add minimal example that reproduces the nanobind leak error

src/nanobind_example/__init__.py

@@ -1 +1 @@
-from .nanobind_example_ext import add
+from .nanobind_example_ext import add, Foo, make_foo

src/nanobind_example_ext.cpp

@@ -4,6 +4,23 @@ namespace nb = nanobind;
 
 using namespace nb::literals;
 
+struct Foo {
+    int a;
+    float b;
+};
+
+Foo makeFoo(int _a, float _b) {
+    Foo foo = {_a, _b};
+    return foo;
+}
+
 NB_MODULE(nanobind_example_ext, m) {
     m.def("add", [](int a, int b) { return a + b; }, "a"_a, "b"_a);
+
+    nb::class_<Foo>(m, "Foo")
+      .def_readonly("a", &Foo::a)
+      .def_readonly("b", &Foo::b);
+
+    m.def("make_foo", &makeFoo, nb::arg("a"), nb::arg("b"));
+
 }

tests/test_autograd.py

@@ -0,0 +1,24 @@
+import torch
+import math
+from nanobind_example import *
+
+# example adapted from https://pytorch.org/tutorials/beginner/examples_autograd/two_layer_net_custom_function.html
+class LegendrePolynomial3(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input, foo):
+        ctx.foo = foo
+        ctx.save_for_backward(input)
+        return 0.5 * (5 * input ** 3 - 3 * input)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, = ctx.saved_tensors
+        return grad_output * 1.5 * (5 * input ** 2 - 1), None
+
+input_tensor = torch.rand(3,3)
+input_tensor.requires_grad_()
+foo = make_foo(1, 0.5)
+
+output = LegendrePolynomial3.apply(input_tensor, foo)
+grad_output = torch.rand_like(output)
+output.backward(grad_output)