Tweeted By @ThomasViehmann

Ever wondered what kind of optimiziations the @PyTorch JIT does to make your scripted RNNs fast? Here is a detailed account of making the JITs fuser (=automatic CUDA kernel generation) deal better with backward graphs (and LSTM backwards in particular): https://t.co/7RIA4PLRAc

— Thomas Viehmann (@ThomasViehmann) March 17, 2019