attn_loss=0

[CeciliaTheBirb]

I also couldn't do

wasserstein_ext = torch.utils.cpp_extension.load_inline("wasserstein", cpp_sources="", cuda_sources=cuda_source,
                                                    extra_cuda_cflags=["--expt-relaxed-constexpr"]   )

so I just commented it out and used the cpu version below in the method sinkstep. However, for some reason during training, the attn_loss is always zero and the results are very bad. how to resolve this issue?

[Wenjiezhao11]

I've encountered the same problem. Have you found a solution?