diff --git a/build/__native_faster_hexbytes.c b/build/__native_faster_hexbytes.c
index e789a56..4c98647 100644
--- a/build/__native_faster_hexbytes.c
+++ b/build/__native_faster_hexbytes.c
@@ -1,22 +1,22 @@
 #ifndef DIFFCHECK_PLACEHOLDER
 #define DIFFCHECK_PLACEHOLDER 0
 #endif
-#include "init.c"
-#include "getargs.c"
-#include "getargsfast.c"
-#include "int_ops.c"
-#include "float_ops.c"
-#include "str_ops.c"
-#include "bytes_ops.c"
-#include "list_ops.c"
-#include "dict_ops.c"
-#include "set_ops.c"
-#include "tuple_ops.c"
-#include "exc_ops.c"
-#include "misc_ops.c"
-#include "generic_ops.c"
-#include "pythonsupport.c"
-#include "function_wrapper.c"
+#include <init.c>
+#include <getargs.c>
+#include <getargsfast.c>
+#include <int_ops.c>
+#include <float_ops.c>
+#include <str_ops.c>
+#include <bytes_ops.c>
+#include <list_ops.c>
+#include <dict_ops.c>
+#include <set_ops.c>
+#include <tuple_ops.c>
+#include <exc_ops.c>
+#include <misc_ops.c>
+#include <generic_ops.c>
+#include <pythonsupport.c>
+#include <function_wrapper.c>
 #include "__native_faster_hexbytes.h"
 #include "__native_internal_faster_hexbytes.h"
 static PyMethodDef faster_hexbytesmodule_methods[] = {
@@ -4115,6 +4115,7 @@ CPyL125: ;
                 }
                 if (exec_faster_hexbytes__mypyc(module) < 0) {
                     Py_DECREF(module);
+                    module = NULL;
                     return NULL;
                 }
                 return module;
diff --git a/build/faster_hexbytes.c b/build/faster_hexbytes.c
index f3a5670..78d8c84 100644
--- a/build/faster_hexbytes.c
+++ b/build/faster_hexbytes.c
@@ -7,7 +7,42 @@ PyMODINIT_FUNC
 PyInit_faster_hexbytes(void)
 {
     PyObject *tmp;
-    if (!(tmp = PyImport_ImportModule("faster_hexbytes__mypyc"))) return NULL;
+    // Use ImportModuleLevel with a non-empty fromlist so Python returns the
+    // leaf submodule via sys.modules lookup, avoiding the top-down getattr
+    // walk done by PyImport_ImportModule. That walk fails when the shim runs
+    // during a parent package's __init__.py, because the parent's submodule
+    // attribute isn't set until after __init__ returns.
+    static PyObject *fromlist = NULL;
+    if (!fromlist) {
+        fromlist = Py_BuildValue("(s)", "*");
+        if (!fromlist) return NULL;
+    }
+    tmp = PyImport_ImportModuleLevel("faster_hexbytes__mypyc", NULL, NULL, fromlist, 0);
+    if (!tmp) return NULL;
+    // Populate cross-group export tables lazily, just before we instantiate
+    // the per-module real init. Deferring this out of the shared lib's own
+    // PyInit keeps separate-mode compiled modules from recursively triggering
+    // sibling-package __init__.py mid-bootstrap.
+    // PyObject_GetOptionalAttrString would be cleaner, but it's 3.13+ and the
+    // shim deliberately #includes only <Python.h>, so we can't pull in the
+    // pythoncapi_compat.h shim from lib-rt here. AttributeError + PyErr_Clear
+    // is the portable form.
+    PyObject *deps_capsule = PyObject_GetAttrString(tmp, "ensure_deps");
+    if (deps_capsule != NULL) {
+        int (*deps_func)(void) = (int (*)(void))PyCapsule_GetPointer(
+            deps_capsule, "faster_hexbytes__mypyc.ensure_deps");
+        Py_DECREF(deps_capsule);
+        if (deps_func == NULL) {
+            Py_DECREF(tmp);
+            return NULL;
+        }
+        if (deps_func() < 0) {
+            Py_DECREF(tmp);
+            return NULL;
+        }
+    } else {
+        PyErr_Clear();
+    }
     PyObject *capsule = PyObject_GetAttrString(tmp, "init_faster_hexbytes");
     Py_DECREF(tmp);
     if (capsule == NULL) return NULL;
diff --git a/build/faster_hexbytes/_utils.c b/build/faster_hexbytes/_utils.c
index 466da50..600e8dd 100644
--- a/build/faster_hexbytes/_utils.c
+++ b/build/faster_hexbytes/_utils.c
@@ -4,7 +4,42 @@ PyMODINIT_FUNC
 PyInit__utils(void)
 {
     PyObject *tmp;
-    if (!(tmp = PyImport_ImportModule("faster_hexbytes__mypyc"))) return NULL;
+    // Use ImportModuleLevel with a non-empty fromlist so Python returns the
+    // leaf submodule via sys.modules lookup, avoiding the top-down getattr
+    // walk done by PyImport_ImportModule. That walk fails when the shim runs
+    // during a parent package's __init__.py, because the parent's submodule
+    // attribute isn't set until after __init__ returns.
+    static PyObject *fromlist = NULL;
+    if (!fromlist) {
+        fromlist = Py_BuildValue("(s)", "*");
+        if (!fromlist) return NULL;
+    }
+    tmp = PyImport_ImportModuleLevel("faster_hexbytes__mypyc", NULL, NULL, fromlist, 0);
+    if (!tmp) return NULL;
+    // Populate cross-group export tables lazily, just before we instantiate
+    // the per-module real init. Deferring this out of the shared lib's own
+    // PyInit keeps separate-mode compiled modules from recursively triggering
+    // sibling-package __init__.py mid-bootstrap.
+    // PyObject_GetOptionalAttrString would be cleaner, but it's 3.13+ and the
+    // shim deliberately #includes only <Python.h>, so we can't pull in the
+    // pythoncapi_compat.h shim from lib-rt here. AttributeError + PyErr_Clear
+    // is the portable form.
+    PyObject *deps_capsule = PyObject_GetAttrString(tmp, "ensure_deps");
+    if (deps_capsule != NULL) {
+        int (*deps_func)(void) = (int (*)(void))PyCapsule_GetPointer(
+            deps_capsule, "faster_hexbytes__mypyc.ensure_deps");
+        Py_DECREF(deps_capsule);
+        if (deps_func == NULL) {
+            Py_DECREF(tmp);
+            return NULL;
+        }
+        if (deps_func() < 0) {
+            Py_DECREF(tmp);
+            return NULL;
+        }
+    } else {
+        PyErr_Clear();
+    }
     PyObject *capsule = PyObject_GetAttrString(tmp, "init_faster_hexbytes____utils");
     Py_DECREF(tmp);
     if (capsule == NULL) return NULL;
diff --git a/build/faster_hexbytes/main.c b/build/faster_hexbytes/main.c
index 54a5b09..cf3af33 100644
--- a/build/faster_hexbytes/main.c
+++ b/build/faster_hexbytes/main.c
@@ -4,7 +4,42 @@ PyMODINIT_FUNC
 PyInit_main(void)
 {
     PyObject *tmp;
-    if (!(tmp = PyImport_ImportModule("faster_hexbytes__mypyc"))) return NULL;
+    // Use ImportModuleLevel with a non-empty fromlist so Python returns the
+    // leaf submodule via sys.modules lookup, avoiding the top-down getattr
+    // walk done by PyImport_ImportModule. That walk fails when the shim runs
+    // during a parent package's __init__.py, because the parent's submodule
+    // attribute isn't set until after __init__ returns.
+    static PyObject *fromlist = NULL;
+    if (!fromlist) {
+        fromlist = Py_BuildValue("(s)", "*");
+        if (!fromlist) return NULL;
+    }
+    tmp = PyImport_ImportModuleLevel("faster_hexbytes__mypyc", NULL, NULL, fromlist, 0);
+    if (!tmp) return NULL;
+    // Populate cross-group export tables lazily, just before we instantiate
+    // the per-module real init. Deferring this out of the shared lib's own
+    // PyInit keeps separate-mode compiled modules from recursively triggering
+    // sibling-package __init__.py mid-bootstrap.
+    // PyObject_GetOptionalAttrString would be cleaner, but it's 3.13+ and the
+    // shim deliberately #includes only <Python.h>, so we can't pull in the
+    // pythoncapi_compat.h shim from lib-rt here. AttributeError + PyErr_Clear
+    // is the portable form.
+    PyObject *deps_capsule = PyObject_GetAttrString(tmp, "ensure_deps");
+    if (deps_capsule != NULL) {
+        int (*deps_func)(void) = (int (*)(void))PyCapsule_GetPointer(
+            deps_capsule, "faster_hexbytes__mypyc.ensure_deps");
+        Py_DECREF(deps_capsule);
+        if (deps_func == NULL) {
+            Py_DECREF(tmp);
+            return NULL;
+        }
+        if (deps_func() < 0) {
+            Py_DECREF(tmp);
+            return NULL;
+        }
+    } else {
+        PyErr_Clear();
+    }
     PyObject *capsule = PyObject_GetAttrString(tmp, "init_faster_hexbytes___main");
     Py_DECREF(tmp);
     if (capsule == NULL) return NULL;
diff --git a/pyproject.toml b/pyproject.toml
index 5d3868f..7fd947c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools", "wheel", "mypy[mypyc]==2.0.0", "hexbytes==1.3.1"]
+requires = ["setuptools", "wheel", "mypy[mypyc]==2.1.0", "hexbytes==1.3.1"]
 build-backend = "setuptools.build_meta"
 
 [tool.autoflake]
diff --git a/setup.py b/setup.py
index aae897b..cddd4e3 100644
--- a/setup.py
+++ b/setup.py
@@ -15,7 +15,7 @@
         "build>=0.9.0",
         "bump_my_version>=0.19.0",
         "ipython",
-        "mypy==2.0.0",
+        "mypy==2.1.0",
         "pre-commit>=3.4.0",
         "tox>=4.0.0",
         "twine",