26 +20,19 @@ void AsStrided::eval(const std::vector<array>& inputs, array& out) {
         "AsStrided must be used with row contiguous arrays only.");
   }
 
-  // Compute the flags given the shape and strides
-  bool row_contiguous = true, col_contiguous = true;
-  size_t r = 1, c = 1;
-  for (int i = strides_.size() - 1, j = 0; i >= 0; i--, j++) {
-    row_contiguous &= (r == strides_[i]) || (shape_[i] == 1);
-    col_contiguous &= (c == strides_[j]) || (shape_[j] == 1);
-    r *= shape_[i];
-    c *= shape_[j];
-  }
+  // Calculate the contiguity based on the given shape and strides
+  auto [ds, rc, cc] = check_contiguity(shape_, strides_);
   auto flags = in.flags();
+
   // TODO: Compute the contiguous flag in a better way cause now we are
   //       unnecessarily strict.
-  flags.contiguous = row_contiguous || col_contiguous;
-  flags.row_contiguous = row_contiguous;
-  flags.col_contiguous = col_contiguous;
+  flags.contiguous = rc || cc;
+  flags.row_contiguous = rc;
+  flags.col_contiguous = cc;
 
-  // There is no easy way to compute the actual data size so we use out.size().
-  // The contiguous flag will almost certainly not be set so no code should
-  // rely on data_size anyway.
-  size_t data_size = out.size();
+  // There is no easy way to compute the actual data size so we use out.size()
+  // when the array is not contiguous.
+  size_t data_size = flags.contiguous ? ds : out.size();
 
   return out.copy_shared_buffer(in, strides_, flags, data_size, offset_);
 }
@@ -270,36 +264,7 @@ void StopGradient::eval(const std::vector<array>& inputs, array& out) {
 
 void Transpose::eval(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
-  Strides out_strides(out.ndim());
-  auto& in = inputs[0];
-  for (int ax = 0; ax < axes_.size(); ++ax) {
-    out_strides[ax] = in.strides()[axes_[ax]];
-  }
-
-  // Conditions for {row/col}_contiguous
-  // - array must be contiguous (no gaps)
-  // - underlying buffer size should have the same size as the array
-  // - cumulative product of shapes is equal to the strides (we can ignore axes
-  //   with size == 1)
-  //   - in the forward direction (column contiguous)
-  //   - in the reverse direction (row contiguous)
-  // - vectors are both row and col contiguous (hence if both row/col are
-  //   true, they stay true)
-  auto flags = in.flags();
-  if (flags.contiguous && in.data_size() == in.size()) {
-    int64_t f_stride = 1;
-    int64_t b_stride = 1;
-    flags.col_contiguous = true;
-    flags.row_contiguous = true;
-    for (int i = 0, ri = out.ndim() - 1; i < out.ndim(); ++i, --ri) {
-      flags.col_contiguous &= (out_strides[i] == f_stride || out.shape(i) == 1);
-      f_stride *= out.shape(i);
-      flags.row_contiguous &=
-          (out_strides[ri] == b_stride || out.shape(ri) == 1);
-      b_stride *= out.shape(ri);
-    }
-  }
-  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
+  transpose(inputs[0], out, axes_);
 }
 
 } // namespace mlx::core

mlx/backend/common/hadamard.h

@@ -99,7 +99,11 @@ inline std::pair<int, int> decompose_hadamard(int n) {
           "[hadamard] Only supports n = m*2^k where m in (1, 12, 20, 28).");
     }
   }
+  if (n > (1 << 26)) {
+    throw std::invalid_argument(
+        "[hadamard] Only supports n = m*2^k where k <= 26");
+  }
   return {n, m};
 }
 
-} // namespace mlx::core
+} // namespace mlx::core

mlx/backend/common/transpose.cpp

@@ -0,0 +1,57 @@
+// Copyright © 2024 Apple Inc.
+
+#include <cassert>
+
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+void transpose(const array& in, array& out, const std::vector<int>& axes) {
+  Strides out_strides(out.ndim());
+  for (int ax = 0; ax < axes.size(); ++ax) {
+    out_strides[ax] = in.strides()[axes[ax]];
+  }
+
+  // Conditions for {row/col}_contiguous
+  // - array must be contiguous (no gaps)
+  // - underlying buffer size should have the same size as the array
+  // - cumulative product of shapes is equal to the strides (we can ignore axes
+  //   with size == 1)
+  //   - in the forward direction (column contiguous)
+  //   - in the reverse direction (row contiguous)
+  // - vectors are both row and col contiguous (hence if both row/col are
+  //   true, they stay true)
+  auto flags = in.flags();
+  if (flags.contiguous && in.data_size() == in.size()) {
+    auto [_, rc, cc] = check_contiguity(out.shape(), out_strides);
+    flags.row_contiguous = rc;
+    flags.col_contiguous = cc;
+  }
+  out.copy_shared_buffer(in, out_strides, flags, in.data_size());
+}
+
+void as_transposed(array& out, const std::vector<int>& axes) {
+  assert(out.data_size() == out.size() && out.flags().contiguous);
+
+  // Calculate the contiguous strides.
+  Strides strides(out.ndim(), 1);
+  for (int i = out.ndim() - 2; i >= 0; i--) {
+    strides[i] = strides[i + 1] * out.shape(i);
+  }
+
+  // Calculate the new strides for transposing.
+  Strides new_strides;
+  new_strides.reserve(out.ndim());
+  for (auto ax : axes) {
+    new_strides.push_back(strides[ax]);
+  }
+
+  auto [ds, rc, cc] = check_contiguity(out.shape(), new_strides);
+  auto flags = out.flags();
+  flags.row_contiguous = rc;
+  flags.col_contiguous = cc;
+
+  out.copy_shared_buffer(out, new_strides, flags, ds);
+}
+
+} // namespace mlx::core

mlx/backend/common/transpose.h

@@ -0,0 +1,12 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+
+namespace mlx::core {
+
+void transpose(const array& in, array& out, const std::vector<int>& axes);
+void as_transposed(array& out, const std::vector<int>& axes);
+
+} // namespace mlx::core

mlx/backend/common/utils.h

@@ -132,6 +132,11 @@ struct ContiguousIterator {
 };
 
 inline auto check_contiguity(const Shape& shape, const Strides& strides) {
+  // Conditions for {row/col}_contiguous
+  // - cumulative product of shapes is equal to the strides (we can ignore axes
+  //   with size == 1)
+  //   - in the forward direction (column contiguous)
+  //   - in the reverse direction (row contiguous)
   size_t no_broadcast_data_size = 1;
   int64_t f_stride = 1;
   int64_t b_stride = 1;

mlx/backend/cpu/CMakeLists.txt

@@ -40,7 +40,8 @@ add_dependencies(mlx cpu_compiled_preamble)
 
 target_sources(
   mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/available.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp

mlx/backend/cpu/available.cpp

@@ -0,0 +1,11 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cpu/available.h"
+
+namespace mlx::core::cpu {
+
+bool is_available() {
+  return true;
+}
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/available.h

@@ -0,0 +1,9 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+namespace mlx::core::cpu {
+
+bool is_available();
+
+} // namespace mlx::core::cpu

mlx/backend/cpu/binary.cpp

@@ -172,9 +172,12 @@ void binary_float(
       case bfloat16:
         binary_op<bfloat16_t, Op>(a, b, out, bopt);
         break;
+      case complex64:
+        binary_op<complex64_t, Op>(a, b, out, bopt);
+        break;
       default:
         throw std::runtime_error(
-            "[binary_float] Only supports non-complex floating point types.");
+            "[binary_float] Only supports floating point types.");
     }
   });
 }

mlx/backend/cpu/compiled.cpp

@@ -40,7 +40,10 @@ struct CompilerCache {
   std::shared_mutex mtx;
 };
 
-static CompilerCache cache{};
+static CompilerCache& cache() {
+  static CompilerCache cache_;
+  return cache_;
+};
 
 // GPU compile is always available if the GPU is available and since we are in
 // this file CPU compile is also available.
@@ -56,14 +59,16 @@ void* compile(
     const std::string& kernel_name,
     const std::function<std::string(void)>& source_builder) {
   {
-    std::shared_lock lock(cache.mtx);
-    if (auto it = cache.kernels.find(kernel_name); it != cache.kernels.end()) {
+    std::shared_lock lock(cache().mtx);
+    if (auto it = cache().kernels.find(kernel_name);
+        it != cache().kernels.end()) {
       return it->second;
     }
   }
 
-  std::unique_lock lock(cache.mtx);
-  if (auto it = cache.kernels.find(kernel_name); it != cache.kernels.end()) {
+  std::unique_lock lock(cache().mtx);
+  if (auto it = cache().kernels.find(kernel_name);
+      it != cache().kernels.end()) {
     return it->second;
   }
   std::string source_code = source_builder();
@@ -120,10 +125,10 @@ void* compile(
   }
 
   // load library
-  cache.libs.emplace_back(shared_lib_path);
+  cache().libs.emplace_back(shared_lib_path);
 
   // Load function
-  void* fun = dlsym(cache.libs.back().lib, kernel_name.c_str());
+  void* fun = dlsym(cache().libs.back().lib, kernel_name.c_str());
   if (!fun) {
     std::ostringstream msg;
     msg << "[Compile::eval_cpu] Failed to load compiled function "
@@ -131,7 +136,7 @@ void* compile(
         << dlerror();
     throw std::runtime_error(msg.str());
   }
-  cache.kernels.insert({kernel_name, fun});
+  cache().kernels.insert({kernel_name, fun});
   return fun;
 }
 

mlx/backend/cpu/scan.cpp

@@ -330,7 +330,8 @@ void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
             reduce_type_, in, out, axis_, reverse_, inclusive_);
         break;
       case complex64:
-        throw std::runtime_error("Scan ops do not support complex types yet");
+        scan_dispatch<complex64_t, complex64_t>(
+            reduce_type_, in, out, axis_, reverse_, inclusive_);
         break;
     }
   });

mlx/backend/cpu/simd/base_simd.h

@@ -88,12 +88,33 @@ DEFAULT_UNARY(expm1, std::expm1)
 DEFAULT_UNARY(floor, std::floor)
 DEFAULT_UNARY(log, std::log)
 DEFAULT_UNARY(log10, std::log10)
-DEFAULT_UNARY(log1p, std::log1p)
 DEFAULT_UNARY(sinh, std::sinh)
 DEFAULT_UNARY(sqrt, std::sqrt)
 DEFAULT_UNARY(tan, std::tan)
 DEFAULT_UNARY(tanh, std::tanh)
 
+template <typename T>
+Simd<T, 1> log1p(Simd<T, 1> in) {
+  if constexpr (is_complex<T>) {
+    auto x = in.value.real();
+    auto y = in.value.imag();
+    auto zabs = std::abs(in.value);
+    auto theta = std::atan2(y, x + 1);
+    if (zabs < 0.5) {
+      auto r = x * (2 + x) + y * y;
+      if (r == 0) { // handle underflow
+        return Simd<T, 1>{T{x, theta}};
+      }
+      return Simd<T, 1>{T{((typeof(x))(0.5)) * std::log1p(r), theta}};
+    } else {
+      auto z0 = std::hypot(x + 1, y);
+      return Simd<T, 1>{T{std::log(z0), theta}};
+    }
+  } else {
+    return Simd<T, 1>{std::log1p(in.value)};
+  }
+}
+
 template <typename T>
 Simd<T, 1> log2(Simd<T, 1> in) {
   if constexpr (is_complex<T>) {

mlx/backend/cuda/CMakeLists.txt

@@ -0,0 +1,57 @@
+# Filename rules in cuda backend:
+#
+# * Use .cu/.cuh if code contains device code, and .cpp/.h if not.
+# * Device-only kernel code should be put in kernels/ subdir.
+# * Files in kernels/ subdir should not include files outside.
+target_sources(
+  mlx
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/event.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/fence.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
+
+target_compile_definitions(mlx PUBLIC MLX_USE_CUDA)
+
+# Enable defining device lambda functions.
+target_compile_options(mlx
+                       PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>")
+
+# 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
+    "75;80"
+    CACHE STRING "CUDA architectures")
+message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
+set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
+                                     "${MLX_CUDA_ARCHITECTURES}")
+
+# Use fixed version of CCCL.
+FetchContent_Declare(
+  cccl
+  URL "https://github.com/NVIDIA/cccl/releases/download/v2.8.1/cccl-v2.8.1.zip")
+FetchContent_MakeAvailable(cccl)
+target_include_directories(mlx PRIVATE BEFORE "${cccl_SOURCE_DIR}/include")
+
+# Use fixed version of NVTX.
+FetchContent_Declare(
+  nvtx3
+  GIT_REPOSITORY https://github.com/NVIDIA/NVTX.git
+  GIT_TAG v3.1.1
+  GIT_SHALLOW TRUE
+  SOURCE_SUBDIR c EXCLUDE_FROM_ALL)
+FetchContent_MakeAvailable(nvtx3)
+target_link_libraries(mlx PUBLIC $<BUILD_INTERFACE:nvtx3-cpp>)
+
+# Make cuda runtime APIs available in non-cuda files.
+find_package(CUDAToolkit REQUIRED)
+target_include_directories(mlx PRIVATE ${CUDAToolkit_INCLUDE_DIRS})
+
+# Suppress nvcc warnings on MLX headers.
+target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
+                                   --diag_suppress=997>)

mlx/backend/cuda/allocator.cpp

@@ -0,0 +1,154 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/allocator.h"
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/backend/cuda/worker.h"
+
+#include <cuda_runtime.h>
+#include <fmt/format.h>
+
+#include <cassert>
+
+namespace mlx::core {
+
+namespace cu {
+
+CudaAllocator::CudaAllocator() {
+  // TODO: Set memory limit for multi-device.
+  size_t free, total;
+  CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
+  memory_limit_ = total * 0.8;
+}
+
+Buffer CudaAllocator::malloc(size_t size) {
+  // TODO: Check memory limit.
+  auto* buf = new CudaBuffer{nullptr, size};
+  cudaError_t err = cudaMallocManaged(&buf->data, size);
+  if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
+    throw std::runtime_error(
+        fmt::format("cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
+  }
+  std::lock_guard lock(mutex_);
+  active_memory_ += size;
+  peak_memory_ = std::max(active_memory_, peak_memory_);
+  return Buffer{buf};
+}
+
+void CudaAllocator::free(Buffer buffer) {
+  auto* buf = static_cast<CudaBuffer*>(buffer.ptr());
+  if (!buf) {
+    return;
+  }
+
+  // If free() is called from a unregistered thread, reschedule the call to
+  // worker.
+  {
+    std::lock_guard lock(worker_mutex_);
+    if (allowed_threads_.count(std::this_thread::get_id()) == 0) {
+      if (!worker_) {
+        worker_.reset(new Worker);
+      }
+      worker_->add_task([buffer]() { allocator().free(buffer); });
+      worker_->end_batch();
+      worker_->commit();
+      return;
+    }
+  }
+
+  size_t size = buf->size;
+  cudaFree(buf->data);
+  delete buf;
+  std::lock_guard lock(mutex_);
+  active_memory_ -= size;
+}
+
+size_t CudaAllocator::size(Buffer buffer) const {
+  auto* buf = static_cast<CudaBuffer*>(buffer.ptr());
+  if (!buf) {
+    return 0;
+  }
+  return buf->size;
+}
+
+void CudaAllocator::register_this_thread() {
+  std::lock_guard lock(worker_mutex_);
+  allowed_threads_.insert(std::this_thread::get_id());
+}
+
+size_t CudaAllocator::get_active_memory() const {
+  return active_memory_;
+}
+
+size_t CudaAllocator::get_peak_memory() const {
+  return peak_memory_;
+}
+
+void CudaAllocator::reset_peak_memory() {
+  std::lock_guard lock(mutex_);
+  peak_memory_ = 0;
+}
+
+size_t CudaAllocator::get_memory_limit() {
+  return memory_limit_;
+}
+
+size_t CudaAllocator::set_memory_limit(size_t limit) {
+  std::lock_guard lock(mutex_);
+  std::swap(limit, memory_limit_);
+  return limit;
+}
+
+CudaAllocator& allocator() {
+  // By creating the |allocator_| on heap, the destructor of CudaAllocator
+  // will not be called on exit and buffers in the cache will be leaked. This
+  // can save some time at program exit.
+  static CudaAllocator* allocator_ = new CudaAllocator;
+  return *allocator_;
+}
+
+} // namespace cu
+
+namespace allocator {
+
+Allocator& allocator() {
+  return cu::allocator();
+}
+
+void* Buffer::raw_ptr() {
+  if (!ptr_) {
+    return nullptr;
+  }
+  return static_cast<cu::CudaBuffer*>(ptr_)->data;
+}
+
+} // namespace allocator
+
+size_t get_active_memory() {
+  return cu::allocator().get_active_memory();
+}
+size_t get_peak_memory() {
+  return cu::allocator().get_peak_memory();
+}
+void reset_peak_memory() {
+  return cu::allocator().reset_peak_memory();
+}
+size_t set_memory_limit(size_t limit) {
+  return cu::allocator().set_memory_limit(limit);
+}
+size_t get_memory_limit() {
+  return cu::allocator().get_memory_limit();
+}
+
+// TODO: Implement buffer cache.
+size_t get_cache_memory() {
+  return 0;
+}
+size_t set_cache_limit(size_t) {
+  return 0;
+}
+size_t set_wired_limit(size_t) {
+  return 0;
+}
+void clear_cache() {}
+
+} // namespace mlx::core

mlx/backend/cuda/allocator.h

@@ -0,0 +1,58 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/allocator.h"
+
+#include <mutex>
+#include <set>
+#include <thread>
+#include <utility>
+
+namespace mlx::core::cu {
+
+class Worker;
+
+using allocator::Buffer;
+
+// Stores cuda-managed unified memory.
+struct CudaBuffer {
+  void* data;
+  size_t size;
+};
+
+class CudaAllocator : public allocator::Allocator {
+ public:
+  Buffer malloc(size_t size) override;
+  void free(Buffer buffer) override;
+  size_t size(Buffer buffer) const override;
+
+  // Register current thread as safe to free buffers.
+  // In cuda freeing a buffer implicitly synchronizes stream, and for threads
+  // that may be waited by gpu stream (for example cpu stream threads), freeing
+  // buffers there would result in dead lock.
+  void register_this_thread();
+
+  size_t get_active_memory() const;
+  size_t get_peak_memory() const;
+  void reset_peak_memory();
+  size_t get_memory_limit();
+  size_t set_memory_limit(size_t limit);
+
+ private:
+  CudaAllocator();
+  friend CudaAllocator& allocator();
+
+  std::mutex worker_mutex_;
+  std::unique_ptr<Worker> worker_;
+  std::set<std::thread::id> allowed_threads_;
+
+  std::mutex mutex_;
+  size_t memory_limit_;
+  size_t active_memory_{0};
+  size_t peak_memory_{0};
+};
+
+CudaAllocator& allocator();
+
+} // namespace mlx::core::cu

mlx/backend/cuda/copy.cpp

@@ -0,0 +1,26 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/gpu/copy.h"
+
+namespace mlx::core {
+
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    const Shape& data_shape,
+    const Strides& strides_in_pre,
+    const Strides& strides_out_pre,
+    int64_t inp_offset,
+    int64_t out_offset,
+    CopyType ctype,
+    const Stream& s,
+    const std::optional<array>& dynamic_i_offset /* = std::nullopt */,
+    const std::optional<array>& dynamic_o_offset /* = std::nullopt */) {
+  throw std::runtime_error("copy_gpu_inplace not implemented in CUDA backend.");
+}
+
+void fill_gpu(const array& val, array& out, const Stream& s) {
+  throw std::runtime_error("fill_gpu not implemented in CUDA backend.");
+}
+
+} // namespace mlx::core

mlx/backend/cuda/device.cpp

@@ -0,0 +1,117 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/worker.h"
+#include "mlx/backend/metal/metal.h"
+
+#include <fmt/format.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+DeviceStream::DeviceStream(Device& device) : device_(device), stream_(device) {}
+
+void DeviceStream::synchronize() {
+  cudaStreamSynchronize(stream_);
+}
+
+cudaStream_t DeviceStream::schedule_cuda_stream() {
+  // TODO: Return a stream that maximizes parallelism.
+  return stream_;
+}
+
+cudaStream_t DeviceStream::last_cuda_stream() {
+  return stream_;
+}
+
+CommandEncoder& DeviceStream::get_encoder() {
+  if (!encoder_) {
+    encoder_ = std::make_unique<CommandEncoder>(*this);
+  }
+  return *encoder_;
+}
+
+Device::Device(int device) : device_(device) {
+  // Validate the requirements of device.
+  int attr = 0;
+  cudaDeviceGetAttribute(&attr, cudaDevAttrConcurrentManagedAccess, device_);
+  if (attr != 1) {
+    throw std::runtime_error(fmt::format(
+        "Device {} does not support synchronization in managed memory.",
+        device_));
+  }
+}
+
+void Device::make_current() {
+  // We need to set/get current CUDA device very frequently, cache it to reduce
+  // actual calls of CUDA APIs. This function assumes single-thread in host.
+  static int current = 0;
+  if (current != device_) {
+    CHECK_CUDA_ERROR(cudaSetDevice(device_));
+    current = device_;
+  }
+}
+
+DeviceStream& Device::get_stream(Stream s) {
+  auto it = streams_.find(s.index);
+  if (it == streams_.end()) {
+    it = streams_.try_emplace(s.index, *this).first;
+  }
+  return it->second;
+}
+
+CommandEncoder::CommandEncoder(DeviceStream& s)
+    : device_(s.device()), stream_(s) {}
+
+void CommandEncoder::add_completed_handler(std::function<void()> task) {
+  worker_.add_task(std::move(task));
+}
+
+void CommandEncoder::end_encoding() {
+  if (!temporaries_.empty()) {
+    add_completed_handler([temporaries = std::move(temporaries_)]() {});
+  }
+
+  // There is no kernel running, run completion handlers immediately.
+  if (!has_gpu_work_) {
+    worker_.consume_in_this_thread();
+    return;
+  }
+  has_gpu_work_ = false;
+
+  // Put completion handlers in a batch.
+  worker_.end_batch();
+
+  // Signaling kernel completion is expensive, delay until enough batches.
+  // TODO: This number is arbitrarily picked, profile for a better stragety.
+  if (worker_.uncommited_batches() > 8) {
+    commit();
+  }
+}
+
+void CommandEncoder::commit() {
+  worker_.commit(stream_.last_cuda_stream());
+}
+
+Device& device(mlx::core::Device device) {
+  static std::unordered_map<int, Device> devices;
+  auto it = devices.find(device.index);
+  if (it == devices.end()) {
+    it = devices.try_emplace(device.index, device.index).first;
+  }
+  return it->second;
+}
+
+DeviceStream& get_stream(Stream s) {
+  return device(s.device).get_stream(s);
+}
+
+CommandEncoder& get_command_encoder(Stream s) {
+  return get_stream(s).get_encoder();
+}
+
+} // namespace cu
+
+} // namespace mlx::core

mlx/backend/cuda/device.h

@@ -0,0 +1,131 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/array.h"
+#include "mlx/backend/cuda/worker.h"
+#include "mlx/stream.h"
+
+#include <thrust/execution_policy.h>
+
+#include <unordered_map>
+
+namespace mlx::core::cu {
+
+class Device;
+class CommandEncoder;
+
+class DeviceStream {
+ public:
+  explicit DeviceStream(Device& device);
+
+  DeviceStream(const DeviceStream&) = delete;
+  DeviceStream& operator=(const DeviceStream&) = delete;
+
+  // Wait until kernels in the stream complete.
+  void synchronize();
+
+  // Return a cuda stream for launching kernels.
+  cudaStream_t schedule_cuda_stream();
+
+  // Return the last cuda stream used.
+  cudaStream_t last_cuda_stream();
+
+  CommandEncoder& get_encoder();
+
+  Device& device() {
+    return device_;
+  }
+
+ private:
+  Device& device_;
+  CudaStream stream_;
+  std::unique_ptr<CommandEncoder> encoder_;
+};
+
+class Device {
+ public:
+  explicit Device(int device);
+
+  Device(const Device&) = delete;
+  Device& operator=(const Device&) = delete;
+
+  // Make this device the current cuda device, required by some cuda calls.
+  void make_current();
+
+  DeviceStream& get_stream(Stream s);
+
+  int cuda_device() const {
+    return device_;
+  }
+
+ private:
+  int device_;
+  std::unordered_map<int, DeviceStream> streams_;
+};
+
+class CommandEncoder {
+ public:
+  explicit CommandEncoder(DeviceStream& stream);
+
+  CommandEncoder(const CommandEncoder&) = delete;
+  CommandEncoder& operator=(const CommandEncoder&) = delete;
+
+  void set_input_array(const array& arr) {}
+  void set_output_array(const array& arr) {}
+
+  void add_temporary(const array& arr) {
+    temporaries_.push_back(arr.data_shared_ptr());
+  }
+
+  void add_completed_handler(std::function<void()> task);
+  void end_encoding();
+  void commit();
+
+  // Schedule a cuda stream for |fun| to launch kernels, and check error
+  // afterwards.
+  template <typename F>
+  void launch_kernel(F&& fun) {
+    launch_kernel(stream_.schedule_cuda_stream(), std::forward<F>(fun));
+  }
+
+  template <typename F>
+  void launch_kernel(cudaStream_t stream, F&& fun) {
+    device_.make_current();
+    fun(stream);
+    check_cuda_error("kernel launch", cudaGetLastError());
+    has_gpu_work_ = true;
+  }
+
+  Device& device() {
+    return device_;
+  }
+
+  DeviceStream& stream() {
+    return stream_;
+  }
+
+  bool has_gpu_work() const {
+    return has_gpu_work_;
+  }
+
+ private:
+  Device& device_;
+  DeviceStream& stream_;
+  Worker worker_;
+  bool has_gpu_work_{false};
+  std::vector<std::shared_ptr<array::Data>> temporaries_;
+};
+
+Device& device(mlx::core::Device device);
+DeviceStream& get_stream(Stream s);
+CommandEncoder& get_command_encoder(Stream s);
+
+// Return an execution policy that does not sync for result.
+// Note that not all thrust APIs support async policy, confirm before using.
+inline auto thrust_policy(cudaStream_t stream) {
+  // TODO: Connect thrust's custom allocator with mlx's allocator.
+  return thrust::cuda::par_nosync.on(stream);
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/dtype_utils.cuh

@@ -0,0 +1,35 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <cuComplex.h>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+
+namespace mlx::core {
+
+// Maps CPU types to CUDA types.
+template <typename T>
+struct CTypeToCudaType {
+  using type = T;
+};
+
+template <>
+struct CTypeToCudaType<float16_t> {
+  using type = __half;
+};
+
+template <>
+struct CTypeToCudaType<bfloat16_t> {
+  using type = __nv_bfloat16;
+};
+
+template <>
+struct CTypeToCudaType<complex64_t> {
+  using type = cuComplex;
+};
+
+template <typename T>
+using cuda_type_t = typename CTypeToCudaType<T>::type;
+
+} // namespace mlx::core

mlx/backend/cuda/eval.cpp

@@ -0,0 +1,68 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/gpu/eval.h"
+#include "mlx/backend/cuda/allocator.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/gpu/available.h"
+#include "mlx/primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core::gpu {
+
+bool is_available() {
+  return true;
+}
+
+void new_stream(Stream s) {
+  // Force initalization of cuda, so cuda runtime get destroyed at last.
+  cudaFree(nullptr);
+  // Ensure the static stream objects get created.
+  cu::get_command_encoder(s);
+  // The main thread is safe to free buffers.
+  cu::allocator().register_this_thread();
+}
+
+void eval(array& arr) {
+  nvtx3::scoped_range r("gpu::eval");
+  auto outputs = arr.outputs();
+  {
+    // If the array is a tracer hold a reference
+    // to its inputs so they don't get donated
+    std::vector<array> inputs;
+    if (arr.is_tracer()) {
+      inputs = arr.inputs();
+    }
+    arr.primitive().eval_gpu(arr.inputs(), outputs);
+  }
+
+  auto& encoder = cu::get_command_encoder(arr.primitive().stream());
+  if (encoder.has_gpu_work()) {
+    // Keep used buffers alive until kernel finishes running.
+    std::unordered_set<std::shared_ptr<array::Data>> buffers;
+    for (auto& in : arr.inputs()) {
+      buffers.insert(in.data_shared_ptr());
+    }
+    for (auto& s : arr.siblings()) {
+      buffers.insert(s.data_shared_ptr());
+    }
+    // Remove the output if it was donated to by an input.
+    if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+      buffers.erase(it);
+    }
+    encoder.add_completed_handler([buffers = std::move(buffers)]() {});
+  }
+  encoder.end_encoding();
+}
+
+void finalize(Stream s) {
+  nvtx3::scoped_range r("gpu::finalize");
+  cu::get_command_encoder(s).commit();
+}
+
+void synchronize(Stream s) {
+  nvtx3::scoped_range r("gpu::synchronize");
+  cu::get_stream(s).synchronize();
+}
+
+} // namespace mlx::core::gpu

mlx/backend/cuda/event.cu

@@ -0,0 +1,265 @@
+// Copyright © 2024 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/event.h"
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/event.h"
+#include "mlx/scheduler.h"
+
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+///////////////////////////////////////////////////////////////////////////////
+// CudaEvent implementations
+///////////////////////////////////////////////////////////////////////////////
+
+// Cuda event managed with RAII.
+class CudaEventHandle {
+ public:
+  CudaEventHandle() {
+    CHECK_CUDA_ERROR(cudaEventCreateWithFlags(
+        &event_, cudaEventDisableTiming | cudaEventBlockingSync));
+  }
+
+  ~CudaEventHandle() {
+    CHECK_CUDA_ERROR(cudaEventDestroy(event_));
+  }
+
+  CudaEventHandle(const CudaEventHandle&) = delete;
+  CudaEventHandle& operator=(const CudaEventHandle&) = delete;
+
+  operator cudaEvent_t() const {
+    return event_;
+  }
+
+ private:
+  cudaEvent_t event_;
+};
+
+CudaEvent::CudaEvent() : event_(std::make_shared<CudaEventHandle>()) {}
+
+void CudaEvent::wait() {
+  nvtx3::scoped_range r("cu::CudaEvent::wait");
+  if (!recorded_) {
+    throw std::runtime_error("Should not wait on a CudaEvent before record.");
+  }
+  cudaEventSynchronize(*event_);
+}
+
+void CudaEvent::wait(cudaStream_t stream) {
+  if (!recorded_) {
+    throw std::runtime_error("Should not wait on a CudaEvent before record.");
+  }
+  cudaStreamWaitEvent(stream, *event_);
+}
+
+void CudaEvent::wait(Stream s) {
+  if (s.device == mlx::core::Device::cpu) {
+    scheduler::enqueue(s, [*this]() mutable { wait(); });
+  } else {
+    wait(cu::get_stream(s).last_cuda_stream());
+  }
+}
+
+void CudaEvent::record(cudaStream_t stream) {
+  cudaEventRecord(*event_, stream);
+  recorded_ = true;
+}
+
+void CudaEvent::record(Stream s) {
+  if (s.device == mlx::core::Device::cpu) {
+    throw std::runtime_error("CudaEvent can not wait on cpu stream.");
+  } else {
+    record(cu::get_stream(s).last_cuda_stream());
+  }
+}
+
+bool CudaEvent::completed() const {
+  return cudaEventQuery(*event_) == cudaSuccess;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// SharedEvent implementations
+///////////////////////////////////////////////////////////////////////////////
+
+namespace {
+
+__host__ __device__ void event_wait(SharedEvent::Atomic* ac, uint64_t value) {
+  uint64_t current;
+  while ((current = ac->load()) < value) {
+    ac->wait(current);
+  }
+}
+
+__host__ __device__ void event_signal(SharedEvent::Atomic* ac, uint64_t value) {
+  ac->store(value);
+  ac->notify_all();
+}
+
+__global__ void event_wait_kernel(SharedEvent::Atomic* ac, uint64_t value) {
+  event_wait(ac, value);
+}
+
+__global__ void event_signal_kernel(SharedEvent::Atomic* ac, uint64_t value) {
+  event_signal(ac, value);
+}
+
+} // namespace
+
+SharedEvent::SharedEvent() {
+  // Allocate cuda::atomic on managed memory.
+  allocator::Buffer buffer = allocator::malloc(sizeof(Atomic));
+  Atomic* ac = static_cast<Atomic*>(buffer.raw_ptr());
+  new (ac) Atomic(0);
+  ac_ = std::shared_ptr<Atomic>(ac, [buffer](Atomic* ptr) {
+    ptr->~Atomic();
+    allocator::free(buffer);
+  });
+}
+
+void SharedEvent::wait(uint64_t value) {
+  nvtx3::scoped_range r("cu::SharedEvent::wait");
+  event_wait(ac_.get(), value);
+}
+
+void SharedEvent::wait(cudaStream_t stream, uint64_t value) {
+  event_wait_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
+}
+
+void SharedEvent::wait(Stream s, uint64_t value) {
+  nvtx3::scoped_range r("cu::SharedEvent::wait(s)");
+  if (s.device == mlx::core::Device::cpu) {
+    scheduler::enqueue(s, [*this, value]() mutable { wait(value); });
+  } else {
+    auto& encoder = get_command_encoder(s);
+    encoder.launch_kernel(
+        encoder.stream().last_cuda_stream(),
+        [this, value](cudaStream_t stream) { wait(stream, value); });
+    encoder.add_completed_handler([ac = ac_]() {});
+    encoder.end_encoding();
+  }
+}
+
+void SharedEvent::signal(uint64_t value) {
+  nvtx3::scoped_range r("cu::SharedEvent::signal");
+  event_signal(ac_.get(), value);
+}
+
+void SharedEvent::signal(cudaStream_t stream, uint64_t value) {
+  event_signal_kernel<<<1, 1, 0, stream>>>(ac_.get(), value);
+}
+
+void SharedEvent::signal(Stream s, uint64_t value) {
+  nvtx3::scoped_range r("cu::SharedEvent::signal(s)");
+  if (s.device == mlx::core::Device::cpu) {
+    scheduler::enqueue(s, [*this, value]() mutable { signal(value); });
+  } else {
+    auto& encoder = get_command_encoder(s);
+    encoder.launch_kernel(
+        encoder.stream().last_cuda_stream(),
+        [this, value](cudaStream_t stream) { signal(stream, value); });
+    encoder.add_completed_handler([ac = ac_]() {});
+    encoder.end_encoding();
+  }
+}
+
+bool SharedEvent::is_signaled(uint64_t value) const {
+  nvtx3::scoped_range r("cu::SharedEvent::is_signaled");
+  return ac_->load() >= value;
+}
+
+uint64_t SharedEvent::value() const {
+  nvtx3::scoped_range r("cu::SharedEvent::value");
+  return ac_->load();
+}
+
+} // namespace cu
+
+///////////////////////////////////////////////////////////////////////////////
+// Event implementations
+///////////////////////////////////////////////////////////////////////////////
+
+namespace {
+
+struct EventImpl {
+  // CudaEvent is preferred when possible because it is fast, however we have
+  // to fallback to SharedEvent in following cases:
+  // 1. the event is used to wait/signal a cpu stream;
+  // 2. signal value other than 1 has been specified.
+  std::unique_ptr<cu::CudaEvent> cuda;
+  std::unique_ptr<cu::SharedEvent> shared;
+
+  bool is_created() const {
+    return cuda || shared;
+  }
+
+  void ensure_created(Stream s, uint64_t signal_value) {
+    if (is_created()) {
+      return;
+    }
+    if (s.device == mlx::core::Device::cpu || signal_value > 1) {
+      nvtx3::mark("Using slow SharedEvent");
+      shared = std::make_unique<cu::SharedEvent>();
+    } else {
+      cuda = std::make_unique<cu::CudaEvent>();
+    }
+  }
+};
+
+} // namespace
+
+Event::Event(Stream s) : stream_(s) {
+  event_ = std::shared_ptr<void>(
+      new EventImpl(), [](void* ptr) { delete static_cast<EventImpl*>(ptr); });
+}
+
+void Event::wait() {
+  auto* event = static_cast<EventImpl*>(event_.get());
+  assert(event->is_created());
+  if (event->cuda) {
+    assert(value() == 1);
+    event->cuda->wait();
+  } else {
+    event->shared->wait(value());
+  }
+}
+
+void Event::wait(Stream s) {
+  auto* event = static_cast<EventImpl*>(event_.get());
+  assert(event->is_created());
+  if (event->cuda) {
+    assert(value() == 1);
+    event->cuda->wait(s);
+  } else {
+    event->shared->wait(s, value());
+  }
+}
+
+void Event::signal(Stream s) {
+  auto* event = static_cast<EventImpl*>(event_.get());
+  event->ensure_created(s, value());
+  if (event->cuda) {
+    assert(value() == 1);
+    event->cuda->record(s);
+  } else {
+    event->shared->signal(s, value());
+  }
+}
+
+bool Event::is_signaled() const {
+  auto* event = static_cast<EventImpl*>(event_.get());
+  if (!event->is_created()) {
+    return false;
+  }
+  if (event->cuda) {
+    assert(value() == 1);
+    return event->cuda->recorded() && event->cuda->completed();
+  } else {
+    return event->shared->is_signaled(value());
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/event.h

@@ -0,0 +1,66 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/stream.h"
+
+#include <cuda_runtime.h>
+#include <cuda/atomic>
+
+#include <memory>
+
+namespace mlx::core::cu {
+
+class CudaEventHandle;
+
+// Wrapper of native cuda event. It can synchronize between GPU streams, or wait
+// on GPU stream in CPU stream, but can not wait on CPU stream.
+class CudaEvent {
+ public:
+  CudaEvent();
+
+  void wait();
+  void wait(cudaStream_t stream);
+  void wait(Stream s);
+  void record(cudaStream_t stream);
+  void record(Stream s);
+
+  // Return whether the recorded kernels have completed. Note that this method
+  // returns true if record() has not been called.
+  bool completed() const;
+
+  bool recorded() const {
+    return recorded_;
+  }
+
+ private:
+  bool recorded_{false};
+  std::shared_ptr<CudaEventHandle> event_;
+};
+
+// Event that can synchronize between CPU and GPU. It is much slower than
+// CudaEvent so the latter should always be preferred when possible.
+class SharedEvent {
+ public:
+  using Atomic = cuda::atomic<uint64_t>;
+
+  SharedEvent();
+
+  void wait(uint64_t value);
+  void wait(cudaStream_t stream, uint64_t value);
+  void wait(Stream s, uint64_t value);
+  void signal(uint64_t value);
+  void signal(cudaStream_t stream, uint64_t value);
+  void signal(Stream s, uint64_t value);
+  bool is_signaled(uint64_t value) const;
+  uint64_t value() const;
+
+  const std::shared_ptr<Atomic>& atomic() const {
+    return ac_;
+  }
+
+ private:
+  std::shared_ptr<Atomic> ac_;
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/fence.cu

@@ -0,0 +1,70 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/event.h"
+#include "mlx/fence.h"
+#include "mlx/scheduler.h"
+
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace {
+
+__host__ __device__ void busy_wait(cuda::atomic<uint64_t>* ac, uint64_t value) {
+  while (true) {
+    // In theory the atomic_thread_fence is not needed, but for CUDA 11 without
+    // it the load() may never return new value.
+    cuda::atomic_thread_fence(cuda::memory_order_seq_cst);
+    uint64_t current = ac->load();
+    if (current >= value) {
+      break;
+    }
+  }
+}
+
+__global__ void busy_wait_kernel(cuda::atomic<uint64_t>* ac, uint64_t value) {
+  busy_wait(ac, value);
+}
+
+} // namespace
+
+struct FenceImpl {
+  uint32_t count;
+  cu::SharedEvent event;
+};
+
+Fence::Fence(Stream s) {
+  fence_ = std::shared_ptr<void>(
+      new FenceImpl{0}, [](void* ptr) { delete static_cast<FenceImpl*>(ptr); });
+}
+
+void Fence::wait(Stream s, const array&) {
+  auto* fence = static_cast<FenceImpl*>(fence_.get());
+  // We can't use SharedEvent::wait because it could hang in CUDA 11, see also:
+  // https://github.com/ml-explore/mlx/issues/2137
+  const auto& ac = fence->event.atomic();
+  if (s.device == mlx::core::Device::cpu) {
+    scheduler::enqueue(s, [ac, count = fence->count]() {
+      nvtx3::scoped_range r("Fence::wait()");
+      busy_wait(ac.get(), count);
+    });
+  } else {
+    nvtx3::scoped_range r("Fence::wait(s)");
+    auto& encoder = cu::get_command_encoder(s);
+    encoder.launch_kernel(
+        encoder.stream().last_cuda_stream(), [&](cudaStream_t stream) {
+          busy_wait_kernel<<<1, 1, 0>>>(ac.get(), fence->count);
+        });
+    encoder.add_completed_handler([ac]() {});
+    encoder.end_encoding();
+  }
+}
+
+void Fence::update(Stream s, const array&) {
+  auto* fence = static_cast<FenceImpl*>(fence_.get());
+  fence->count++;
+  fence->event.signal(s, fence->count);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/kernels/arange.cuh

@@ -0,0 +1,15 @@
+// Copyright © 2025 Apple Inc.
+
+namespace mlx::core::cu {
+
+template <typename T>
+struct Arange {
+  const T start;
+  const T step;
+
+  __device__ T operator()(uint32_t i) const {
+    return start + i * step;
+  }
+};
+
+} // namespace mlx::core::cu

mlx/backend/cuda/kernels/fp16_math.cuh

@@ -0,0 +1,107 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <cuda_fp16.h>
+#include <cuda/std/limits>
+#include <cuda/std/type_traits>
+
+namespace mlx::core::cu {
+
+///////////////////////////////////////////////////////////////////////////////
+// Missing C++ operator overrides for CUDA 7.
+///////////////////////////////////////////////////////////////////////////////
+
+#if CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
+
+#define MLX_DEFINE_BF16_OP(OP)                                           \
+  __forceinline__ __device__ __nv_bfloat16 operator OP(                  \
+      __nv_bfloat16 x, __nv_bfloat16 y) {                                \
+    return __float2bfloat16(__bfloat162float(x) OP __bfloat162float(y)); \
+  }
+
+#define MLX_DEFINE_BF16_CMP(OP)                                          \
+  __forceinline__ __device__ bool operator OP(                           \
+      __nv_bfloat16 x, __nv_bfloat16 y) {                                \
+    return __float2bfloat16(__bfloat162float(x) OP __bfloat162float(y)); \
+  }
+
+MLX_DEFINE_BF16_OP(+)
+MLX_DEFINE_BF16_OP(-)
+MLX_DEFINE_BF16_OP(*)
+MLX_DEFINE_BF16_OP(/)
+MLX_DEFINE_BF16_CMP(>)
+MLX_DEFINE_BF16_CMP(<)
+MLX_DEFINE_BF16_CMP(>=)
+MLX_DEFINE_BF16_CMP(<=)
+
+#undef MLX_DEFINE_BF16_OP
+#undef MLX_DEFINE_BF16_CMP
+
+#endif // CUDART_VERSION < 12000 && __CUDA_ARCH__ < 800
+
+///////////////////////////////////////////////////////////////////////////////
+// Additional C++ operator overrides between half types and native types.
+///////////////////////////////////////////////////////////////////////////////
+
+template <typename T, typename U>
+constexpr bool is_integral_except =
+    cuda::std::is_integral_v<T> && !cuda::std::is_same_v<T, U>;
+
+template <typename T, typename U>
+constexpr bool is_arithmetic_except =
+    cuda::std::is_arithmetic_v<T> && !cuda::std::is_same_v<T, U>;
+
+#define MLX_DEFINE_HALF_OP(HALF, HALF2FLOAT, FLOAT2HALF, OP)          \
+  template <                                                          \
+      typename T,                                                     \
+      typename = cuda::std::enable_if_t<is_integral_except<T, HALF>>> \
+  __forceinline__ __device__ HALF operator OP(HALF x, T y) {          \
+    return FLOAT2HALF(HALF2FLOAT(x) OP static_cast<float>(y));        \
+  }                                                                   \
+  template <                                                          \
+      typename T,                                                     \
+      typename = cuda::std::enable_if_t<is_integral_except<T, HALF>>> \
+  __forceinline__ __device__ HALF operator OP(T x, HALF y) {          \
+    return FLOAT2HALF(static_cast<float>(x) OP HALF2FLOAT(y));        \
+  }
+
+#define MLX_DEFINE_HALF_CMP(HALF, HALF2FLOAT, OP)                       \
+  template <                                                            \
+      typename T,                                                       \
+      typename = cuda::std::enable_if_t<is_arithmetic_except<T, HALF>>> \
+  __forceinline__ __device__ bool operator OP(HALF x, T y) {            \
+    return HALF2FLOAT(x) OP static_cast<float>(y);                      \
+  }                                                                     \
+  template <                                                            \
+      typename T,                                                       \
+      typename = cuda::std::enable_if_t<is_arithmetic_except<T, HALF>>> \
+  __forceinline__ __device__ bool operator OP(T x, HALF y) {            \
+    return static_cast<float>(y) OP HALF2FLOAT(x);                      \
+  }
+
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, +)
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, -)
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, *)
+MLX_DEFINE_HALF_OP(__half, __half2float, __float2half, /)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, +)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, -)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, *)
+MLX_DEFINE_HALF_OP(__nv_bfloat16, __bfloat162float, __float2bfloat16, /)
+MLX_DEFINE_HALF_CMP(__half, __half2float, <)
+MLX_DEFINE_HALF_CMP(__half, __half2float, >)
+MLX_DEFINE_HALF_CMP(__half, __half2float, <=)
+MLX_DEFINE_HALF_CMP(__half, __half2float, >=)
+MLX_DEFINE_HALF_CMP(__half, __half2float, ==)
+MLX_DEFINE_HALF_CMP(__half, __half2float, !=)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, <)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, >)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, <=)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, >=)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, ==)
+MLX_DEFINE_HALF_CMP(__nv_bfloat16, __bfloat162float, !=)
+
+#undef MLX_DEFINE_HALF_OP
+#undef MLX_DEFINE_HALF_CMP
+
+} // namespace mlx::core::cu

mlx/backend/cuda/primitives.cu

@@ -0,0 +1,163 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/dtype_utils.cuh"
+#include "mlx/backend/cuda/kernels/arange.cuh"
+#include "mlx/backend/cuda/kernels/fp16_math.cuh"
+#include "mlx/distributed/primitives.h"
+#include "mlx/dtype_utils.h"
+#include "mlx/fast_primitives.h"
+#include "mlx/primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+#include <thrust/device_ptr.h>
+#include <thrust/transform.h>
+
+#include <cassert>
+
+namespace mlx::core {
+
+void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Arange::eval_gpu");
+  assert(inputs.size() == 0);
+  out.set_data(allocator::malloc(out.nbytes()));
+  if (out.size() == 0) {
+    return;
+  }
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_output_array(out);
+  encoder.launch_kernel([&, this](cudaStream_t stream) {
+    MLX_SWITCH_INT_FLOAT_TYPES_CHECKED(out.dtype(), "Arange", CTYPE, {
+      using OutType = cuda_type_t<CTYPE>;
+      CTYPE step =
+          static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_);
+      thrust::transform(
+          cu::thrust_policy(stream),
+          thrust::counting_iterator<uint32_t>(0),
+          thrust::counting_iterator<uint32_t>(out.data_size()),
+          thrust::device_pointer_cast(out.data<OutType>()),
+          cu::Arange<OutType>{
+              static_cast<OutType>(start_), static_cast<OutType>(step)});
+    });
+  });
+}
+
+#define NO_GPU_MULTI(func)                                             \
+  void func::eval_gpu(                                                 \
+      const std::vector<array>& inputs, std::vector<array>& outputs) { \
+    throw std::runtime_error(#func " has no CUDA implementation.");    \
+  }
+
+#define NO_GPU(func)                                                  \
+  void func::eval_gpu(const std::vector<array>& inputs, array& out) { \
+    throw std::runtime_error(#func " has no CUDA implementation.");   \
+  }
+
+NO_GPU(Abs)
+NO_GPU(Add)
+NO_GPU(AddMM)
+NO_GPU(ArcCos)
+NO_GPU(ArcCosh)
+NO_GPU(ArcSin)
+NO_GPU(ArcSinh)
+NO_GPU(ArcTan)
+NO_GPU(ArcTan2)
+NO_GPU(ArcTanh)
+NO_GPU(ArgPartition)
+NO_GPU(ArgReduce)
+NO_GPU(ArgSort)
+NO_GPU(BitwiseBinary)
+NO_GPU(BitwiseInvert)
+NO_GPU(BlockMaskedMM)
+NO_GPU(Ceil)
+NO_GPU_MULTI(Compiled)
+NO_GPU(Conjugate)
+NO_GPU(Convolution)
+NO_GPU(Cos)
+NO_GPU(Cosh)
+NO_GPU(Divide)
+NO_GPU_MULTI(DivMod)
+NO_GPU(DynamicSlice)
+NO_GPU(DynamicSliceUpdate)
+NO_GPU(Remainder)
+NO_GPU(Equal)
+NO_GPU(Erf)
+NO_GPU(ErfInv)
+NO_GPU(Exp)
+NO_GPU(Expm1)
+NO_GPU(FFT)
+NO_GPU(Floor)
+NO_GPU(Gather)
+NO_GPU(GatherAxis)
+NO_GPU(GatherMM)
+NO_GPU(GatherQMM)
+NO_GPU(Greater)
+NO_GPU(GreaterEqual)
+NO_GPU(Hadamard)
+NO_GPU(Imag)
+NO_GPU(Less)
+NO_GPU(LessEqual)
+NO_GPU(Load)
+NO_GPU(Log)
+NO_GPU(Log1p)
+NO_GPU(LogicalNot)
+NO_GPU(LogicalAnd)
+NO_GPU(LogicalOr)
+NO_GPU(LogAddExp)
+NO_GPU(LogSumExp)
+NO_GPU_MULTI(LUF)
+NO_GPU(Matmul)
+NO_GPU(Maximum)
+NO_GPU(Minimum)
+NO_GPU(Multiply)
+NO_GPU(Negative)
+NO_GPU(NotEqual)
+NO_GPU(Partition)
+NO_GPU(Power)
+NO_GPU_MULTI(QRF)
+NO_GPU(QuantizedMatmul)
+NO_GPU(RandomBits)
+NO_GPU(Real)
+NO_GPU(Reduce)
+NO_GPU(Round)
+NO_GPU(Scan)
+NO_GPU(Scatter)
+NO_GPU(ScatterAxis)
+NO_GPU(Select)
+NO_GPU(Sigmoid)
+NO_GPU(Sign)
+NO_GPU(Sin)
+NO_GPU(Sinh)
+NO_GPU(SliceUpdate)
+NO_GPU(Softmax)
+NO_GPU(Sort)
+NO_GPU(Square)
+NO_GPU(Sqrt)
+NO_GPU(Subtract)
+NO_GPU_MULTI(SVD)
+NO_GPU(Tan)
+NO_GPU(Tanh)
+NO_GPU(Inverse)
+NO_GPU(Cholesky)
+NO_GPU_MULTI(Eigh)
+
+namespace fast {
+NO_GPU_MULTI(LayerNorm)
+NO_GPU_MULTI(LayerNormVJP)
+NO_GPU_MULTI(RMSNorm)
+NO_GPU_MULTI(RMSNormVJP)
+NO_GPU_MULTI(RoPE)
+NO_GPU(ScaledDotProductAttention)
+NO_GPU_MULTI(AffineQuantize)
+NO_GPU_MULTI(CustomKernel)
+} // namespace fast
+
+namespace distributed {
+NO_GPU_MULTI(AllReduce)
+NO_GPU_MULTI(AllGather)
+NO_GPU_MULTI(Send)
+NO_GPU_MULTI(Recv)
+} // namespace distributed
+
+} // namespace mlx::core

mlx/backend/cuda/slicing.cpp

@@ -0,0 +1,15 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/gpu/slicing.h"
+
+namespace mlx::core {
+
+void concatenate_gpu(
+    const std::vector<array>& inputs,
+    array& out,
+    int axis,
+    const Stream& s) {
+  throw std::runtime_error("concatenate_gpu not implemented in CUDA backend.");
+}
+
+} // namespace mlx::core

mlx/backend/cuda/utils.cpp

@@ -0,0 +1,26 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/backend/cuda/device.h"
+
+#include <fmt/format.h>
+
+namespace mlx::core {
+
+CudaStream::CudaStream(cu::Device& device) {
+  device.make_current();
+  CHECK_CUDA_ERROR(cudaStreamCreateWithFlags(&stream_, cudaStreamNonBlocking));
+}
+
+CudaStream::~CudaStream() {
+  CHECK_CUDA_ERROR(cudaStreamDestroy(stream_));
+}
+
+void check_cuda_error(const char* name, cudaError_t err) {
+  if (err != cudaSuccess) {
+    throw std::runtime_error(
+        fmt::format("{} failed: {}", name, cudaGetErrorString(err)));
+  }
+}
+
+} // namespace mlx::core

mlx/backend/cuda/utils.h

@@ -0,0 +1,36 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <cuda_runtime.h>
+
+namespace mlx::core {
+
+namespace cu {
+class Device;
+}
+
+// Cuda stream managed with RAII.
+class CudaStream {
+ public:
+  explicit CudaStream(cu::Device& device);
+  ~CudaStream();
+
+  CudaStream(const CudaStream&) = delete;
+  CudaStream& operator=(const CudaStream&) = delete;
+
+  operator cudaStream_t() const {
+    return stream_;
+  }
+
+ private:
+  cudaStream_t stream_;
+};
+
+// Throw exception if the cuda API does not succeed.
+void check_cuda_error(const char* name, cudaError_t err);
+
+// The macro version that prints the command that failed.
+#define CHECK_CUDA_ERROR(cmd) check_cuda_error(#cmd, (cmd))
+
+} // namespace mlx::core

mlx/backend/cuda/worker.cpp

@@ -0,0 +1,90 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/worker.h"
+#include "mlx/backend/cuda/allocator.h"
+#include "mlx/backend/cuda/device.h"
+
+namespace mlx::core::cu {
+
+Worker::Worker()
+    : signal_stream_(device(mlx::core::Device::gpu)),
+      worker_(&Worker::thread_fn, this) {}
+
+Worker::~Worker() {
+  {
+    std::lock_guard lock(worker_mutex_);
+    stop_ = true;
+  }
+  worker_event_.signal(batch_ + 1);
+  worker_.join();
+}
+
+void Worker::add_task(std::function<void()> task) {
+  pending_tasks_.push_back(std::move(task));
+}
+
+void Worker::consume_in_this_thread() {
+  for (auto& task : pending_tasks_) {
+    task();
+  }
+  pending_tasks_.clear();
+}
+
+void Worker::end_batch() {
+  batch_++;
+  {
+    std::lock_guard lock(worker_mutex_);
+    worker_tasks_[batch_] = std::move(pending_tasks_);
+  }
+  uncommited_batches_++;
+}
+
+void Worker::commit() {
+  if (uncommited_batches_ == 0) {
+    return;
+  }
+  uncommited_batches_ = 0;
+  worker_event_.signal(batch_);
+}
+
+void Worker::commit(cudaStream_t stream) {
+  if (uncommited_batches_ == 0) {
+    return;
+  }
+  uncommited_batches_ = 0;
+  // Signal the |worker_event_| in |signal_stream_| after the kernels in
+  // |stream_| finish running.
+  signal_event_.record(stream);
+  signal_event_.wait(signal_stream_);
+  worker_event_.signal(signal_stream_, batch_);
+}
+
+void Worker::thread_fn() {
+  // The worker thread is safe to free buffers.
+  allocator().register_this_thread();
+
+  while (!stop_) {
+    uint64_t batch = worker_event_.value();
+    Tasks tasks;
+    {
+      std::lock_guard lock(worker_mutex_);
+      // Move tasks in signaled batches.
+      auto end = worker_tasks_.upper_bound(batch);
+      for (auto it = worker_tasks_.begin(); it != end; ++it) {
+        if (tasks.empty()) {
+          tasks = std::move(it->second);
+        } else {
+          std::move(
+              it->second.begin(), it->second.end(), std::back_inserter(tasks));
+        }
+      }
+      worker_tasks_.erase(worker_tasks_.begin(), end);
+    }
+    for (auto& task : tasks) {
+      task();
+    }
+    worker_event_.wait(batch + 1);
+  }
+}
+
+} // namespace mlx::core::cu

mlx/backend/cuda/worker.h

@@ -0,0 +1,68 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/cuda/event.h"
+#include "mlx/backend/cuda/utils.h"
+
+#include <functional>
+#include <map>
+#include <mutex>
+#include <thread>
+
+namespace mlx::core::cu {
+
+// Run tasks in worker thread, synchronized with cuda stream.
+class Worker {
+ public:
+  Worker();
+  ~Worker();
+
+  Worker(const Worker&) = delete;
+  Worker& operator=(const Worker&) = delete;
+
+  // Add a pending |task| that will run when consumed or commited.
+  void add_task(std::function<void()> task);
+
+  // Run pending tasks immediately in current thread.
+  void consume_in_this_thread();
+
+  // Put pending tasks in a batch.
+  void end_batch();
+
+  // Inform worker thread to run current batches now.
+  void commit();
+
+  // Inform worker thread to run current batches after kernels in |stream|
+  // finish running.
+  void commit(cudaStream_t stream);
+
+  // Return how many batches have been added but not committed yet.
+  size_t uncommited_batches() const {
+    return uncommited_batches_;
+  }
+
+ private:
+  void thread_fn();
+
+  uint64_t batch_{0};
+  size_t uncommited_batches_{0};
+
+  // Cuda stream and event for signaling kernel completion.
+  CudaStream signal_stream_;
+  CudaEvent signal_event_;
+
+  // Worker thread.
+  SharedEvent worker_event_;
+  std::thread worker_;
+  std::mutex worker_mutex_;
+  bool stop_{false};
+
+  // Tasks are put in |pending_tasks_| first, and then moved to
+  // |worker_tasks_| when end_batch() is called.
+  using Tasks = std::vector<std::function<void()>>;
+  Tasks pending_tasks_;
+  std::map<uint64_t, Tasks> worker_tasks_;
+};
+
+} // namespace mlx::core::cu

mlx/backend/gpu/CMakeLists.txt

@@ -0,0 +1,5 @@
+target_sources(
+  mlx
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/copy.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp)

mlx/backend/gpu/available.h

@@ -0,0 +1,9 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+namespace mlx::core::gpu {
+
+bool is_available();
+
+} // namespace mlx::core::gpu

mlx/backend/gpu/copy.cpp

@@ -0,0 +1,49 @@
+// Copyright © 2023-2024 Apple Inc.
+
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/primitives.h"
+
+#include <cassert>
+
+namespace mlx::core {
+
+void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s) {
+  bool donated = set_copy_output_data(in, out, ctype);
+  if (donated && in.dtype() == out.dtype()) {
+    // If the output has the same type as the input then there is nothing to
+    // copy, just use the buffer.
+    return;
+  }
+  if (ctype == CopyType::GeneralGeneral) {
+    ctype = CopyType::General;
+  }
+  copy_gpu_inplace(in, out, ctype, s);
+}
+
+void copy_gpu(const array& in, array& out, CopyType ctype) {
+  copy_gpu(in, out, ctype, out.primitive().stream());
+}
+
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    CopyType ctype,
+    const Stream& s) {
+  assert(in.shape() == out.shape());
+  return copy_gpu_inplace(
+      in, out, in.shape(), in.strides(), out.strides(), 0, 0, ctype, s);
+}
+
+void copy_gpu_inplace(
+    const array& in,
+    array& out,
+    const Strides& i_strides,
+    int64_t i_offset,
+    CopyType ctype,
+    const Stream& s) {
+  assert(in.shape() == out.shape());
+  return copy_gpu_inplace(
+      in, out, in.shape(), i_strides, out.strides(), i_offset, 0, ctype, s);
+}
+
+} // namespace mlx::core

mlx/backend/gpu/copy.h

@@ -5,6 +5,8 @@
 #include "mlx/backend/common/copy.h"
 #include "mlx/stream.h"
 
+#include <optional>
+
 namespace mlx::core {
 
 // Generic copy inplace

mlx/backend/gpu/eval.h

@@ -8,14 +8,11 @@
 #include "mlx/array.h"
 #include "mlx/stream.h"
 
-namespace mlx::core::metal {
+namespace mlx::core::gpu {
 
 void new_stream(Stream stream);
-
-std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool();
-
 void eval(array& arr);
 void finalize(Stream s);
 void synchronize(Stream s);
 
-} // namespace mlx::core::metal
+} // namespace mlx::core::gpu

mlx/backend/gpu/primitives.cpp

@@ -0,0 +1,222 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/primitives.h"
+#include "mlx/backend/common/utils.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/backend/gpu/slicing.h"
+
+#include <cassert>
+
+#define MLX_PROFILER_RANGE(message)
+
+namespace mlx::core {
+
+namespace {
+
+void reshape(const array& in, array& out, Stream s) {
+  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
+  if (copy_necessary) {
+    out.set_data(allocator::malloc(out.nbytes()));
+    copy_gpu_inplace(
+        in,
+        out,
+        in.shape(),
+        in.strides(),
+        make_contiguous_strides(in.shape()),
+        0,
+        0,
+        CopyType::General,
+        s);
+  } else {
+    shared_buffer_reshape(in, out_strides, out);
+  }
+}
+
+} // namespace
+
+void AsStrided::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("AsStrided::eval_gpu");
+  eval(inputs, out);
+}
+
+void AsType::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("AsType::eval_gpu");
+  CopyType ctype =
+      inputs[0].flags().contiguous ? CopyType::Vector : CopyType::General;
+  copy_gpu(inputs[0], out, ctype);
+}
+
+void Broadcast::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Broadcast::eval_gpu");
+  eval(inputs, out);
+}
+
+void BroadcastAxes::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("BroadcastAxes::eval_gpu");
+  eval(inputs, out);
+}
+
+void Concatenate::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Concatenate::eval_gpu");
+  concatenate_gpu(inputs, out, axis_, stream());
+}
+
+void Contiguous::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Contiguous::eval_gpu");
+  assert(inputs.size() == 1);
+  auto& in = inputs[0];
+  constexpr size_t extra_bytes = 16384;
+  if (in.buffer_size() <= out.nbytes() + extra_bytes &&
+      (in.flags().row_contiguous ||
+       (allow_col_major_ && in.flags().col_contiguous))) {
+    out.copy_shared_buffer(in);
+  } else {
+    out.set_data(allocator::malloc(out.nbytes()));
+    copy_gpu_inplace(
+        in,
+        out,
+        in.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+        stream());
+  }
+}
+
+void Copy::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Copy::eval_gpu");
+  eval(inputs, out);
+}
+
+void CustomTransforms::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  MLX_PROFILER_RANGE("CustomTransforms::eval_gpu");
+  eval(inputs, outputs);
+}
+
+void Depends::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  MLX_PROFILER_RANGE("Depends::eval_gpu");
+  eval(inputs, outputs);
+}
+
+void ExpandDims::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("ExpandDims::eval_gpu");
+  eval(inputs, out);
+}
+
+void Full::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Full::eval_gpu");
+  auto in = inputs[0];
+  CopyType ctype;
+  if (in.data_size() == 1) {
+    ctype = CopyType::Scalar;
+  } else if (in.flags().contiguous) {
+    ctype = CopyType::Vector;
+  } else {
+    ctype = CopyType::General;
+  }
+  copy_gpu(in, out, ctype);
+}
+
+void Flatten::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Flatten::eval_gpu");
+  reshape(inputs[0], out, stream());
+}
+
+void NumberOfElements::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("NumberOfElements::eval_gpu");
+  eval(inputs, out);
+}
+
+void Pad::eval_gpu(const std::vector<array>& inputs, array& out) {
+  // Inputs must be base input array and scalar val array
+  assert(inputs.size() == 2);
+  auto& in = inputs[0];
+  auto& val = inputs[1];
+
+  // Padding value must be a scalar
+  assert(val.size() == 1);
+
+  // Padding value, input and output must be of the same type
+  assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
+
+  pad_gpu(in, val, out, axes_, low_pad_size_, stream());
+}
+
+void Reshape::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Reshape::eval_gpu");
+  reshape(inputs[0], out, stream());
+}
+
+void Split::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  MLX_PROFILER_RANGE("Split::eval_gpu");
+  eval(inputs, outputs);
+}
+
+void Slice::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Slice::eval_gpu");
+  assert(inputs.size() == 1);
+  if (out.size() == 0) {
+    out.set_data(nullptr);
+    return;
+  }
+
+  auto& in = inputs[0];
+  slice_gpu(in, out, start_indices_, strides_, stream());
+}
+
+void Squeeze::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Squeeze::eval_gpu");
+  eval(inputs, out);
+}
+
+void StopGradient::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("StopGradient::eval_gpu");
+  eval(inputs, out);
+}
+
+void Transpose::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Transpose::eval_gpu");
+  eval(inputs, out);
+}
+
+void Unflatten::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("Unflatten::eval_gpu");
+  reshape(inputs[0], out, stream());
+}
+
+void View::eval_gpu(const std::vector<array>& inputs, array& out) {
+  MLX_PROFILER_RANGE("View::eval_gpu");
+  auto& in = inputs[0];
+  auto ibytes = size_of(in.dtype());
+  auto obytes = size_of(out.dtype());
+  // Conditions for buffer copying (disjunction):
+  // - type size is the same
+  // - type size is smaller and the last axis is contiguous
+  // - the entire array is row contiguous
+  if (ibytes == obytes || (obytes < ibytes && in.strides().back() == 1) ||
+      in.flags().row_contiguous) {
+    auto strides = in.strides();
+    for (int i = 0; i < static_cast<int>(strides.size()) - 1; ++i) {
+      strides[i] *= ibytes;
+      strides[i] /= obytes;
+    }
+    out.copy_shared_buffer(
+        in, strides, in.flags(), in.data_size() * ibytes / obytes);
+  } else {
+    auto tmp = array(in.shape(), in.dtype(), nullptr, {});
+    tmp.set_data(allocator::malloc(tmp.nbytes()));
+    copy_gpu_inplace(in, tmp, CopyType::General, stream());
+
+    auto flags = out.flags();
+    flags.contiguous = true;
+    flags.row_contiguous = true;
+    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
+    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
+    out.copy_shared_buffer(tmp, out.strides(), flags, out.size());
+  }
+}
+
+} // namespace mlx::core

mlx/backend/gpu/slicing.cpp

@@ -0,0 +1,44 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/slicing.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/backend/gpu/slicing.h"
+
+namespace mlx::core {
+
+void slice_gpu(
+    const array& in,
+    array& out,
+    const Shape& start_indices,
+    const Shape& strides,
+    const Stream& s) {
+  slice(in, out, start_indices, strides);
+}
+
+void pad_gpu(
+    const array& in,
+    const array& val,
+    array& out,
+    const std::vector<int>& axes,
+    const Shape& low_pad_size,
+    const Stream& s) {
+  // Fill output with val
+  fill_gpu(val, out, s);
+
+  // Find offset for start of input values
+  size_t data_offset = 0;
+  for (int i = 0; i < axes.size(); i++) {
+    auto ax = axes[i] < 0 ? out.ndim() + axes[i] : axes[i];
+    data_offset += out.strides()[ax] * low_pad_size[i];
+  }
+
+  // Extract slice from output where input will be pasted
+  array out_slice(in.shape(), out.dtype(), nullptr, {});
+  out_slice.copy_shared_buffer(
+      out, out.strides(), out.flags(), out_slice.size(), data_offset);
+
+  // Copy input values into the slice
+  copy_gpu_inplace(in, out_slice, CopyType::GeneralGeneral, s);
+}
+
+} // namespace mlx::core

mlx/backend/metal/CMakeLists.txt

@@ -93,6 +93,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/distributed.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/event.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/fence.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/fft.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/hadamard.cpp

mlx/backend/metal/allocator.cpp

@@ -1,7 +1,6 @@
 // Copyright © 2023-2024 Apple Inc.
 #include "mlx/backend/metal/allocator.h"
 #include "mlx/backend/metal/metal.h"
-#include "mlx/backend/metal/metal_impl.h"
 #include "mlx/backend/metal/resident.h"
 #include "mlx/memory.h"
 

mlx/backend/metal/binary.cpp

@@ -90,7 +90,7 @@ void binary_op_gpu_inplace(
     work_per_thread = large ? 4 : 2;
   } else {
     large = out.data_size() > UINT32_MAX;
-    work_per_thread = 1;
+    work_per_thread = get_work_per_thread(a.dtype());
   }
   std::string kernel_name =
       get_kernel_name(bopt, op, a, large, shape.size(), work_per_thread);
@@ -137,13 +137,20 @@ void binary_op_gpu_inplace(
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   } else {
     // Launch a 1D or 2D grid of threads
-    size_t nthreads = out.data_size();
+    size_t nthreads = ceildiv(out.data_size(), work_per_thread);
     if (thread_group_size > nthreads) {
       thread_group_size = nthreads;
     }
+
     MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
-    MTL::Size grid_dims = large ? get_2d_grid_dims(out.shape(), out.strides())
-                                : MTL::Size(nthreads, 1, 1);
+    MTL::Size grid_dims;
+    if (large) {
+      compute_encoder.set_bytes<int64_t>(out.data_size(), arg_idx++);
+      grid_dims = get_2d_grid_dims(out.shape(), out.strides(), work_per_thread);
+    } else {
+      compute_encoder.set_bytes<int>(out.data_size(), arg_idx++);
+      grid_dims = MTL::Size(nthreads, 1, 1);
+    }
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   }
 }

mlx/backend/metal/compiled.cpp

@@ -64,6 +64,7 @@ inline void build_kernel(
         cnt++);
   }
 
+  std::string idx_type = use_big_index ? "int64_t" : "uint";
   if (add_indices) {
     os += fmt::format(
         "    constant const int64_t* in_strides [[buffer({0})]],\n", cnt++);
@@ -83,6 +84,9 @@ inline void build_kernel(
         "    constant const int64_t* output_strides [[buffer({0})]],\n", cnt++);
     os += fmt::format(
         "    constant const int* output_shape [[buffer({0})]],\n", cnt++);
+  } else {
+    os += fmt::format(
+        "    constant const {0}& size [[buffer({1})]],\n", idx_type, cnt++);
   }
   if (dynamic_dims) {
     os += fmt::format("    constant const int& ndim [[buffer({0})]],\n", cnt++);
@@ -92,13 +96,14 @@ inline void build_kernel(
   os += "    uint3 pos [[thread_position_in_grid]],\n";
   os += "    uint3 grid [[threads_per_grid]]) {\n";
 
-  std::string idx_type = use_big_index ? "int64_t" : "uint";
+  os += fmt::format("  constexpr int N_ = {0};\n", work_per_thread);
   if (contiguous && use_big_index) {
     // This is only used for contiguous kernels which don't have
     // a third grid dimension
-    os += "  int64_t index = pos.x + grid.x * int64_t(pos.y);\n";
+    os += "  int64_t index = N_ * (pos.x + grid.x * int64_t(pos.y));\n";
+  } else if (contiguous) {
+    os += "  uint index = N_ * pos.x;\n";
   } else if (work_per_thread > 1) {
-    os += fmt::format("  constexpr int N_ = {0};\n", work_per_thread);
     os += fmt::format(
         "  int xshape = output_shape[{0}];\n",
         dynamic_dims ? "ndim - 1" : std::to_string(ndim - 1));
@@ -110,6 +115,9 @@ inline void build_kernel(
         "  {0} index = pos.x + grid.x * (pos.y + {0}(grid.y) * pos.z);\n",
         idx_type);
   }
+  if (work_per_thread > 1 && contiguous) {
+    os += "  for (int i = 0; i < N_ && index < size; ++i) {\n";
+  }
 
   // Read constant / contiguous inputs in tmps
   std::vector<array> nc_inputs;
@@ -193,7 +201,7 @@ inline void build_kernel(
   }
 
   // Open per-thread loop
-  if (work_per_thread > 1) {
+  if (work_per_thread > 1 && !contiguous) {
     os +=
         "  for (int i = 0; i < N_ && (int(N_ * pos.x) + i) < xshape; ++i) {\n";
   }
@@ -272,6 +280,7 @@ void Compiled::eval_gpu(
   auto& s = stream();
   auto& d = metal::device(s.device);
   auto lib = d.get_library(kernel_lib_, [&]() {
+    int work_per_thread = get_work_per_thread(outputs_[0].dtype());
     std::string kernel = metal::utils();
     concatenate(
         kernel, metal::unary_ops(), metal::binary_ops(), metal::ternary_ops());
@@ -284,7 +293,9 @@ void Compiled::eval_gpu(
         constant_ids_,
         /* contiguous = */ true,
         /* ndim = */ 0,
-        /* dynamic_dims = */ false);
+        /* dynamic_dims = */ false,
+        /* use_big_index = */ false,
+        /* work_per_thread = */ work_per_thread);
     build_kernel(
         kernel,
         kernel_lib_ + "_contiguous_large",
@@ -295,7 +306,8 @@ void Compiled::eval_gpu(
         /* contiguous = */ true,
         /* ndim = */ 0,
         /* dynamic_dims = */ false,
-        /* use_big_index = */ true);
+        /* use_big_index = */ true,
+        /* work_per_thread = */ work_per_thread);
     for (int i = 1; i < 8; i++) {
       build_kernel(
           kernel,
@@ -468,6 +480,13 @@ void Compiled::eval_gpu(
   if (!contiguous) {
     compute_encoder.set_vector_bytes(strides[0], cnt++);
     compute_encoder.set_vector_bytes(shape, cnt++);
+  } else {
+    auto size = outputs[0].data_size();
+    if (large) {
+      compute_encoder.set_bytes<int64_t>(size, cnt++);
+    } else {
+      compute_encoder.set_bytes<int>(size, cnt++);
+    }
   }
 
   // Put the number of dims in if it is dynamic
@@ -477,12 +496,13 @@ void Compiled::eval_gpu(
 
   // Launch the kernel
   if (contiguous) {
-    size_t nthreads = outputs[0].data_size();
+    int work_per_thread = get_work_per_thread(outputs[0].dtype());
+    size_t nthreads = ceildiv(outputs[0].data_size(), work_per_thread);
     MTL::Size group_dims(
         std::min(nthreads, kernel->maxTotalThreadsPerThreadgroup()), 1, 1);
-
     MTL::Size grid_dims = large
-        ? get_2d_grid_dims(outputs[0].shape(), outputs[0].strides())
+        ? get_2d_grid_dims(
+              outputs[0].shape(), outputs[0].strides(), work_per_thread)
         : MTL::Size(nthreads, 1, 1);
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   } else {

mlx/backend/metal/conv.cpp

@@ -5,7 +5,7 @@
 #include <numeric>
 #include <sstream>
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/kernels/defines.h"

mlx/backend/metal/copy.cpp

@@ -1,35 +1,15 @@
 // Copyright © 2023-2024 Apple Inc.
 
-#include <sstream>
-
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/utils.h"
-#include "mlx/primitives.h"
 
 namespace mlx::core {
 
 constexpr int MAX_COPY_SPECIALIZED_DIMS = 3;
 
-void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s) {
-  bool donated = set_copy_output_data(in, out, ctype);
-  if (donated && in.dtype() == out.dtype()) {
-    // If the output has the same type as the input then there is nothing to
-    // copy, just use the buffer.
-    return;
-  }
-  if (ctype == CopyType::GeneralGeneral) {
-    ctype = CopyType::General;
-  }
-  copy_gpu_inplace(in, out, ctype, s);
-}
-
-void copy_gpu(const array& in, array& out, CopyType ctype) {
-  copy_gpu(in, out, ctype, out.primitive().stream());
-}
-
 void copy_gpu_inplace(
     const array& in,
     array& out,
@@ -104,6 +84,8 @@ void copy_gpu_inplace(
             "[Copy::eval_gpu] Dynamic output offset requires GeneralGeneral copy");
       }
     }
+  } else {
+    work_per_thread = get_work_per_thread(in.dtype());
   }
   concatenate(kernel_name, "_copy", type_to_name(in), type_to_name(out));
   auto kernel = dynamic ? get_dynamic_copy_kernel(d, kernel_name, in, out)
@@ -165,39 +147,23 @@ void copy_gpu_inplace(
     MTL::Size grid_dims = MTL::Size(dim0, dim1, rest);
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   } else {
-    size_t nthreads = out.data_size();
+    size_t nthreads = ceildiv(out.data_size(), work_per_thread);
     if (thread_group_size > nthreads) {
       thread_group_size = nthreads;
     }
     MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
-    MTL::Size grid_dims = large ? get_2d_grid_dims(out.shape(), out.strides())
-                                : MTL::Size(nthreads, 1, 1);
+    MTL::Size grid_dims;
+    if (large) {
+      compute_encoder.set_bytes<int64_t>(out.data_size(), 2);
+      grid_dims = get_2d_grid_dims(out.shape(), out.strides(), work_per_thread);
+    } else {
+      compute_encoder.set_bytes<int>(out.data_size(), 2);
+      grid_dims = MTL::Size(nthreads, 1, 1);
+    }
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   }
 }
 
-void copy_gpu_inplace(
-    const array& in,
-    array& out,
-    CopyType ctype,
-    const Stream& s) {
-  assert(in.shape() == out.shape());
-  return copy_gpu_inplace(
-      in, out, in.shape(), in.strides(), out.strides(), 0, 0, ctype, s);
-}
-
-void copy_gpu_inplace(
-    const array& in,
-    array& out,
-    const Strides& i_strides,
-    int64_t i_offset,
-    CopyType ctype,
-    const Stream& s) {
-  assert(in.shape() == out.shape());
-  return copy_gpu_inplace(
-      in, out, in.shape(), i_strides, out.strides(), i_offset, 0, ctype, s);
-}
-
 void fill_gpu(const array& val, array& out, const Stream& s) {
   if (out.size() == 0) {
     return;
@@ -214,14 +180,21 @@ void fill_gpu(const array& val, array& out, const Stream& s) {
   compute_encoder.set_input_array(val, 0);
   compute_encoder.set_output_array(out, 1);
 
+  int work_per_thread = get_work_per_thread(val.dtype());
   auto thread_group_size = kernel->maxTotalThreadsPerThreadgroup();
-  size_t nthreads = out.data_size();
+  size_t nthreads = ceildiv(out.data_size(), work_per_thread);
   if (thread_group_size > nthreads) {
     thread_group_size = nthreads;
   }
   MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
-  MTL::Size grid_dims = large ? get_2d_grid_dims(out.shape(), out.strides())
-                              : MTL::Size(nthreads, 1, 1);
+  MTL::Size grid_dims;
+  if (large) {
+    compute_encoder.set_bytes<int64_t>(out.data_size(), 2);
+    grid_dims = get_2d_grid_dims(out.shape(), out.strides(), work_per_thread);
+  } else {
+    compute_encoder.set_bytes<int>(out.data_size(), 2);
+    grid_dims = MTL::Size(nthreads, 1, 1);
+  }
   compute_encoder.dispatch_threads(grid_dims, group_dims);
 }
 

mlx/backend/metal/custom_kernel.cpp

@@ -1,6 +1,6 @@
 // Copyright © 2024 Apple Inc.
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/jit/includes.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/fast_primitives.h"

mlx/backend/metal/device.cpp

@@ -1,20 +1,20 @@
 // Copyright © 2023-2024 Apple Inc.
 
 #include <cstdlib>
+#include <filesystem>
 #include <sstream>
 
-#include <sys/sysctl.h>
-
 #define NS_PRIVATE_IMPLEMENTATION
 #define CA_PRIVATE_IMPLEMENTATION
 #define MTL_PRIVATE_IMPLEMENTATION
 
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/metal.h"
-#include "mlx/backend/metal/metal_impl.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/utils.h"
 
+namespace fs = std::filesystem;
+
 namespace mlx::core::metal {
 
 namespace {
@@ -66,8 +66,8 @@ MTL::Library* try_load_bundle(
   if (bundle != nullptr) {
     std::string resource_path =
         std::string(bundle->resourceURL()->fileSystemRepresentation()) + "/" +
-        lib_name + ".metallib" auto [lib, error] =
-            load_library_from_path(device, resource_path.c_str());
+        lib_name + ".metallib";
+    auto [lib, error] = load_library_from_path(device, resource_path.c_str());
     if (lib) {
       return lib;
     }
@@ -79,12 +79,18 @@ MTL::Library* try_load_bundle(
 // Firstly, search for the metallib in the same path as this binary
 std::pair<MTL::Library*, NS::Error*> load_colocated_library(
     MTL::Device* device,
-    const std::string& lib_name) {
-  std::string lib_path = get_colocated_mtllib_path(lib_name);
-  if (lib_path.size() != 0) {
-    return load_library_from_path(device, lib_path.c_str());
+    const std::string& relative_path) {
+  std::string binary_dir = get_binary_directory();
+  if (binary_dir.size() == 0) {
+    return {nullptr, nullptr};
   }
-  return {nullptr, nullptr};
+
+  auto path = fs::path(binary_dir) / relative_path;
+  if (!path.has_extension()) {
+    path.replace_extension(".metallib");
+  }
+
+  return load_library_from_path(device, path.c_str());
 }
 
 std::pair<MTL::Library*, NS::Error*> load_swiftpm_library(
@@ -99,7 +105,7 @@ std::pair<MTL::Library*, NS::Error*> load_swiftpm_library(
   auto bundles = NS::Bundle::allBundles();
   for (int i = 0, c = (int)bundles->count(); i < c; i++) {
     auto bundle = reinterpret_cast<NS::Bundle*>(bundles->object(i));
-    library = try_load_bundle(device, bundle->resourceURL());
+    library = try_load_bundle(device, bundle->resourceURL(), lib_name);
     if (library != nullptr) {
       return {library, nullptr};
     }
@@ -109,33 +115,34 @@ std::pair<MTL::Library*, NS::Error*> load_swiftpm_library(
 }
 
 MTL::Library* load_default_library(MTL::Device* device) {
-  NS::Error *error1, *error2, *error3;
+  NS::Error* error[4];
   MTL::Library* lib;
   // First try the colocated mlx.metallib
-  std::tie(lib, error1) = load_colocated_library(device, "mlx");
+  std::tie(lib, error[0]) = load_colocated_library(device, "mlx");
+  if (lib) {
+    return lib;
+  }
+
+  std::tie(lib, error[1]) = load_colocated_library(device, "Resources/mlx");
   if (lib) {
     return lib;
   }
 
   // Then try default.metallib in a SwiftPM bundle if we have one
-  std::tie(lib, error2) = load_swiftpm_library(device, "default");
+  std::tie(lib, error[2]) = load_swiftpm_library(device, "default");
   if (lib) {
     return lib;
   }
 
   // Finally try default_mtllib_path
-  std::tie(lib, error3) = load_library_from_path(device, default_mtllib_path);
+  std::tie(lib, error[3]) = load_library_from_path(device, default_mtllib_path);
   if (!lib) {
     std::ostringstream msg;
     msg << "Failed to load the default metallib. ";
-    if (error1 != nullptr) {
-      msg << error1->localizedDescription()->utf8String() << " ";
-    }
-    if (error2 != nullptr) {
-      msg << error2->localizedDescription()->utf8String() << " ";
-    }
-    if (error3 != nullptr) {
-      msg << error3->localizedDescription()->utf8String() << " ";
+    for (int i = 0; i < 4; i++) {
+      if (error[i] != nullptr) {
+        msg << error[i]->localizedDescription()->utf8String() << " ";
+      }
     }
     throw std::runtime_error(msg.str());
   }
@@ -156,6 +163,7 @@ MTL::Library* load_library(
           << error->localizedDescription()->utf8String();
       throw std::runtime_error(msg.str());
     }
+    return lib;
   }
 
   // We have been given a path so try to load from lib_path / lib_name.metallib
@@ -168,6 +176,7 @@ MTL::Library* load_library(
           << "> with error " << error->localizedDescription()->utf8String();
       throw std::runtime_error(msg.str());
     }
+    return lib;
   }
 
   // Try to load the colocated library
@@ -188,8 +197,8 @@ MTL::Library* load_library(
 
   std::ostringstream msg;
   msg << "Failed to load the metallib " << lib_name << ".metallib. "
-      << "We attempted to load it from <" << get_colocated_mtllib_path(lib_name)
-      << ">";
+      << "We attempted to load it from <" << get_binary_directory() << "/"
+      << lib_name << ".metallib" << ">";
 #ifdef SWIFTPM_BUNDLE
   msg << " and from the Swift PM bundle.";
 #endif
@@ -760,42 +769,4 @@ std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool() {
       NS::AutoreleasePool::alloc()->init(), dtor);
 }
 
-void new_stream(Stream stream) {
-  if (stream.device == mlx::core::Device::gpu) {
-    device(stream.device).new_queue(stream.index);
-  }
-}
-
-const std::unordered_map<std::string, std::variant<std::string, size_t>>&
-device_info() {
-  auto init_device_info = []()
-      -> std::unordered_map<std::string, std::variant<std::string, size_t>> {
-    auto pool = new_scoped_memory_pool();
-    auto raw_device = device(default_device()).mtl_device();
-    auto name = std::string(raw_device->name()->utf8String());
-    auto arch = std::string(raw_device->architecture()->name()->utf8String());
-
-    size_t memsize = 0;
-    size_t length = sizeof(memsize);
-    sysctlbyname("hw.memsize", &memsize, &length, NULL, 0);
-
-    size_t rsrc_limit = 0;
-    sysctlbyname("iogpu.rsrc_limit", &rsrc_limit, &length, NULL, 0);
-    if (rsrc_limit == 0) {
-      rsrc_limit = 499000;
-    }
-
-    return {
-        {"device_name", name},
-        {"architecture", arch},
-        {"max_buffer_length", raw_device->maxBufferLength()},
-        {"max_recommended_working_set_size",
-         raw_device->recommendedMaxWorkingSetSize()},
-        {"memory_size", memsize},
-        {"resource_limit", rsrc_limit}};
-  };
-  static auto device_info_ = init_device_info();
-  return device_info_;
-}
-
 } // namespace mlx::core::metal

mlx/backend/metal/device.h

@@ -21,18 +21,14 @@ namespace mlx::core::metal {
 
 // Note, this function must be left inline in a header so that it is not
 // dynamically linked.
-inline std::string get_colocated_mtllib_path(const std::string& lib_name) {
+inline std::string get_binary_directory() {
   Dl_info info;
-  std::string mtllib_path;
-  std::string lib_ext = lib_name + ".metallib";
-
-  int success = dladdr((void*)get_colocated_mtllib_path, &info);
+  std::string directory;
+  int success = dladdr((void*)get_binary_directory, &info);
   if (success) {
-    auto mtllib = fs::path(info.dli_fname).remove_filename() / lib_ext;
-    mtllib_path = mtllib.c_str();
+    directory = fs::path(info.dli_fname).remove_filename().c_str();
   }
-
-  return mtllib_path;
+  return directory;
 }
 
 using MTLFCList =
@@ -270,4 +266,6 @@ class Device {
 
 Device& device(mlx::core::Device);
 
+std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool();
+
 } // namespace mlx::core::metal

mlx/backend/metal/distributed.cpp

@@ -4,7 +4,7 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/distributed/ops.h"

mlx/backend/metal/eval.cpp

@@ -0,0 +1,102 @@
+// Copyright © 2023-2024 Apple Inc.
+#include <memory>
+
+#include "mlx/backend/gpu/available.h"
+#include "mlx/backend/gpu/eval.h"
+#include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/utils.h"
+#include "mlx/primitives.h"
+#include "mlx/scheduler.h"
+
+namespace mlx::core::gpu {
+
+bool is_available() {
+  return true;
+}
+
+void new_stream(Stream stream) {
+  if (stream.device == mlx::core::Device::gpu) {
+    metal::device(stream.device).new_queue(stream.index);
+  }
+}
+
+inline void check_error(MTL::CommandBuffer* cbuf) {
+  if (cbuf->status() == MTL::CommandBufferStatusError) {
+    std::ostringstream msg;
+    msg << "[METAL] Command buffer execution failed: "
+        << cbuf->error()->localizedDescription()->utf8String();
+    throw std::runtime_error(msg.str());
+  }
+}
+
+void eval(array& arr) {
+  auto pool = metal::new_scoped_memory_pool();
+  auto s = arr.primitive().stream();
+  auto& d = metal::device(s.device);
+  auto command_buffer = d.get_command_buffer(s.index);
+
+  auto outputs = arr.outputs();
+  {
+    // If the array is a tracer hold a reference
+    // to its inputs so they don't get donated
+    std::vector<array> inputs;
+    if (arr.is_tracer()) {
+      inputs = arr.inputs();
+    }
+
+    debug_set_primitive_buffer_label(command_buffer, arr.primitive());
+    arr.primitive().eval_gpu(arr.inputs(), outputs);
+  }
+  std::unordered_set<std::shared_ptr<array::Data>> buffers;
+  for (auto& in : arr.inputs()) {
+    buffers.insert(in.data_shared_ptr());
+  }
+  for (auto& s : arr.siblings()) {
+    buffers.insert(s.data_shared_ptr());
+  }
+  // Remove the output if it was donated to by an input
+  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
+    buffers.erase(it);
+  }
+
+  if (d.command_buffer_needs_commit(s.index)) {
+    d.end_encoding(s.index);
+    scheduler::notify_new_task(s);
+    command_buffer->addCompletedHandler(
+        [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
+          scheduler::notify_task_completion(s);
+          check_error(cbuf);
+        });
+    d.commit_command_buffer(s.index);
+    d.get_command_buffer(s.index);
+  } else {
+    command_buffer->addCompletedHandler(
+        [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
+          check_error(cbuf);
+        });
+  }
+}
+
+void finalize(Stream s) {
+  auto pool = metal::new_scoped_memory_pool();
+  auto& d = metal::device(s.device);
+  auto cb = d.get_command_buffer(s.index);
+  d.end_encoding(s.index);
+  cb->addCompletedHandler([s](MTL::CommandBuffer* cbuf) { check_error(cbuf); });
+  d.commit_command_buffer(s.index);
+  d.get_command_buffer(s.index);
+}
+
+void synchronize(Stream s) {
+  auto pool = metal::new_scoped_memory_pool();
+  auto& d = metal::device(s.device);
+  auto cb = d.get_command_buffer(s.index);
+  cb->retain();
+  d.end_encoding(s.index);
+  d.commit_command_buffer(s.index);
+  cb->waitUntilCompleted();
+  check_error(cb);
+  cb->release();
+}
+
+} // namespace mlx::core::gpu

mlx/backend/metal/event.cpp

@@ -2,7 +2,6 @@
 
 #include "mlx/event.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/metal_impl.h"
 #include "mlx/scheduler.h"
 
 namespace mlx::core {

mlx/backend/metal/fence.cpp

@@ -1,7 +1,6 @@
 // Copyright © 2024 Apple Inc.
 #include "mlx/fence.h"
 #include "mlx/backend/metal/device.h"
-#include "mlx/backend/metal/metal_impl.h"
 #include "mlx/scheduler.h"
 #include "mlx/utils.h"
 
@@ -139,7 +138,7 @@ void Fence::update(Stream stream, const array& x) {
   compute_encoder.set_compute_pipeline_state(kernel);
   compute_encoder.set_input_array(x, 0);
   compute_encoder.set_bytes(nthreads, 1);
-  compute_encoder.dispatch_threadgroups(group_dims, grid_dims);
+  compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
 
   // Barrier on previous kernels
   compute_encoder.barrier();

mlx/backend/metal/fft.cpp

@@ -1,16 +1,18 @@
 // Copyright © 2024 Apple Inc.
 #include <cassert>
 #include <complex>
+#include <iostream>
 #include <map>
 #include <numeric>
 #include <set>
 
 #include "mlx/3rdparty/pocketfft.h"
+#include "mlx/backend/common/transpose.h"
 #include "mlx/backend/common/utils.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/backend/gpu/slicing.h"
 #include "mlx/backend/metal/binary.h"
-#include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/kernels.h"
-#include "mlx/backend/metal/slicing.h"
 #include "mlx/backend/metal/unary.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/utils.h"
@@ -27,7 +29,7 @@ using MTLFC = std::tuple<const void*, MTL::DataType, NS::UInteger>;
 // For strided reads/writes, coalesce at least this many complex64s
 #define MIN_COALESCE_WIDTH 4
 
-inline const std::vector<int> supported_radices() {
+inline constexpr std::array<int, 9> supported_radices() {
   // Ordered by preference in decomposition.
   return {13, 11, 8, 7, 6, 5, 4, 3, 2};
 }
@@ -49,6 +51,35 @@ std::vector<int> prime_factors(int n) {
   return factors;
 }
 
+int next_fast_n(int n) {
+  return next_power_of_2(n);
+}
+
+std::vector<int> stockham_decompose(int n) {
+  auto radices = supported_radices();
+  std::vector<int> steps(radices.size(), 0);
+  int orig_n = n;
+
+  for (int i = 0; i < radices.size(); i++) {
+    int radix = radices[i];
+
+    // Manually tuned radices for powers of 2
+    if (is_power_of_2(orig_n) && orig_n < 512 && radix > 4) {
+      continue;
+    }
+
+    while (n % radix == 0) {
+      steps[i] += 1;
+      n /= radix;
+      if (n == 1) {
+        return steps;
+      }
+    }
+  }
+
+  return {};
+}
+
 struct FourStepParams {
   bool required = false;
   bool first_step = true;
@@ -65,9 +96,10 @@ void fft_op(
     bool real,
     const FourStepParams four_step_params,
     bool inplace,
+    metal::Device& d,
     const Stream& s);
 
-struct FFTPlan {
+struct OldFFTPlan {
   int n = 0;
   // Number of steps for each radix in the Stockham decomposition
   std::vector<int> stockham;
@@ -82,9 +114,104 @@ struct FFTPlan {
   int n2 = 0;
 };
 
-int next_fast_n(int n) {
-  return next_power_of_2(n);
-}
+class FFTPlan {
+ public:
+  enum FFTType {
+    UNSUPPORTED,
+    NOOP,
+    STOCKHAM,
+    RADER,
+    BLUESTEIN,
+    MULTIUPLOAD_BLUESTEIN,
+    SMALL_FOUR_STEP,
+    LARGE_FOUR_STEP
+  };
+
+  FFTPlan(int n) : n_(n) {
+    // NOOP
+    if (n == 1) {
+      type_ = NOOP;
+    }
+
+    // Too large for Stockham so do four step fft for powers of 2
+    else if (n > MAX_STOCKHAM_FFT_SIZE && is_power_of_2(n)) {
+      if (n <= 1 << 20) {
+        type_ = SMALL_FOUR_STEP;
+        n2_ = n > 65536 ? 1024 : 64;
+        n1_ = n / n2_;
+        steps1_ = stockham_decompose(n1_);
+        steps2_ = stockham_decompose(n2_);
+      } else {
+        type_ = LARGE_FOUR_STEP;
+      }
+    }
+
+    // Too large and not power of 2 so do multi-upload Bluestein fft
+    else if (n > MAX_STOCKHAM_FFT_SIZE) {
+      type_ = MULTIUPLOAD_BLUESTEIN;
+      bluestein_n_ = next_fast_n(2 * n - 1);
+    }
+
+    // Stockham fft
+    else if (auto steps = stockham_decompose(n); steps.size() > 0) {
+      type_ = STOCKHAM;
+      steps_ = steps;
+    }
+
+    // Add rader but for now simply fall back to bluestein when stockham not
+    // posssible
+    else if (n > MAX_BLUESTEIN_FFT_SIZE) {
+      type_ = MULTIUPLOAD_BLUESTEIN;
+      bluestein_n_ = next_fast_n(2 * n - 1);
+    } else {
+      type_ = BLUESTEIN;
+      bluestein_n_ = next_fast_n(2 * n - 1);
+      steps_ = stockham_decompose(bluestein_n_);
+    }
+  }
+
+  FFTType type() const {
+    return type_;
+  }
+
+  int size() const {
+    return n_;
+  }
+
+  const std::vector<int>& steps() const {
+    return steps_;
+  }
+
+  int first_size() const {
+    return n1_;
+  }
+
+  const std::vector<int>& first_steps() const {
+    return steps1_;
+  }
+
+  int second_size() const {
+    return n2_;
+  }
+
+  const std::vector<int>& second_steps() const {
+    return steps2_;
+  }
+
+  int bluestein_size() const {
+    return bluestein_n_;
+  }
+
+ private:
+  int n_;
+  FFTType type_;
+  std::vector<int> steps_;
+  int n1_;
+  std::vector<int> steps1_;
+  int n2_;
+  std::vector<int> steps2_;
+  int bluestein_n_;
+};
 
 std::vector<int> plan_stockham_fft(int n) {
   auto radices = supported_radices();
@@ -110,15 +237,12 @@ std::vector<int> plan_stockham_fft(int n) {
   throw std::runtime_error("Unplannable");
 }
 
-FFTPlan plan_fft(int n) {
+OldFFTPlan plan_fft(int n) {
   auto radices = supported_radices();
-  std::set<int> radices_set(radices.begin(), radices.end());
 
-  FFTPlan plan;
+  OldFFTPlan plan;
   plan.n = n;
   plan.rader = std::vector<int>(radices.size(), 0);
-  auto factors = prime_factors(n);
-  int remaining_n = n;
 
   // Four Step FFT when N is too large for shared mem.
   if (n > MAX_STOCKHAM_FFT_SIZE && is_power_of_2(n)) {
@@ -128,16 +252,20 @@ FFTPlan plan_fft(int n) {
     plan.n2 = n > 65536 ? 1024 : 64;
     plan.n1 = n / plan.n2;
     return plan;
-  } else if (n > MAX_STOCKHAM_FFT_SIZE) {
+  }
+
+  if (n > MAX_STOCKHAM_FFT_SIZE) {
     // Otherwise we use a multi-upload Bluestein's
     plan.four_step = true;
     plan.bluestein_n = next_fast_n(2 * n - 1);
     return plan;
   }
 
+  int remaining_n = n;
+  auto factors = prime_factors(n);
   for (int factor : factors) {
     // Make sure the factor is a supported radix
-    if (radices_set.find(factor) == radices_set.end()) {
+    if (std::find(radices.begin(), radices.end(), factor) == radices.end()) {
       // We only support a single Rader factor currently
       // TODO(alexbarron) investigate weirdness with large
       // Rader sizes -- possibly a compiler issue?
@@ -154,7 +282,7 @@ FFTPlan plan_fft(int n) {
       for (int rf : rader_factors) {
         // We don't nest Rader's algorithm so if `factor - 1`
         // isn't Stockham decomposable we give up and do Bluestein's.
-        if (radices_set.find(rf) == radices_set.end()) {
+        if (std::find(radices.begin(), radices.end(), rf) == radices.end()) {
           plan.four_step = n > MAX_BLUESTEIN_FFT_SIZE;
           plan.bluestein_n = next_fast_n(2 * n - 1);
           plan.stockham = plan_stockham_fft(plan.bluestein_n);
@@ -172,7 +300,7 @@ FFTPlan plan_fft(int n) {
   return plan;
 }
 
-int compute_elems_per_thread(FFTPlan plan) {
+int compute_elems_per_thread(OldFFTPlan plan) {
   // Heuristics for selecting an efficient number
   // of threads to use for a particular mixed-radix FFT.
   auto n = plan.n;
@@ -355,9 +483,11 @@ void multi_upload_bluestein_fft(
     size_t axis,
     bool inverse,
     bool real,
-    FFTPlan& plan,
+    OldFFTPlan& plan,
     std::vector<array>& copies,
     const Stream& s) {
+  auto& d = metal::device(s.device);
+
   // TODO(alexbarron) Implement fused kernels for mutli upload bluestein's
   // algorithm
   int n = inverse ? out.shape(axis) : in.shape(axis);
@@ -420,6 +550,7 @@ void multi_upload_bluestein_fft(
       /*real=*/false,
       FourStepParams(),
       /*inplace=*/false,
+      d,
       s);
   copies.push_back(pad_temp1);
 
@@ -435,6 +566,7 @@ void multi_upload_bluestein_fft(
       /* real= */ false,
       FourStepParams(),
       /*inplace=*/true,
+      d,
       s);
 
   int offset = plan.bluestein_n - (2 * n - 1);
@@ -480,7 +612,7 @@ void four_step_fft(
     size_t axis,
     bool inverse,
     bool real,
-    FFTPlan& plan,
+    OldFFTPlan& plan,
     std::vector<array>& copies,
     const Stream& s,
     bool in_place) {
@@ -493,7 +625,15 @@ void four_step_fft(
     auto temp_shape = (real && inverse) ? out.shape() : in.shape();
     array temp(temp_shape, complex64, nullptr, {});
     fft_op(
-        in, temp, axis, inverse, real, four_step_params, /*inplace=*/false, s);
+        in,
+        temp,
+        axis,
+        inverse,
+        real,
+        four_step_params,
+        /*inplace=*/false,
+        d,
+        s);
     four_step_params.first_step = false;
     fft_op(
         temp,
@@ -503,6 +643,7 @@ void four_step_fft(
         real,
         four_step_params,
         /*inplace=*/in_place,
+        d,
         s);
     copies.push_back(temp);
   } else {
@@ -518,9 +659,8 @@ void fft_op(
     bool real,
     const FourStepParams four_step_params,
     bool inplace,
+    metal::Device& d,
     const Stream& s) {
-  auto& d = metal::device(s.device);
-
   size_t n = out.dtype() == float32 ? out.shape(axis) : in.shape(axis);
   if (n == 1) {
     out.copy_shared_buffer(in);
@@ -755,57 +895,517 @@ void fft_op(
   d.add_temporaries(std::move(copies), s.index);
 }
 
-void fft_op(
+inline int compute_elems_per_thread(int n, const std::vector<int>& steps) {
+  auto radices = supported_radices();
+  std::set<int> used_radices;
+  for (int i = 0; i < steps.size(); i++) {
+    if (steps[i] > 0) {
+      used_radices.insert(radices[i % radices.size()]);
+    }
+  }
+
+  // Manual tuning for 7/11/13
+  if (used_radices.find(7) != used_radices.end() &&
+      (used_radices.find(11) != used_radices.end() ||
+       used_radices.find(13) != used_radices.end())) {
+    return 7;
+  } else if (
+      used_radices.find(11) != used_radices.end() &&
+      used_radices.find(13) != used_radices.end()) {
+    return 11;
+  }
+
+  // TODO(alexbarron) Some really weird stuff is going on
+  // for certain `elems_per_thread` on large composite n.
+  // Possibly a compiler issue?
+  if (n == 3159)
+    return 13;
+  if (n == 3645)
+    return 5;
+  if (n == 3969)
+    return 7;
+  if (n == 1982)
+    return 5;
+
+  if (used_radices.size() == 1) {
+    return *(used_radices.begin());
+  }
+  if (used_radices.size() == 2 &&
+      (used_radices.find(11) != used_radices.end() ||
+       used_radices.find(13) != used_radices.end())) {
+    return std::accumulate(used_radices.begin(), used_radices.end(), 0) / 2;
+  }
+
+  // In all other cases use the second smallest radix.
+  return *(++used_radices.begin());
+}
+
+inline array ensure_fastest_moving_axis(
+    const array& x,
+    int axis,
+    metal::Device& d,
+    const Stream& s) {
+  // The axis is already with a stride of 1 so check that we have no overlaps
+  // and broadcasting and avoid the copy.
+  if (x.strides(axis) == 1) {
+    // This is a fairly strict test perhaps consider relaxing it in the future.
+    if (x.flags().row_contiguous || x.flags().col_contiguous) {
+      return x;
+    }
+  }
+
+  // To make it the fastest moving axis simply transpose it, then copy it and
+  // then transpose it back.
+
+  // Transpose it
+  std::vector<int> axes(x.ndim(), 0);
+  for (int ax = 0; ax < axes.size(); ax++) {
+    axes[ax] = (ax < axis) ? ax : ax + 1;
+  }
+  axes.back() = axis;
+  Shape xtshape;
+  xtshape.reserve(axes.size());
+  for (auto ax : axes) {
+    xtshape.push_back(x.shape(ax));
+  }
+  array xt(xtshape, x.dtype(), nullptr, {});
+  transpose(x, xt, axes);
+
+  // Copy it
+  array xtc(xt.shape(), x.dtype(), nullptr, {});
+  copy_gpu(
+      xt,
+      xtc,
+      xt.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+      s);
+  d.add_temporary(xtc, s.index);
+
+  // Transpose it
+  for (int ax = 0; ax < axes.size(); ax++) {
+    axes[ax] = (ax < axis) ? ax : ((ax == axis) ? axes.size() - 1 : ax - 1);
+  }
+  array y(x.shape(), x.dtype(), nullptr, {});
+  transpose(xtc, y, axes);
+
+  return y;
+}
+
+inline void prepare_output_array(const array& in, array& out, int axis) {
+  // Prepare the output array such that it matches the input in terms of
+  // stride ordering. Namely we might have moved `axis` around in the `in`
+  // array. We must do the same in `out`. The difference is that we don't have
+  // to copy anything because `out` contains garbage at the moment.
+
+  if (in.flags().row_contiguous && out.flags().row_contiguous) {
+    return;
+  }
+
+  std::vector<int> axes(out.ndim(), 0);
+  for (int ax = 0; ax < axes.size(); ax++) {
+    axes[ax] = (ax < axis) ? ax : ax + 1;
+  }
+  axes.back() = axis;
+  as_transposed(out, axes);
+}
+
+void fft_stockham_inplace(
+    const FFTPlan& plan,
+    const array& in_,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    metal::Device& d,
+    const Stream& s) {
+  // Prepare the input and output arrays such that `axis` has stride 1.
+  // Possibly copy the input but never the output as it doesn't have anything
+  // useful in it yet.
+  array in = ensure_fastest_moving_axis(in_, axis, d, s);
+  prepare_output_array(in, out, axis);
+
+  // Prepare the arguments for stockham fft
+  int n = plan.size();
+  bool power_of_2 = is_power_of_2(n);
+  int total_batch_size =
+      out.dtype() == float32 ? out.size() / n : in.size() / n;
+  auto& steps = plan.steps();
+  int elems_per_thread = compute_elems_per_thread(n, steps);
+  int threads_per_fft = ceildiv(n, elems_per_thread);
+  int tg_batch_size = std::max(MIN_THREADGROUP_MEM_SIZE / n, 1);
+  int tg_mem_size = next_power_of_2(tg_batch_size * n);
+  int batch_size = ceildiv(total_batch_size, tg_batch_size);
+  batch_size = real ? ceildiv(batch_size, 2) : batch_size; // 2 RFFTs at once
+  std::vector<MTLFC> func_consts = {
+      {&inverse, MTL::DataType::DataTypeBool, 0},
+      {&power_of_2, MTL::DataType::DataTypeBool, 1},
+      {&elems_per_thread, MTL::DataType::DataTypeInt, 2}};
+  for (int i = 0; i < steps.size(); i++) {
+    func_consts.emplace_back(&steps[i], MTL::DataType::DataTypeInt, 4 + i);
+  }
+
+  // Get the kernel
+  auto in_type = in.dtype() == float32 ? "float" : "float2";
+  auto out_type = out.dtype() == float32 ? "float" : "float2";
+  std::string hash_name;
+  std::string kname;
+  kname.reserve(64);
+  hash_name.reserve(64);
+  concatenate(kname, "fft_mem_", tg_mem_size, "_", in_type, "_", out_type);
+  concatenate(hash_name, kname, "_n", n, "_inv_", inverse);
+  auto template_def =
+      get_template_definition(kname, "fft", tg_mem_size, in_type, out_type);
+  auto kernel = get_fft_kernel(d, kname, hash_name, func_consts, template_def);
+
+  // Launch it
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder.set_compute_pipeline_state(kernel);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_output_array(out, 1);
+  compute_encoder.set_bytes(n, 2);
+  compute_encoder.set_bytes(total_batch_size, 3);
+
+  MTL::Size group_dims(1, tg_batch_size, threads_per_fft);
+  MTL::Size grid_dims(batch_size, tg_batch_size, threads_per_fft);
+  compute_encoder.dispatch_threads(grid_dims, group_dims);
+}
+
+void fft_four_step_inplace(
+    const FFTPlan& plan,
+    const array& in_,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    metal::Device& d,
+    const Stream& s) {
+  // Prepare the input and output arrays such that `axis` has stride 1.
+  // Possibly copy the input but never the output as it doesn't have anything
+  // useful in it yet.
+  array in = ensure_fastest_moving_axis(in_, axis, d, s);
+  prepare_output_array(in, out, axis);
+
+  // Also prepare the intermediate array for the four-step fft which is
+  // implemented with 2 kernel calls.
+  array intermediate(
+      (real && inverse) ? out.shape() : in.shape(), complex64, nullptr, {});
+  intermediate.set_data(allocator::malloc(intermediate.nbytes()));
+  prepare_output_array(in, intermediate, axis);
+  d.add_temporary(intermediate, s.index);
+
+  // Make the two calls
+  for (int step = 0; step < 2; step++) {
+    // Create the parameters
+    int n1 = plan.first_size();
+    int n2 = plan.second_size();
+    int n = (step == 0) ? n1 : n2;
+    bool power_of_2 = true;
+    int total_batch_size =
+        out.dtype() == float32 ? out.size() / n : in.size() / n;
+    auto& steps = (step == 0) ? plan.first_steps() : plan.second_steps();
+    int elems_per_thread = compute_elems_per_thread(n, steps);
+    int threads_per_fft = ceildiv(n, elems_per_thread);
+    int tg_batch_size =
+        std::max(MIN_THREADGROUP_MEM_SIZE / n, MIN_COALESCE_WIDTH);
+    int tg_mem_size = next_power_of_2(tg_batch_size * n);
+    int batch_size = ceildiv(total_batch_size, tg_batch_size);
+    std::vector<MTLFC> func_consts = {
+        {&inverse, MTL::DataType::DataTypeBool, 0},
+        {&power_of_2, MTL::DataType::DataTypeBool, 1},
+        {&elems_per_thread, MTL::DataType::DataTypeInt, 2}};
+    for (int i = 0; i < steps.size(); i++) {
+      func_consts.emplace_back(&steps[i], MTL::DataType::DataTypeInt, 4 + i);
+    }
+
+    // Get the kernel
+    auto to_type = [](const array& x) {
+      return x.dtype() == float32 ? "float" : "float2";
+    };
+    auto in_type = step == 0 ? to_type(in) : to_type(intermediate);
+    auto out_type = step == 0 ? to_type(intermediate) : to_type(out);
+    std::string hash_name;
+    std::string kname;
+    kname.reserve(64);
+    hash_name.reserve(64);
+    concatenate(
+        kname,
+        "four_step_mem_",
+        tg_mem_size,
+        "_",
+        in_type,
+        "_",
+        out_type,
+        "_",
+        step,
+        (real ? "_true" : "_false"));
+    concatenate(hash_name, kname, "_n", n, "_inv_", inverse);
+    auto template_def = get_template_definition(
+        kname, "four_step_fft", tg_mem_size, in_type, out_type, step, real);
+    auto kernel =
+        get_fft_kernel(d, kname, hash_name, func_consts, template_def);
+
+    // Launch it
+    auto& compute_encoder = d.get_command_encoder(s.index);
+    compute_encoder.set_compute_pipeline_state(kernel);
+    compute_encoder.set_input_array((step == 0) ? in : intermediate, 0);
+    compute_encoder.set_output_array((step == 0) ? intermediate : out, 1);
+    compute_encoder.set_bytes(n1, 2);
+    compute_encoder.set_bytes(n2, 3);
+    compute_encoder.set_bytes(total_batch_size, 4);
+
+    MTL::Size group_dims(1, tg_batch_size, threads_per_fft);
+    MTL::Size grid_dims(batch_size, tg_batch_size, threads_per_fft);
+    compute_encoder.dispatch_threads(grid_dims, group_dims);
+  }
+}
+
+void fft_bluestein(
+    const FFTPlan& plan,
+    const array& in_,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    metal::Device& d,
+    const Stream& s) {
+  // Prepare the input and output arrays such that `axis` has stride 1.
+  // Possibly copy the input but never the output as it doesn't have anything
+  // useful in it yet.
+  array in = ensure_fastest_moving_axis(in_, axis, d, s);
+  prepare_output_array(in, out, axis);
+
+  // Prepare the arguments for bluestein fft
+  int n = plan.bluestein_size();
+  bool power_of_2 = true;
+  int total_batch_size = out.dtype() == float32 ? out.size() / plan.size()
+                                                : in.size() / plan.size();
+  auto& steps = plan.steps();
+  int elems_per_thread = compute_elems_per_thread(n, steps);
+  int threads_per_fft = ceildiv(n, elems_per_thread);
+  int tg_batch_size = std::max(MIN_THREADGROUP_MEM_SIZE / n, 1);
+  int tg_mem_size = next_power_of_2(tg_batch_size * n);
+  int batch_size = ceildiv(total_batch_size, tg_batch_size);
+  batch_size = real ? ceildiv(batch_size, 2) : batch_size; // 2 RFFTs at once
+  std::vector<MTLFC> func_consts = {
+      {&inverse, MTL::DataType::DataTypeBool, 0},
+      {&power_of_2, MTL::DataType::DataTypeBool, 1},
+      {&elems_per_thread, MTL::DataType::DataTypeInt, 2}};
+  for (int i = 0; i < steps.size(); i++) {
+    func_consts.emplace_back(&steps[i], MTL::DataType::DataTypeInt, 4 + i);
+  }
+
+  // Get the kernel
+  auto in_type = in.dtype() == float32 ? "float" : "float2";
+  auto out_type = out.dtype() == float32 ? "float" : "float2";
+  std::string hash_name;
+  std::string kname;
+  kname.reserve(64);
+  hash_name.reserve(64);
+  concatenate(
+      kname, "bluestein_fft_mem_", tg_mem_size, "_", in_type, "_", out_type);
+  concatenate(hash_name, kname, "_n", n, "_inv_", inverse);
+  auto template_def = get_template_definition(
+      kname, "bluestein_fft", tg_mem_size, in_type, out_type);
+  auto kernel = get_fft_kernel(d, kname, hash_name, func_consts, template_def);
+
+  // Get the bluestein constants
+  auto [w_k, w_q] =
+      compute_bluestein_constants(plan.size(), plan.bluestein_size());
+  d.add_temporary(w_k, s.index);
+  d.add_temporary(w_q, s.index);
+
+  // Launch it
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder.set_compute_pipeline_state(kernel);
+  compute_encoder.set_input_array(in, 0);
+  compute_encoder.set_output_array(out, 1);
+  compute_encoder.set_input_array(w_q, 2);
+  compute_encoder.set_input_array(w_k, 3);
+  compute_encoder.set_bytes(plan.size(), 4);
+  compute_encoder.set_bytes(n, 5);
+  compute_encoder.set_bytes(total_batch_size, 6);
+
+  MTL::Size group_dims(1, tg_batch_size, threads_per_fft);
+  MTL::Size grid_dims(batch_size, tg_batch_size, threads_per_fft);
+  compute_encoder.dispatch_threads(grid_dims, group_dims);
+}
+
+void fft_multi_upload_bluestein(
+    const FFTPlan& plan,
+    const array& in_,
+    array& out,
+    size_t axis,
+    bool inverse,
+    bool real,
+    metal::Device& d,
+    const Stream& s) {
+  // Get Bluestein's constants using the CPU (this is done in the submission
+  // thread which is pretty bad).
+  auto [w_k, w_q] =
+      compute_bluestein_constants(plan.size(), plan.bluestein_size());
+  d.add_temporary(w_k, s.index);
+  d.add_temporary(w_q, s.index);
+
+  // Prepare the input
+  auto in_shape = inverse ? out.shape() : in_.shape();
+  array in(std::move(in_shape), complex64, nullptr, {});
+  if (real && !inverse) {
+    copy_gpu(
+        in_,
+        in,
+        in_.flags().row_contiguous ? CopyType::Vector : CopyType::General,
+        s);
+    d.add_temporary(in, s.index);
+  } else if (real && inverse) {
+    int back_offset = plan.size() % 2 == 0 ? 2 : 1;
+    auto slice_shape = in.shape();
+    slice_shape[axis] -= back_offset;
+    array slice_temp(slice_shape, complex64, nullptr, {});
+    array conj_temp(in.shape(), complex64, nullptr, {});
+    Shape rstarts(in.ndim(), 0);
+    Shape rstrides(in.ndim(), 1);
+    rstarts[axis] = in.shape(axis) - back_offset;
+    rstrides[axis] = -1;
+    unary_op_gpu({in_}, conj_temp, "Conjugate", s);
+    slice_gpu(in_, slice_temp, rstarts, rstrides, s);
+    concatenate_gpu({conj_temp, slice_temp}, in, (int)axis, s);
+    d.add_temporary(conj_temp, s.index);
+  } else if (inverse) {
+    unary_op_gpu({in_}, in, "Conjugate", s);
+    d.add_temporary(in, s.index);
+  } else {
+    in.copy_shared_buffer(in_);
+  }
+
+  // Multiply with
+  Strides b_strides(in.ndim(), 0);
+  b_strides[axis] = 1;
+  array w_k_broadcast(in.shape(), complex64, nullptr, {});
+  w_k_broadcast.copy_shared_buffer(w_k, b_strides, {}, w_k.data_size());
+  array x(in.shape(), complex64, nullptr, {});
+  binary_op_gpu({in, w_k_broadcast}, x, "Multiply", s);
+  d.add_temporary(x, s.index);
+
+  // Pad
+  auto padded_shape = out.shape();
+  padded_shape[axis] = plan.bluestein_size();
+  array padded_x(padded_shape, complex64, nullptr, {});
+  auto zero = array(complex64_t{0.0f, 0.0f});
+  pad_gpu(x, zero, padded_x, {(int)axis}, {0}, s);
+  d.add_temporary(zero, s.index);
+  d.add_temporary(padded_x, s.index);
+
+  // First fft
+}
+
+void fft_op_inplace(
     const array& in,
     array& out,
     size_t axis,
     bool inverse,
     bool real,
-    bool inplace,
+    metal::Device& d,
     const Stream& s) {
-  fft_op(in, out, axis, inverse, real, FourStepParams(), inplace, s);
+  // Get the FFT size and plan it
+  auto plan =
+      FFTPlan(out.dtype() == float32 ? out.shape(axis) : in.shape(axis));
+
+  switch (plan.type()) {
+    case FFTPlan::NOOP:
+      std::cout << "--------------> 1-size FFT <-----------------" << std::endl;
+      break;
+    case FFTPlan::STOCKHAM:
+      return fft_stockham_inplace(plan, in, out, axis, inverse, real, d, s);
+    case FFTPlan::SMALL_FOUR_STEP:
+      return fft_four_step_inplace(plan, in, out, axis, inverse, real, d, s);
+    case FFTPlan::BLUESTEIN:
+      return fft_bluestein(plan, in, out, axis, inverse, real, d, s);
+    case FFTPlan::UNSUPPORTED: {
+      std::string msg;
+      concatenate(msg, "FFT of size ", plan.size(), " not supported");
+      throw std::runtime_error(msg);
+    }
+    default:
+      std::cout << "----- NYI ----" << std::endl;
+      break;
+  }
 }
 
-void nd_fft_op(
+void nd_fft_op_inplace(
     const array& in,
     array& out,
     const std::vector<size_t>& axes,
     bool inverse,
     bool real,
+    metal::Device& d,
     const Stream& s) {
-  // Perform ND FFT on GPU as a series of 1D FFTs
-  auto temp_shape = inverse ? in.shape() : out.shape();
-  array temp1(temp_shape, complex64, nullptr, {});
-  array temp2(temp_shape, complex64, nullptr, {});
-  std::vector<array> temp_arrs = {temp1, temp2};
-  for (int i = axes.size() - 1; i >= 0; i--) {
-    int reverse_index = axes.size() - i - 1;
-    // For 5D and above, we don't want to reallocate our two temporary arrays
-    bool inplace = reverse_index >= 3 && i != 0;
-    // Opposite order for fft vs ifft
-    int index = inverse ? reverse_index : i;
-    size_t axis = axes[index];
-    // Mirror np.fft.(i)rfftn and perform a real transform
-    // only on the final axis.
-    bool step_real = (real && index == axes.size() - 1);
-    auto step_shape = inverse ? out.shape(axis) : in.shape(axis);
-    const array& in_arr = i == axes.size() - 1 ? in : temp_arrs[1 - i % 2];
-    array& out_arr = i == 0 ? out : temp_arrs[i % 2];
-    fft_op(in_arr, out_arr, axis, inverse, step_real, inplace, s);
-  }
+  // We are going to make and possibly reuse some intermediate arrays that will
+  // hold the intermediate fft results.
+  auto shape = inverse ? in.shape() : out.shape();
+  std::vector<array> intermediates;
+  intermediates.reserve(2);
 
-  auto& d = metal::device(s.device);
-  d.add_temporaries(std::move(temp_arrs), s.index);
+  // Utility to return either in or one of the intermediates.
+  auto get_input_array = [&](int step) -> const array& {
+    // The first step so use the input array
+    if (step == 0) {
+      return in;
+    }
+
+    return intermediates[(step - 1) % 2];
+  };
+
+  // Utility to return either out or one of the intermediates. It also informs
+  // us if we should allocate memory for that output or there is already some
+  // allocated.
+  auto get_output_array = [&](int step) -> array& {
+    // It is the final step so return the output array
+    if (step == axes.size() - 1) {
+      return out;
+    }
+
+    // We already have made an array that we can use so go ahead and use it and
+    // don't reallocate the memory.
+    if (step % 2 < intermediates.size()) {
+      return intermediates[step % 2];
+    }
+
+    array x(shape, complex64, nullptr, {});
+    x.set_data(allocator::malloc(x.nbytes()));
+    intermediates.emplace_back(std::move(x));
+    d.add_temporary(intermediates.back(), s.index);
+
+    return intermediates.back();
+  };
+
+  // Perform ND FFT on GPU as a series of 1D FFTs
+  for (int step = 0; step < axes.size(); step++) {
+    auto x = get_input_array(step);
+    auto y = get_output_array(step);
+    auto step_axis = axes[inverse ? step : axes.size() - step - 1];
+    auto step_real = real && (inverse ? step == axes.size() - 1 : step == 0);
+    fft_op_inplace(x, y, step_axis, inverse, step_real, d, s);
+  }
 }
 
 void FFT::eval_gpu(const std::vector<array>& inputs, array& out) {
   auto& s = stream();
+  auto& d = metal::device(s.device);
   auto& in = inputs[0];
 
+  // The FFT ops above have the *_inplace suffix. This means that the memory
+  // needs to be already allocated in the output array. Similar to
+  // copy_gpu_inplace and so on.
+  //
+  // Even though we allocate the memory, we do not necessarily want the
+  // contiguous strides so the *_inplace ops may change the strides and flags
+  // of the array but will not reallocate the memory.
+
+  out.set_data(allocator::malloc(out.nbytes()));
+
   if (axes_.size() > 1) {
-    nd_fft_op(in, out, axes_, inverse_, real_, s);
+    nd_fft_op_inplace(in, out, axes_, inverse_, real_, d, s);
   } else {
-    fft_op(in, out, axes_[0], inverse_, real_, /*inplace=*/false, s);
+    fft_op_inplace(in, out, axes_[0], inverse_, real_, d, s);
   }
 }
 

mlx/backend/metal/hadamard.cpp

@@ -1,11 +1,9 @@
 // Copyright © 2024 Apple Inc.
 
-#include <map>
-
-#include "mlx/backend/common/compiled.h"
 #include "mlx/backend/common/hadamard.h"
+#include "mlx/backend/common/compiled.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/jit/includes.h"
 #include "mlx/backend/metal/kernels.h"
@@ -15,7 +13,6 @@
 namespace mlx::core {
 
 constexpr int MAX_HADAMARD_THREADS_PER_GROUP = 256;
-constexpr int MAX_HADAMARD_BYTES = 32768; // 32KB
 
 std::string gen_hadamard_codelet(int m) {
   // Generate a O(m^2) hadamard codelet for a given M
@@ -60,121 +57,142 @@ std::string gen_hadamard_codelet(int m) {
   return source.str();
 }
 
-void Hadamard::eval_gpu(const std::vector<array>& inputs, array& out) {
-  auto& s = stream();
+void hadamard_mn_contiguous(
+    const array& x,
+    array& y,
+    int m,
+    int n1,
+    int n2,
+    float scale,
+    metal::Device& d,
+    const Stream& s) {
+  int n = n1 * n2;
+  int read_width_n1 = n1 == 2 ? 2 : 4;
+  int read_width_n2 = n2 == 2 ? 2 : 4;
+  int read_width_m = (n == 2 || m == 28) ? 2 : 4;
+  int max_radix_1 = std::min(n1, 16);
+  int max_radix_2 = std::min(n2, 16);
+  float scale_n1 = 1.0;
+  float scale_n2 = (m == 1) ? scale : 1.0;
+  float scale_m = scale;
 
-  auto& in = inputs[0];
+  // n2 is a row contiguous power of 2 hadamard transform
+  MTL::Size group_dims_n2(n2 / max_radix_2, 1, 1);
+  MTL::Size grid_dims_n2(n2 / max_radix_2, x.size() / n2, 1);
 
-  std::vector<array> copies;
-  // Only support the last axis for now
-  int axis = in.ndim() - 1;
-  auto check_input = [&copies, &s](const array& x) {
-    // TODO(alexbarron) pass strides to kernel to relax this constraint
-    bool no_copy = x.flags().row_contiguous;
-    if (no_copy) {
-      return x;
-    } else {
-      copies.push_back(array(x.shape(), x.dtype(), nullptr, {}));
-      copy_gpu(x, copies.back(), CopyType::General, s);
-      return copies.back();
+  // n1 is a strided power of 2 hadamard transform with stride n2
+  MTL::Size group_dims_n1(n1 / max_radix_1, 1, 1);
+  MTL::Size grid_dims_n1(n1 / max_radix_1, x.size() / n, n2);
+
+  // m is a strided hadamard transform with stride n = n1 * n2
+  MTL::Size group_dims_m(
+      std::min(n / read_width_m, MAX_HADAMARD_THREADS_PER_GROUP), 1, 1);
+  MTL::Size grid_dims_m(
+      group_dims_m.width, x.size() / m / read_width_m / group_dims_m.width, 1);
+
+  // Make the kernel
+  std::string kname;
+  kname.reserve(32);
+  concatenate(kname, "hadamard_", n * m, "_", type_to_name(x));
+  auto lib = d.get_library(kname, [&]() {
+    std::string kernel;
+    concatenate(
+        kernel,
+        metal::utils(),
+        gen_hadamard_codelet(m),
+        metal::hadamard(),
+        get_template_definition(
+            "n2" + kname,
+            "hadamard_n",
+            get_type_string(x.dtype()),
+            n2,
+            max_radix_2,
+            read_width_n2));
+    if (n1 > 1) {
+      kernel += get_template_definition(
+          "n1" + kname,
+          "hadamard_n",
+          get_type_string(x.dtype()),
+          n1,
+          max_radix_1,
+          read_width_n1,
+          n2);
     }
-  };
-  const array& in_contiguous = check_input(in);
-
-  if (in_contiguous.is_donatable()) {
-    out.copy_shared_buffer(in_contiguous);
-  } else {
-    out.set_data(allocator::malloc(out.nbytes()));
-  }
-
-  int n, m;
-  std::tie(n, m) = decompose_hadamard(in.shape(axis));
-
-  if (n * (int)size_of(in.dtype()) > MAX_HADAMARD_BYTES) {
-    throw std::invalid_argument(
-        "[hadamard] For n = m*2^k, 2^k > 8192 for FP32 or 2^k > 16384 for FP16/BF16 NYI");
-  }
-
-  int max_radix = std::min(n, 16);
-  // Use read_width 2 for m = 28 to avoid register spilling
-  int read_width = (n == 2 || m == 28) ? 2 : 4;
-
-  std::ostringstream kname;
-  kname << "hadamard_" << n * m << "_" << type_to_name(out);
-  auto kernel_name = kname.str();
-  auto& d = metal::device(s.device);
-  const auto& lib_name = kernel_name;
-  auto lib = d.get_library(lib_name, [&]() {
-    std::ostringstream kernel_source;
-    auto codelet = gen_hadamard_codelet(m);
-    kernel_source << metal::utils() << codelet << metal::hadamard();
-    kernel_source << get_template_definition(
-        "n" + kernel_name,
-        "hadamard_n",
-        get_type_string(in.dtype()),
-        n,
-        max_radix,
-        read_width);
-    kernel_source << get_template_definition(
-        "m" + kernel_name,
-        "hadamard_m",
-        get_type_string(in.dtype()),
-        n,
-        m,
-        read_width);
-    return kernel_source.str();
+    if (m > 1) {
+      kernel += get_template_definition(
+          "m" + kname,
+          "hadamard_m",
+          get_type_string(x.dtype()),
+          n,
+          m,
+          read_width_m);
+    }
+    return kernel;
   });
 
-  int batch_size = in.size() / n;
-  int threads_per = n / max_radix;
-
-  auto& compute_encoder = d.get_command_encoder(s.index);
-
-  auto launch_hadamard = [&](const array& in,
-                             array& out,
-                             const std::string& kernel_name,
-                             float scale) {
-    auto kernel = d.get_kernel(kernel_name, lib);
-    assert(threads_per <= kernel->maxTotalThreadsPerThreadgroup());
-
+  // Launch the strided transform for n1
+  if (n1 > 1) {
+    auto& compute_encoder = d.get_command_encoder(s.index);
+    auto kernel = d.get_kernel("n1" + kname, lib);
     compute_encoder.set_compute_pipeline_state(kernel);
-    compute_encoder.set_input_array(in, 0);
-    compute_encoder.set_output_array(out, 1);
-    compute_encoder.set_bytes(scale, 2);
-
-    MTL::Size group_dims = MTL::Size(1, threads_per, 1);
-    MTL::Size grid_dims = MTL::Size(batch_size, threads_per, 1);
-    compute_encoder.dispatch_threads(grid_dims, group_dims);
-  };
-
-  if (m > 1) {
-    // When m is greater than 1, we decompose the
-    // computation into two uploads to the GPU:
-    //
-    // e.g. len(x) = 12*4 = 48, m = 12, n = 4
-    //
-    // y = h48 @ x
-    //
-    // Upload 1:
-    // tmp = a.reshape(12, 4) @ h4
-    //
-    // Upload 2:
-    // y = h12 @ tmp
-    array temp(in.shape(), in.dtype(), nullptr, {});
-    temp.set_data(allocator::malloc(temp.nbytes()));
-    copies.push_back(temp);
-
-    launch_hadamard(in_contiguous, temp, "n" + kernel_name, 1.0);
-
-    // Metal sometimes reports 256 max threads per group for hadamard_m kernel
-    threads_per = std::min(n / read_width, MAX_HADAMARD_THREADS_PER_GROUP);
-    batch_size = in.size() / m / read_width / threads_per;
-    launch_hadamard(temp, out, "m" + kernel_name, scale_);
-  } else {
-    launch_hadamard(in_contiguous, out, "n" + kernel_name, scale_);
+    compute_encoder.set_input_array(x, 0);
+    compute_encoder.set_output_array(y, 1);
+    compute_encoder.set_bytes(scale_n1, 2);
+    compute_encoder.dispatch_threads(grid_dims_n1, group_dims_n1);
   }
 
-  d.add_temporaries(std::move(copies), s.index);
+  // Launch the transform for n2
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  auto kernel = d.get_kernel("n2" + kname, lib);
+  compute_encoder.set_compute_pipeline_state(kernel);
+  compute_encoder.set_input_array(n1 > 1 ? y : x, 0);
+  compute_encoder.set_output_array(y, 1);
+  compute_encoder.set_bytes(scale_n2, 2);
+  compute_encoder.dispatch_threads(grid_dims_n2, group_dims_n2);
+
+  // Launch the strided transform for m
+  if (m > 1) {
+    auto kernel = d.get_kernel("m" + kname, lib);
+    compute_encoder.set_compute_pipeline_state(kernel);
+    compute_encoder.set_input_array(y, 0);
+    compute_encoder.set_output_array(y, 1);
+    compute_encoder.set_bytes(scale_m, 2);
+    compute_encoder.dispatch_threads(grid_dims_m, group_dims_m);
+  }
+}
+
+void Hadamard::eval_gpu(const std::vector<array>& inputs, array& out) {
+  auto& s = stream();
+  auto& d = metal::device(s.device);
+  auto& in = inputs[0];
+
+  // Split the hadamard transform so that all of them work on vectors smaller
+  // than 8192 elements.
+  //
+  // We decompose it in the following way:
+  //
+  // n = m * n1 * n2 = m * 2^k1 * 2^k2
+  //
+  // where m is in (1, 12, 20, 28) and n1 and n2 <= 8192
+  auto [n, m] = decompose_hadamard(in.shape().back());
+  int n1 = 1, n2 = n;
+  if (n > 8192) {
+    for (n2 = 2; n2 * n2 < n; n2 *= 2) {
+    }
+    n1 = n / n2;
+  }
+
+  if (in.flags().row_contiguous) {
+    if (in.is_donatable()) {
+      out.copy_shared_buffer(in);
+    } else {
+      out.set_data(allocator::malloc(out.nbytes()));
+    }
+    hadamard_mn_contiguous(in, out, m, n1, n2, scale_, d, s);
+  } else {
+    copy_gpu(in, out, CopyType::General, s);
+    hadamard_mn_contiguous(out, out, m, n1, n2, scale_, d, s);
+  }
 }
 
 } // namespace mlx::core

mlx/backend/metal/indexing.cpp

@@ -2,7 +2,7 @@
 #include <fmt/format.h>
 
 #include "mlx/backend/common/compiled.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/jit/includes.h"
 #include "mlx/backend/metal/jit/indexing.h"

mlx/backend/metal/kernels/binary.h

@@ -9,64 +9,85 @@ template <typename T, typename U, typename Op>
   c[index] = Op()(a[0], b[0]);
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_sv(
     device const T* a,
     device const T* b,
     device U* c,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  c[index] = Op()(a[0], b[index]);
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    c[index + i] = Op()(a[0], b[index + i]);
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vs(
     device const T* a,
     device const T* b,
     device U* c,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  c[index] = Op()(a[index], b[0]);
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    c[index + i] = Op()(a[index + i], b[0]);
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vv(
     device const T* a,
     device const T* b,
     device U* c,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  c[index] = Op()(a[index], b[index]);
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    c[index + i] = Op()(a[index + i], b[index + i]);
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_sv2(
     device const T* a,
     device const T* b,
     device U* c,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  int64_t offset = index.x + grid_dim.x * int64_t(index.y);
-  c[offset] = Op()(a[0], b[offset]);
+  int64_t offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    c[offset + i] = Op()(a[0], b[offset + i]);
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vs2(
     device const T* a,
     device const T* b,
     device U* c,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  int64_t offset = index.x + grid_dim.x * int64_t(index.y);
-  c[offset] = Op()(a[offset], b[0]);
+  int64_t offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    c[offset + i] = Op()(a[offset + i], b[0]);
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vv2(
     device const T* a,
     device const T* b,
     device U* c,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  int64_t offset = index.x + grid_dim.x * int64_t(index.y);
-  c[offset] = Op()(a[offset], b[offset]);
+  int64_t offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    c[offset + i] = Op()(a[offset + i], b[offset + i]);
+  }
 }
 
 template <typename T, typename U, typename Op, typename IdxT = int64_t>

mlx/backend/metal/kernels/binary.metal

@@ -71,6 +71,7 @@ instantiate_binary_types_bool(Less)
 instantiate_binary_types_bool(LessEqual)
 instantiate_binary_types_bool(NotEqual)
 instantiate_binary_float(LogAddExp)
+instantiate_binary_all(LogAddExp, complex64, complex64_t, complex64_t)
 instantiate_binary_types(Maximum)
 instantiate_binary_types(Minimum)
 instantiate_binary_types(Multiply)

mlx/backend/metal/kernels/binary_ops.h

@@ -130,6 +130,24 @@ struct LogAddExp {
         ? maxval
         : (maxval + log1p(metal::exp(minval - maxval)));
   };
+
+  complex64_t operator()(complex64_t x, complex64_t y) {
+    if (metal::isnan(x.real) || metal::isnan(x.imag) || metal::isnan(y.real) ||
+        metal::isnan(y.imag)) {
+      return metal::numeric_limits<float>::quiet_NaN();
+    }
+    constexpr float inf = metal::numeric_limits<float>::infinity();
+    complex64_t maxval = x > y ? x : y;
+    complex64_t minval = x < y ? x : y;
+    if (minval.real == -inf || maxval.real == inf)
+      return maxval;
+    float m = metal::exp(minval.real - maxval.real);
+    complex64_t dexp{
+        m * metal::cos(minval.imag - maxval.imag),
+        m * metal::sin(minval.imag - maxval.imag),
+    };
+    return maxval + log1p(dexp);
+  }
 };
 
 struct Maximum {

mlx/backend/metal/kernels/binary_two.h

@@ -12,82 +12,103 @@ template <typename T, typename U, typename Op>
   d[index] = out[1];
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_sv(
     device const T* a,
     device const T* b,
     device U* c,
     device U* d,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  auto out = Op()(a[0], b[index]);
-  c[index] = out[0];
-  d[index] = out[1];
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    auto out = Op()(a[0], b[index + i]);
+    c[index + i] = out[0];
+    d[index + i] = out[1];
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vs(
     device const T* a,
     device const T* b,
     device U* c,
     device U* d,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  auto out = Op()(a[index], b[0]);
-  c[index] = out[0];
-  d[index] = out[1];
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    auto out = Op()(a[index + i], b[0]);
+    c[index + i] = out[0];
+    d[index + i] = out[1];
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vv(
     device const T* a,
     device const T* b,
     device U* c,
     device U* d,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  auto out = Op()(a[index], b[index]);
-  c[index] = out[0];
-  d[index] = out[1];
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    auto out = Op()(a[index + i], b[index + i]);
+    c[index + i] = out[0];
+    d[index + i] = out[1];
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_sv2(
     device const T* a,
     device const T* b,
     device U* c,
     device U* d,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  auto offset = index.x + grid_dim.x * int64_t(index.y);
-  auto out = Op()(a[0], b[offset]);
-  c[offset] = out[0];
-  d[offset] = out[1];
+  auto offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    auto out = Op()(a[0], b[offset + i]);
+    c[offset + i] = out[0];
+    d[offset + i] = out[1];
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vs2(
     device const T* a,
     device const T* b,
     device U* c,
     device U* d,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  auto offset = index.x + grid_dim.x * int64_t(index.y);
-  auto out = Op()(a[offset], b[0]);
-  c[offset] = out[0];
-  d[offset] = out[1];
+  auto offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    auto out = Op()(a[offset + i], b[0]);
+    c[offset + i] = out[0];
+    d[offset + i] = out[1];
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void binary_vv2(
     device const T* a,
     device const T* b,
     device U* c,
     device U* d,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  auto offset = index.x + grid_dim.x * int64_t(index.y);
-  auto out = Op()(a[offset], b[offset]);
-  c[offset] = out[0];
-  d[offset] = out[1];
+  auto offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    auto out = Op()(a[offset + i], b[offset + i]);
+    c[offset + i] = out[0];
+    d[offset + i] = out[1];
+  }
 }
 
 template <typename T, typename U, typename Op, typename IdxT = int64_t>

mlx/backend/metal/kernels/copy.h

@@ -1,39 +1,53 @@
 // Copyright © 2024 Apple Inc.
 
-template <typename T, typename U>
+template <typename T, typename U, int N = WorkPerThread<T>::n>
 [[kernel]] void copy_s(
     device const T* src [[buffer(0)]],
     device U* dst [[buffer(1)]],
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  dst[index] = static_cast<U>(src[0]);
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    dst[index + i] = static_cast<U>(src[0]);
+  }
 }
 
-template <typename T, typename U>
+template <typename T, typename U, int N = WorkPerThread<T>::n>
 [[kernel]] void copy_v(
     device const T* src [[buffer(0)]],
     device U* dst [[buffer(1)]],
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  dst[index] = static_cast<U>(src[index]);
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    dst[index + i] = static_cast<U>(src[index + i]);
+  }
 }
 
-template <typename T, typename U>
+template <typename T, typename U, int N = WorkPerThread<T>::n>
 [[kernel]] void copy_s2(
     device const T* src [[buffer(0)]],
     device U* dst [[buffer(1)]],
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  auto offset = index.x + grid_dim.x * int64_t(index.y);
-  dst[offset] = static_cast<U>(src[0]);
+  auto offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    dst[offset + i] = static_cast<U>(src[0]);
+  }
 }
 
-template <typename T, typename U>
+template <typename T, typename U, int N = WorkPerThread<T>::n>
 [[kernel]] void copy_v2(
     device const T* src [[buffer(0)]],
     device U* dst [[buffer(1)]],
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  auto offset = index.x + grid_dim.x * int64_t(index.y);
-  dst[offset] = static_cast<U>(src[offset]);
+  auto offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    dst[offset + i] = static_cast<U>(src[offset + i]);
+  }
 }
 
 template <typename T, typename U, typename IdxT = int64_t>

mlx/backend/metal/kernels/fft/readwrite.h

@@ -10,7 +10,7 @@ For many sizes, GPU FFTs are memory bandwidth bound so
 read/write performance is important.
 
 Where possible, we read 128 bits sequentially in each thread,
-coalesced with accesses from adajcent threads for optimal performance.
+coalesced with accesses from adjacent threads for optimal performance.
 
 We implement specialized reading/writing for:
   - FFT

mlx/backend/metal/kernels/hadamard.h

@@ -26,7 +26,7 @@ METAL_FUNC void radix_func(thread float* x) {
   }
 }
 
-template <typename T, int N, int max_radix, int read_width>
+template <typename T, int N, int max_radix, int read_width, int stride = 1>
 [[kernel]] void hadamard_n(
     const device T* in [[buffer(0)]],
     device T* out [[buffer(1)]],
@@ -46,18 +46,25 @@ template <typename T, int N, int max_radix, int read_width>
   constexpr short logFinal = logN % logR;
   constexpr short final_radix = 1 << (logFinal);
 
-  int batch_idx = elem.x * N;
-  short i = elem.y;
+  int batch_idx = elem.y * N * stride + elem.z;
+  short i = elem.x;
 
   threadgroup T buf[N];
 
   // Read values from device
-  STEEL_PRAGMA_UNROLL
-  for (short j = 0; j < max_radix / read_width; j++) {
-    short index = j * read_width * num_threads + i * read_width;
+  if (stride == 1) {
     STEEL_PRAGMA_UNROLL
-    for (short r = 0; r < read_width; r++) {
-      buf[index + r] = in[batch_idx + index + r];
+    for (short j = 0; j < max_radix / read_width; j++) {
+      short index = j * read_width * num_threads + i * read_width;
+      STEEL_PRAGMA_UNROLL
+      for (short r = 0; r < read_width; r++) {
+        buf[index + r] = in[batch_idx + index + r];
+      }
+    }
+  } else {
+    STEEL_PRAGMA_UNROLL
+    for (short j = 0; j < max_radix; j++) {
+      buf[j * num_threads + i] = in[batch_idx + (j * num_threads + i) * stride];
     }
   }
 
@@ -113,12 +120,20 @@ template <typename T, int N, int max_radix, int read_width>
   }
 
   // Write values to device
-  STEEL_PRAGMA_UNROLL
-  for (short j = 0; j < max_radix / read_width; j++) {
-    short index = j * read_width * num_threads + i * read_width;
+  if (stride == 1) {
     STEEL_PRAGMA_UNROLL
-    for (short r = 0; r < read_width; r++) {
-      out[batch_idx + index + r] = T(buf[index + r] * scale);
+    for (short j = 0; j < max_radix / read_width; j++) {
+      short index = j * read_width * num_threads + i * read_width;
+      STEEL_PRAGMA_UNROLL
+      for (short r = 0; r < read_width; r++) {
+        out[batch_idx + index + r] = T(buf[index + r] * scale);
+      }
+    }
+  } else {
+    STEEL_PRAGMA_UNROLL
+    for (short j = 0; j < max_radix; j++) {
+      out[batch_idx + (j * num_threads + i) * stride] =
+          buf[j * num_threads + i];
     }
   }
 }

mlx/backend/metal/kernels/quantized.h

@@ -1008,11 +1008,11 @@ METAL_FUNC void qmm_t_impl(
 
   auto wl = (const device uint8_t*)w;
 
-  x += y_row * K;
+  x += y_row * static_cast<int64_t>(K);
   wl += y_col * K_w;
   scales += y_col * K_g;
   biases += y_col * K_g;
-  y += y_row * N + y_col;
+  y += y_row * static_cast<int64_t>(N) + y_col;
 
   // Make the x loader and mma operation
   const short num_els = min(BM, M - y_row);
@@ -1132,11 +1132,11 @@ METAL_FUNC void qmm_n_impl(
   // Set the block
   const int y_row = tid.y * BM;
   const int y_col = tid.x * BN;
-  x += y_row * K;
+  x += y_row * static_cast<int64_t>(K);
   wl += y_col * bytes_per_pack / pack_factor;
   scales += y_col / group_size;
   biases += y_col / group_size;
-  y += y_row * N + y_col;
+  y += y_row * static_cast<int64_t>(N) + y_col;
 
   // Make the x loader and mma operation
   const short num_els = min(BM, M - y_row);

mlx/backend/metal/kernels/scan.metal

@@ -104,4 +104,5 @@ instantiate_scan_helper(min_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumMi
 instantiate_scan_helper(min_complex64_complex64, complex64_t, complex64_t, CumMin, 2)
 instantiate_scan_helper(logaddexp_float16_float16,     half,        half,        CumLogaddexp, 4)
 instantiate_scan_helper(logaddexp_float32_float32,     float,       float,       CumLogaddexp, 4)
-instantiate_scan_helper(logaddexp_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumLogaddexp, 4) // clang-format on
+instantiate_scan_helper(logaddexp_bfloat16_bfloat16,   bfloat16_t,  bfloat16_t,  CumLogaddexp, 4)
+instantiate_scan_helper(logaddexp_complex64_complex64, complex64_t, complex64_t, CumLogaddexp, 2) // clang-format on

mlx/backend/metal/kernels/sdpa_vector.h

@@ -56,9 +56,9 @@ template <typename T, int D, int V = D>
   const int head_idx = tid.x;
   const int q_seq_idx = tid.y;
   const int kv_head_idx = head_idx / gqa_factor;
-  const int o_offset = tpg.x * q_seq_idx + head_idx;
+  const int o_offset = head_idx * tpg.y + q_seq_idx;
   const int q_offset =
-      query_transposed ? o_offset : head_idx * tpg.y + q_seq_idx;
+      query_transposed ? tpg.x * q_seq_idx + head_idx : o_offset;
   queries += q_offset * D + simd_lid * qk_per_thread;
   keys += kv_head_idx * k_head_stride + simd_gid * k_seq_stride +
       simd_lid * qk_per_thread;
@@ -213,9 +213,9 @@ template <typename T, int D, int V = D>
   const int block_idx = tid.z;
   const int head_idx = tid.x;
   const int q_seq_idx = tid.y;
-  const int o_offset = tpg.x * q_seq_idx + head_idx;
+  const int o_offset = head_idx * tpg.y + q_seq_idx;
   const int q_offset =
-      query_transposed ? o_offset : head_idx * tpg.y + q_seq_idx;
+      query_transposed ? tpg.x * q_seq_idx + head_idx : o_offset;
   const int kv_head_idx = head_idx / gqa_factor;
 
   queries += q_offset * D + simd_lid * qk_per_thread;
@@ -358,8 +358,8 @@ template <typename T, int D>
   // Adjust positions
   const int head_idx = tid.x;
   const int q_seq_idx = tid.y;
-  const int n_heads = tpg.x;
-  const int q_offset = n_heads * q_seq_idx + head_idx;
+  const int q_offset = head_idx * tpg.y + q_seq_idx;
+  ;
   partials += q_offset * blocks * D + simd_gid * D + simd_lid * elem_per_thread;
   sums += q_offset * blocks;
   maxs += q_offset * blocks;

mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.h

@@ -95,7 +95,7 @@ template <
 
   Q += tidl.z * params->Q_strides[0] + // Batch
       tidl.y * params->Q_strides[1] + // Head
-      tidl.x * BQ * params->Q_strides[2]; // Seqeunce
+      tidl.x * BQ * params->Q_strides[2]; // Sequence
 
   ulong kv_head_idx = int(tid.y) / params->gqa_factor;
   K += tidl.z * params->K_strides[0] + // Batch
@@ -106,7 +106,7 @@ template <
 
   O += tidl.z * params->O_strides[0] + // Batch
       tidl.y * params->O_strides[1] + // Head
-      tidl.x * BQ * params->O_strides[2]; // Seqeunce
+      tidl.x * BQ * params->O_strides[2]; // Sequence
 
   if (has_mask) {
     mask += tidl.z * mask_params->M_strides[0] + // Batch

mlx/backend/metal/kernels/steel/attn/loader.h

@@ -113,7 +113,7 @@ struct BlockLoader {
         tmp_val[j] = src[(tmp_idx[j] ? i * src_ld + j : 0)];
       }
 
-      // Zero out uneeded values
+      // Zero out unneeded values
       STEEL_PRAGMA_UNROLL
       for (short j = 0; j < vec_size; j++) {
         tmp_val[j] = tmp_idx[j] ? tmp_val[j] : T(0);
@@ -240,7 +240,7 @@ struct BlockLoaderT {
         tmp_val[j] = src[(tmp_idx[j] ? i * src_ld + j : 0)];
       }
 
-      // Zero out uneeded values
+      // Zero out unneeded values
       STEEL_PRAGMA_UNROLL
       for (short j = 0; j < vec_size; j++) {
         tmp_val[j] = tmp_idx[j] ? tmp_val[j] : T(0);

mlx/backend/metal/kernels/steel/conv/kernels/steel_conv_general.h

@@ -141,7 +141,7 @@ implicit_gemm_conv_2d_general(
 
   // Store results to device memory
   {
-    // Adjust for simdgroup and thread locatio
+    // Adjust for simdgroup and thread location
     int offset_m = c_row + mma_op.sm;
     int offset_n = c_col + mma_op.sn;
     C += offset_n;

mlx/backend/metal/kernels/steel/gemm/loader.h

@@ -113,7 +113,7 @@ struct BlockLoader {
         tmp_val[j] = src[(tmp_idx[j] ? i * src_ld + j : 0)];
       }
 
-      // Zero out uneeded values
+      // Zero out unneeded values
       STEEL_PRAGMA_UNROLL
       for (short j = 0; j < vec_size; j++) {
         tmp_val[j] = tmp_idx[j] ? tmp_val[j] : T(0);

mlx/backend/metal/kernels/ternary.h

@@ -1,25 +1,32 @@
 // Copyright © 2024 Apple Inc.
 
-template <typename T, typename Op>
+template <typename T, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void ternary_v(
     device const bool* a,
     device const T* b,
     device const T* c,
     device T* d,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  d[index] = Op()(a[index], b[index], c[index]);
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    d[index + i] = Op()(a[index + i], b[index + i], c[index + i]);
+  }
 }
 
-template <typename T, typename Op>
+template <typename T, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void ternary_v2(
     device const bool* a,
     device const T* b,
     device const T* c,
     device T* d,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  auto offset = index.x + grid_dim.x * int64_t(index.y);
-  d[offset] = Op()(a[offset], b[offset], c[offset]);
+  auto offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    d[offset + i] = Op()(a[offset + i], b[offset + i], c[offset + i]);
+  }
 }
 
 template <typename T, typename Op, typename IdxT = int64_t>

mlx/backend/metal/kernels/unary.h

@@ -1,21 +1,28 @@
 // Copyright © 2024 Apple Inc.
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void unary_v(
     device const T* in,
     device U* out,
+    constant uint& size,
     uint index [[thread_position_in_grid]]) {
-  out[index] = Op()(in[index]);
+  index *= N;
+  for (int i = 0; i < N && (index + i) < size; ++i) {
+    out[index + i] = Op()(in[index + i]);
+  }
 }
 
-template <typename T, typename U, typename Op>
+template <typename T, typename U, typename Op, int N = WorkPerThread<T>::n>
 [[kernel]] void unary_v2(
     device const T* in,
     device U* out,
+    constant int64_t& size,
     uint2 index [[thread_position_in_grid]],
     uint2 grid_dim [[threads_per_grid]]) {
-  auto offset = index.x + grid_dim.x * int64_t(index.y);
-  out[offset] = Op()(in[offset]);
+  auto offset = N * (index.x + grid_dim.x * int64_t(index.y));
+  for (int i = 0; i < N && (offset + i) < size; ++i) {
+    out[offset + i] = Op()(in[offset + i]);
+  }
 }
 
 template <

mlx/backend/metal/kernels/unary.metal

@@ -77,6 +77,7 @@ instantiate_unary_all_same(Cos, complex64, complex64_t)
 instantiate_unary_all_same(Cosh, complex64, complex64_t)
 instantiate_unary_all_same(Exp, complex64, complex64_t)
 instantiate_unary_all_same(Log, complex64, complex64_t)
+instantiate_unary_all_same(Log1p, complex64, complex64_t)
 instantiate_unary_all_same(Log2, complex64, complex64_t)
 instantiate_unary_all_same(Log10, complex64, complex64_t)
 instantiate_unary_all_same(Negative, complex64, complex64_t)

mlx/backend/metal/kernels/utils.h

@@ -15,6 +15,14 @@
 
 typedef half float16_t;
 
+// Work per thread values for different types. The values here are expected to
+// match get_work_per_thread in mlx/backend/metal/utils.h
+template <typename U>
+struct WorkPerThread {
+  static_assert(sizeof(U) <= 8, "Type too large");
+  static constexpr int constant n = 8 / sizeof(U);
+};
+
 ///////////////////////////////////////////////////////////////////////////////
 // Type limits utils
 ///////////////////////////////////////////////////////////////////////////////
@@ -328,6 +336,23 @@ inline bfloat16_t log1p(bfloat16_t x) {
   return bfloat16_t(x * (metal::log(xp1) / (xp1 - 1.0f)));
 }
 
+inline complex64_t log1p(complex64_t in) {
+  float x = in.real;
+  float y = in.imag;
+  float zabs = metal::precise::sqrt(x * x + y * y);
+  float theta = metal::atan2(y, x + 1);
+  if (zabs < 0.5f) {
+    float r = x * (2 + x) + y * y;
+    if (r == 0) { // handle underflow
+      return {x, theta};
+    }
+    return {0.5f * log1p(r), theta};
+  } else {
+    auto z0 = metal::sqrt((x + 1) * (x + 1) + y * y);
+    return {metal::log(z0), theta};
+  }
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // SIMD shuffle ops
 ///////////////////////////////////////////////////////////////////////////////

mlx/backend/metal/logsumexp.cpp

@@ -1,7 +1,7 @@
 // Copyright © 2023-2024 Apple Inc.
 #include <algorithm>
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/utils.h"

mlx/backend/metal/matmul.cpp

@@ -7,7 +7,7 @@
 
 #include "mlx/backend/common/broadcasting.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/kernels/defines.h"

mlx/backend/metal/metal.cpp

@@ -1,11 +1,11 @@
 // Copyright © 2023-2024 Apple Inc.
 #include <memory>
 
+#include <sys/sysctl.h>
+
 #include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/metal.h"
 #include "mlx/backend/metal/utils.h"
-#include "mlx/primitives.h"
-#include "mlx/scheduler.h"
-#include "mlx/utils.h"
 
 namespace mlx::core::metal {
 
@@ -13,85 +13,6 @@ bool is_available() {
   return true;
 }
 
-inline void check_error(MTL::CommandBuffer* cbuf) {
-  if (cbuf->status() == MTL::CommandBufferStatusError) {
-    std::ostringstream msg;
-    msg << "[METAL] Command buffer execution failed: "
-        << cbuf->error()->localizedDescription()->utf8String();
-    throw std::runtime_error(msg.str());
-  }
-}
-
-void eval(array& arr) {
-  auto pool = new_scoped_memory_pool();
-  auto s = arr.primitive().stream();
-  auto& d = metal::device(s.device);
-  auto command_buffer = d.get_command_buffer(s.index);
-
-  auto outputs = arr.outputs();
-  {
-    // If the array is a tracer hold a reference
-    // to its inputs so they don't get donated
-    std::vector<array> inputs;
-    if (arr.is_tracer()) {
-      inputs = arr.inputs();
-    }
-
-    debug_set_primitive_buffer_label(command_buffer, arr.primitive());
-    arr.primitive().eval_gpu(arr.inputs(), outputs);
-  }
-  std::unordered_set<std::shared_ptr<array::Data>> buffers;
-  for (auto& in : arr.inputs()) {
-    buffers.insert(in.data_shared_ptr());
-  }
-  for (auto& s : arr.siblings()) {
-    buffers.insert(s.data_shared_ptr());
-  }
-  // Remove the output if it was donated to by an input
-  if (auto it = buffers.find(arr.data_shared_ptr()); it != buffers.end()) {
-    buffers.erase(it);
-  }
-
-  if (d.command_buffer_needs_commit(s.index)) {
-    d.end_encoding(s.index);
-    scheduler::notify_new_task(s);
-    command_buffer->addCompletedHandler(
-        [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
-          scheduler::notify_task_completion(s);
-          check_error(cbuf);
-        });
-    d.commit_command_buffer(s.index);
-    d.get_command_buffer(s.index);
-  } else {
-    command_buffer->addCompletedHandler(
-        [s, buffers = std::move(buffers)](MTL::CommandBuffer* cbuf) {
-          check_error(cbuf);
-        });
-  }
-}
-
-void finalize(Stream s) {
-  auto pool = new_scoped_memory_pool();
-  auto& d = metal::device(s.device);
-  auto cb = d.get_command_buffer(s.index);
-  d.end_encoding(s.index);
-  cb->addCompletedHandler([s](MTL::CommandBuffer* cbuf) { check_error(cbuf); });
-  d.commit_command_buffer(s.index);
-  d.get_command_buffer(s.index);
-}
-
-void synchronize(Stream s) {
-  auto pool = new_scoped_memory_pool();
-  auto& d = metal::device(s.device);
-  auto cb = d.get_command_buffer(s.index);
-  cb->retain();
-  d.end_encoding(s.index);
-  d.commit_command_buffer(s.index);
-  cb->waitUntilCompleted();
-  check_error(cb);
-  cb->release();
-}
-
 void start_capture(std::string path, id object) {
   auto pool = new_scoped_memory_pool();
 
@@ -128,4 +49,36 @@ void stop_capture() {
   manager->stopCapture();
 }
 
+const std::unordered_map<std::string, std::variant<std::string, size_t>>&
+device_info() {
+  auto init_device_info = []()
+      -> std::unordered_map<std::string, std::variant<std::string, size_t>> {
+    auto pool = new_scoped_memory_pool();
+    auto raw_device = device(default_device()).mtl_device();
+    auto name = std::string(raw_device->name()->utf8String());
+    auto arch = std::string(raw_device->architecture()->name()->utf8String());
+
+    size_t memsize = 0;
+    size_t length = sizeof(memsize);
+    sysctlbyname("hw.memsize", &memsize, &length, NULL, 0);
+
+    size_t rsrc_limit = 0;
+    sysctlbyname("iogpu.rsrc_limit", &rsrc_limit, &length, NULL, 0);
+    if (rsrc_limit == 0) {
+      rsrc_limit = 499000;
+    }
+
+    return {
+        {"device_name", name},
+        {"architecture", arch},
+        {"max_buffer_length", raw_device->maxBufferLength()},
+        {"max_recommended_working_set_size",
+         raw_device->recommendedMaxWorkingSetSize()},
+        {"memory_size", memsize},
+        {"resource_limit", rsrc_limit}};
+  };
+  static auto device_info_ = init_device_info();
+  return device_info_;
+}
+
 } // namespace mlx::core::metal

mlx/backend/metal/metal.h

@@ -2,11 +2,10 @@
 
 #pragma once
 
+#include <string>
 #include <unordered_map>
 #include <variant>
 
-#include "mlx/array.h"
-
 namespace mlx::core::metal {
 
 /* Check if the Metal backend is available. */

mlx/backend/metal/no_metal.cpp

@@ -0,0 +1,22 @@
+// Copyright © 2025 Apple Inc.
+
+#include <stdexcept>
+
+#include "mlx/backend/metal/metal.h"
+
+namespace mlx::core::metal {
+
+bool is_available() {
+  return false;
+}
+
+void start_capture(std::string) {}
+void stop_capture() {}
+
+const std::unordered_map<std::string, std::variant<std::string, size_t>>&
+device_info() {
+  throw std::runtime_error(
+      "[metal::device_info] Cannot get device info without metal backend");
+};
+
+} // namespace mlx::core::metal

mlx/backend/metal/normalization.cpp

@@ -1,7 +1,7 @@
 // Copyright © 2024 Apple Inc.
 #include <algorithm>
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels/defines.h"
 #include "mlx/backend/metal/reduce.h"

mlx/backend/metal/primitives.cpp

@@ -7,10 +7,10 @@
 #include "mlx/backend/common/compiled.h"
 #include "mlx/backend/common/slicing.h"
 #include "mlx/backend/common/utils.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/backend/gpu/slicing.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
-#include "mlx/backend/metal/slicing.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/primitives.h"
 #include "mlx/scheduler.h"
@@ -25,25 +25,6 @@ void arange_set_scalars(T start, T next, metal::CommandEncoder& enc) {
   enc.set_bytes(step, 1);
 }
 
-void reshape(const array& in, array& out, Stream s) {
-  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
-  if (copy_necessary) {
-    out.set_data(allocator::malloc(out.nbytes()));
-    copy_gpu_inplace(
-        in,
-        out,
-        in.shape(),
-        in.strides(),
-        make_contiguous_strides(in.shape()),
-        0,
-        0,
-        CopyType::General,
-        s);
-  } else {
-    shared_buffer_reshape(in, out_strides, out);
-  }
-}
-
 static array compute_dynamic_offset(
     const array& indices,
     const Strides& strides,
@@ -226,105 +207,10 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   }
 }
 
-void AsType::eval_gpu(const std::vector<array>& inputs, array& out) {
-  CopyType ctype =
-      inputs[0].flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy_gpu(inputs[0], out, ctype);
-}
-
-void AsStrided::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void Broadcast::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void BroadcastAxes::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void Concatenate::eval_gpu(const std::vector<array>& inputs, array& out) {
-  concatenate_gpu(inputs, out, axis_, stream());
-}
-
-void Contiguous::eval_gpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  auto& in = inputs[0];
-  constexpr size_t extra_bytes = 16384;
-  if (in.buffer_size() <= out.nbytes() + extra_bytes &&
-      (in.flags().row_contiguous ||
-       (allow_col_major_ && in.flags().col_contiguous))) {
-    out.copy_shared_buffer(in);
-  } else {
-    copy_gpu(in, out, CopyType::General);
-  }
-}
-
-void Copy::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void CustomTransforms::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  eval(inputs, outputs);
-}
-
-void Depends::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  eval(inputs, outputs);
-}
-
-void Full::eval_gpu(const std::vector<array>& inputs, array& out) {
-  auto in = inputs[0];
-  CopyType ctype;
-  if (in.data_size() == 1) {
-    ctype = CopyType::Scalar;
-  } else if (in.flags().contiguous) {
-    ctype = CopyType::Vector;
-  } else {
-    ctype = CopyType::General;
-  }
-  copy_gpu(in, out, ctype);
-}
-
-void ExpandDims::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void Flatten::eval_gpu(const std::vector<array>& inputs, array& out) {
-  reshape(inputs[0], out, stream());
-}
-
-void Unflatten::eval_gpu(const std::vector<array>& inputs, array& out) {
-  reshape(inputs[0], out, stream());
-}
-
 void Load::eval_gpu(const std::vector<array>& inputs, array& out) {
   throw std::runtime_error("[Load::eval_gpu] Not implemented.");
 }
 
-void NumberOfElements::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void Pad::eval_gpu(const std::vector<array>& inputs, array& out) {
-  // Inputs must be base input array and scalar val array
-  assert(inputs.size() == 2);
-  auto& in = inputs[0];
-  auto& val = inputs[1];
-
-  // Padding value must be a scalar
-  assert(val.size() == 1);
-
-  // Padding value, input and output must be of the same type
-  assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
-
-  pad_gpu(in, val, out, axes_, low_pad_size_, stream());
-}
-
 void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
 
@@ -370,27 +256,6 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   compute_encoder.dispatch_threads(grid_dims, group_dims);
 }
 
-void Reshape::eval_gpu(const std::vector<array>& inputs, array& out) {
-  reshape(inputs[0], out, stream());
-}
-
-void Split::eval_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs) {
-  eval(inputs, outputs);
-}
-
-void Slice::eval_gpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 1);
-  if (out.size() == 0) {
-    out.set_data(nullptr);
-    return;
-  }
-
-  auto& in = inputs[0];
-  slice_gpu(in, out, start_indices_, strides_, stream());
-}
-
 void DynamicSlice::eval_gpu(const std::vector<array>& inputs, array& out) {
   if (out.size() == 0) {
     out.set_data(nullptr);
@@ -492,18 +357,6 @@ void SliceUpdate::eval_gpu(const std::vector<array>& inputs, array& out) {
       /* const Stream& s = */ stream());
 }
 
-void Squeeze::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void StopGradient::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
-void Transpose::eval_gpu(const std::vector<array>& inputs, array& out) {
-  eval(inputs, out);
-}
-
 void QRF::eval_gpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
@@ -537,35 +390,4 @@ void LUF::eval_gpu(
   throw std::runtime_error("[LUF::eval_gpu] Metal LU factorization NYI.");
 }
 
-void View::eval_gpu(const std::vector<array>& inputs, array& out) {
-  auto& in = inputs[0];
-  auto ibytes = size_of(in.dtype());
-  auto obytes = size_of(out.dtype());
-  // Conditions for buffer copying (disjunction):
-  // - type size is the same
-  // - type size is smaller and the last axis is contiguous
-  // - the entire array is row contiguous
-  if (ibytes == obytes || (obytes < ibytes && in.strides().back() == 1) ||
-      in.flags().row_contiguous) {
-    auto strides = in.strides();
-    for (int i = 0; i < static_cast<int>(strides.size()) - 1; ++i) {
-      strides[i] *= ibytes;
-      strides[i] /= obytes;
-    }
-    out.copy_shared_buffer(
-        in, strides, in.flags(), in.data_size() * ibytes / obytes);
-  } else {
-    auto tmp = array(in.shape(), in.dtype(), nullptr, {});
-    tmp.set_data(allocator::malloc(tmp.nbytes()));
-    copy_gpu_inplace(in, tmp, CopyType::General, stream());
-
-    auto flags = out.flags();
-    flags.contiguous = true;
-    flags.row_contiguous = true;
-    auto max_dim = std::max_element(out.shape().begin(), out.shape().end());
-    flags.col_contiguous = out.size() <= 1 || out.size() == *max_dim;
-    out.copy_shared_buffer(tmp, out.strides(), flags, out.size());
-  }
-}
-
 } // namespace mlx::core

mlx/backend/metal/quantized.cpp

@@ -4,7 +4,7 @@
 
 #include "mlx/backend/common/broadcasting.h"
 #include "mlx/backend/common/compiled.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/reduce.h"

mlx/backend/metal/reduce.cpp

@@ -3,7 +3,7 @@
 #include <algorithm>
 #include <cassert>
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/kernels/defines.h"

mlx/backend/metal/resident.cpp

@@ -1,7 +1,6 @@
 // Copyright © 2024 Apple Inc.
 
 #include "mlx/backend/metal/resident.h"
-#include "mlx/backend/metal/metal_impl.h"
 
 namespace mlx::core::metal {
 

mlx/backend/metal/rope.cpp

@@ -1,5 +1,5 @@
 // Copyright © 2023-2024 Apple Inc.
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/utils.h"
 #include "mlx/fast_primitives.h"
 

mlx/backend/metal/scaled_dot_product_attention.cpp

@@ -2,7 +2,7 @@
 #include <sstream>
 
 #include "mlx/backend/common/compiled.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 
 #include "mlx/backend/metal/kernels/steel/attn/params.h"
@@ -154,9 +154,9 @@ void sdpa_vector(
   int gqa_factor = q.shape(1) / k.shape(1);
   int N = k.shape(2);
   int B = q.shape(0) * q.shape(1);
-  size_t k_head_stride = k.strides()[1];
+  size_t k_head_stride = k.shape(1) == 1 ? k.strides(0) : k.strides(1);
   size_t k_seq_stride = k.strides()[2];
-  size_t v_head_stride = v.strides()[1];
+  size_t v_head_stride = v.shape(1) == 1 ? v.strides(0) : v.strides(1);
   size_t v_seq_stride = v.strides()[2];
 
   MTL::Size group_dims(1024, 1, 1);
@@ -199,11 +199,10 @@ void sdpa_vector(
   if (has_mask) {
     auto& m = *mask;
     compute_encoder.set_input_array(m, 11 + float_mask);
-    auto nd = m.ndim();
-    int32_t kv_seq_stride =
-        nd >= 1 && m.shape(-1) > 1 ? m.strides()[nd - 1] : 0;
-    int32_t q_seq_stride = nd >= 2 && m.shape(-2) > 1 ? m.strides()[nd - 2] : 0;
-    int32_t head_stride = nd >= 3 && m.shape(-3) > 1 ? m.strides()[nd - 3] : 0;
+    int32_t kv_seq_stride = m.shape(3) > 1 ? m.strides(3) : 0;
+    int32_t q_seq_stride = m.shape(2) > 1 ? m.strides(2) : 0;
+    int32_t head_stride =
+        m.shape(1) > 1 ? m.strides(1) : (m.shape(0) > 1 ? m.strides(0) : 0);
     compute_encoder.set_bytes(kv_seq_stride, 13);
     compute_encoder.set_bytes(q_seq_stride, 14);
     compute_encoder.set_bytes(head_stride, 15);
@@ -238,9 +237,10 @@ void sdpa_vector_2pass(
   int N = k.shape(2);
   int blocks = 32;
   int B = q.shape(0) * q.shape(1);
-  size_t k_head_stride = k.strides()[1];
+
+  size_t k_head_stride = k.shape(1) == 1 ? k.strides(0) : k.strides(1);
   size_t k_seq_stride = k.strides()[2];
-  size_t v_head_stride = v.strides()[1];
+  size_t v_head_stride = v.shape(1) == 1 ? v.strides(0) : v.strides(1);
   size_t v_seq_stride = v.strides()[2];
   MTL::Size group_dims(8 * 32, 1, 1);
   MTL::Size grid_dims(B, q.shape(2), blocks);
@@ -302,11 +302,10 @@ void sdpa_vector_2pass(
   if (has_mask) {
     auto& m = *mask;
     compute_encoder.set_input_array(m, 13 + float_mask);
-    auto nd = m.ndim();
-    int32_t kv_seq_stride =
-        nd >= 1 && m.shape(-1) > 1 ? m.strides()[nd - 1] : 0;
-    int32_t q_seq_stride = nd >= 2 && m.shape(-2) > 1 ? m.strides()[nd - 2] : 0;
-    int32_t head_stride = nd >= 3 && m.shape(-3) > 1 ? m.strides()[nd - 3] : 0;
+    int32_t kv_seq_stride = m.shape(3) > 1 ? m.strides(3) : 0;
+    int32_t q_seq_stride = m.shape(2) > 1 ? m.strides(2) : 0;
+    int32_t head_stride =
+        m.shape(1) > 1 ? m.strides(1) : (m.shape(0) > 1 ? m.strides(0) : 0);
     compute_encoder.set_bytes(kv_seq_stride, 15);
     compute_encoder.set_bytes(q_seq_stride, 16);
     compute_encoder.set_bytes(head_stride, 17);
@@ -368,18 +367,6 @@ void ScaledDotProductAttention::eval_gpu(
     }
   };
 
-  // Checks if arr is row contiguous or the sequence and head dimension are
-  // transposed
-  auto is_contiguous_or_head_seq_transposed = [](const array& arr) {
-    if (arr.flags().row_contiguous) {
-      return true;
-    }
-    auto& strides = arr.strides();
-    auto& shape = arr.shape();
-    return (strides[3] == 1) && (strides[2] == shape[3] * shape[1]) &&
-        (strides[1] == shape[3]) && (strides[0] == strides[2] * shape[2]);
-  };
-
   // Checks that the headdim dimension has stride 1.
   auto is_matrix_contiguous = [](const array& arr) {
     return arr.strides(-1) == 1;
@@ -387,30 +374,58 @@ void ScaledDotProductAttention::eval_gpu(
 
   // We are in vector mode ie single query
   if (q_pre.shape(2) <= 8) {
-    const auto& q = copy_unless(is_contiguous_or_head_seq_transposed, q_pre);
-    const auto& k = copy_unless(is_matrix_contiguous, k_pre);
-    const auto& v = copy_unless(is_matrix_contiguous, v_pre);
+    auto q_copy_unless = [](const array& arr) {
+      if (arr.flags().row_contiguous) {
+        return true;
+      }
+      auto& strides = arr.strides();
+      auto& shape = arr.shape();
+      if (shape[0] == 1 || shape[1] == 1) {
+        // If either the batch or head dimension is a singleton, the other can
+        // be transposed with the sequence dimension
+        auto bidx = shape[0] == 1 ? 1 : 0;
+        return (strides[3] == 1) && (strides[2] == shape[3] * shape[bidx]) &&
+            (strides[bidx] == shape[3]);
+      }
+      return false;
+    };
+
+    auto kv_copy_unless = [](const array& arr) {
+      // keys and values should be copied if:
+      // - the last dimension is not contiguous
+      // - the batch and head dim are not contiguous
+      auto& strides = arr.strides();
+      auto& shape = arr.shape();
+      if (strides.back() != 1) {
+        return false;
+      }
+      if (shape[0] == 1 || shape[1] == 1) {
+        return true;
+      }
+      return (strides[0] == strides[1] * shape[1]);
+    };
+
+    const auto& q = copy_unless(q_copy_unless, q_pre);
+    const auto& k = copy_unless(kv_copy_unless, k_pre);
+    const auto& v = copy_unless(kv_copy_unless, v_pre);
 
     // Donate the query if possible
-    if (q.is_donatable() && (q.shape(2) == 1 || !q.flags().row_contiguous) &&
-        q.size() == o.size()) {
+    if (q.is_donatable() && q.flags().row_contiguous && q.size() == o.size()) {
       o.copy_shared_buffer(q);
     } else {
-      if (o.shape(2) == 1) {
-        o.set_data(allocator::malloc(o.nbytes()));
-      } else {
-        auto strides = o.strides();
-        strides[2] = o.shape(1) * o.shape(3);
-        strides[1] = o.shape(3);
-        auto flags = q.flags();
-        flags.row_contiguous = q.shape(1) == 1;
-        o.set_data(
-            allocator::malloc(o.nbytes()), o.size(), std::move(strides), flags);
-      }
+      o.set_data(allocator::malloc(o.nbytes()));
     }
 
-    auto mask =
-        inputs.size() > 3 ? std::optional<array>{inputs[3]} : std::nullopt;
+    auto mask_copy_unless = [&q](const array& arr) {
+      auto& strides = arr.strides();
+      auto& shape = arr.shape();
+      return arr.flags().row_contiguous || q.shape(0) == 1 || q.shape(1) == 1 ||
+          (strides[0] == strides[1] * shape[1]);
+    };
+
+    auto mask = inputs.size() > 3
+        ? std::optional<array>{copy_unless(mask_copy_unless, inputs[3])}
+        : std::nullopt;
 
     // We route to the 2 pass fused attention if
     // - The device is large and the sequence length long

mlx/backend/metal/scan.cpp

@@ -3,7 +3,7 @@
 #include <cassert>
 #include <sstream>
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/utils.h"

mlx/backend/metal/slicing.cpp

@@ -2,21 +2,12 @@
 
 #include <numeric>
 
-#include "mlx/backend/common/slicing.h"
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/backend/gpu/slicing.h"
 #include "mlx/backend/metal/device.h"
 
 namespace mlx::core {
 
-void slice_gpu(
-    const array& in,
-    array& out,
-    const Shape& start_indices,
-    const Shape& strides,
-    const Stream& s) {
-  slice(in, out, start_indices, strides);
-}
-
 void concatenate_gpu(
     const std::vector<array>& inputs,
     array& out,
@@ -48,30 +39,4 @@ void concatenate_gpu(
   }
 }
 
-void pad_gpu(
-    const array& in,
-    const array& val,
-    array& out,
-    const std::vector<int>& axes,
-    const Shape& low_pad_size,
-    const Stream& s) {
-  // Fill output with val
-  fill_gpu(val, out, s);
-
-  // Find offset for start of input values
-  size_t data_offset = 0;
-  for (int i = 0; i < axes.size(); i++) {
-    auto ax = axes[i] < 0 ? out.ndim() + axes[i] : axes[i];
-    data_offset += out.strides()[ax] * low_pad_size[i];
-  }
-
-  // Extract slice from output where input will be pasted
-  array out_slice(in.shape(), out.dtype(), nullptr, {});
-  out_slice.copy_shared_buffer(
-      out, out.strides(), out.flags(), out_slice.size(), data_offset);
-
-  // Copy input values into the slice
-  copy_gpu_inplace(in, out_slice, CopyType::GeneralGeneral, s);
-}
-
 } // namespace mlx::core

mlx/backend/metal/softmax.cpp

@@ -1,7 +1,7 @@
 // Copyright © 2023-2024 Apple Inc.
 #include <algorithm>
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/kernels/defines.h"

mlx/backend/metal/sort.cpp

@@ -2,7 +2,7 @@
 
 #include <algorithm>
 
-#include "mlx/backend/metal/copy.h"
+#include "mlx/backend/gpu/copy.h"
 #include "mlx/backend/metal/device.h"
 #include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/utils.h"

mlx/backend/metal/ternary.cpp

@@ -45,7 +45,7 @@ void ternary_op_gpu_inplace(
     work_per_thread = large ? 4 : 2;
   } else {
     large = out.data_size() > INT32_MAX;
-    work_per_thread = 1;
+    work_per_thread = get_work_per_thread(b.dtype());
   }
   std::string kernel_name;
   if (topt == TernaryOpType::General) {
@@ -106,13 +106,19 @@ void ternary_op_gpu_inplace(
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   } else {
     // Launch a 1D or 2D grid of threads
-    size_t nthreads = out.data_size();
+    size_t nthreads = ceildiv(out.data_size(), work_per_thread);
     if (thread_group_size > nthreads) {
       thread_group_size = nthreads;
     }
     MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
-    MTL::Size grid_dims = large ? get_2d_grid_dims(out.shape(), out.strides())
-                                : MTL::Size(nthreads, 1, 1);
+    MTL::Size grid_dims;
+    if (large) {
+      compute_encoder.set_bytes<int64_t>(out.data_size(), 4);
+      grid_dims = get_2d_grid_dims(out.shape(), out.strides(), work_per_thread);
+    } else {
+      compute_encoder.set_bytes<int>(out.data_size(), 4);
+      grid_dims = MTL::Size(nthreads, 1, 1);
+    }
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   }
 }

mlx/backend/metal/unary.cpp

@@ -34,18 +34,19 @@ void unary_op_gpu_inplace(
   };
   auto [shape, strides] = maybe_collapse();
   int ndim = shape.size();
-  size_t nthreads = contig ? in.data_size() : in.size();
   bool large;
   if (!contig) {
     large = in.data_size() > INT32_MAX || out.size() > INT32_MAX;
   } else {
     large = in.data_size() > UINT32_MAX;
   }
-  int work_per_thread = !contig && large ? 4 : 1;
+  int work_per_thread;
   std::string kernel_name;
   if (contig) {
+    work_per_thread = get_work_per_thread(in.dtype());
     kernel_name = (large ? "v2" : "v");
   } else {
+    work_per_thread = large ? 4 : 1;
     kernel_name = "gn" + std::to_string(work_per_thread);
     if (large) {
       kernel_name += "large";
@@ -75,12 +76,20 @@ void unary_op_gpu_inplace(
     MTL::Size grid_dims = MTL::Size(dim0, dim1, rest);
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   } else {
+    size_t nthreads = ceildiv(in.data_size(), work_per_thread);
     if (thread_group_size > nthreads) {
       thread_group_size = nthreads;
     }
+
     MTL::Size group_dims = MTL::Size(thread_group_size, 1, 1);
-    MTL::Size grid_dims = large ? get_2d_grid_dims(out.shape(), out.strides())
-                                : MTL::Size(nthreads, 1, 1);
+    MTL::Size grid_dims;
+    if (large) {
+      compute_encoder.set_bytes<int64_t>(in.data_size(), 2);
+      grid_dims = get_2d_grid_dims(out.shape(), out.strides(), work_per_thread);
+    } else {
+      compute_encoder.set_bytes<int>(in.data_size(), 2);
+      grid_dims = MTL::Size(nthreads, 1, 1);
+    }
     compute_encoder.dispatch_threads(grid_dims, group_dims);
   }
 }

mlx/backend/metal/utils.h

@@ -84,4 +84,12 @@ void concatenate(std::string& acc, T first, Args... args) {
   concatenate(acc, args...);
 }
 
+inline int get_work_per_thread(Dtype dtype) {
+  return std::max(1, 8 / dtype.size());
+}
+
+inline size_t ceildiv(size_t n, size_t m) {
+  return (n + m - 1) / m;
+}
+
 } // namespace mlx::core

mlx/backend/no_cpu/CMakeLists.txt

@@ -1,6 +1,7 @@
 target_sources(
   mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/available.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/../cpu/eval.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/../cpu/encoder.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp)