Hengrui Zhang

See this issue #1

The paper samples nodes from the graph then compute the loss.

This issue occurs only when using `import hadamard_cuda`. When using ```python import torch.utils.cpp_extension hadamard_cuda = torch.utils.cpp_extension.load( name='hadamard_cuda', sources=[ 'hadamard_cuda.cpp', 'hadamard_cuda_kernel.cu', ], extra_cuda_cflags=['-O2'], verbose=False ) ``` The output is correct.

Meet the same issue.