Revert "Add magic TORCH_MAKE_PYBIND_ENUM_FASTER macro (#163527)"

This reverts commit 50c0550f5a5b1e35885d892081a7d5115d8b4489.

Reverted https://github.com/pytorch/pytorch/pull/163527 on behalf of https://github.com/swolchok due to breaking import torch in debug builds, see #164297 ([comment](https://github.com/pytorch/pytorch/pull/163527#issuecomment-3361919142))

tools/autograd/templates/python_enum_tag.cpp

@@ -4,9 +4,6 @@
 #include <ATen/core/enum_tag.h>
 
 namespace py = pybind11;
-
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::Tag)
-
 namespace torch {
     namespace autograd {
     void initEnumTag(PyObject* module) {

torch/csrc/Module.cpp

@@ -145,14 +145,6 @@
 
 namespace py = pybind11;
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::native::ConvBackend)
-TORCH_MAKE_PYBIND_ENUM_FASTER(sdp::SDPBackend)
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::LinalgBackend)
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::BlasBackend)
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::ROCmFABackend)
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::native::BatchNormBackend)
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::impl::TorchFunctionDisabledState)
-
 static PyObject* module;
 
 static THPGenerator* THPDefaultCPUGenerator = nullptr;

torch/csrc/autograd/init.cpp

@@ -44,9 +44,6 @@
 using torch::impl::py_context_manager;
 using torch::impl::py_context_manager_DEPRECATED;
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::autograd::CreationMeta)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10::DeviceType)
-
 namespace {
 
 struct DisableFuncTorch {

torch/csrc/cuda/shared/cudart.cpp

@@ -10,11 +10,6 @@
 #include <c10/cuda/CUDAException.h>
 #include <c10/cuda/CUDAGuard.h>
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(cudaError_t)
-#if !defined(USE_ROCM) && defined(CUDA_VERSION) && CUDA_VERSION < 12000
-TORCH_MAKE_PYBIND_ENUM_FASTER(cudaOutputMode_t)
-#endif
-
 namespace torch::cuda::shared {
 
 #ifdef USE_ROCM

torch/csrc/cuda/shared/cudnn.cpp

@@ -13,8 +13,6 @@ using version_tuple = std::tuple<size_t, size_t, size_t>;
 #ifdef USE_CUDNN
 #include <cudnn.h>
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(cudnnRNNMode_t)
-
 namespace {
 
 version_tuple getCompileVersion() {

torch/csrc/distributed/c10d/init.cpp

@@ -66,14 +66,6 @@
 
 #include <torch/custom_class.h>
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10d::BuiltinCommHookType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10d::DebugLevel)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10d::ReduceOp::RedOpType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10d::ProcessGroup::BackendType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10d::OpType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10d::WorkResult)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10d::ErrorType)
-
 namespace {
 
 #ifdef USE_C10D_NCCL

torch/csrc/dynamo/init.cpp

@@ -19,8 +19,6 @@ static struct PyModuleDef _module =
 
 PYBIND11_MAKE_OPAQUE(std::vector<uint8_t>)
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(FrameAction)
-
 namespace torch::dynamo {
 
 #if IS_PYTHON_3_11_PLUS

torch/csrc/functorch/init.cpp

@@ -7,7 +7,6 @@
 #include <ATen/FunctionalTensorWrapper.h>
 #include <ATen/WrapDimUtils.h>
 #include <torch/csrc/functorch/init.h>
-#include <torch/csrc/utils/pybind.h>
 #include <torch/csrc/utils/python_raii.h>
 #include <torch/python.h>
 
@@ -25,9 +24,6 @@
 
 // This file contains functorch's Python bindings.
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::functorch::TransformType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::functorch::RandomnessType)
-
 namespace torch::functorch::impl {
 
 using namespace at::functorch;

torch/csrc/jit/python/init.cpp

@@ -121,9 +121,6 @@
 #include <tuple>
 #include <utility>
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(MobileOptimizerType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10::SchemaArgType)
-
 namespace torch::jit {
 
 using c10::AliasInfo;

torch/csrc/jit/python/script_init.cpp

@@ -75,9 +75,6 @@
 
 #include <fmt/format.h>
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(
-    torch::jit::logging::LockingLogger::AggregationType)
-
 namespace torch::jit {
 
 using ::c10::Argument;

torch/csrc/monitor/python_init.cpp

@@ -17,8 +17,6 @@
 #include <torch/csrc/monitor/events.h>
 #include <torch/csrc/monitor/python_init.h>
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::monitor::Aggregation)
-
 namespace pybind11::detail {
 template <>
 struct type_caster<torch::monitor::data_value_t> {

torch/csrc/onnx/init.cpp

@@ -28,11 +28,6 @@
 #include <torch/csrc/jit/passes/onnx/shape_type_inference.h>
 #include <torch/csrc/jit/passes/onnx/unpack_quantized_weights.h>
 #include <torch/csrc/jit/serialization/export.h>
-#include <torch/csrc/utils/pybind.h>
-
-TORCH_MAKE_PYBIND_ENUM_FASTER(ONNX_NAMESPACE::TensorProto_DataType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::onnx::OperatorExportTypes)
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::onnx::TrainingMode)
 
 namespace torch::onnx {
 

torch/csrc/profiler/python/init.cpp

@@ -12,12 +12,6 @@
 #include <torch/csrc/profiler/standalone/execution_trace_observer.h>
 #include <torch/csrc/utils/pybind.h>
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(at::RecordScope)
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::profiler::impl::ProfilerState)
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::profiler::impl::ActiveProfilerType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::profiler::impl::ActivityType)
-TORCH_MAKE_PYBIND_ENUM_FASTER(torch::profiler::impl::EventType)
-
 struct THPCapturedTraceback {
   PyObject_HEAD
   std::shared_ptr<torch::CapturedTraceback> data;

torch/csrc/utils/pybind.h

@@ -29,28 +29,6 @@ PYBIND11_DECLARE_HOLDER_TYPE(T, c10::intrusive_ptr<T>, true)
 PYBIND11_DECLARE_HOLDER_TYPE(T, c10::SingletonOrSharedTypePtr<T>)
 PYBIND11_DECLARE_HOLDER_TYPE(T, c10::SingletonTypePtr<T>, true)
 
-// pybind11 3.x's type_caster_enum_type handles both py::native_enum
-// and py::enum_. py::native_enum is preferred, so it gets checked
-// first. We still use lots of py::enum_ because we don't (yet?)
-// require pybind11 3.x, and possibly because the difference is
-// user-visible. Putting TORCH_MAKE_PYBIND_ENUM_FASTER(T) at global
-// scope before using py::enum_<T> will cause pybind function calls
-// that pass arguments of type T to go faster (16% at time of writing,
-// but they are quite slow currently and the savings is a fixed cost,
-// so the percentage may be higher after other optimizations for
-// py::enum_ happen).
-#ifdef PYBIND11_HAS_NATIVE_ENUM
-#define TORCH_MAKE_PYBIND_ENUM_FASTER(T)                                    \
-  namespace pybind11::detail {                                              \
-  template <>                                                               \
-  struct type_caster_enum_type_enabled<T, void> : std::false_type {};       \
-  template <>                                                               \
-  struct type_caster_enum_type_enabled<const T, void> : std::false_type {}; \
-  } // namespace pybind11::detail
-#else // PYBIND11_HAS_NATIVE_ENUM
-#define TORCH_MAKE_PYBIND_ENUM_FASTER(T)
-#endif // PYBIND11_HAS_NATIVE_ENUM
-
 namespace pybind11::detail {
 
 // torch.Tensor <-> at::Tensor conversions (without unwrapping)

torch/csrc/utils/python_dispatch.cpp

@@ -34,9 +34,6 @@
 
 namespace py = pybind11;
 
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10::DispatchKey)
-TORCH_MAKE_PYBIND_ENUM_FASTER(c10::impl::TorchDispatchModeKey)
-
 namespace torch::impl::dispatch {
 
 // Global storage for leaked Python filenames to ensure they remain valid