motus
diff --git a/‎doxygen/doxygen.config
100755100644 b/‎doxygen/doxygen.config
100755100644
diff --git a/‎libtorch/1_libtorch_classes.md
+14 b/‎libtorch/1_libtorch_classes.md
+14
diff --git a/‎libtorch/structat_1_1Type__inherit__graph.png
164 KB b/‎libtorch/structat_1_1Type__inherit__graph.png
164 KB
diff --git a/‎th/1_trace_blas.md
-7 b/‎th/1_trace_blas.md
-7
@@ -0,0 +1,14 @@
+
+# The origin of PyTorch methods
+
+At the top of high-level libtorch C++ API is the `at::Tensor` class. It provides a comprehensive set
+of methods for tensor storage management, data initialization, auto-differentiation, as well as all
+basic tensor operations. It is defined at
+[ATen/core/Tensor.h](https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/core/Tensor.h)
+
+All actual work is delegated from `Tensor` to the classes `TensorImpl` and `Type`.
+
+
+Here's how the `Type` hierarchy looks like for CUDA:
+
+![libtorch Type hierarchy](structat_1_1Type__inherit__graph.png)