> But these computation graphs are similar every time. For now I'm testing one-after-another because failure of previous (overflow, too-small-transfer, etc) gi

> But these computation graphs are similar every time. For now I'm testing one-after-another because failure of previous (overflow, too-small-transfer, etc) gives a hint about which part can be changed in the graph. I've been thinking about vectorizing over multiple guesses at once, but don't see a fast reliable way to merge several graphs (also that induces significant burden on other parts of the project - autograd is just one component).