How to make the CPU block instead of spin on cudaMemcpyAsync/cudaStreamSynchronize

I don’t think it’s currently possible to change this flag in PyTorch and an older feature request was discussed here. However, I don’t know the status of it.