Wow. Thank you. I thought that PyTorch would use the GPU automatically. Apparently not. I got 6 second wall time via %time with a K80 in Colab. CPU version with PyTorch is about 10x slower. (Intel(R) Xeon(R) CPU @ 2.30GHz). Will update.
FYI, following this link said the referrer was spamming- probably Medium’s fault. I was able to continue on. Span and live time are very interesting concepts. Great article, Will. I totally agree with your thesis. All the best!