some of the for loops on range are done with prange instead, and maybe run the ensembles with only one job.

Hard to tell without benchmarks but I doubt that this approach would be faster. Parallelizing the outermost loop is usually the fastest (unless there are only a few steps in that loop => won't use all cores). Moreover if there's inherent sequential part in the trees, it would hurt the performances.

The most obvious parallelization would be to parallelize the split finding procedure over each feature since they're independent.

But as Jeremie noted that could lead to over subscription, typically for forests.

I feel like the only benefit would be to grow single trees.

I was meeting with people at JHU this week who are working on this:
https://github.com/neurodata/RerF
and they build several trees in parallel on the same core, I think, and said it was super fast. I haven't looked into it, though.

I would tag this as hard. it requires a pretty solid knowledge of the trees, and some cython code.

(BTW, "intermediate" and "moderate" are synonyms to me, assuming a "easy / something_in_between / hard" scale.)

Can I work on the benchmark?
The idea is to measure the time taken for the master and a new version which replaces range with prange?