Ensure CRAN checks only use 2 cores max

[njtierney]

From CRAN team, we get:

Yes, we see, e.g., i a local run,

   Re-building vignettes had CPU time 3.8 times elapsed time

so clearly greta uses more than 2 cores´by default.

Please fix and resubmit.

Need to understand where and how this is happening to get greta back onto CRAN

[njtierney]

See Heather Turner's very helpful answer here: https://forum.posit.co/t/cpu-time-xx-times-elapsed-time/210116/2

[njtierney]

I tried to resolve this in #793 but ultimately despite my best efforts, trying to set the number of cores, setting n_cores to be 2, chains = 2, and even setting environment variables in vignette building, but all to no avail.

At the moment it seems that the computation use is not necessarily respected by greta/tensorflow/reticulate.

One option is to add some tests/R cmd check's for 2 cores only, to simulate CRAN machines. We currently do this for single core windows machines, e.g., see: https://github.com/greta-dev/greta/blob/main/.github/workflows/R-CMD-check.yaml#L123-L127

@goldingn do you have any suggestions here on resolving this? I am currently thinking that perhaps our fastest way onto CRAN would be use_article(), which essentially will remove vignettes from CRAN, but put them onto pkgdown...but that might be a problem because of #363

[goldingn]

Have you tried setting environment variables and tf config settings, similar to this? https://stackoverflow.com/questions/75344648/limit-number-of-cores-used-on-server-for-tensorflow-2-and-keras

Failing that, can't vignettes just be precompiled rather than being built on CRAN? Or is that no longer possible? That would avoid interaction with pkgdown.

You might be also able to precompute the figures and then load them when compiling on cran, with some creative use of eval (FALSE when running on CRAN) and echo

[njtierney]

Yes - that will be the next thing to try! It may also be an issue related to rstudio/tensorflow - see rstudio/tensorflow#625