On Friday I woke up realizing that I could not take the weekend off as I’d planned, even though I’d already worked over the previous weekend, because I had to turn on the CUDA flag on our new inference service. Otherwise, we’d be wasting GPUs over the weekend, and wouldn’t be ready for an important meeting next week.

It took me all Friday and all of today, literally until midnight, but I was finally able to run the inference with CUDA enabled.

I don’t feel like it could have gone much faster. Even having learned from a few mistakes (This PyTorch is only compatible with THAT version of CUDA), I think it still would have taken one and a half days to do.

To “turn on CUDA”.

A big part of learning to be effective in tech is to learn to navigate overly complicated processes using various heuristics (this feels like it’s taking too long, I’m going to backtrack).

Don’t beat yourself up. Move slower and more methodically. Draw a map and ask yourself where in the process you think you are exactly and what the obstacle is, and write those down. Take extensive notes, especially on paper, as you try to understand things. As you systematize your understanding, distill it into checklists and scripts.

And stretch. And stretch. And hydrate. And hydrate.