Decide on the development environment

[charlesfrye]

Bare metal or containers? Cloud-native, cloud-friendly, or local? Are M1 Macs in scope? Is Colab a target?

[charlesfrye]

I am bullish on cloud-native development now.

GitHub Codespaces just became generally available. You can develop in a CPU-backed environment for up to 60 hours a month. I expect there to be really nice integrations with GitHub Actions as well.

GPU-backed Codespaces are not GA, but you can ask for access here. I expect them to become GA by the time the next class is taught.

Alternatively, we could use Modal. I've found the development experience very smooth -- e.g. you can run a IPython REPL inside of your programs executing on the cloud with just a single (non-Modal) line of code.

In either case, but especially with Codespaces, we'll want to make the best possible use of devcontainers. We can use Dockerfiles to configure the containers where possible to allow for development outside of VSCode.

[EdAbati]

I really like the idea of using containers: it will be easier to move from Dockerfile to devcontainers to Codespaces/VSCode.

Cloud-native, cloud-friendly, or local?

I am a bit unsure. On one side, I understand that cloud development tools (like Codespaces) are very handy. On the other, when looking into these tools, I am often worried about how much the final cost of usage would be, and I end up developing locally. (At least at the start, when writing code and running small tests, local laptops are often enough.)
Anyway, if we stick with containers it should be easier to move from local to cloud development, right?

Are M1 Macs in scope?

If we go with containers, I am not sure with could fully support M1 Mac. For example, is there a way to connect to their GPU via Docker? (But as a M1 user, I hope that they will be supported at some point 😇 )

Is Colab a target?

I have used Colab mostly to quickly interact with a public notebooks (e.g. the course labs). I never used it for 'full' development. My opinion is that it should not be priority num 1.

[charlesfrye]

I am often worried about how much the final cost of usage would be

At least in the context of the demo project, I think we can keep much of the development within the free tier of Codespaces. For CPU-only development, ie everything except training models, the free tier is very generous.

Modal (https://modal.com) provides ~15 hours of free GPU compute to cover what training we need.