diff.patch

diff --git a/torch/csrc/distributed/Module.cpp b/torch/csrc/distributed/Module.cpp
index a985509..293a4e1 100644
--- a/torch/csrc/distributed/Module.cpp
+++ b/torch/csrc/distributed/Module.cpp
@@ -186,8 +186,8 @@ THDTensorDescriptor THDPModule_makeDescriptor(PyObject *obj)
   PyObject *type = (PyObject*)Py_TYPE(obj);
 #define REGISTER_TH_DESCRIPTOR(TYPE, REAL)                                           \
   if (type == THP##TYPE##Class)                                                \
-    return at::CPU(REAL).unsafeTensorFromTH(((THP##TYPE*)obj)->cdata, true);
-    /* return THDTensorDescriptor_newFromTH##TYPE(((THP##TYPE*)obj)->cdata); */
+    return THDTensorDescriptor_newFromTH##TYPE(((THP##TYPE*)obj)->cdata);
+    /* return at::CPU(REAL).unsafeTensorFromTH(((THP##TYPE*)obj)->cdata, true); */
   REGISTER_TH_DESCRIPTOR(DoubleTensor, at::kDouble);
   REGISTER_TH_DESCRIPTOR(FloatTensor, at::kFloat);
   REGISTER_TH_DESCRIPTOR(LongTensor, at::kLong);
@@ -199,12 +199,12 @@ THDTensorDescriptor THDPModule_makeDescriptor(PyObject *obj)
 #ifdef WITH_CUDA
 #define REGISTER_THC_DESCRIPTOR(TYPE, REAL)                                           \
   if (type == THCP##TYPE##Class)                                                \
-    return at::CUDA(REAL).unsafeTensorFromTH(((THP##TYPE*)obj)->cdata, true);
-    /* return THDTensorDescriptor_newFromTHCuda##TYPE((THCuda##TYPE*)(((torch::THPVoidTensor*)obj)->cdata)); */
+    return THDTensorDescriptor_newFromTHCuda##TYPE((THCuda##TYPE*)(((torch::THPVoidTensor*)obj)->cdata));
+    /* return at::CUDA(REAL).unsafeTensorFromTH(((THP##TYPE*)obj)->cdata, true); */
   REGISTER_THC_DESCRIPTOR(DoubleTensor, at::kDouble);
   if (type == THCPFloatTensorClass)
-    return at::CUDA(at::kFloat).unsafeTensorFromTH((THCudaTensor*)(((torch::THPVoidTensor*)obj)->cdata), true);
-    /* return THDTensorDescriptor_newFromTHCudaFloatTensor((THCudaTensor*)(((torch::THPVoidTensor*)obj)->cdata)); */
+    return THDTensorDescriptor_newFromTHCudaFloatTensor((THCudaTensor*)(((torch::THPVoidTensor*)obj)->cdata));
+    /* return at::CUDA(at::kFloat).unsafeTensorFromTH((THCudaTensor*)(((torch::THPVoidTensor*)obj)->cdata), true); */
   REGISTER_THC_DESCRIPTOR(LongTensor, at::kLong);
   REGISTER_THC_DESCRIPTOR(IntTensor, at::kInt);
   REGISTER_THC_DESCRIPTOR(ShortTensor, at::kShort);