fix(build): torch 1.8 with cuda 11.3 string_view patch

CMakeLists.txt

@@ -747,6 +747,7 @@ if (USE_TORCH)
     ${PYTORCH_PATCHES_PATH}/pytorch_16_new_logger.patch
     ${PYTORCH_PATCHES_PATH}/pytorch_16_use_new_logger.patch
     ${PYTORCH_PATCHES_PATH}/pytorch_14_namespace.patch
+    ${PYTORCH_PATCHES_PATH}/pytorch_18_string_view.patch
   )
 
   message(STATUS "Configuring libtorch")

patches/pytorch/pytorch_18_string_view.patch

@@ -0,0 +1,13 @@
+diff --git a/c10/util/string_view.h b/c10/util/string_view.h
+index 2958887afc..140521553e 100644
+--- a/c10/util/string_view.h
++++ b/c10/util/string_view.h
+@@ -581,7 +581,7 @@ class basic_string_view final {
+   constexpr bool equals_(basic_string_view rhs) const {
+     // We don't use string_view::compare() here but implement it manually because
+     // only looking at equality allows for more optimized code.
+-#if defined(__GNUC__)
++#if defined(__GNUC__) && !defined(__CUDACC__)
+     return size() == rhs.size() && 0 == __builtin_memcmp(data(), rhs.data(), size());
+ #elif __cpp_constexpr >= 201304
+     // if we are in C++14, write it iteratively. This is faster than the recursive C++11 implementation below.