revert sort + flaky test

mlx/backend/metal/sort.cpp

@@ -21,8 +21,6 @@ void single_block_sort(
     int bn,
     int tn,
     bool argsort) {
-  out.set_data(allocator::malloc(out.nbytes()));
-
   // Prepare shapes
   int n_rows = in.size() / in.shape(axis);
 
@@ -158,6 +156,9 @@ void multi_block_sort(
   dev_idxs_1.set_data(allocator::malloc(dev_idxs_1.nbytes()));
   block_partitions.set_data(allocator::malloc(block_partitions.nbytes()));
 
+  std::vector<array> copies = {
+      dev_vals_0, dev_vals_1, dev_idxs_0, dev_idxs_1, block_partitions};
+
   // Prepare command encoder
   auto& compute_encoder = d.get_command_encoder(s.index);
 
@@ -249,17 +250,25 @@ void multi_block_sort(
       compute_encoder.dispatch_threadgroups(grid_dims, group_dims);
     }
   }
-  out.copy_shared_buffer(
+
-      argsort ? dev_idxs_out : dev_vals_out,
+  // Copy outputs with appropriate strides
+  auto strides = out.strides();
+  for (int ax = axis + 1; ax < strides.size(); ax++) {
+    strides[ax] *= out.shape(axis);
+  }
+  strides[axis] = 1;
+  copy_gpu_inplace(
+      (argsort) ? dev_idxs_out : dev_vals_out,
+      out,
+      out.shape(),
+      strides,
       out.strides(),
-      out.flags(),
+      0,
-      out.data_size());
+      0,
-  d.add_temporaries(
+      (axis == in.ndim() - 1) ? CopyType::Vector : CopyType::General,
-      {dev_vals_in,
+      s);
-       dev_idxs_in,
+
-       argsort ? dev_vals_in : dev_idxs_in,
+  d.add_temporaries(std::move(copies), s.index);
-       block_partitions},
-      s.index);
 }
 
 void gpu_merge_sort(
@@ -309,6 +318,8 @@ void gpu_merge_sort(
 void ArgSort::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
 
+  out.set_data(allocator::malloc(out.nbytes()));
+
   auto& s = stream();
   auto& d = metal::device(s.device);
   auto& in = inputs[0];
@@ -319,6 +330,8 @@ void ArgSort::eval_gpu(const std::vector<array>& inputs, array& out) {
 void Sort::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
 
+  out.set_data(allocator::malloc(out.nbytes()));
+
   auto& s = stream();
   auto& d = metal::device(s.device);
   auto& in = inputs[0];
@@ -330,6 +343,8 @@ void ArgPartition::eval_gpu(const std::vector<array>& inputs, array& out) {
   // We direct arg partition to sort for now
   assert(inputs.size() == 1);
 
+  out.set_data(allocator::malloc(out.nbytes()));
+
   auto& s = stream();
   auto& d = metal::device(s.device);
   auto& in = inputs[0];
@@ -341,6 +356,8 @@ void Partition::eval_gpu(const std::vector<array>& inputs, array& out) {
   // We direct partition to sort for now
   assert(inputs.size() == 1);
 
+  out.set_data(allocator::malloc(out.nbytes()));
+
   auto& s = stream();
   auto& d = metal::device(s.device);
   auto& in = inputs[0];

python/tests/test_vmap.py

@@ -634,6 +634,7 @@ class TestVmap(mlx_tests.MLXTestCase):
         self.assertEqual(fy.shape, (4, 5, 6, 7))
 
     def test_leaks(self):
+        mx.synchronize()
         if mx.metal.is_available():
             mem_pre = mx.get_active_memory()
         else: