dil always own whole storage whne create aten tensor if dil_tensor is int8

Author: XiaobingSuper

intel_pytorch_extension_py/__init__.py

@@ -15,6 +15,8 @@ def __init__(self, mixed_dtype = torch.bfloat16, configure_file = None):
         self.dtype = mixed_dtype
         self.configure_file = configure_file
 
+        if self.dtype != torch.bfloat16:
+            core.clear_indicators()
         # for int8 path, if user give a exited configure file, load it.
         if self.configure_file != None and self.dtype != torch.bfloat16:
             if os.path.exists(self.configure_file) and os.stat(self.configure_file).st_size != 0:

torch_ipex/csrc/auto_opt_config.h

@@ -98,6 +98,10 @@ class AutoOptConfig {
     }
   }
 
+  inline void clear_indicators() {
+    indicators_.clear();
+  }
+
   inline void add_indicators() {
     num_ops_id = 0;
     indicators_.clear();

torch_ipex/csrc/cpu/dbl/Common.cpp

@@ -230,8 +230,8 @@ at::Tensor gen_aten_tensor_by(dil::tensor&& dil_tensor) {
     // And even the dil tensor is plain format, it also cannot be shared with cpu buffer.
     shade_data_context->mix_prec_type = cpu::MIX_PREC_TYPE::MIX_BF16_FP32;
     at_data_type = at::kFloat;
-  } else if (check_auto_mix_int8_fp32() && (dil_tensor_type == dil::data_type::s8 || dil_tensor_type == dil::data_type::u8)) {
-    // If the user enables auto-mix-precision, then the aten tensor should always be float.
+  } else if ((dil_tensor_type == dil::data_type::s8 || dil_tensor_type == dil::data_type::u8)) {
+    // If the dil_tensor is int8 or unint8, then the aten tensor should always be float.
     // And even the dil tensor is plain format, it also cannot be shared with cpu buffer.
     shade_data_context->mix_prec_type = cpu::MIX_PREC_TYPE::MIX_INT8_FP32;
     at_data_type = at::kFloat;

torch_ipex/csrc/init_python_bindings.cpp

@@ -147,9 +147,12 @@ void InitIpexModuleBindings(py::module m) {
   m.def("get_int8_calibration", []() { return AutoOptConfig::singleton().get_int8_calibration(); });
   m.def("calibration_reset", []() { AutoOptConfig::singleton().calibration_reset(); });
   m.def("add_indicators", []() { AutoOptConfig::singleton().add_indicators(); });
+  // clear indicators for case having many scopes which have different structure
+  m.def("clear_indicators", []() { AutoOptConfig::singleton().clear_indicators(); }); 
   m.def("get_int8_configures", []() {
       py::list output_list;
       auto indicators = AutoOptConfig::singleton().get_indicators();
+      IPEX_CHECK(indicators.size() > 0, "can't load a empty indicators, please first do calibration step");
       for (auto indicator: indicators) {
         py::dict d;
         d["id"] = indicator.get_indicator_id();
@@ -169,7 +172,7 @@ void InitIpexModuleBindings(py::module m) {
       }
       return output_list; } );
   m.def("load_indicators_file", [](const py::list& l) {
-      IPEX_CHECK(py::len(l) > 0, "can't load a empty configures, please first do calibration setp"); 
+      IPEX_CHECK(py::len(l) > 0, "can't load a empty configures, please first do calibration step");
       std::vector<Indicator> indicators;
       for (py::handle i : l) {
         int64_t id = py::cast<std::int64_t>(i["id"]);