Use cuda

src/pystencils_autodiff/backends/_torch_native.py

@@ -62,7 +62,8 @@ def create_autograd_function(autodiff_obj, use_cuda):
                 kwargs[field.name] = torch.zeros(
                     field.shape,
                     dtype=numpy_dtype_to_torch(field.dtype.numpy_dtype),
-                    device=next(chain(args, kwargs.values())).device)
+                    device='cuda' if use_cuda else 'cpu')  # use device of tensor
+
         output_tensors = OrderedDict({f.name:
                                       field_to_tensor_dict.get(f, kwargs[f.name])
                                       for f in autodiff_obj.forward_output_fields})

src/pystencils_autodiff/backends/astnodes.py

@@ -55,7 +55,7 @@ class Header(JinjaCppFile):
 
 class TorchTensorDestructuring(DestructuringBindingsForFieldClass):
     CLASS_TO_MEMBER_DICT = {
-        FieldPointerSymbol: "data_ptr<{dtype}>()",
+        FieldPointerSymbol: "data<{dtype}>()",
         FieldShapeSymbol: "size({dim})",
         FieldStrideSymbol: "strides()[{dim}]"
     }