add reference counting for get_autocast_dtype (intel#238)

* add reference counting for get_autocast_dtype

* fix test error

Author: XiaobingSuper

torch_ipex/csrc/init_python_bindings.cpp

@@ -52,8 +52,9 @@ void InitIpexModuleBindings(py::module m) {
   // ipex amp autocast
   m.def("get_autocast_dtype", []() {
     at::ScalarType current_dtype = torch_ipex::autocast::get_autocast_dtype();
-    return py::reinterpret_steal<py::object>(
-        (PyObject*)torch::getTHPDtype(current_dtype));
+    auto dtype = (PyObject*)torch::getTHPDtype(current_dtype);
+    Py_INCREF(dtype);
+    return py::reinterpret_steal<py::object>(dtype);
   });
   m.def("set_autocast_dtype", [](py::object dtype) {
     at::ScalarType target_dtype =

torch_ipex_py/quantization/quantization_utils.py

@@ -1,6 +1,7 @@
 import torch
 import functools
 import warnings
+import copy
 import numpy as np
 import intel_extension_for_pytorch._C as core
 from .. import conf
@@ -167,14 +168,16 @@ def decorate_autocast(*args, **kwargs):
 def convert(model, conf, inputs):
     # pre-conver model's parameters dtype if it has conv, linear
     # and Embedding for bfloat16 path.
+    model_ = model
     if torch.is_autocast_cpu_enabled() and core.get_autocast_dtype() == torch.bfloat16:
-        model = utils._convert_module_data_type(model, torch.bfloat16)
+        model_ = utils._convert_module_data_type(copy.deepcopy(model), torch.bfloat16)
+
     core.disable_jit_opt()
     core._jit_set_llga_enabled(True)
     torch._C._jit_set_profiling_mode(True)
     torch._C._jit_set_profiling_executor(True)
     with torch.no_grad(), _quantization_int8():
-        trace_model = torch.jit.trace(model, inputs, check_trace=False)
+        trace_model = torch.jit.trace(model_, inputs, check_trace=False)
     trace_model = torch.jit.freeze(trace_model)
 
     return trace_model