Respect default dtype (#3798)

* Respect default dtype

Prior to the linked PR, the python API never actually passed nullopt
and so the `OrFloat` path was never tested.

* Trigger CI

* Remove .torch_pin

Author: peterbell10

torch_xla/csrc/aten_xla_type.cpp

@@ -84,10 +84,6 @@ torch::lazy::BackendDevice GetXlaDeviceOrCurrent(
   return xla_device_opt ? *xla_device_opt : GetCurrentDevice();
 }
 
-at::ScalarType GetScalarTypeOrFloat(c10::optional<at::ScalarType> scalar_type) {
-  return scalar_type ? *scalar_type : at::ScalarType::Float;
-}
-
 bool IsOperationOnType(const c10::optional<at::ScalarType>& opt_dtype,
                        at::ScalarType tensor_type, at::ScalarType type) {
   if (opt_dtype && *opt_dtype == type) {
@@ -1286,7 +1282,7 @@ at::Tensor XLANativeFunctions::empty(
   // s_copy_().
   return bridge::AtenFromXlaTensor(XLATensor::full(
       XlaHelpers::I64List(size), 0, GetXlaDeviceOrCurrent(device),
-      GetScalarTypeOrFloat(dtype)));
+      at::dtype_or_default(dtype)));
 }
 
 at::Tensor XLANativeFunctions::empty_symint(
@@ -1689,7 +1685,7 @@ at::Tensor XLANativeFunctions::linspace(const at::Scalar& start,
   }
 
   return bridge::AtenFromXlaTensor(
-      XLATensor::linspace(start, end, steps, GetScalarTypeOrFloat(dtype),
+      XLATensor::linspace(start, end, steps, at::dtype_or_default(dtype),
                           GetXlaDeviceOrCurrent(device)));
 }