diff --git a/advanced_source/cpp_extension.rst b/advanced_source/cpp_extension.rst
index 9fe1db38d7a..265b3774f88 100644
--- a/advanced_source/cpp_extension.rst
+++ b/advanced_source/cpp_extension.rst
@@ -1009,6 +1009,18 @@ simpler::
 
   lltm = load(name='lltm', sources=['lltm_cuda.cpp', 'lltm_cuda_kernel.cu'])
 
+Note that when constructing :func:`CUDAExtension`, you might have to pass
+``extra_compile_args`` to :func:`CUDAExtension`, to avoid collisions
+between half operator overloading in pytorch and cuda headers. (More on this
+issue can be found here: 
+https://github.com/pytorch/pytorch/pull/10301#issuecomment-416773333)::
+
+  extra_compile_args = {'cxx':[],
+                        'nvcc':['-DCUDA_HAS_FP16=1',
+                        '-D__CUDA_NO_HALF_OPERATORS__',
+                        '-D__CUDA_NO_HALF_CONVERSIONS__',
+                        '-D__CUDA_NO_HALF2_OPERATORS__']}
+
 Performance Comparison
 **********************