Cycle leak break (#1856)

* detect and break leaks in custom function

* detect and break leaks in custom function

python/src/CMakeLists.txt

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

python/src/export.cpp

@@ -60,8 +60,56 @@ validate_and_extract_inputs(
   return {args_, kwargs_};
 }
 
-auto wrap_export_function(const nb::callable& fun) {
+int py_function_exporter_tp_traverse(
-  return [fun](
+    PyObject* self,
+    visitproc visit,
+    void* arg);
+
+class PyFunctionExporter {
+ public:
+  PyFunctionExporter(mx::FunctionExporter exporter, nb::handle dep)
+      : exporter_(std::move(exporter)), dep_(dep) {}
+  ~PyFunctionExporter() {
+    nb::gil_scoped_acquire gil;
+  }
+  PyFunctionExporter(const PyFunctionExporter&) = delete;
+  PyFunctionExporter& operator=(const PyFunctionExporter&) = delete;
+  PyFunctionExporter& operator=(const PyFunctionExporter&&) = delete;
+  PyFunctionExporter(PyFunctionExporter&& other)
+      : exporter_(std::move(other.exporter_)), dep_(std::move(other.dep_)) {}
+
+  void close() {
+    exporter_.close();
+  }
+  void operator()(
+      const std::vector<mx::array>& args,
+      const std::map<std::string, mx::array>& kwargs) {
+    exporter_(args, kwargs);
+  }
+
+  friend int py_function_exporter_tp_traverse(PyObject*, visitproc, void*);
+
+ private:
+  mx::FunctionExporter exporter_;
+  nb::handle dep_;
+};
+
+int py_function_exporter_tp_traverse(
+    PyObject* self,
+    visitproc visit,
+    void* arg) {
+  auto* p = nb::inst_ptr<PyFunctionExporter>(self);
+  Py_VISIT(p->dep_.ptr());
+  Py_VISIT(Py_TYPE(self));
+  return 0;
+}
+
+PyType_Slot py_function_exporter_slots[] = {
+    {Py_tp_traverse, (void*)py_function_exporter_tp_traverse},
+    {0, 0}};
+
+auto wrap_export_function(nb::callable fun) {
+  return [fun = std::move(fun)](
              const std::vector<mx::array>& args_,
              const std::map<std::string, mx::array>& kwargs_) {
     auto kwargs = nb::dict();
@@ -173,21 +221,21 @@ void init_export(nb::module_& m) {
           >>> out = fn((a, b), {"x": x, "y": y}[0]
       )pbdoc");
 
-  nb::class_<mx::FunctionExporter>(
+  nb::class_<PyFunctionExporter>(
       m,
       "FunctionExporter",
+      nb::type_slots(py_function_exporter_slots),
       R"pbdoc(
        A context managing class for exporting multiple traces of the same
        function to a file.
 
        Make an instance of this class by calling fun:`mx.exporter`.
       )pbdoc")
-      .def("close", &mx::FunctionExporter::close)
+      .def("close", &PyFunctionExporter::close)
-      .def(
+      .def("__enter__", [](PyFunctionExporter& exporter) { return &exporter; })
-          "__enter__", [](mx::FunctionExporter& exporter) { return &exporter; })
       .def(
           "__exit__",
-          [](mx::FunctionExporter& exporter,
+          [](PyFunctionExporter& exporter,
              const std::optional<nb::object>&,
              const std::optional<nb::object>&,
              const std::optional<nb::object>&) { exporter.close(); },
@@ -196,7 +244,7 @@ void init_export(nb::module_& m) {
           "traceback"_a = nb::none())
       .def(
           "__call__",
-          [](mx::FunctionExporter& exporter,
+          [](PyFunctionExporter& exporter,
              const nb::args& args,
              const nb::kwargs& kwargs) {
             auto [args_, kwargs_] =
@@ -206,8 +254,9 @@ void init_export(nb::module_& m) {
 
   m.def(
       "exporter",
-      [](const std::string& file, const nb::callable& fun, bool shapeless) {
+      [](const std::string& file, nb::callable fun, bool shapeless) {
-        return mx::exporter(file, wrap_export_function(fun), shapeless);
+        return PyFunctionExporter{
+            mx::exporter(file, wrap_export_function(fun), shapeless), fun};
       },
       "file"_a,
       "fun"_a,

python/src/mlx.cpp

@@ -7,6 +7,7 @@
 
 namespace nb = nanobind;
 
+void init_mlx_func(nb::module_&);
 void init_array(nb::module_&);
 void init_device(nb::module_&);
 void init_stream(nb::module_&);
@@ -28,6 +29,7 @@ NB_MODULE(core, m) {
   nb::module_::import_("mlx._os_warning");
   nb::set_leak_warnings(false);
 
+  init_mlx_func(m);
   init_device(m);
   init_stream(m);
   init_array(m);

python/src/mlx_func.cpp

@@ -0,0 +1,108 @@
+// Copyright © 2025 Apple Inc.
+
+#include "python/src/mlx_func.h"
+
+// A garbage collected function which wraps nb::cpp_function
+// See https://github.com/wjakob/nanobind/discussions/919
+
+struct gc_func {
+  PyObject_HEAD
+      // Vector call implementation that forwards calls to nanobind
+      PyObject* (*vectorcall)(PyObject*, PyObject* const*, size_t, PyObject*);
+  // The function itself
+  PyObject* func;
+  // A non-owning reference to dependencies owned by 'func'
+  std::vector<PyObject*> deps;
+};
+
+int gc_func_tp_traverse(PyObject* self, visitproc visit, void* arg) {
+  gc_func* w = (gc_func*)self;
+  Py_VISIT(w->func);
+  for (auto d : w->deps) {
+    Py_VISIT(d);
+  }
+  Py_VISIT(Py_TYPE(self));
+  return 0;
+};
+
+int gc_func_tp_clear(PyObject* self) {
+  gc_func* w = (gc_func*)self;
+  Py_CLEAR(w->func);
+  return 0;
+}
+
+PyObject* gc_func_get_doc(PyObject* self, void*) {
+  return PyObject_GetAttrString(((gc_func*)self)->func, "__doc__");
+}
+
+PyObject* gc_func_get_sig(PyObject* self, void*) {
+  return PyObject_GetAttrString(((gc_func*)self)->func, "__nb_signature__");
+}
+
+PyObject* gc_func_vectorcall(
+    PyObject* self,
+    PyObject* const* args,
+    size_t nargs,
+    PyObject* kwnames) {
+  return PyObject_Vectorcall(((gc_func*)self)->func, args, nargs, kwnames);
+}
+
+void gc_func_dealloc(PyObject* self) {
+  PyObject_GC_UnTrack(self);
+  Py_XDECREF(((gc_func*)self)->func);
+  PyObject_GC_Del(self);
+}
+
+static PyMemberDef gc_func_members[] = {
+    {"__vectorcalloffset__",
+     T_PYSSIZET,
+     (Py_ssize_t)offsetof(gc_func, vectorcall),
+     READONLY,
+     nullptr},
+    {nullptr, 0, 0, 0, nullptr}};
+
+static PyGetSetDef gc_func_getset[] = {
+    {"__doc__", gc_func_get_doc, nullptr, nullptr, nullptr},
+    {"__nb_signature__", gc_func_get_sig, nullptr, nullptr, nullptr},
+    {nullptr, nullptr, nullptr, nullptr, nullptr}};
+
+static PyObject* gc_func_getattro(PyObject* self, PyObject* name_) {
+  gc_func* w = (gc_func*)self;
+  auto f = PyCFunction(PyType_GetSlot(Py_TYPE(w->func), Py_tp_getattro));
+  return f(w->func, name_);
+}
+
+// Table of custom type slots we want to install
+PyType_Slot gc_func_slots[] = {
+    {Py_tp_traverse, (void*)gc_func_tp_traverse},
+    {Py_tp_clear, (void*)gc_func_tp_clear},
+    {Py_tp_getset, (void*)gc_func_getset},
+    {Py_tp_getattro, (void*)gc_func_getattro},
+    {Py_tp_members, (void*)gc_func_members},
+    {Py_tp_call, (void*)PyVectorcall_Call},
+    {Py_tp_dealloc, (void*)gc_func_dealloc},
+    {0, 0}};
+
+static PyType_Spec gc_func_spec = {
+    /* .name = */ "mlx.gc_func",
+    /* .basicsize = */ (int)sizeof(gc_func),
+    /* .itemsize = */ 0,
+    /* .flags = */ Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HAVE_GC | NB_HAVE_VECTORCALL,
+    /* .slots = */ gc_func_slots};
+
+static PyTypeObject* gc_func_tp = nullptr;
+
+nb::callable mlx_func(nb::object func, std::vector<PyObject*> deps) {
+  gc_func* r = (gc_func*)PyType_GenericAlloc(gc_func_tp, 0);
+  r->func = func.inc_ref().ptr();
+  r->deps = std::move(deps);
+  r->vectorcall = gc_func_vectorcall;
+  return nb::steal<nb::callable>((PyObject*)r);
+}
+
+void init_mlx_func(nb::module_& m) {
+  gc_func_tp = (PyTypeObject*)PyType_FromSpec(&gc_func_spec);
+  if (!gc_func_tp) {
+    nb::raise("Could not register MLX function type.");
+  }
+}

python/src/mlx_func.h

@@ -0,0 +1,24 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <vector>
+
+#include <nanobind/nanobind.h>
+#include <nanobind/stl/function.h>
+
+namespace nb = nanobind;
+using namespace nb::literals;
+
+nb::callable mlx_func(nb::object func, std::vector<PyObject*> deps);
+
+template <typename F, typename... Deps>
+nb::callable mlx_func(F func, Deps&&... deps) {
+  return mlx_func(
+      nb::cpp_function(std::move(func)), std::vector<PyObject*>{deps.ptr()...});
+}
+
+template <typename... Deps>
+nb::callable mlx_func(nb::object func, Deps&&... deps) {
+  return mlx_func(std::move(func), std::vector<PyObject*>{deps.ptr()...});
+}

python/src/transforms.cpp

@@ -19,6 +19,7 @@
 #include "mlx/transforms.h"
 #include "mlx/transforms_impl.h"
 #include "mlx/utils.h"
+#include "python/src/mlx_func.h"
 #include "python/src/trees.h"
 
 namespace mx = mlx::core;
@@ -543,9 +544,9 @@ struct PyCompiledFun {
 
     tree_cache().erase(fun_id);
     mx::detail::compile_erase(fun_id);
-    fun.release().dec_ref();
+    fun.reset();
-    captured_inputs.release().dec_ref();
+    captured_inputs.reset();
-    captured_outputs.release().dec_ref();
+    captured_outputs.reset();
   }
 };
 
@@ -555,7 +556,7 @@ class PyCheckpointedFun {
   ~PyCheckpointedFun() {
     nb::gil_scoped_acquire gil;
 
-    fun_.release().dec_ref();
+    fun_.reset();
   }
 
   struct InnerFunction {
@@ -573,8 +574,8 @@ class PyCheckpointedFun {
     ~InnerFunction() {
       nb::gil_scoped_acquire gil;
 
-      fun_.release().dec_ref();
+      fun_.reset();
-      args_structure_.release().dec_ref();
+      args_structure_.reset();
     }
 
     std::vector<mx::array> operator()(const std::vector<mx::array>& inputs) {
@@ -609,6 +610,10 @@ class PyCheckpointedFun {
   nb::callable fun_;
 };
 
+int py_custom_function_tp_traverse(PyObject* self, visitproc visit, void* arg);
+
+int py_custom_function_tp_clear(PyObject* self);
+
 /**
  * PyCustomFunction is the class that implements the python decorator
  * `mx.custom_function`.
@@ -641,17 +646,7 @@ class PyCustomFunction {
   PyCustomFunction(nb::callable fun) : fun_(std::move(fun)) {}
   ~PyCustomFunction() {
     nb::gil_scoped_acquire gil;
-
+    reset();
-    fun_.release().dec_ref();
-    if (vjp_fun_.has_value()) {
-      (*vjp_fun_).release().dec_ref();
-    }
-    if (jvp_fun_.has_value()) {
-      (*jvp_fun_).release().dec_ref();
-    }
-    if (vmap_fun_.has_value()) {
-      (*vmap_fun_).release().dec_ref();
-    }
   }
 
   struct InnerFunction {
@@ -669,10 +664,10 @@ class PyCustomFunction {
     ~InnerFunction() {
       nb::gil_scoped_acquire gil;
 
-      fun_.release().dec_ref();
+      fun_.reset();
-      input_structure_.release().dec_ref();
+      input_structure_.reset();
       if (output_structure_.use_count() == 1) {
-        output_structure_->release().dec_ref();
+        output_structure_->reset();
       }
     }
 
@@ -703,10 +698,10 @@ class PyCustomFunction {
     ~InnerVJPFunction() {
       nb::gil_scoped_acquire gil;
 
-      vjp_fun_.release().dec_ref();
+      vjp_fun_.reset();
-      input_structure_.release().dec_ref();
+      input_structure_.reset();
       if (output_structure_.use_count() == 1) {
-        output_structure_->release().dec_ref();
+        output_structure_->reset();
       }
     }
 
@@ -746,8 +741,8 @@ class PyCustomFunction {
     ~InnerJVPFunction() {
       nb::gil_scoped_acquire gil;
 
-      jvp_fun_.release().dec_ref();
+      jvp_fun_.reset();
-      input_structure_.release().dec_ref();
+      input_structure_.reset();
     }
 
     std::vector<mx::array> operator()(
@@ -801,8 +796,8 @@ class PyCustomFunction {
     ~InnerVmapFunction() {
       nb::gil_scoped_acquire gil;
 
-      vmap_fun_.release().dec_ref();
+      vmap_fun_.reset();
-      input_structure_.release().dec_ref();
+      input_structure_.reset();
     }
 
     std::pair<std::vector<mx::array>, std::vector<int>> operator()(
@@ -904,6 +899,20 @@ class PyCustomFunction {
     vmap_fun_ = vmap_fun;
     return *this;
   }
+  void reset() {
+    fun_.reset();
+    if (vjp_fun_.has_value()) {
+      (*vjp_fun_).reset();
+    }
+    if (jvp_fun_.has_value()) {
+      (*jvp_fun_).reset();
+    }
+    if (vmap_fun_.has_value()) {
+      (*vmap_fun_).reset();
+    }
+  }
+
+  friend int py_custom_function_tp_traverse(PyObject*, visitproc, void*);
 
  private:
   std::optional<InnerVJPFunction> make_vjp_function(
@@ -940,10 +949,40 @@ class PyCustomFunction {
   std::optional<nb::callable> vmap_fun_;
 };
 
+int py_custom_function_tp_traverse(PyObject* self, visitproc visit, void* arg) {
+  auto* p = nb::inst_ptr<PyCustomFunction>(self);
+  nb::handle v = nb::find(p->fun_);
+  Py_VISIT(v.ptr());
+  if (p->vjp_fun_.has_value()) {
+    nb::handle v = nb::find(*(p->vjp_fun_));
+    Py_VISIT(v.ptr());
+  }
+  if (p->jvp_fun_.has_value()) {
+    nb::handle v = nb::find(*(p->jvp_fun_));
+    Py_VISIT(v.ptr());
+  }
+  if (p->vmap_fun_.has_value()) {
+    nb::handle v = nb::find(*(p->vmap_fun_));
+    Py_VISIT(v.ptr());
+  }
+  Py_VISIT(Py_TYPE(self));
+  return 0;
+}
+int py_custom_function_tp_clear(PyObject* self) {
+  auto* p = nb::inst_ptr<PyCustomFunction>(self);
+  p->reset();
+  return 0;
+}
+PyType_Slot py_custom_function_slots[] = {
+    {Py_tp_traverse, (void*)py_custom_function_tp_traverse},
+    {Py_tp_clear, (void*)py_custom_function_tp_clear},
+    {0, 0}};
+
 void init_transforms(nb::module_& m) {
   nb::class_<PyCustomFunction>(
       m,
       "custom_function",
+      nb::type_slots(py_custom_function_slots),
       R"pbdoc(
       Set up a function for custom gradient and vmap definitions.
 
@@ -1224,8 +1263,10 @@ void init_transforms(nb::module_& m) {
          const StrOrSet& argnames) {
         auto [argnums_vec, argnames_set] =
             validate_argnums_argnames(argnums, argnames);
-        return nb::cpp_function(py_value_and_grad(
+        return mlx_func(
-            fun, argnums_vec, argnames_set, "[value_and_grad]", false));
+            py_value_and_grad(
+                fun, argnums_vec, argnames_set, "[value_and_grad]", false),
+            fun);
       },
       "fun"_a,
       "argnums"_a = nb::none(),
@@ -1290,9 +1331,11 @@ void init_transforms(nb::module_& m) {
             validate_argnums_argnames(argnums, argnames);
         auto fn =
             py_value_and_grad(fun, argnums_vec, argnames_set, "[grad]", true);
-        return nb::cpp_function([fn](nb::args& args, nb::kwargs& kwargs) {
+        return mlx_func(
+            [fn = std::move(fn)](nb::args& args, nb::kwargs& kwargs) {
               return fn(args, kwargs).second;
-        });
+            },
+            fun);
       },
       "fun"_a,
       "argnums"_a = nb::none(),
@@ -1324,7 +1367,8 @@ void init_transforms(nb::module_& m) {
       [](const nb::callable& fun,
          const nb::object& in_axes,
          const nb::object& out_axes) {
-        return nb::cpp_function(py_vmap(fun, in_axes, out_axes));
+        return mlx_func(
+            py_vmap(fun, in_axes, out_axes), fun, in_axes, out_axes);
       },
       "fun"_a,
       "in_axes"_a = 0,
@@ -1379,11 +1423,15 @@ void init_transforms(nb::module_& m) {
             d.is_none() ? "MLX compiled function." : nb::cast<std::string>(d);
 
         auto sig_str = sig.str();
-        return nb::cpp_function(
+        return mlx_func(
+            nb::cpp_function(
                 PyCompiledFun{fun, inputs, outputs, shapeless},
                 nb::name(name.c_str()),
                 nb::sig(sig_str.c_str()),
-            doc.c_str());
+                doc.c_str()),
+            fun,
+            inputs,
+            outputs);
       },
       "fun"_a,
       "inputs"_a = nb::none(),
@@ -1435,7 +1483,7 @@ void init_transforms(nb::module_& m) {
       )pbdoc");
   m.def(
       "checkpoint",
-      [](nb::callable fun) { return nb::cpp_function(PyCheckpointedFun{fun}); },
+      [](nb::callable fun) { return mlx_func(PyCheckpointedFun{fun}, fun); },
       "fun"_a);
 
   // Register static Python object cleanup before the interpreter exits

python/tests/test_autograd.py

@@ -1,5 +1,6 @@
 # Copyright © 2023 Apple Inc.
 
+import gc
 import unittest
 
 import mlx.core as mx
@@ -737,6 +738,38 @@ class TestAutograd(mlx_tests.MLXTestCase):
         expected[4:-5:-2] = tan_b
         self.assertTrue(mx.allclose(grad, expected))
 
+    def test_leaks(self):
+        for transform in [
+            mx.grad,
+            mx.value_and_grad,
+            mx.custom_function,
+            mx.checkpoint,
+        ]:
+            if mx.metal.is_available():
+                mem_pre = mx.metal.get_active_memory()
+            else:
+                mem_pre = 0
+
+            def outer():
+                d = {}
+
+                def f(x):
+                    return d["x"]
+
+                d["f"] = transform(f)
+                d["x"] = mx.array([0] * 1000)
+
+            for _ in range(5):
+                outer()
+                gc.collect()
+
+            if mx.metal.is_available():
+                mem_post = mx.metal.get_active_memory()
+            else:
+                mem_post = 0
+
+            self.assertEqual(mem_pre, mem_post)
+
 
 if __name__ == "__main__":
     unittest.main()

python/tests/test_compile.py

@@ -1,5 +1,6 @@
 # Copyright © 2023-2024 Apple Inc.
 
+import gc
 import io
 import unittest
 from functools import partial
@@ -926,6 +927,32 @@ class TestCompile(mlx_tests.MLXTestCase):
         self.assertEqual(out[0].shape, (3, 1, 4, 2))
         self.assertEqual(out[1].shape, (2, 2, 5))
 
+    def test_leaks(self):
+        if mx.metal.is_available():
+            mem_pre = mx.metal.get_active_memory()
+        else:
+            mem_pre = 0
+
+        def outer():
+            d = {}
+
+            def f(x):
+                return d["x"]
+
+            d["f"] = mx.compile(f)
+            d["x"] = mx.array([0] * 1000)
+
+        for _ in range(5):
+            outer()
+            gc.collect()
+
+        if mx.metal.is_available():
+            mem_post = mx.metal.get_active_memory()
+        else:
+            mem_post = 0
+
+        self.assertEqual(mem_pre, mem_post)
+
 
 if __name__ == "__main__":
     unittest.main()

python/tests/test_export_import.py

@@ -1,5 +1,6 @@
 # Copyright © 2024 Apple Inc.
 
+import gc
 import os
 import tempfile
 import unittest
@@ -239,6 +240,33 @@ class TestExportImport(mlx_tests.MLXTestCase):
         constants_size = constant.nbytes + 8192
         self.assertTrue(os.path.getsize(path) < constants_size)
 
+    def test_leaks(self):
+        path = os.path.join(self.test_dir, "fn.mlxfn")
+        if mx.metal.is_available():
+            mem_pre = mx.metal.get_active_memory()
+        else:
+            mem_pre = 0
+
+        def outer():
+            d = {}
+
+            def f(x):
+                return d["x"]
+
+            d["f"] = mx.exporter(path, f)
+            d["x"] = mx.array([0] * 1000)
+
+        for _ in range(5):
+            outer()
+            gc.collect()
+
+        if mx.metal.is_available():
+            mem_post = mx.metal.get_active_memory()
+        else:
+            mem_post = 0
+
+        self.assertEqual(mem_pre, mem_post)
+
 
 if __name__ == "__main__":
     unittest.main()

python/tests/test_vmap.py

@@ -1,5 +1,6 @@
 # Copyright © 2023-2024 Apple Inc.
 
+import gc
 import unittest
 
 import mlx.core as mx
@@ -608,6 +609,32 @@ class TestVmap(mlx_tests.MLXTestCase):
         self.assertEqual(fx.shape, (5, 6, 7))
         self.assertEqual(fy.shape, (4, 5, 6, 7))
 
+    def test_leaks(self):
+        if mx.metal.is_available():
+            mem_pre = mx.metal.get_active_memory()
+        else:
+            mem_pre = 0
+
+        def outer():
+            d = {}
+
+            def f(x):
+                return d["x"]
+
+            d["f"] = mx.vmap(f)
+            d["x"] = mx.array([0] * 1000)
+
+        for _ in range(5):
+            outer()
+            gc.collect()
+
+        if mx.metal.is_available():
+            mem_post = mx.metal.get_active_memory()
+        else:
+            mem_post = 0
+
+        self.assertEqual(mem_pre, mem_post)
+
 
 if __name__ == "__main__":
     unittest.main()