Merge b3c1aaafd2 into b8022c578a

benchmarks/python/comparative/bench_mlx.py

@@ -224,6 +224,13 @@ def relu6(x):
     mx.eval(y)
 
 
+def relu_squared(x):
+    y = x
+    for i in range(100):
+        y = nn.relu_squared(y)
+    mx.eval(y)
+
+
 def softplus(x):
     y = x
     for i in range(100):
@@ -458,6 +465,9 @@ if __name__ == "__main__":
     elif args.benchmark == "relu6":
         print(bench(relu6, x))
 
+    elif args.benchmark == "relu_squared":
+        print(bench(relu_squared, x))
+
     elif args.benchmark == "celu":
         print(bench(celu, x))
 

benchmarks/python/comparative/bench_torch.py

@@ -157,6 +157,15 @@ def relu6(x):
     sync_if_needed(x)
 
 
+@torch.no_grad()
+def relu_squared(x):
+    y = x
+    for i in range(100):
+        y = torch.nn.functional.relu(y)
+        y = torch.square(y)
+    sync_if_needed(x)
+
+
 @torch.no_grad()
 def softplus(x):
     y = x
@@ -407,6 +416,9 @@ if __name__ == "__main__":
     elif args.benchmark == "relu6":
         print(bench(relu6, x))
 
+    elif args.benchmark == "relu_squared":
+        print(bench(relu_squared, x))
+
     elif args.benchmark == "softplus":
         print(bench(softplus, x))
 

benchmarks/python/comparative/compare.py

@@ -207,6 +207,8 @@ if __name__ == "__main__":
     compare_filtered("elu --size 32x16x1024 --cpu")
     compare_filtered("relu6 --size 32x16x1024")
     compare_filtered("relu6 --size 32x16x1024 --cpu")
+    compare_filtered("relu_squared --size 32x16x1024")
+    compare_filtered("relu_squared --size 32x16x1024 --cpu")
     compare_filtered("softplus --size 32x16x1024")
     compare_filtered("softplus --size 32x16x1024 --cpu")
     compare_filtered("celu --size 32x16x1024")

docs/src/python/nn/functions.rst

@@ -28,6 +28,7 @@ simple functions.
    prelu
    relu
    relu6
+   relu_squared
    selu
    sigmoid
    silu

docs/src/python/nn/layers.rst

@@ -51,6 +51,7 @@ Layers
    RMSNorm
    ReLU
    ReLU6
+   ReLUSquared
    RNN
    RoPE
    SELU

mlx/backend/cuda/CMakeLists.txt

@@ -8,6 +8,7 @@ target_sources(
   PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/binary.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/binary_two.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/compiled.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/copy.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/copy/copy_contiguous.cu

mlx/backend/cuda/binary.cu

@@ -125,13 +125,12 @@ constexpr bool supports_binary_op() {
 template <typename Op>
 void binary_op_gpu_inplace(
     const std::vector<array>& inputs,
-    std::vector<array>& outputs,
+    array& out,
     std::string_view op,
     const Stream& s) {
   assert(inputs.size() > 1);
   const auto& a = inputs[0];
   const auto& b = inputs[1];
-  auto& out = outputs[0];
   if (out.size() == 0) {
     return;
   }
@@ -146,7 +145,6 @@ void binary_op_gpu_inplace(
         if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
           using InType = cuda_type_t<CTYPE_IN>;
           using OutType = cuda_type_t<CTYPE_OUT>;
-
           auto bopt = get_binary_op_type(a, b);
           if (bopt == BinaryOpType::General) {
             auto [shape, strides] = collapse_contiguous_dims(a, b, out);
@@ -219,20 +217,6 @@ void binary_op_gpu_inplace(
   });
 }
 
-template <typename Op>
-void binary_op_gpu(
-    const std::vector<array>& inputs,
-    std::vector<array>& outputs,
-    std::string_view op,
-    const Stream& s) {
-  auto& a = inputs[0];
-  auto& b = inputs[1];
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, outputs[0], bopt);
-  set_binary_op_output_data(a, b, outputs[1], bopt);
-  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
-}
-
 template <typename Op>
 void binary_op_gpu(
     const std::vector<array>& inputs,
@@ -243,8 +227,7 @@ void binary_op_gpu(
   auto& b = inputs[1];
   auto bopt = get_binary_op_type(a, b);
   set_binary_op_output_data(a, b, out, bopt);
-  std::vector<array> outputs{out};
-  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
+  binary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 
 #define BINARY_GPU(func)                                                 \
@@ -254,14 +237,6 @@ void binary_op_gpu(
     binary_op_gpu<cu::func>(inputs, out, get_primitive_string(this), s); \
   }
 
-#define BINARY_GPU_MULTI(func)                                               \
-  void func::eval_gpu(                                                       \
-      const std::vector<array>& inputs, std::vector<array>& outputs) {       \
-    nvtx3::scoped_range r(#func "::eval_gpu");                               \
-    auto& s = outputs[0].primitive().stream();                               \
-    binary_op_gpu<cu::func>(inputs, outputs, get_primitive_string(this), s); \
-  }
-
 BINARY_GPU(Add)
 BINARY_GPU(ArcTan2)
 BINARY_GPU(Divide)

mlx/backend/cuda/binary_two.cu

@@ -0,0 +1,248 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/common/binary.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/device/binary_ops.cuh"
+#include "mlx/backend/cuda/device/cucomplex_math.cuh"
+#include "mlx/backend/cuda/kernel_utils.cuh"
+#include "mlx/dtype_utils.h"
+#include "mlx/primitives.h"
+
+#include <cooperative_groups.h>
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace cu {
+
+namespace cg = cooperative_groups;
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void
+binary_ss(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto out = Op{}(a[0], b[0]);
+    out_a[0] = out[0];
+    out_b[0] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void
+binary_sv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto out = Op{}(a[0], b[index]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void
+binary_vs(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto out = Op{}(a[index], b[0]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void
+binary_vv(const In* a, const In* b, Out* out_a, Out* out_b, IdxT size) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto out = Op{}(a[index], b[index]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT, int NDIM>
+__global__ void binary_g_nd(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size,
+    const __grid_constant__ cuda::std::array<int32_t, NDIM> shape,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> a_strides,
+    const __grid_constant__ cuda::std::array<int64_t, NDIM> b_strides) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [a_idx, b_idx] = elem_to_loc_nd<NDIM>(
+        index, shape.data(), a_strides.data(), b_strides.data());
+    auto out = Op{}(a[a_idx], b[b_idx]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out, typename IdxT>
+__global__ void binary_g(
+    const In* a,
+    const In* b,
+    Out* out_a,
+    Out* out_b,
+    IdxT size,
+    const __grid_constant__ Shape shape,
+    const __grid_constant__ Strides a_strides,
+    const __grid_constant__ Strides b_strides,
+    int ndim) {
+  IdxT index = cg::this_grid().thread_rank();
+  if (index < size) {
+    auto [a_idx, b_idx] = elem_to_loc_4d(
+        index, shape.data(), a_strides.data(), b_strides.data(), ndim);
+    auto out = Op{}(a[a_idx], b[b_idx]);
+    out_a[index] = out[0];
+    out_b[index] = out[1];
+  }
+}
+
+template <typename Op, typename In, typename Out>
+constexpr bool supports_binary_op() {
+  if (std::is_same_v<Op, DivMod>) {
+    return std::is_same_v<In, Out> &&
+        (std::is_integral_v<Out> || is_floating_v<Out>);
+  }
+  return false;
+}
+
+} // namespace cu
+
+template <typename Op>
+void binary_op_gpu_inplace(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    std::string_view op,
+    const Stream& s) {
+  assert(inputs.size() > 1);
+  const auto& a = inputs[0];
+  const auto& b = inputs[1];
+  auto& out_a = outputs[0];
+  auto& out_b = outputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out_a, bopt);
+  set_binary_op_output_data(a, b, out_b, bopt);
+
+  if (out_a.size() == 0) {
+    return;
+  }
+
+  auto& encoder = cu::get_command_encoder(s);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out_a);
+  encoder.set_output_array(out_b);
+  encoder.launch_kernel([&](cudaStream_t stream) {
+    MLX_SWITCH_ALL_TYPES(a.dtype(), CTYPE_IN, {
+      MLX_SWITCH_ALL_TYPES(out_a.dtype(), CTYPE_OUT, {
+        if constexpr (cu::supports_binary_op<Op, CTYPE_IN, CTYPE_OUT>()) {
+          using InType = cuda_type_t<CTYPE_IN>;
+          using OutType = cuda_type_t<CTYPE_OUT>;
+
+          auto bopt = get_binary_op_type(a, b);
+          if (bopt == BinaryOpType::General) {
+            auto [shape, strides] = collapse_contiguous_dims(a, b, out_a);
+            auto& a_strides = strides[0];
+            auto& b_strides = strides[1];
+            bool large = a.data_size() > INT32_MAX ||
+                b.data_size() > INT32_MAX || out_a.data_size() > INT32_MAX;
+            MLX_SWITCH_BOOL(large, LARGE, {
+              using IdxT = std::conditional_t<LARGE, int64_t, int32_t>;
+              int ndim = shape.size();
+              if (ndim <= 3) {
+                MLX_SWITCH_1_2_3(ndim, NDIM, {
+                  auto kernel =
+                      &cu::binary_g_nd<Op, InType, OutType, IdxT, NDIM>;
+                  auto [num_blocks, block_dims] =
+                      get_launch_args(kernel, out_a, large);
+                  kernel<<<num_blocks, block_dims, 0, stream>>>(
+                      a.data<InType>(),
+                      b.data<InType>(),
+                      out_a.data<OutType>(),
+                      out_b.data<OutType>(),
+                      out_a.size(),
+                      const_param<NDIM>(shape),
+                      const_param<NDIM>(a_strides),
+                      const_param<NDIM>(b_strides));
+                });
+              } else {
+                auto kernel = cu::binary_g<Op, InType, OutType, IdxT>;
+                auto [num_blocks, block_dims] =
+                    get_launch_args(kernel, out_a, large);
+                kernel<<<num_blocks, block_dims, 0, stream>>>(
+                    a.data<InType>(),
+                    b.data<InType>(),
+                    out_a.data<OutType>(),
+                    out_b.data<OutType>(),
+                    out_a.size(),
+                    const_param(shape),
+                    const_param(a_strides),
+                    const_param(b_strides),
+                    ndim);
+              }
+            });
+          } else {
+            MLX_SWITCH_BOOL(out_a.data_size() > UINT32_MAX, LARGE, {
+              using IdxT = std::conditional_t<LARGE, int64_t, uint32_t>;
+              auto kernel = cu::binary_ss<Op, InType, OutType, IdxT>;
+              if (bopt == BinaryOpType::ScalarVector) {
+                kernel = cu::binary_sv<Op, InType, OutType, IdxT>;
+              } else if (bopt == BinaryOpType::VectorScalar) {
+                kernel = cu::binary_vs<Op, InType, OutType, IdxT>;
+              } else if (bopt == BinaryOpType::VectorVector) {
+                kernel = cu::binary_vv<Op, InType, OutType, IdxT>;
+              }
+              auto [num_blocks, block_dims] = get_launch_args(
+                  kernel,
+                  out_a.data_size(),
+                  out_a.shape(),
+                  out_a.strides(),
+                  LARGE);
+              kernel<<<num_blocks, block_dims, 0, stream>>>(
+                  a.data<InType>(),
+                  b.data<InType>(),
+                  out_a.data<OutType>(),
+                  out_b.data<OutType>(),
+                  out_a.data_size());
+            });
+          }
+        } else {
+          throw std::runtime_error(fmt::format(
+              "Can not do binary op {} on inputs of {} with result of {}.",
+              op,
+              dtype_to_string(a.dtype()),
+              dtype_to_string(out_a.dtype())));
+        }
+      });
+    });
+  });
+}
+
+template <typename Op>
+void binary_op_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs,
+    std::string_view op,
+    const Stream& s) {
+  auto& a = inputs[0];
+  auto& b = inputs[1];
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, outputs[0], bopt);
+  set_binary_op_output_data(a, b, outputs[1], bopt);
+  binary_op_gpu_inplace<Op>(inputs, outputs, op, s);
+}
+
+void DivMod::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("DivMod::eval_gpu");
+  auto& s = outputs[0].primitive().stream();
+  binary_op_gpu<cu::DivMod>(inputs, outputs, get_primitive_string(this), s);
+}
+
+} // namespace mlx::core

mlx/backend/cuda/device/binary_ops.cuh

@@ -22,7 +22,7 @@ struct FloorDivide {
     if constexpr (cuda::std::is_integral_v<T>) {
       return x / y;
     } else {
-      return trunc(x / y);
+      return truncf(x / y);
     }
   }
 };
@@ -132,7 +132,7 @@ struct LogAddExp {
           cuda::std::numeric_limits<float>::quiet_NaN(),
           cuda::std::numeric_limits<float>::quiet_NaN()};
     }
-    constexpr float inf = cuda::std::numeric_limits<float>::infinity();
+    float inf = cuda::std::numeric_limits<float>::infinity();
     auto maxval = x > y ? x : y;
     auto minval = x < y ? x : y;
     if (cuCrealf(minval) == -inf || cuCrealf(maxval) == inf)

mlx/backend/cuda/device/config.h

@@ -5,7 +5,7 @@
 #pragma once
 
 // The maximum dimensions of shape/strides passed as kernel parameters.
-#define MAX_NDIM 8
+#define MAX_NDIM 10
 
 // All existing NVIDIA hardware has a fixed 32 warp size. Though a built-in
 // warpSize variable exists, using it would prevent compile-time optimizations.

mlx/backend/cuda/primitives.cu

@@ -71,10 +71,8 @@ bool fast::ScaledDotProductAttention::use_fallback(
     throw std::runtime_error(#func " has no CUDA implementation.");   \
   }
 
-NO_GPU(ArgPartition)
 NO_GPU(BlockMaskedMM)
 NO_GPU(Convolution)
-NO_GPU_MULTI(DivMod)
 NO_GPU(DynamicSlice)
 NO_GPU(DynamicSliceUpdate)
 NO_GPU(FFT)
@@ -83,7 +81,6 @@ NO_GPU(GatherQMM)
 NO_GPU(Hadamard)
 NO_GPU(Load)
 NO_GPU_MULTI(LUF)
-NO_GPU(Partition)
 NO_GPU_MULTI(QRF)
 NO_GPU(QuantizedMatmul)
 NO_GPU(Scan)

mlx/backend/cuda/sort.cu

@@ -86,7 +86,6 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
     axis += in.ndim();
   }
   int nsort = in.shape(axis);
-  int nsegments = in.data_size() / nsort;
   int last_dim = in.ndim() - 1;
 
   // If we are not sorting the innermost dimension of a contiguous array,
@@ -100,7 +99,11 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
     out = array(allocator::malloc(out.nbytes()), in.shape(), out.dtype());
     encoder.add_temporary(out);
   } else {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(
+        allocator::malloc(in.data_size() * out.itemsize()),
+        in.data_size(),
+        in.strides(),
+        in.flags());
   }
 
   encoder.launch_kernel([&](cudaStream_t stream) {
@@ -134,7 +137,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
               indices.data<uint32_t>(),
               out.data<uint32_t>(),
               in.data_size(),
-              nsegments,
+              in.data_size() / nsort,
               offsets,
               offsets + 1,
               stream);
@@ -144,7 +147,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
               in.data<Type>(),
               out.data<Type>(),
               in.data_size(),
-              nsegments,
+              in.data_size() / nsort,
               offsets,
               offsets + 1,
               stream);
@@ -177,4 +180,14 @@ void Sort::eval_gpu(const std::vector<array>& inputs, array& out) {
   gpu_sort(stream(), inputs[0], out, axis_, false);
 }
 
+void ArgPartition::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("ArgPartition::eval_gpu");
+  gpu_sort(stream(), inputs[0], out, axis_, true);
+}
+
+void Partition::eval_gpu(const std::vector<array>& inputs, array& out) {
+  nvtx3::scoped_range r("Partition::eval_gpu");
+  gpu_sort(stream(), inputs[0], out, axis_, false);
+}
+
 } // namespace mlx::core

python/mlx/nn/layers/__init__.py

@@ -16,6 +16,7 @@ from mlx.nn.layers.activations import (
     PReLU,
     ReLU,
     ReLU6,
+    ReLUSquared,
     Sigmoid,
     SiLU,
     Softmax,
@@ -41,6 +42,7 @@ from mlx.nn.layers.activations import (
     prelu,
     relu,
     relu6,
+    relu_squared,
     selu,
     sigmoid,
     silu,

python/mlx/nn/layers/activations.py

@@ -71,6 +71,17 @@ def relu6(x):
     return mx.minimum(mx.maximum(x, 0), 6.0)
 
 
+@partial(mx.compile, shapeless=True)
+def relu_squared(x):
+    r"""Applies the Rectified Linear Unit squared.
+
+    Applies :math:`\max(x, 0)^2` element wise.
+
+    Reference: https://arxiv.org/abs/2109.08668v2
+    """
+    return relu(x).square()
+
+
 @partial(mx.compile, shapeless=True)
 def softmax(x, axis=-1):
     r"""Applies the Softmax function.
@@ -420,6 +431,18 @@ class ReLU6(Module):
     """
 
 
+@_make_activation_module(relu_squared)
+class ReLUSquared(Module):
+    r"""Applies the Rectified Linear Unit squared.
+
+    Applies :math:`\max(x, 0)^2` element wise.
+
+    Reference: https://arxiv.org/abs/2109.08668v2
+
+    See :func:`relu_squared` for the functional equivalent.
+    """
+
+
 @_make_activation_module(softmax)
 class Softmax(Module):
     r"""Applies the Softmax function.

python/tests/cuda_skip.py

@@ -1,10 +1,8 @@
 cuda_skip = {
     "TestArray.test_api",
-    "TestAutograd.test_update_state",
     "TestBF16.test_arg_reduction_ops",
     "TestBF16.test_reduction_ops",
     "TestBlas.test_complex_gemm",
-    "TestCompile.test_compile_dynamic_dims",
     "TestEinsum.test_ellipses",
     "TestEinsum.test_opt_einsum_test_cases",
     "TestLoad.test_load_f8_e4m3",
@@ -14,24 +12,14 @@ cuda_skip = {
     "TestLayers.test_quantized_embedding",
     "TestLayers.test_sin_pe",
     "TestLayers.test_upsample",
-    "TestOps.test_array_equal",
     "TestOps.test_complex_ops",
     "TestOps.test_dynamic_slicing",
     "TestOps.test_softmax",
-    "TestOps.test_sort",
-    "TestOps.test_tile",
     "TestReduce.test_axis_permutation_sums",
     "TestReduce.test_dtypes",
     "TestReduce.test_expand_sums",
     "TestReduce.test_many_reduction_axes",
     "TestUpsample.test_torch_upsample",
-    # DivMod NYI
-    "TestOps.test_divmod",
-    "TestEval.test_multi_output_eval_during_transform",
-    # Partition NYI
-    "TestAutograd.test_topk_grad",
-    "TestOps.test_argpartition",
-    "TestOps.test_partition",
     # Block masked matmul NYI
     "TestBlas.test_block_masked_matmul",
     # Gather matmul NYI

python/tests/test_nn.py

@@ -855,6 +855,13 @@ class TestLayers(mlx_tests.MLXTestCase):
         self.assertEqual(y.shape, (3,))
         self.assertEqual(y.dtype, mx.float32)
 
+    def test_relu_squared(self):
+        x = mx.array([-1.0, 0.0, 1.0, 2.0, 3.0])
+        y = nn.relu_squared(x)
+        self.assertTrue(mx.array_equal(y, mx.array([0.0, 0.0, 1.0, 4.0, 9.0])))
+        self.assertEqual(y.shape, (5,))
+        self.assertEqual(y.dtype, mx.float32)
+
     def test_leaky_relu(self):
         x = mx.array([1.0, -1.0, 0.0])
         y = nn.leaky_relu(x)