Making a note here that I don't think the existing cuml implementation has the ability to tune the percentage of explained variance.
For example, in sklearn we can set 0 < n_components < 1 where the user can select a percentage of the explained variance to recover and the n_components is automatically determined by the algorithm in order to satisfy that.

We will have to build that piece out since it doesn't exist in the current implementation.

Tuning is not what's being asked for. Exposing the explained variance is what's being requested (it is used for tuning / selecting the number of components but that's something the user does, not something we need to do).

We are missing verbose here, no?

verbose was a unused parameter, removed it from pca.hpp in 99f32fc

Remove commented code

Addressed in 074fd96

Small thing but important for API consistency: params config is passed by value here and in all other functions in the PR. The cuVS convention afaik is const params& for example in kmeans::fit:

void fit(raft::resources const& handle,
         const cuvs::cluster::kmeans::params& params,
         raft::device_matrix_view<const float, int> X,
         std::optional<raft::device_vector_view<const float, int>> sample_weight,
         raft::device_matrix_view<float, int> centroids,
         raft::host_scalar_view<float> inertia,
         raft::host_scalar_view<int> n_iter);

This won't affect performance or correctness at all, but for consistency I would suggest changing to const params& config throughout. Applies to all 8 overloads in this header plus the detail layer.

Addressed in c7c52a7

Both basicTest and advancedTest set n_components = n_col, so the inverse transform should perfectly reconstruct the input.

Consider adding a test with n_components < n_col that verifies the reconstruction error is bounded but non-zero, this would confirm the dimensionality reduction is actually working, not just passing data through unchanged and be a useful test case.

Addressed in 948882c