Replace pybind11 with nanobind in the OQC module (#1391)

**Context:** OQC was still using pybind11, which made it incompatible
with the stable ABI.

**Description of the Change:** Replace pybind11 with nanobind in the OQC
module.

**Benefits:** Conforming with the stable ABI.

Makefile

@@ -192,6 +192,7 @@ wheel:
 	cp $(RT_BUILD_DIR)/lib/librt_capi.* $(MK_DIR)/frontend/catalyst/lib
 	cp $(RT_BUILD_DIR)/lib/backend/*.toml $(MK_DIR)/frontend/catalyst/lib/backend
 	cp $(OQC_BUILD_DIR)/librtd_oqc* $(MK_DIR)/frontend/catalyst/lib
+	cp $(OQC_BUILD_DIR)/oqc_python_module.so $(MK_DIR)/frontend/catalyst/lib
 	cp $(OQC_BUILD_DIR)/backend/*.toml $(MK_DIR)/frontend/catalyst/lib/backend
 	cp $(OQD_BUILD_DIR)/librtd_oqd* $(MK_DIR)/frontend/catalyst/lib
 	cp $(OQD_BUILD_DIR)/backend/*.toml $(MK_DIR)/frontend/catalyst/lib/backend

doc/releases/changelog-dev.md

@@ -15,6 +15,7 @@
 * Replace pybind11 with nanobind for C++/Python bindings in the frontend and in the runtime.
   [(#1173)](https://github.com/PennyLaneAI/catalyst/pull/1173)
   [(#1293)](https://github.com/PennyLaneAI/catalyst/pull/1293)
+  [(#1391)](https://github.com/PennyLaneAI/catalyst/pull/1391)
 
   Nanobind has been developed as a natural successor to the pybind11 library and offers a number of
   [advantages](https://nanobind.readthedocs.io/en/latest/why.html#major-additions), in particular,

frontend/catalyst/third_party/oqc/src/CMakeLists.txt

@@ -10,18 +10,39 @@ set(oqc_backend_dir "${OQC_BUILD_DIR}/backend")
 set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
-# Locate Python & PyBind11
-# The optional component is only used for the C++ test suite (to spin up their its interpreter),
+# nanobind suggests including these lines to configure CMake to perform an optimized release build
+# by default unless another build type is specified. Without this addition, binding code may run
+# slowly and produce large binaries.
+# See https://nanobind.readthedocs.io/en/latest/building.html#preliminaries
+if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Choose the type of build." FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+# Locate Python
+# The optional component is only used for the C++ test suite (to spin up its own interpreter),
 # and requires libpython.so to be available on the system.
 find_package(Python REQUIRED
     COMPONENTS Interpreter Development.Module
-    OPTIONAL_COMPONENTS Development.Embed
+    OPTIONAL_COMPONENTS Development.Embed Development.SABIModule
 )
+
+# Locate nanobind
 execute_process(
-    COMMAND "${Python_EXECUTABLE}" -c "import pybind11; print(pybind11.get_cmake_dir())"
-    OUTPUT_VARIABLE pybind11_DIR OUTPUT_STRIP_TRAILING_WHITESPACE
+    COMMAND "${Python_EXECUTABLE}" -c "import nanobind; print(nanobind.cmake_dir())"
+    OUTPUT_VARIABLE nanobind_DIR OUTPUT_STRIP_TRAILING_WHITESPACE
 )
-find_package(pybind11 CONFIG REQUIRED)
+find_package(nanobind CONFIG REQUIRED)
+
+# Create the Python `oqc_python_module` module
+# Target the stable ABI for Python 3.12+, which reduces the number of binary wheels that must be
+# built (`STABLE_ABI` does nothing on older Python versions).
+nanobind_add_module(oqc_python_module STABLE_ABI oqc_python_module.cpp)
+
+# Use a consistant suffix ".so" rather than, e.g. ".abi3.so" (when using the Stable ABI) or
+# ".cpython-3xx-darwin.so". Doing so simplifies the process to locate it when calling
+# `dlopen(OQC_PY)` in frontend/catalyst/third_party/oqc/src/OQCRunner.hpp.
+set_target_properties(oqc_python_module PROPERTIES SUFFIX ".so")
 
 message(STATUS "Building the OQC device.")
 
@@ -39,7 +60,6 @@ set(OQC_LIBRARIES
 set_target_properties(rtd_oqc PROPERTIES BUILD_RPATH "$ORIGIN/../utils")
 target_link_directories(rtd_oqc PRIVATE ${runtime_lib})
 
-pybind11_add_module(oqc_python_module oqc_python_module.cpp)
 target_include_directories(oqc_python_module PRIVATE ${runtime_includes})
 
 add_dependencies(rtd_oqc oqc_python_module)

frontend/catalyst/third_party/oqc/src/oqc_python_module.cpp

@@ -12,9 +12,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <pybind11/eval.h>
-#include <pybind11/pybind11.h>
 #include <string>
+#include <vector>
+
+#include <nanobind/eval.h>
+#include <nanobind/nanobind.h>
+#include <nanobind/stl/string.h>
 
 #include "Exception.hpp"
 
@@ -46,28 +49,34 @@ except Exception as e:
 [[gnu::visibility("default")]] void counts(const char *_circuit, const char *_device, size_t shots,
                                            size_t num_qubits, const char *_kwargs, void *_vector)
 {
-    namespace py = pybind11;
-    using namespace py::literals;
+    namespace nb = nanobind;
+    using namespace nb::literals;
 
-    py::gil_scoped_acquire lock;
+    nb::gil_scoped_acquire lock;
 
-    auto locals = py::dict("circuit"_a = _circuit, "device"_a = _device, "kwargs"_a = _kwargs,
-                           "shots"_a = shots, "msg"_a = "");
+    nb::dict locals;
+    locals["circuit"] = _circuit;
+    locals["device"] = _device;
+    locals["kwargs"] = _kwargs;
+    locals["shots"] = shots;
+    locals["msg"] = "";
 
-    py::exec(program, py::globals(), locals);
+    // Evaluate in scope of main module
+    nb::object scope = nb::module_::import_("__main__").attr("__dict__");
+    nb::exec(nb::str(program.c_str()), scope, locals);
 
-    auto &&msg = locals["msg"].cast<std::string>();
+    auto msg = nb::cast<std::string>(locals["msg"]);
     RT_FAIL_IF(!msg.empty(), msg.c_str());
 
-    py::dict results = locals["counts"];
+    nb::dict results = locals["counts"];
 
     std::vector<size_t> *counts_value = reinterpret_cast<std::vector<size_t> *>(_vector);
     for (auto item : results) {
         auto key = item.first;
         auto value = item.second;
-        counts_value->push_back(value.cast<size_t>());
+        counts_value->push_back(nb::cast<size_t>(value));
     }
     return;
 }
 
-PYBIND11_MODULE(oqc_python_module, m) { m.doc() = "oqc"; }
+NB_MODULE(oqc_python_module, m) { m.doc() = "oqc"; }

frontend/catalyst/third_party/oqc/src/tests/CMakeLists.txt

@@ -32,6 +32,13 @@ target_include_directories(runner_tests_oqc PRIVATE
 
 target_link_directories(runner_tests_oqc PRIVATE ${runtime_lib})
 
+# Locate PyBind11
+execute_process(
+    COMMAND "${Python_EXECUTABLE}" -c "import pybind11; print(pybind11.get_cmake_dir())"
+    OUTPUT_VARIABLE pybind11_DIR OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+find_package(pybind11 CONFIG REQUIRED)
+
 target_link_libraries(runner_tests_oqc PRIVATE
     Catch2::Catch2
     pybind11::embed

frontend/catalyst/utils/CMakeLists.txt

@@ -17,6 +17,7 @@ endif()
 # Locate Python & nanobind
 find_package(Python REQUIRED
     COMPONENTS Interpreter Development.Module NumPy
+    OPTIONAL_COMPONENTS Development.SABIModule
 )
 execute_process(
     COMMAND "${Python_EXECUTABLE}" -c "import nanobind; print(nanobind.cmake_dir())"

runtime/CMakeLists.txt

@@ -67,7 +67,7 @@ FetchContent_Declare(
 # and requires libpython.so to be available on the system.
 find_package(Python REQUIRED
     COMPONENTS Interpreter Development.Module
-    OPTIONAL_COMPONENTS Development.Embed
+    OPTIONAL_COMPONENTS Development.Embed Development.SABIModule
 )
 
 if(RUNTIME_ENABLE_WARNINGS)

runtime/lib/backend/openqasm/CMakeLists.txt

@@ -25,7 +25,7 @@ execute_process(
 )
 find_package(nanobind CONFIG REQUIRED)
 
-# Create the Python `catalyst_callback_registry` module
+# Create the Python `openqasm_python_module` module
 # Target the stable ABI for Python 3.12+, which reduces the number of binary wheels that must be
 # built (`STABLE_ABI` does nothing on older Python versions).
 nanobind_add_module(openqasm_python_module STABLE_ABI openqasm_python_module.cpp)